Compare commits

..
Author SHA1 Message Date
bsodfather d80eda7329 Merge pull request 'Story 4.4 — безопасные повторные попытки и отмена durable Task' (#112) from delivery/epic4-story-4.4 into main
CI / UI Checks (push) Successful in 11s
CI / Rust Checks (push) Successful in 27m33s
CI / Rust Feature Validation (push) Skipped
CI / Frontend E2E (push) Skipped
CI / Community Image Smoke (push) Skipped
CI / Required Checks (push) Successful in 1s
CI / Deploy (push) Skipped
Community Release / release (push) Successful in 44m29s
Merge pull request 'Story 4.4 — безопасные повторные попытки и отмена durable Task' (#112) from delivery/epic4-story-4.4 into main
2026-10-30 12:46:29 +00:00
bsodfather a3057de51e feat(epic4): безопасные повторы и отмена durable Task (Story 4.4)
CI / UI Checks (pull_request) Successful in 8s
CI / Rust Checks (pull_request) Successful in 20m46s
CI / Rust Feature Validation (pull_request) Skipped
CI / Frontend E2E (pull_request) Skipped
CI / Community Image Smoke (pull_request) Skipped
CI / Required Checks (pull_request) Successful in 1s
CI / Deploy (pull_request) Skipped
Durable per-outbound attempt до отправки запроса, bounded retry/read,
server-owned verified writes, честный cancel/unknown, deadline/fencing,
additive schema v29. Staged renames task.rs/tasks.rs разложены на модули.

Includes disclosed cosmetic parity exception: единственная правка вне
принятой дельты — удалена лишняя пустая строка в конце
scripts/ci-story-checks.sh (EOF blank line); story-код не менялся.
2026-10-30 12:21:57 +00:00
bsodfather 8c02380762 fix(ui): временно ограничить npm build-стадию UI TLS 1.2 на CI-хосте
CI / UI Checks (push) Successful in 9s
CI / Rust Checks (push) Successful in 25m23s
CI / Rust Feature Validation (push) Skipped
CI / Frontend E2E (push) Skipped
CI / Community Image Smoke (push) Skipped
CI / Required Checks (push) Successful in 3s
CI / Deploy (push) Skipped
Merge pull request #111 'fix(ui): временно ограничить npm build-стадию UI TLS 1.2 на CI-хосте' from delivery/ui-npm-tls12-cap
2026-10-30 10:51:38 +00:00
hermes_dev 3b5fcc51cc fix(ui): pin npm build stage to TLS 1.2 on the CI host uplink
CI / UI Checks (pull_request) Successful in 9s
CI / Rust Checks (pull_request) Successful in 20m41s
CI / Rust Feature Validation (pull_request) Skipped
CI / Frontend E2E (pull_request) Skipped
CI / Community Image Smoke (pull_request) Skipped
CI / Required Checks (pull_request) Successful in 1s
CI / Deploy (pull_request) Skipped
The CI host uplink (192.168.1.108) drops TLS 1.3 ClientHello carrying the
hybrid X25519MLKEM768 key share, so registry.npmjs.org is unreachable from
the seeded node:22-alpine build stage while the same host answers over TLS
1.2. RUN npm ci hung until the build was cancelled.

Cap only the UI build stage at TLS 1.2 via NODE_OPTIONS=--tls-max-v1.2.
Certificates, SRI and the lockfile are untouched; no mirror, fallback or
cache-substituted packages. Remove the line once TLS 1.3 works from the CI
host again.
2026-10-30 09:56:19 +00:00
bsodfather 433987413e Merge pull request 'Story 4.3 — исполнять и восстанавливать durable Task через PG lease/fence' (#110) from delivery/epic4-story-4.3 into main
CI / UI Checks (push) Successful in 8s
CI / Rust Checks (push) Successful in 26m20s
CI / Rust Feature Validation (push) Skipped
CI / Frontend E2E (push) Skipped
CI / Community Image Smoke (push) Skipped
CI / Required Checks (push) Successful in 2s
CI / Deploy (push) Skipped
Community Release / release (push) Failing after 31m2s
2026-10-29 23:08:04 +00:00
bsodfather 5488eb73ec chore(migrations): зафиксировать каноническую последовательность v26–v28
CI / Rust Checks (pull_request) Successful in 22m6s
CI / Rust Feature Validation (pull_request) Skipped
CI / Frontend E2E (pull_request) Skipped
CI / Community Image Smoke (pull_request) Skipped
CI / Required Checks (pull_request) Successful in 1s
CI / Deploy (pull_request) Skipped
CI / UI Checks (pull_request) Successful in 12s
Обязательный CI-гейт «Check canonical migration contract» сравнивает
`crank-migrate plan` с docs/schemas/migration-sequence.json. Принятая дельта
Story 4.3 добавила миграции v26–v28, но снимок контракта не был регенерирован,
поэтому exact-head CI падал на точном PR head с contract_drift (plan.compare).

Файл получен машинной регенерацией из того же дерева:
`crank-migrate plan > docs/schemas/migration-sequence.json`; проверка
`crank-migrate plan --check` возвращает {"status":"contract_current"}.
Содержимое меняется только добавлением трёх записей 26/27/28; существующие
записи 1–25 и их checksum не изменены, поведение продукта не меняется.
2026-10-29 22:43:05 +00:00
bsodfather f9f8c625d1 feat(epic4): исполнять durable Task через PG lease/fence и восстанавливать после сбоев (Story 4.3)
CI / UI Checks (pull_request) Successful in 15s
CI / Rust Checks (pull_request) Failing after 4m58s
CI / Rust Feature Validation (pull_request) Skipped
CI / Frontend E2E (pull_request) Skipped
CI / Community Image Smoke (pull_request) Skipped
CI / Required Checks (pull_request) Failing after 1s
CI / Deploy (pull_request) Skipped
mcp-server запускает сохранённые фоновые задачи: захват по PostgreSQL lease с
fence-токеном, зашифрованный input и повторная проверка доступа до сетевого
вызова, checkpoint с AttemptID до dispatch и восстановление только по PG.
Устаревший worker не может зафиксировать результат.

Добавлены миграции v26–v28 (V25 не меняется), guard их отсутствия в старой
схеме, метрики/трассы попыток (attempt_trace) и остановка до отправки запроса:
стоп в окне после checkpoint и до permit завершает задачу как
failed(shutdown_before_dispatch) без сетевых вызовов, соседние queued-задачи
остаются доступными. Retry-попытки получают отдельные дочерние spans с durable
current_attempt_id. Shared admission/fairness из Story 4.1 сохранены.

Проверки: durable_tasks 13/13, stages 11/11 (retry 5), runtime lib 47,
bin shutdown 1, registry tasks 16, admission 11, migrations 58 (независимая QA
и исходный runtime-прогон), точный exact-head CI остаётся обязательным.
2026-10-29 22:26:52 +00:00
bsodfather 6df6ffac4e Merge pull request 'Story 4.2 — подтверждать durable Task после атомарного PostgreSQL commit (FR-22, FR-23)' (#109) from delivery/epic4-story-4.2 into main
CI / UI Checks (push) Successful in 13s
CI / Rust Checks (push) Successful in 21m12s
CI / Rust Feature Validation (push) Skipped
CI / Frontend E2E (push) Skipped
CI / Community Image Smoke (push) Skipped
CI / Required Checks (push) Successful in 1s
CI / Deploy (push) Skipped
Community Release / release (push) Successful in 36m42s
Merge accepted Story 4.2 durable Task acceptance
2026-10-27 07:33:30 +00:00
bsodfather 2219a7217d feat(epic4): подтверждать durable Task после атомарного PG-commit (Story 4.2)
CI / UI Checks (pull_request) Successful in 8s
CI / Rust Checks (pull_request) Successful in 19m34s
CI / Rust Feature Validation (pull_request) Skipped
CI / Frontend E2E (pull_request) Skipped
CI / Community Image Smoke (pull_request) Skipped
CI / Required Checks (pull_request) Successful in 1s
CI / Deploy (pull_request) Skipped
Application entry выдаёт Task ID только после атомарного PostgreSQL commit
Task+intent+capacity; replay/conflict, owner scope, bounded envelope,
terminal CAS и fail-closed readiness проверены на реальном PG.

Также включён принятый план перехода на rmcp (только документ, миграция не
запускается).
2026-10-27 00:09:36 +00:00
bsodfather cc612516a2 Merge pull request 'Story 4.1 — публиковать execution mode и применять admission limits (FR-21, FR-24)' (#104) from delivery/epic4-story-4.1 into main
CI / Rust Checks (push) Successful in 20m52s
CI / Rust Feature Validation (push) Skipped
CI / Frontend E2E (push) Skipped
CI / UI Checks (push) Successful in 10s
CI / Community Image Smoke (push) Skipped
CI / Required Checks (push) Successful in 1s
CI / Deploy (push) Skipped
2026-10-26 21:38:06 +00:00
bsodfather 599dc1867b Merge pull request 'ci: безопасно интегрировать принятый Crank и отделить production deploy' (#105) from ops/separate-production-deploy into main
CI / UI Checks (push) Successful in 12s
CI / Rust Checks (push) Successful in 16m0s
CI / Rust Feature Validation (push) Skipped
CI / Frontend E2E (push) Skipped
CI / Community Image Smoke (push) Skipped
CI / Required Checks (push) Successful in 2s
CI / Deploy (push) Skipped
2026-10-26 21:36:09 +00:00
bsodfather e80d30ddb0 feat(epic4): публиковать execution mode и применять общий admission (Story 4.1, FR-21, FR-24)
CI / UI Checks (pull_request) Successful in 11s
CI / Rust Checks (pull_request) Successful in 12m45s
CI / Rust Feature Validation (pull_request) Skipped
CI / Frontend E2E (pull_request) Skipped
CI / Community Image Smoke (pull_request) Skipped
CI / Required Checks (pull_request) Successful in 1s
CI / Deploy (pull_request) Skipped
Community Release / release (push) Successful in 40m35s
Tool Draft фиксирует synchronous/background/adaptive, deadline и fallback;
runtime применяет единый снимок политики (concurrency, очередь, approval,
deadline) до внешнего действия. Насыщение отклоняется typed overload с
retry guidance, mandatory safety store fail closed, background/adaptive
используют опубликованный synchronous/unsupported fallback.

Refs: #29
2026-10-26 21:08:08 +00:00
128 changed files with 16865 additions and 290 deletions

No files matched your search

Generated
+1
View File
@@ -826,6 +826,7 @@ dependencies = [
name = "crank-registry"
version = "0.3.1"
dependencies = [
"async-trait",
"crank-artifacts",
"crank-core",
"crank-mapping",
+99 -87
View File
@@ -2,6 +2,10 @@
mod admin_auth_command;
#[path = "crank_migrate/db_connect.rs"]
mod db_connect;
#[path = "crank_migrate/reporting.rs"]
mod reporting;
use reporting::{active_rotation, ensure_no_options, read_required_key_file, rotation_json};
use crank_config::{ConfigSource, parse_migrator};
use crank_registry::{
@@ -11,7 +15,7 @@ use crank_registry::{
};
use crank_runtime::SecretCrypto;
use serde_json::json;
use std::{path::Path, process::ExitCode};
use std::process::ExitCode;
use time::OffsetDateTime;
const MASTER_KEY_ROTATION_PAGE_SIZE: i64 = 1_000;
#[tokio::main]
@@ -765,6 +769,59 @@ async fn process_rotation_batch(
processed += 1;
}
}
let mut after_task_id: Option<String> = None;
loop {
let tasks = registry
.list_tasks_for_master_key_epoch_page(
rotation.source_epoch,
after_task_id.as_deref(),
MASTER_KEY_ROTATION_PAGE_SIZE,
)
.await
.map_err(CliError::from_registry)?;
if tasks.is_empty() {
break;
}
for task in tasks {
after_task_id = Some(task.task_id.clone());
if task.target_epoch == Some(rotation.target_epoch) {
continue;
}
if max_versions.is_some_and(|limit| processed as usize >= limit) {
return Ok(processed);
}
let plaintext = current_crypto
.decrypt_for_epoch(&task.key_version, task.epoch, &task.ciphertext)
.map_err(|_| {
CliError::new(
"master_key_rotation_verification_failed",
"master_key.rotation",
"rerun_rotation_or_abort",
)
})?;
let target_ciphertext = target_crypto.encrypt(&plaintext).map_err(|_| {
CliError::new(
"master_key_rotation_verification_failed",
"master_key.rotation",
"rerun_rotation_or_abort",
)
})?;
let now = OffsetDateTime::now_utc();
registry
.stage_master_key_rotation_task_ciphertext(
&rotation.id,
&task.task_id,
rotation.source_epoch,
&target_ciphertext,
target_crypto.key_version(),
rotation.target_epoch,
&now,
)
.await
.map_err(CliError::from_registry)?;
processed += 1;
}
}
Ok(processed)
}
@@ -818,6 +875,47 @@ async fn verify_staged_targets(
verified += 1;
}
}
let mut after_task_id: Option<String> = None;
loop {
let tasks = registry
.list_target_tasks_for_master_key_rotation_page(
rotation.target_epoch,
after_task_id.as_deref(),
MASTER_KEY_ROTATION_PAGE_SIZE,
)
.await
.map_err(CliError::from_registry)?;
if tasks.is_empty() {
break;
}
for task in tasks {
after_task_id = Some(task.task_id.clone());
let ciphertext = task.target_ciphertext.as_deref().ok_or_else(|| {
CliError::new(
"master_key_rotation_verification_failed",
"master_key.rotation",
"rerun_rotation_or_abort",
)
})?;
let key_version = task.target_key_version.as_deref().ok_or_else(|| {
CliError::new(
"master_key_rotation_verification_failed",
"master_key.rotation",
"rerun_rotation_or_abort",
)
})?;
target_crypto
.decrypt_for_epoch(key_version, rotation.target_epoch, ciphertext)
.map_err(|_| {
CliError::new(
"master_key_rotation_verification_failed",
"master_key.rotation",
"rerun_rotation_or_abort",
)
})?;
verified += 1;
}
}
if verified != rotation.total_secret_versions {
return Err(CliError::new(
"master_key_rotation_verification_failed",
@@ -877,89 +975,3 @@ fn decrypt_current(
)
})
}
async fn active_rotation(
registry: &PostgresRegistry,
allowed_states: &[&str],
) -> Result<Option<MasterKeyRotationRecord>, CliError> {
let status = registry
.master_key_rotation_status()
.await
.map_err(CliError::from_registry)?;
Ok(status
.rotations
.into_iter()
.find(|rotation| allowed_states.contains(&rotation.state.as_str())))
}
fn rotation_json(rotation: &MasterKeyRotationRecord) -> serde_json::Value {
json!({
"id": rotation.id,
"source_epoch": rotation.source_epoch,
"target_epoch": rotation.target_epoch,
"state": rotation.state,
"backup_ref": rotation.backup_ref.as_ref().map(|_| "configured"),
"checkpoint_secret_id": rotation.checkpoint_secret_id,
"total_secret_versions": rotation.total_secret_versions,
"processed_secret_versions": rotation.processed_secret_versions,
"verified_secret_versions": rotation.verified_secret_versions,
"failure_code": rotation.failure_code,
})
}
fn read_required_key_file(path: Option<&str>) -> Result<String, CliError> {
let path = path.ok_or_else(|| {
CliError::new("invalid_command", "cli.arguments", "run_master_key_status")
})?;
if path.len() > 512 || path.bytes().any(|byte| byte.is_ascii_control()) {
return Err(CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
));
}
let metadata = std::fs::metadata(Path::new(path)).map_err(|_| {
CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
)
})?;
if !metadata.is_file() || metadata.len() > 16_384 {
return Err(CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
));
}
let value = std::fs::read_to_string(Path::new(path)).map_err(|_| {
CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
)
})?;
if value.trim().is_empty() {
return Err(CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
));
}
Ok(value)
}
fn ensure_no_options(options: &MasterKeyOptions) -> Result<(), CliError> {
if options.current_key_file.is_some()
|| options.target_key_file.is_some()
|| options.backup_ref.is_some()
|| options.rotation_id.is_some()
|| options.max_versions.is_some()
{
return Err(CliError::new(
"invalid_command",
"cli.arguments",
"run_master_key_status",
));
}
Ok(())
}
@@ -0,0 +1,98 @@
//! Leaf/reporting helpers of the migrator CLI.
//!
//! Extracted from `crank-migrate.rs` without behaviour changes so the binary entry stays within
//! the code-health line limit.
use std::path::Path;
use crank_registry::{MasterKeyRotationRecord, PostgresRegistry};
use serde_json::json;
use super::{CliError, MasterKeyOptions};
pub(super) async fn active_rotation(
registry: &PostgresRegistry,
allowed_states: &[&str],
) -> Result<Option<MasterKeyRotationRecord>, CliError> {
let status = registry
.master_key_rotation_status()
.await
.map_err(CliError::from_registry)?;
Ok(status
.rotations
.into_iter()
.find(|rotation| allowed_states.contains(&rotation.state.as_str())))
}
pub(super) fn rotation_json(rotation: &MasterKeyRotationRecord) -> serde_json::Value {
json!({
"id": rotation.id,
"source_epoch": rotation.source_epoch,
"target_epoch": rotation.target_epoch,
"state": rotation.state,
"backup_ref": rotation.backup_ref.as_ref().map(|_| "configured"),
"checkpoint_secret_id": rotation.checkpoint_secret_id,
"total_secret_versions": rotation.total_secret_versions,
"processed_secret_versions": rotation.processed_secret_versions,
"verified_secret_versions": rotation.verified_secret_versions,
"failure_code": rotation.failure_code,
})
}
pub(super) fn read_required_key_file(path: Option<&str>) -> Result<String, CliError> {
let path = path.ok_or_else(|| {
CliError::new("invalid_command", "cli.arguments", "run_master_key_status")
})?;
if path.len() > 512 || path.bytes().any(|byte| byte.is_ascii_control()) {
return Err(CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
));
}
let metadata = std::fs::metadata(Path::new(path)).map_err(|_| {
CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
)
})?;
if !metadata.is_file() || metadata.len() > 16_384 {
return Err(CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
));
}
let value = std::fs::read_to_string(Path::new(path)).map_err(|_| {
CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
)
})?;
if value.trim().is_empty() {
return Err(CliError::new(
"master_key_input_invalid",
"master_key.input",
"verify_operator_input",
));
}
Ok(value)
}
pub(super) fn ensure_no_options(options: &MasterKeyOptions) -> Result<(), CliError> {
if options.current_key_file.is_some()
|| options.target_key_file.is_some()
|| options.backup_ref.is_some()
|| options.rotation_id.is_some()
|| options.max_versions.is_some()
{
return Err(CliError::new(
"invalid_command",
"cli.arguments",
"run_master_key_status",
));
}
Ok(())
}
+1
View File
@@ -181,6 +181,7 @@ mod tests {
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
},
+18 -1
View File
@@ -3,7 +3,7 @@ use std::{io, net::SocketAddr, process::ExitCode, time::Duration};
use admin_api::{
app::build_app,
auth::{AuthSettings, BootstrapAdminConfig},
pool_metrics::spawn_postgres_pool_metrics,
pool_metrics::{spawn_admission_metrics, spawn_postgres_pool_metrics},
reconciliation::{open_reconciliation_store, spawn_artifact_reconciliation},
service::AdminServiceBuilder,
state::AppState,
@@ -169,6 +169,7 @@ async fn run(
};
if metrics_enabled {
spawn_postgres_pool_metrics(registry.pool().clone());
spawn_admission_metrics(registry.pool().clone());
}
let base_url = config
.runtime
@@ -214,6 +215,16 @@ async fn run(
.with_limits(runtime_limits)
.with_response_cache(cache_stores.response.clone())
.with_coordination_store(cache_stores.coordination.clone())
.with_admission(std::sync::Arc::new(
crank_registry::PostgresExecutionAdmission::new(
registry.pool().clone(),
crank_registry::AdmissionLimits {
global_limit: runtime_limits.max_concurrent_unary,
..crank_registry::AdmissionLimits::default()
},
crank_core::AdmissionSchedulerConfig::default(),
)?,
))
.build();
let identity_provider =
PasswordIdentityProvider::new(registry.clone(), auth_settings.password_pepper.clone());
@@ -321,6 +332,12 @@ fn preflight_config(config: &AdminProcessConfig) -> Result<(), crank_config::Con
config.runtime.max_concurrent_sessions,
)
.map_err(|_| invalid("runtime.limits"))?;
crank_registry::AdmissionLimits {
global_limit: config.runtime.max_concurrent_unary,
..crank_registry::AdmissionLimits::default()
}
.validate()
.map_err(|_| invalid("runtime.admission"))?;
runtime_cache_config(config).map_err(|_| invalid("cache"))?;
RequestRateLimitConfig::new(
config.rate_limit.requests_per_second,
+14
View File
@@ -11,3 +11,17 @@ pub fn spawn_postgres_pool_metrics(pool: PgPool) -> tokio::task::JoinHandle<()>
}
})
}
/// Story 4.1: samples the authoritative admission queue so the oldest-waiting age grows (and returns
/// to zero) without waiting for a new request.
pub fn spawn_admission_metrics(pool: PgPool) -> tokio::task::JoinHandle<()> {
tokio::spawn(async move {
let mut interval = tokio::time::interval(Duration::from_secs(5));
loop {
interval.tick().await;
if let Ok(snapshot) = crank_registry::admission_snapshot(&pool).await {
crank_observability::record_admission_oldest_waiting(snapshot.oldest_waiting_ms);
}
}
})
}
+27 -2
View File
@@ -37,13 +37,22 @@ pub async fn readiness(State(state): State<AppState>) -> impl IntoResponse {
} else {
"not_ready"
};
let execution_admission = if checks.execution_admission {
"ready"
} else {
"not_ready"
};
if checks.is_ready() {
(
StatusCode::OK,
Json(json!({
"service": "admin-api",
"status": "ready",
"checks": { "postgres": postgres, "artifact_storage": artifact_storage }
"checks": {
"postgres": postgres,
"artifact_storage": artifact_storage,
"execution_admission": execution_admission
}
})),
)
} else {
@@ -52,7 +61,11 @@ pub async fn readiness(State(state): State<AppState>) -> impl IntoResponse {
Json(json!({
"service": "admin-api",
"status": "not_ready",
"checks": { "postgres": postgres, "artifact_storage": artifact_storage }
"checks": {
"postgres": postgres,
"artifact_storage": artifact_storage,
"execution_admission": execution_admission
}
})),
)
}
@@ -67,6 +80,7 @@ mod tests {
let checks = ReadinessChecks {
postgres: false,
artifact_storage: true,
execution_admission: true,
};
assert!(!checks.is_ready());
assert_eq!(
@@ -92,6 +106,7 @@ mod tests {
let checks = ReadinessChecks {
postgres: true,
artifact_storage: false,
execution_admission: true,
};
assert!(!checks.is_ready());
assert_eq!(
@@ -111,4 +126,14 @@ mod tests {
"not_ready"
);
}
#[test]
fn readiness_checks_identify_an_admission_state_failure() {
let checks = ReadinessChecks {
postgres: true,
artifact_storage: true,
execution_admission: false,
};
assert!(!checks.is_ready());
}
}
+9 -3
View File
@@ -55,8 +55,8 @@ use crate::{
storage::LocalArtifactStorage,
};
use operation_validation::{
validate_approval_policy, validate_execution_timeout, validate_idempotency_policy,
validate_protocol_target, validate_response_cache_policy,
validate_approval_policy, validate_execution_policy, validate_execution_timeout,
validate_idempotency_policy, validate_protocol_target, validate_response_cache_policy,
};
#[derive(Clone)]
@@ -104,11 +104,12 @@ pub use crate::dto::*;
pub struct ReadinessChecks {
pub postgres: bool,
pub artifact_storage: bool,
pub execution_admission: bool,
}
impl ReadinessChecks {
pub const fn is_ready(self) -> bool {
self.postgres && self.artifact_storage
self.postgres && self.artifact_storage && self.execution_admission
}
}
@@ -150,6 +151,8 @@ pub(super) struct CredentialAuditRecord<'a> {
impl AdminService {
pub async fn readiness(&self) -> ReadinessChecks {
let postgres = self.registry.ping().await.is_ok();
// Story 4.1: the mandatory admission state is a readiness dependency of every execution.
let execution_admission = self.registry.execution_admission_health().await.is_ok();
let store = Arc::clone(&self.artifact_store);
let artifact_storage = tokio::task::spawn_blocking(move || store.check_health())
.await
@@ -157,6 +160,7 @@ impl AdminService {
ReadinessChecks {
postgres,
artifact_storage,
execution_admission,
}
}
@@ -725,6 +729,7 @@ impl AdminService {
validate_response_cache_policy(&payload.target, &payload.execution_config)?;
validate_idempotency_policy(&payload.target, &payload.execution_config)?;
validate_approval_policy(&payload.execution_config)?;
validate_execution_policy(&payload.execution_config)?;
payload.input_mapping.validate_paths()?;
payload.output_mapping.validate_paths()?;
Ok(())
@@ -738,6 +743,7 @@ impl AdminService {
validate_response_cache_policy(&operation.target, &operation.execution_config)?;
validate_idempotency_policy(&operation.target, &operation.execution_config)?;
validate_approval_policy(&operation.execution_config)?;
validate_execution_policy(&operation.execution_config)?;
operation.input_mapping.validate_paths()?;
operation.output_mapping.validate_paths()?;
Ok(())
+1
View File
@@ -441,6 +441,7 @@ fn demo_rest_operation_payload() -> OperationPayload {
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
},
+1
View File
@@ -635,6 +635,7 @@ impl AdminService {
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
},
@@ -610,6 +610,7 @@ impl AdminService {
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref,
headers: BTreeMap::new(),
},
@@ -160,6 +160,23 @@ pub(super) fn validate_approval_policy(
Ok(())
}
pub(super) fn validate_execution_policy(
execution_config: &crank_core::ExecutionConfig,
) -> Result<(), ApiError> {
let Some(policy) = execution_config.execution_policy.as_ref() else {
return Ok(());
};
if let Err(error) = policy.validate() {
return Err(ApiError::validation_with_context(
error.to_string(),
json!({
"field": "execution_config.execution_policy.deadline_ms",
}),
));
}
Ok(())
}
#[cfg(test)]
mod tests {
use std::collections::BTreeMap;
@@ -171,8 +188,8 @@ mod tests {
};
use super::{
validate_approval_policy, validate_execution_timeout, validate_idempotency_policy,
validate_response_cache_policy,
validate_approval_policy, validate_execution_policy, validate_execution_timeout,
validate_idempotency_policy, validate_response_cache_policy,
};
fn cacheable_execution_config() -> ExecutionConfig {
@@ -183,6 +200,7 @@ mod tests {
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
}
@@ -201,6 +219,7 @@ mod tests {
}),
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
}
@@ -365,4 +384,36 @@ mod tests {
"approval ttl must be between 1 and 300 seconds"
);
}
#[test]
fn accepts_published_execution_policy_and_absent_legacy_policy() {
let mut config = cacheable_execution_config();
validate_execution_policy(&config).unwrap();
config.execution_policy = Some(crank_core::ToolExecutionPolicy {
mode: crank_core::ToolExecutionMode::Adaptive,
deadline_ms: 120_000,
no_tasks_fallback: crank_core::NoTasksFallback::Unsupported,
});
validate_execution_policy(&config).unwrap();
}
#[test]
fn rejects_zero_and_overflowing_execution_policy_deadline() {
let mut config = cacheable_execution_config();
config.execution_policy = Some(crank_core::ToolExecutionPolicy {
mode: crank_core::ToolExecutionMode::Synchronous,
deadline_ms: 0,
no_tasks_fallback: crank_core::NoTasksFallback::Synchronous,
});
let error = validate_execution_policy(&config).unwrap_err();
assert!(matches!(error, crate::error::ApiError::Validation { .. }));
config.execution_policy = Some(crank_core::ToolExecutionPolicy {
mode: crank_core::ToolExecutionMode::Background,
deadline_ms: crank_core::MAX_TOOL_DEADLINE_MS + 1,
no_tasks_fallback: crank_core::NoTasksFallback::Synchronous,
});
assert!(validate_execution_policy(&config).is_err());
}
}
@@ -495,6 +495,7 @@ pub(super) fn test_operation_payload(base_url: &str, name: &str) -> OperationPay
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
},
+2 -1
View File
@@ -29,7 +29,7 @@ sha2.workspace = true
sqlx.workspace = true
thiserror.workspace = true
time.workspace = true
tokio = { workspace = true, features = ["sync"] }
tokio = { workspace = true, features = ["sync", "signal"] }
tracing.workspace = true
tracing-subscriber.workspace = true
uuid.workspace = true
@@ -45,5 +45,6 @@ opentelemetry_sdk.workspace = true
prost.workspace = true
reqwest.workspace = true
tower.workspace = true
tokio = { workspace = true, features = ["net", "time"] }
tracing-opentelemetry.workspace = true
metrics-util = "0.20.4"
+193 -11
View File
@@ -1,7 +1,7 @@
use std::{io, process::ExitCode, time::Duration};
use crank_community_mcp::{
auth::CommunityMachineCredentialVerifier, build_app_with_artifacts,
auth::CommunityMachineCredentialVerifier, build_app_with_artifacts_and_tasks,
session::PostgresTransportSessionStore,
};
use crank_config::{
@@ -14,13 +14,14 @@ use crank_observability::{
OtlpTraceConfig, RedactionLimits, SentryConfig, ServiceIdentity, capture_critical_error,
};
use crank_registry::{
MASTER_KEY_CIPHER_CONTRACT, MasterKeyIdentityCandidate, PostgresPoolConfig, PostgresRegistry,
AdmissionLimits, MASTER_KEY_CIPHER_CONTRACT, MasterKeyIdentityCandidate,
PostgresExecutionAdmission, PostgresPoolConfig, PostgresRegistry, PostgresTaskStore,
};
use crank_runtime::{
RequestRateLimitConfig, RequestRateLimiter, RuntimeCacheConfig, RuntimeCacheStores,
RuntimeLimits, SecretCrypto,
};
use mcp_server::pool_metrics::spawn_postgres_pool_metrics;
use mcp_server::pool_metrics::{spawn_admission_metrics, spawn_postgres_pool_metrics};
use sqlx::postgres::PgConnectOptions;
use tokio::net::TcpListener;
use tracing::info;
@@ -175,6 +176,7 @@ async fn run(
};
if metrics_enabled {
spawn_postgres_pool_metrics(registry.pool().clone());
spawn_admission_metrics(registry.pool().clone());
}
let session_store = PostgresTransportSessionStore::from_pool(registry.pool().clone()).await?;
let secret_crypto =
@@ -190,12 +192,55 @@ async fn run(
.with_limits(runtime_limits)
.with_response_cache(cache_stores.response.clone())
.with_coordination_store(cache_stores.coordination.clone())
.with_admission(execution_admission(
registry.pool().clone(),
runtime_limits,
)?)
.build();
// Published static Resources are read from the same immutable content-addressed root the
// admin-api publishes into. The compatibility constructor keeps an unprovisioned/unsafe root
// fail-closed per read instead of preventing startup; nothing is ever written from this process.
let artifact_store = crank_artifacts::ArtifactStore::new(config.storage_root.clone());
let app = build_app_with_artifacts(
// Story 4.2/4.3: the durable Task application entry shares the registry pool, the published
// catalog refresh interval and the exact admission limits of the runtime; the public wire stays
// synchronous/unsupported while a fenced in-process worker executes accepted Tasks.
let task_store = std::sync::Arc::new(PostgresTaskStore::new(
registry.pool().clone(),
admission_limits(runtime_limits),
crank_core::AdmissionSchedulerConfig::default(),
)?);
let tasks = std::sync::Arc::new(crank_community_mcp::TaskApplication::new(
registry.clone(),
Duration::from_millis(config.refresh_ms),
cache_stores.coordination.clone(),
crank_runtime::TaskService::with_secret_crypto(
task_store.clone(),
std::sync::Arc::new(secret_crypto.clone()),
),
));
// Story 4.3: the durable Task worker runs in this process over the same PostgreSQL store,
// published catalog, executor and encrypted-input authority as the accepting application.
let task_handler =
std::sync::Arc::new(crank_community_mcp::TaskWorkerHandler::from_application(
&tasks,
registry.clone(),
runtime.clone(),
secret_crypto.clone(),
));
let task_worker = std::sync::Arc::new(crank_runtime::TaskWorker::new(
task_store,
task_handler,
crank_runtime::TaskWorkerConfig {
owner: format!("mcp-server-{}", std::process::id()),
..crank_runtime::TaskWorkerConfig::default()
},
)?);
let (task_shutdown_tx, task_shutdown_rx) = tokio::sync::watch::channel(false);
let task_worker_handle = tokio::spawn({
let worker = task_worker.clone();
async move { worker.run_until_shutdown(task_shutdown_rx).await }
});
let app = build_app_with_artifacts_and_tasks(
registry,
Duration::from_millis(config.refresh_ms),
config.runtime.base_url.clone(),
@@ -211,6 +256,7 @@ async fn run(
std::sync::Arc::new(CommunityMachineCredentialVerifier),
runtime_limits.max_concurrent_sessions,
artifact_store,
tasks,
);
let listener = TcpListener::bind(config.bind_addr).await?;
@@ -231,15 +277,77 @@ async fn run(
info!(name: "mcp.server.listening", bind_address = %config.bind_addr, "mcp-server listening");
*startup_completed = true;
if let Some(metrics_server) = metrics_server {
tokio::select! {
result = axum::serve(listener, app) => result?,
result = metrics_server.serve() => result?,
let serve_result: Result<(), Box<dyn std::error::Error>> = async {
if let Some(metrics_server) = metrics_server {
tokio::select! {
result = serve_with_worker_shutdown(app, listener, shutdown_signal(), task_shutdown_tx.clone()) => result?,
result = metrics_server.serve() => result?,
}
} else {
serve_with_worker_shutdown(app, listener, shutdown_signal(), task_shutdown_tx.clone())
.await?;
}
} else {
axum::serve(listener, app).await?;
Ok(())
}
.await;
// Stop new claims; active attempts checkpoint/drain or are cancelled inside the bounded worker
// shutdown timeout (<=30s), and queued PostgreSQL work is left for recovery.
task_shutdown_tx.send(true).ok();
if tokio::time::timeout(Duration::from_secs(30), task_worker_handle)
.await
.is_err()
{
tracing::warn!(
name: "mcp.task_worker.shutdown_timeout",
"durable Task worker did not stop within the story shutdown budget"
);
}
serve_result
}
/// Gracefully stop the durable Task worker as soon as shutdown is requested, while still
/// draining in-flight HTTP connections before returning.
async fn serve_with_worker_shutdown<S>(
app: axum::Router,
listener: tokio::net::TcpListener,
signal: S,
worker_stop: tokio::sync::watch::Sender<bool>,
) -> std::io::Result<()>
where
S: std::future::Future<Output = ()> + Send + 'static,
{
let worker_stop_graceful = worker_stop.clone();
let graceful = async move {
signal.await;
let _ = worker_stop_graceful.send(true);
};
let result = axum::serve(listener, app)
.with_graceful_shutdown(graceful)
.await;
let _ = worker_stop.send(true);
result
}
/// Process shutdown trigger for the HTTP surface and the durable Task worker.
async fn shutdown_signal() {
let ctrl_c = async {
let _ = tokio::signal::ctrl_c().await;
};
#[cfg(unix)]
let terminate = async {
if let Ok(mut signal) =
tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate())
{
signal.recv().await;
}
};
#[cfg(not(unix))]
let terminate = std::future::pending::<()>();
tokio::select! {
_ = ctrl_c => {}
_ = terminate => {}
}
Ok(())
}
fn init_observability(
@@ -287,6 +395,25 @@ fn init_observability(
)?)
}
fn admission_limits(runtime_limits: RuntimeLimits) -> AdmissionLimits {
AdmissionLimits {
global_limit: runtime_limits.max_concurrent_unary,
..AdmissionLimits::default()
}
}
fn execution_admission(
pool: sqlx::PgPool,
runtime_limits: RuntimeLimits,
) -> Result<std::sync::Arc<dyn crank_core::ExecutionAdmission>, Box<dyn std::error::Error>> {
let store = PostgresExecutionAdmission::new(
pool,
admission_limits(runtime_limits),
crank_core::AdmissionSchedulerConfig::default(),
)?;
Ok(std::sync::Arc::new(store))
}
fn preflight_config(config: &McpProcessConfig) -> Result<(), crank_config::ConfigError> {
let invalid = |field| crank_config::ConfigError::single(DiagnosticCode::InvalidType, field);
database_options(&config.database).map_err(|_| invalid("database.source"))?;
@@ -307,6 +434,12 @@ fn preflight_config(config: &McpProcessConfig) -> Result<(), crank_config::Confi
config.runtime.max_concurrent_sessions,
)
.map_err(|_| invalid("runtime.limits"))?;
crank_registry::AdmissionLimits {
global_limit: config.runtime.max_concurrent_unary,
..crank_registry::AdmissionLimits::default()
}
.validate()
.map_err(|_| invalid("runtime.admission"))?;
runtime_cache_config(config).map_err(|_| invalid("cache"))?;
RequestRateLimitConfig::new(
config.rate_limit.requests_per_second,
@@ -460,3 +593,52 @@ fn runtime_cache_config(
.map(|value| value.expose_secret().to_owned()),
)
}
#[cfg(test)]
mod tests {
use super::*;
use axum::{Router, routing::get};
use std::time::Duration;
use tokio::sync::{oneshot, watch};
async fn slow_handler() -> &'static str {
tokio::time::sleep(Duration::from_millis(600)).await;
"ok"
}
#[tokio::test]
async fn worker_stops_before_http_drain_completes() {
let app = Router::new().route("/slow", get(slow_handler));
let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
let addr = listener.local_addr().unwrap();
let (stop_tx, mut stop_rx) = watch::channel(false);
let (signal_tx, signal_rx) = oneshot::channel::<()>();
let handle = tokio::spawn(serve_with_worker_shutdown(
app,
listener,
async {
let _ = signal_rx.await;
},
stop_tx,
));
let client = reqwest::Client::new();
let request = client.get(format!("http://{}/slow", addr));
let in_flight = tokio::spawn(async move { request.send().await });
tokio::time::sleep(Duration::from_millis(150)).await;
let _ = signal_tx.send(());
tokio::time::timeout(Duration::from_millis(250), stop_rx.changed())
.await
.expect("worker stop signal should arrive before HTTP drain")
.expect("watch changed");
assert!(*stop_rx.borrow());
let response = in_flight.await.unwrap().unwrap();
assert!(response.status().is_success());
let serve_result = handle.await.unwrap();
assert!(serve_result.is_ok());
}
}
+14
View File
@@ -11,3 +11,17 @@ pub fn spawn_postgres_pool_metrics(pool: PgPool) -> tokio::task::JoinHandle<()>
}
})
}
/// Story 4.1: samples the authoritative admission queue so the oldest-waiting age grows (and returns
/// to zero) without waiting for a new request.
pub fn spawn_admission_metrics(pool: PgPool) -> tokio::task::JoinHandle<()> {
tokio::spawn(async move {
let mut interval = tokio::time::interval(Duration::from_secs(5));
loop {
interval.tick().await;
if let Ok(snapshot) = crank_registry::admission_snapshot(&pool).await {
crank_observability::record_admission_oldest_waiting(snapshot.oldest_waiting_ms);
}
}
})
}
+874
View File
@@ -0,0 +1,874 @@
//! ATDD Story 4.1 (FR-21/FR-24): the runtime asks one authoritative store for admission.
//!
//! Real PostgreSQL state, two independent `RuntimeExecutor` instances (the multi-process shape) and a
//! counted adapter: one shared global budget, typed overload with retry guidance, typed unsupported
//! fallback, deadline that includes queue waiting and zero outbound calls on every refusal.
//!
//! Lives in the `mcp-server` app crate because the runtime must not depend on the registry crate; this
//! app already wires both and this is the real deployment shape.
use std::{
collections::BTreeMap,
sync::{
Arc, Mutex,
atomic::{AtomicUsize, Ordering},
},
time::{Duration, Instant},
};
use async_trait::async_trait;
use crank_core::{
AdapterResponse, AdmissionMode, AdmissionRequest, AdmissionScope, ApprovalEvidence,
ConfirmationPolicy, ExecutionAdmission, ExecutionConfig, ExecutionErrorCode, ExecutionMode,
ExecutionOrigin, ExecutionStage, HttpMethod, InvocationSource, NoTasksFallback, Operation,
OperationId, OperationSafetyClass, OperationSafetyPolicy, OperationSecurityLevel,
OperationStatus, OutcomeCertainty, Protocol, ProtocolAdapter, RestTarget,
RuntimeRequestContext, Target, ToolDescription, ToolExecutionMode, ToolExecutionPolicy,
WorkspaceId,
};
use crank_mapping::MappingSet;
use crank_registry::{AdmissionLimits, PostgresExecutionAdmission};
use crank_runtime::{RuntimeExecutor, RuntimeExecutorBuilder, RuntimeLimits};
use crank_schema::{Schema, SchemaKind};
use serde_json::json;
use sqlx::PgPool;
use time::OffsetDateTime;
struct CountingAdapter {
calls: Arc<AtomicUsize>,
current: Arc<AtomicUsize>,
peak: Arc<AtomicUsize>,
}
#[async_trait]
impl ProtocolAdapter for CountingAdapter {
fn protocol(&self) -> Protocol {
Protocol::Rest
}
fn supports_mode(&self, mode: ExecutionMode) -> bool {
mode == ExecutionMode::Unary
}
async fn invoke_unary(
&self,
_target: &Target,
_prepared: &crank_core::PreparedRequest,
context: &RuntimeRequestContext,
) -> Result<AdapterResponse, crank_core::ProtocolAdapterError> {
context.mark_dispatch_started();
self.calls.fetch_add(1, Ordering::SeqCst);
let current = self.current.fetch_add(1, Ordering::SeqCst) + 1;
self.peak.fetch_max(current, Ordering::SeqCst);
tokio::time::sleep(Duration::from_millis(150)).await;
self.current.fetch_sub(1, Ordering::SeqCst);
Ok(AdapterResponse {
status_code: 200,
headers: BTreeMap::new(),
body: json!({"ok": true}),
data: json!({"ok": true}),
})
}
}
fn schema(kind: SchemaKind) -> Schema {
Schema {
kind,
format: None,
description: None,
required: false,
nullable: false,
default_value: None,
fields: BTreeMap::new(),
items: None,
enum_values: Vec::new(),
variants: Vec::new(),
}
}
fn operation(policy: Option<ToolExecutionPolicy>) -> Operation<Schema, MappingSet> {
Operation {
id: OperationId::new("op_admission_budget"),
name: "admission_budget".to_owned(),
display_name: "Admission budget".to_owned(),
category: "general".to_owned(),
protocol: Protocol::Rest,
security_level: OperationSecurityLevel::Standard,
status: OperationStatus::Published,
version: 7,
target: Target::Rest(RestTarget {
base_url: "https://api.example.com".to_owned(),
method: HttpMethod::Get,
path_template: "/v1/items".to_owned(),
static_headers: BTreeMap::new(),
request_contract: None,
response_contract: Vec::new(),
multipart_field_mime_types: BTreeMap::new(),
}),
input_schema: schema(SchemaKind::Object),
output_schema: schema(SchemaKind::Object),
input_mapping: MappingSet::default(),
output_mapping: MappingSet::default(),
execution_config: ExecutionConfig {
timeout_ms: 5_000,
execution_policy: policy,
..ExecutionConfig::default()
},
tool_description: ToolDescription {
title: "Admission".to_owned(),
description: "Admission budget probe".to_owned(),
tags: Vec::new(),
examples: Vec::new(),
},
samples: None,
generated_draft: None,
config_export: None,
wizard_state: None,
created_at: OffsetDateTime::now_utc(),
updated_at: OffsetDateTime::now_utc(),
published_at: None,
}
}
async fn execute(
executor: &RuntimeExecutor,
operation: &crank_runtime::RuntimeOperation,
input: &serde_json::Value,
) -> Result<serde_json::Value, crank_core::ExecutionFailure> {
let workspace = WorkspaceId::new("ws_test");
let context = crank_runtime::RuntimeRequestContext::from_correlation(
&crank_core::CorrelationContext::generate(),
)
.with_metering_context(workspace.clone(), None, InvocationSource::AdminTestRun);
execute_in(executor, operation, input, &context).await
}
async fn execute_in(
executor: &RuntimeExecutor,
operation: &crank_runtime::RuntimeOperation,
input: &serde_json::Value,
context: &crank_runtime::RuntimeRequestContext,
) -> Result<serde_json::Value, crank_core::ExecutionFailure> {
let workspace = context
.metering_context()
.map(|metering| metering.workspace_id.clone())
.unwrap_or_else(|| WorkspaceId::new("ws_test"));
let request = crank_runtime::RuntimeExecutionRequest::try_new(
&workspace,
ExecutionOrigin::AdminDraft,
None,
operation,
input,
crank_runtime::ExecutionAuthorization::Authorized,
None,
context,
Instant::now() + Duration::from_secs(30),
)
.map_err(|_| {
crank_core::ExecutionFailure::new(
ExecutionErrorCode::RuntimeInternal,
crank_core::CorrelationContext::new(
context.request_id.clone(),
context.trace_context.clone(),
),
)
})?;
executor
.execute_outcome(request)
.await
.map(|success| success.output)
}
async fn schema_pool(schema_url: &str) -> PgPool {
let pool = PgPool::connect(schema_url)
.await
.expect("admission schema pool");
crank_registry::MigrationAuthority::apply(&pool)
.await
.expect("migration authority applies the admission schema");
pool
}
fn admission_store(pool: PgPool, global: usize, queue: usize) -> Arc<PostgresExecutionAdmission> {
Arc::new(
PostgresExecutionAdmission::new(
pool,
AdmissionLimits {
global_limit: global,
agent_limit: 8,
tool_limit: 8,
queue_limit: queue,
},
crank_core::AdmissionSchedulerConfig::default(),
)
.expect("valid admission limits"),
)
}
fn executor(
store: Arc<PostgresExecutionAdmission>,
calls: Arc<AtomicUsize>,
current: Arc<AtomicUsize>,
peak: Arc<AtomicUsize>,
) -> RuntimeExecutor {
RuntimeExecutorBuilder::new()
.register_adapter(Arc::new(CountingAdapter {
calls,
current,
peak,
}))
.with_limits(RuntimeLimits::try_new(8, 2).expect("local limits"))
.with_admission(store)
.build()
}
#[tokio::test]
async fn two_executors_share_one_global_budget() {
let schema_url = crank_test_support::postgres_schema_url("admission_budget").await;
let pool = schema_pool(&schema_url).await;
let store = admission_store(pool.clone(), 2, 8);
let calls = Arc::new(AtomicUsize::new(0));
let current = Arc::new(AtomicUsize::new(0));
let peak = Arc::new(AtomicUsize::new(0));
let first = executor(
Arc::clone(&store),
Arc::clone(&calls),
Arc::clone(&current),
Arc::clone(&peak),
);
let second = executor(
Arc::clone(&store),
Arc::clone(&calls),
Arc::clone(&current),
Arc::clone(&peak),
);
let operation: crank_runtime::RuntimeOperation = operation(None).into();
let mut tasks = Vec::new();
for index in 0..4 {
let executor = if index % 2 == 0 {
first.clone()
} else {
second.clone()
};
let operation = operation.clone();
tasks.push(tokio::spawn(async move {
execute(&executor, &operation, &json!({})).await
}));
}
let mut completed = 0;
for task in tasks {
match task.await.expect("execution task") {
Ok(_) => completed += 1,
Err(failure) => panic!("admission must grant every request: {failure:?}"),
}
}
assert_eq!(completed, 4);
assert_eq!(calls.load(Ordering::SeqCst), 4);
assert!(
peak.load(Ordering::SeqCst) <= 2,
"peak concurrency {} exceeded the shared global limit",
peak.load(Ordering::SeqCst)
);
assert_eq!(store.snapshot().await.unwrap().executing, 0);
}
#[tokio::test]
async fn saturation_returns_typed_overload_without_an_outbound_call() {
let schema_url = crank_test_support::postgres_schema_url("admission_overload").await;
let pool = schema_pool(&schema_url).await;
let store = admission_store(pool, 1, 1);
let calls = Arc::new(AtomicUsize::new(0));
let current = Arc::new(AtomicUsize::new(0));
let peak = Arc::new(AtomicUsize::new(0));
let runtime = executor(
Arc::clone(&store),
Arc::clone(&calls),
Arc::clone(&current),
Arc::clone(&peak),
);
// Another process already holds the only global permit.
let held = match store
.admit(AdmissionRequest {
scope: AdmissionScope {
workspace_id: "ws_test".to_owned(),
agent_key: "admin".to_owned(),
operation_id: "op_admission_budget".to_owned(),
operation_version: 7,
},
mode: AdmissionMode::Synchronous,
policy_mode: ToolExecutionMode::Synchronous,
no_tasks_fallback: NoTasksFallback::Synchronous,
approval: ApprovalEvidence::NotRequired,
fingerprint: "fp_holder".to_owned(),
expires_at_unix_ms: crank_core::current_unix_ms() + 10_000,
})
.await
.expect("admission store is reachable")
{
crank_core::AdmissionOutcome::Permitted(permit) => permit,
other => panic!("expected a permit, got {other:?}"),
};
let waiting_operation: crank_runtime::RuntimeOperation = operation(Some(ToolExecutionPolicy {
mode: ToolExecutionMode::Synchronous,
deadline_ms: 250,
no_tasks_fallback: NoTasksFallback::Synchronous,
}))
.into();
let waiting_runtime = runtime.clone();
let first_waiting_operation = waiting_operation.clone();
let waiting = tokio::spawn(async move {
execute(&waiting_runtime, &first_waiting_operation, &json!({})).await
});
// Let the first request occupy the bounded queue before the next one arrives.
while store.snapshot().await.unwrap().waiting == 0 {
tokio::time::sleep(Duration::from_millis(5)).await;
}
let overloaded = execute(&runtime, &waiting_operation, &json!({})).await;
let failure = overloaded.expect_err("a full queue must refuse the request");
assert_eq!(
failure.error_code(),
ExecutionErrorCode::ExecutionOverloaded
);
assert_eq!(failure.stage(), ExecutionStage::Admission);
assert_eq!(failure.retry_after_ms(), Some(250));
// The queued request hits its absolute deadline instead of dispatching. That refusal happened
// inside admission (nothing was ever dispatched), so the client must see an admission-stage
// overload with retry guidance — not an `upstream_timeout`.
let expired = waiting
.await
.expect("waiting task")
.expect_err("deadline must refuse");
assert_eq!(
expired.error_code(),
ExecutionErrorCode::ExecutionOverloaded,
"queue waiting counts against the published deadline"
);
assert_eq!(expired.stage(), ExecutionStage::Admission);
assert_eq!(expired.retry_after_ms(), Some(250));
assert_eq!(expired.outcome_certainty(), OutcomeCertainty::Certain);
assert_eq!(
calls.load(Ordering::SeqCst),
0,
"no outbound call on refusal"
);
assert!(store.release(&held).await.unwrap());
}
#[tokio::test]
async fn background_policy_without_a_task_transport_is_typed_unsupported() {
let schema_url = crank_test_support::postgres_schema_url("admission_unsupported").await;
let pool = schema_pool(&schema_url).await;
let store = admission_store(pool, 2, 4);
let calls = Arc::new(AtomicUsize::new(0));
let current = Arc::new(AtomicUsize::new(0));
let peak = Arc::new(AtomicUsize::new(0));
let runtime = executor(
Arc::clone(&store),
Arc::clone(&calls),
Arc::clone(&current),
Arc::clone(&peak),
);
let operation: crank_runtime::RuntimeOperation = operation(Some(ToolExecutionPolicy {
mode: ToolExecutionMode::Background,
deadline_ms: 2_000,
no_tasks_fallback: NoTasksFallback::Unsupported,
}))
.into();
let failure = execute(&runtime, &operation, &json!({}))
.await
.expect_err("no Task transport exists in Story 4.1");
assert_eq!(
failure.error_code(),
ExecutionErrorCode::ExecutionModeUnsupported
);
assert_eq!(calls.load(Ordering::SeqCst), 0);
let snapshot = store.snapshot().await.unwrap();
assert_eq!(snapshot.waiting, 0);
assert_eq!(snapshot.executing, 0);
}
#[tokio::test]
async fn background_policy_with_a_synchronous_fallback_keeps_the_sync_response_shape() {
let schema_url = crank_test_support::postgres_schema_url("admission_fallback").await;
let pool = schema_pool(&schema_url).await;
let store = admission_store(pool, 2, 4);
let calls = Arc::new(AtomicUsize::new(0));
let current = Arc::new(AtomicUsize::new(0));
let peak = Arc::new(AtomicUsize::new(0));
let runtime = executor(
Arc::clone(&store),
Arc::clone(&calls),
Arc::clone(&current),
Arc::clone(&peak),
);
let operation: crank_runtime::RuntimeOperation = operation(Some(ToolExecutionPolicy {
mode: ToolExecutionMode::Adaptive,
deadline_ms: 2_000,
no_tasks_fallback: NoTasksFallback::Synchronous,
}))
.into();
let output = execute(&runtime, &operation, &json!({}))
.await
.expect("adaptive resolves through the published synchronous fallback");
// The response shape is unchanged by admission: an operation without output mapping still
// returns the same mapped output object it returned before Story 4.1.
assert_eq!(output, json!({}));
assert_eq!(calls.load(Ordering::SeqCst), 1);
assert_eq!(store.snapshot().await.unwrap().executing, 0);
}
#[tokio::test]
async fn admission_store_outage_fails_closed_before_dispatch() {
let schema_url = crank_test_support::postgres_schema_url("admission_outage").await;
let pool = schema_pool(&schema_url).await;
let store = admission_store(pool.clone(), 2, 4);
let calls = Arc::new(AtomicUsize::new(0));
let current = Arc::new(AtomicUsize::new(0));
let peak = Arc::new(AtomicUsize::new(0));
let runtime = executor(
Arc::clone(&store),
Arc::clone(&calls),
Arc::clone(&current),
Arc::clone(&peak),
);
// The mandatory state disappears after the store was built: every execution must fail closed.
sqlx::query("drop table execution_admission_entries")
.execute(&pool)
.await
.expect("drop the mandatory admission state");
let operation: crank_runtime::RuntimeOperation = operation(None).into();
let failure = execute(&runtime, &operation, &json!({}))
.await
.expect_err("a missing mandatory store must fail closed");
assert_eq!(
failure.error_code(),
ExecutionErrorCode::SafetyStoreUnavailable
);
assert_eq!(calls.load(Ordering::SeqCst), 0);
assert!(store.health().await.is_err());
}
/// Admission store that records the snapshot it was asked to decide on (Story 4.1, AC2/F2).
struct RecordingAdmission {
inner: Arc<PostgresExecutionAdmission>,
requests: Mutex<Vec<AdmissionRequest>>,
permits: AtomicUsize,
releases: AtomicUsize,
}
impl RecordingAdmission {
fn new(inner: Arc<PostgresExecutionAdmission>) -> Self {
Self {
inner,
requests: Mutex::new(Vec::new()),
permits: AtomicUsize::new(0),
releases: AtomicUsize::new(0),
}
}
fn requests(&self) -> Vec<AdmissionRequest> {
self.requests.lock().expect("admission log").clone()
}
fn permits(&self) -> usize {
self.permits.load(Ordering::SeqCst)
}
fn releases(&self) -> usize {
self.releases.load(Ordering::SeqCst)
}
}
#[async_trait]
impl ExecutionAdmission for RecordingAdmission {
async fn health(&self) -> Result<(), crank_core::AdmissionStoreError> {
self.inner.health().await
}
async fn admit(
&self,
request: AdmissionRequest,
) -> Result<crank_core::AdmissionOutcome, crank_core::AdmissionStoreError> {
self.requests
.lock()
.expect("admission log")
.push(request.clone());
let outcome = self.inner.admit(request).await?;
if matches!(outcome, crank_core::AdmissionOutcome::Permitted(_)) {
self.permits.fetch_add(1, Ordering::SeqCst);
}
Ok(outcome)
}
async fn release(
&self,
permit: &crank_core::AdmissionPermit,
) -> Result<bool, crank_core::AdmissionStoreError> {
let released = self.inner.release(permit).await?;
if released {
self.releases.fetch_add(1, Ordering::SeqCst);
}
Ok(released)
}
async fn cancel(
&self,
ticket_id: u64,
owner_token: &str,
) -> Result<bool, crank_core::AdmissionStoreError> {
self.inner.cancel(ticket_id, owner_token).await
}
async fn snapshot(
&self,
) -> Result<crank_core::AdmissionQueueSnapshot, crank_core::AdmissionStoreError> {
self.inner.snapshot().await
}
}
fn confirmation_operation(version: u32) -> crank_runtime::RuntimeOperation {
let mut source = operation(None);
source.id = OperationId::new("op_admission_confirmation");
source.name = "admission_confirmation".to_owned();
source.version = version;
let Target::Rest(target) = &mut source.target;
target.method = HttpMethod::Delete;
source.execution_config.safety = Some(OperationSafetyPolicy {
class: OperationSafetyClass::Destructive,
confirmation: Some(ConfirmationPolicy { ttl_ms: 60_000 }),
});
source.into()
}
fn confirmation_runtime(
admission: Arc<RecordingAdmission>,
coordination: Option<Arc<dyn crank_core::CoordinationStateStore>>,
calls: Arc<AtomicUsize>,
) -> RuntimeExecutor {
let store: Arc<dyn ExecutionAdmission> = admission;
let builder = RuntimeExecutorBuilder::new()
.register_adapter(Arc::new(CountingAdapter {
calls,
current: Arc::new(AtomicUsize::new(0)),
peak: Arc::new(AtomicUsize::new(0)),
}))
.with_limits(RuntimeLimits::try_new(8, 2).expect("local limits"))
.with_admission(store);
match coordination {
Some(coordination) => builder.with_coordination_store(coordination).build(),
None => builder.build(),
}
}
fn confirmation_context() -> crank_runtime::RuntimeRequestContext {
crank_runtime::RuntimeRequestContext::from_correlation(
&crank_core::CorrelationContext::generate(),
)
.with_response_cache_scope("ws_test", "agent_a")
.with_metering_context(
WorkspaceId::new("ws_test"),
None,
InvocationSource::AdminTestRun,
)
}
#[tokio::test]
async fn confirmation_requiring_call_takes_a_permit_only_with_verified_approval() {
let schema_url = crank_test_support::postgres_schema_url("admission_confirmation").await;
let pool = schema_pool(&schema_url).await;
let recording = Arc::new(RecordingAdmission::new(Arc::new(
PostgresExecutionAdmission::new(
pool,
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
crank_core::AdmissionSchedulerConfig::default(),
)
.expect("valid admission limits"),
)));
let coordination: Arc<dyn crank_core::CoordinationStateStore> =
Arc::new(crank_runtime::InMemoryCoordinationStateStore::default());
let calls = Arc::new(AtomicUsize::new(0));
let runtime = confirmation_runtime(
Arc::clone(&recording),
Some(Arc::clone(&coordination)),
Arc::clone(&calls),
);
let operation = confirmation_operation(3);
let context = confirmation_context();
let input = json!({ "reason": "remove the stale item" });
// 1. No approval yet: the existing challenge is returned and no execution capacity is taken.
let challenge = execute_in(&runtime, &operation, &input, &context)
.await
.expect_err("an unapproved destructive call must only stage confirmation");
assert_eq!(
challenge.error_code(),
ExecutionErrorCode::ConfirmationRequired
);
let token = challenge
.confirmation()
.expect("expected the confirmation challenge")
.token()
.to_owned();
assert_eq!(
recording.permits(),
0,
"an unapproved request must not take an execution permit"
);
assert_eq!(calls.load(Ordering::SeqCst), 0);
let requests = recording.requests();
assert_eq!(
requests.len(),
1,
"the one admission snapshot is consulted once"
);
assert_eq!(
requests[0].approval,
ApprovalEvidence::NotGranted,
"admission must see the real approval state, not a fixed NotRequired"
);
assert_eq!(requests[0].scope.operation_version, 3);
let snapshot = recording.snapshot().await.expect("admission snapshot");
assert_eq!((snapshot.executing, snapshot.waiting), (0, 0));
// 2. An approval issued for another published version never opens this execution.
let other_version = confirmation_operation(4);
let foreign = execute_in(
&runtime,
&other_version,
&input,
&context.clone().with_confirmation_token(token.clone()),
)
.await
.expect_err("a foreign approval must not authorise this version");
assert_eq!(
foreign.error_code(),
ExecutionErrorCode::ConfirmationInvalid
);
assert_eq!(recording.permits(), 0);
assert_eq!(
recording.requests().len(),
1,
"an unusable approval is refused before any capacity request"
);
// 3. An approval is bound to its exact input as well.
let mismatched_input = json!({ "reason": "remove another item" });
let mismatched = execute_in(
&runtime,
&operation,
&mismatched_input,
&context.clone().with_confirmation_token(token.clone()),
)
.await
.expect_err("a confirmation token must not authorise different input");
assert_eq!(
mismatched.error_code(),
ExecutionErrorCode::ConfirmationInvalid
);
assert_eq!(recording.permits(), 0);
assert_eq!(recording.requests().len(), 1);
// 4. The matching approval passes admission with truthful evidence and executes exactly once.
let executed = execute_in(
&runtime,
&operation,
&input,
&context.clone().with_confirmation_token(token.clone()),
)
.await
.expect("a verified approval executes");
assert!(executed.is_object());
assert_eq!(calls.load(Ordering::SeqCst), 1);
let requests = recording.requests();
assert_eq!(requests.len(), 2);
assert_eq!(requests[1].approval, ApprovalEvidence::Granted);
assert_eq!(requests[1].scope.operation_version, 3);
assert_eq!(recording.permits(), 1);
assert_eq!(
recording.releases(),
1,
"the permit is released after execution"
);
assert_eq!(recording.snapshot().await.expect("snapshot").executing, 0);
// 5. Replay of the consumed confirmation is refused before capacity and without a side effect.
let replay = execute_in(
&runtime,
&operation,
&input,
&context.clone().with_confirmation_token(token),
)
.await
.expect_err("a confirmation token is single-use");
assert_eq!(replay.error_code(), ExecutionErrorCode::ConfirmationInvalid);
assert_eq!(calls.load(Ordering::SeqCst), 1);
assert_eq!(recording.permits(), 1);
assert_eq!(
recording.requests().len(),
2,
"a replayed approval never reaches admission"
);
// 6. The trusted approved-recovery path still passes the same gate with real evidence.
let approved = execute_in(
&runtime,
&operation,
&json!({ "reason": "approved by an operator" }),
&context.with_approval_granted(),
)
.await;
assert!(approved.is_ok());
assert_eq!(calls.load(Ordering::SeqCst), 2);
let requests = recording.requests();
assert_eq!(requests.len(), 3);
assert_eq!(requests[2].approval, ApprovalEvidence::Granted);
assert_eq!(recording.permits(), 2);
assert_eq!(recording.snapshot().await.expect("snapshot").executing, 0);
}
#[tokio::test]
async fn mandatory_safety_store_outage_fails_closed_before_admission() {
let schema_url = crank_test_support::postgres_schema_url("admission_confirmation_outage").await;
let pool = schema_pool(&schema_url).await;
let recording = Arc::new(RecordingAdmission::new(Arc::new(
PostgresExecutionAdmission::new(
pool,
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
crank_core::AdmissionSchedulerConfig::default(),
)
.expect("valid admission limits"),
)));
let calls = Arc::new(AtomicUsize::new(0));
// Confirmation is mandatory for this Tool, but no safety store is wired: the request must fail
// closed instead of reserving capacity and hoping for an in-memory fallback.
let runtime = confirmation_runtime(Arc::clone(&recording), None, Arc::clone(&calls));
let operation = confirmation_operation(3);
let failure = execute_in(
&runtime,
&operation,
&json!({ "reason": "no store" }),
&confirmation_context(),
)
.await
.expect_err("a mandatory safety store outage must fail closed");
assert_eq!(
failure.error_code(),
ExecutionErrorCode::SafetyStoreUnavailable
);
assert_eq!(calls.load(Ordering::SeqCst), 0);
assert_eq!(recording.permits(), 0);
let after_missing_store = recording.requests().len();
// The same Tool with a claimed approval and a *failing* safety store never reaches the capacity
// store at all: the approval claim stays unverified, so no permit and no outbound call exist.
let failing = confirmation_runtime(
Arc::clone(&recording),
Some(Arc::new(UnavailableSafetyStore)),
Arc::clone(&calls),
);
let outage = execute_in(
&failing,
&operation,
&json!({ "reason": "failing store" }),
&confirmation_context().with_confirmation_token("ct_unavailable"),
)
.await
.expect_err("an unreachable safety store must refuse the claimed approval");
assert_eq!(
outage.error_code(),
ExecutionErrorCode::SafetyStoreUnavailable
);
assert_eq!(
recording.requests().len(),
after_missing_store,
"an unverified approval never becomes a capacity request"
);
assert_eq!(recording.permits(), 0);
assert_eq!(calls.load(Ordering::SeqCst), 0);
}
/// Coordination store whose mandatory backend is down: every operation fails closed.
struct UnavailableSafetyStore;
impl UnavailableSafetyStore {
fn error() -> crank_core::CacheStoreError {
crank_core::CacheStoreError::Unavailable {
message: "test backend unavailable".to_owned(),
}
}
}
#[async_trait]
impl crank_core::CoordinationStateStore for UnavailableSafetyStore {
async fn get_value(
&self,
_scope: crank_core::CacheScope,
_key: &str,
) -> Result<Option<crank_core::CoordinationStateValue>, crank_core::CacheStoreError> {
Err(Self::error())
}
async fn put_value(
&self,
_scope: crank_core::CacheScope,
_key: &str,
_value: crank_core::CoordinationStateValue,
_ttl: Duration,
) -> Result<(), crank_core::CacheStoreError> {
Err(Self::error())
}
async fn delete_value(
&self,
_scope: crank_core::CacheScope,
_key: &str,
) -> Result<(), crank_core::CacheStoreError> {
Err(Self::error())
}
async fn take_value(
&self,
_scope: crank_core::CacheScope,
_key: &str,
) -> Result<Option<crank_core::CoordinationStateValue>, crank_core::CacheStoreError> {
Err(Self::error())
}
async fn reserve_value(
&self,
_scope: crank_core::CacheScope,
_key: &str,
_value: crank_core::CoordinationStateValue,
_ttl: Duration,
) -> Result<crank_core::CoordinationStateReservation, crank_core::CacheStoreError> {
Err(Self::error())
}
async fn compare_and_set_value(
&self,
_scope: crank_core::CacheScope,
_key: &str,
_expected: &crank_core::CoordinationStateValue,
_value: crank_core::CoordinationStateValue,
_ttl: Duration,
) -> Result<bool, crank_core::CacheStoreError> {
Err(Self::error())
}
}
+1
View File
@@ -1,6 +1,7 @@
mod integration {
mod catalog_access;
mod common;
mod durable_tasks;
mod era_isolation;
mod execution_certainty;
mod execution_stages;
@@ -621,6 +621,7 @@ pub(super) fn test_operation(base_url: &str, name: &str) -> Operation<Schema, Ma
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
},
@@ -0,0 +1,555 @@
//! Story 4.4 ATDD: no outbound send exists without a committed durable attempt, and an elapsed
//! accepted deadline fails before dispatch.
//!
//! Both oracles drive the production composition (application entry → PG store → runtime executor →
//! real local HTTP upstream); only the upstream and the checkpoint failure are test-controlled.
use super::*;
use axum::Router;
use axum::http::StatusCode;
use axum::routing::post;
use crank_community_mcp::TaskWorkerHandler;
use crank_core::{
AgentOperationBinding, PlatformApiKeyScope, TaskState, ToolExecutionMode, ToolExecutionPolicy,
};
use crank_runtime::{SecretCrypto, TaskWorker, TaskWorkerConfig};
use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};
use time::OffsetDateTime;
async fn composition(
registry: &PostgresRegistry,
upstream: &str,
policy: ToolExecutionPolicy,
caller_name: &str,
scoped_key: &str,
) -> (
Arc<PostgresTaskStore>,
crank_community_mcp::TaskApplication,
SecretCrypto,
crank_community_mcp::TaskPrincipal,
) {
let mut operation = common::test_operation(upstream, TOOL_NAME);
operation.execution_config.execution_policy = Some(policy);
registry
.create_operation(&common::test_workspace_id(), &operation, Some("alice"))
.await
.unwrap();
publish(registry, &operation.id, 1).await;
common::publish_agent_with_bindings(
registry,
AGENT_SLUG,
vec![AgentOperationBinding {
operation_id: operation.id.clone(),
operation_version: 1,
tool_name: operation.name.clone(),
tool_title: operation.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
}],
)
.await;
let caller = principal(
registry,
caller_name,
&[PlatformApiKeyScope::Read, PlatformApiKeyScope::Write],
)
.await;
let crypto = SecretCrypto::new(TEST_MASTER_KEY).unwrap();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &OffsetDateTime::now_utc(),
})
.await
.expect("test master-key identity registered");
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration"),
);
let tasks = TaskApplication::new(
registry.clone(),
Duration::ZERO,
Arc::new(InMemoryCoordinationStateStore::default()),
TaskService::with_secret_crypto(store.clone(), Arc::new(crypto.clone())),
);
let _ = scoped_key;
(store, tasks, crypto, caller)
}
fn worker_for(
registry: &PostgresRegistry,
store: &Arc<PostgresTaskStore>,
tasks: &TaskApplication,
crypto: &SecretCrypto,
owner: &str,
) -> TaskWorker {
let runtime = crank_runtime::community_with_outbound_policy(
crank_runtime::OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
)
.build();
let handler = Arc::new(TaskWorkerHandler::from_application(
tasks,
registry.clone(),
runtime,
crypto.clone(),
));
let config = TaskWorkerConfig {
owner: owner.to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(50),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(5),
..TaskWorkerConfig::default()
};
TaskWorker::new(store.clone(), handler, config).expect("valid worker")
}
async fn state_of(registry: &PostgresRegistry, task_id: crank_core::TaskId) -> Option<String> {
sqlx::query_scalar("select state from tasks where task_id = $1")
.bind(task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap()
}
/// AC1: a dispatch checkpoint that cannot commit produces zero outbound calls and no attempt row.
#[tokio::test]
async fn checkpoint_commit_failure_never_reaches_upstream() {
let registry = common::test_registry().await;
let calls = Arc::new(AtomicUsize::new(0));
let upstream_app = Router::new().route(
"/crm/leads",
post({
let calls = Arc::clone(&calls);
move || {
let calls = Arc::clone(&calls);
async move {
calls.fetch_add(1, Ordering::SeqCst);
(StatusCode::OK, axum::Json(json!({"id": "lead-1"})))
}
}
}),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, upstream_app).await.unwrap() });
let upstream = format!("http://{address}");
let (store, tasks, crypto, caller) = composition(
&registry,
&upstream,
background_policy(),
"attempt-guards-checkpoint",
"guards-checkpoint-key",
)
.await;
let acceptance = tasks
.accept_background(
&caller,
TOOL_NAME,
None,
"guards-checkpoint-key",
&json!({"email": "user@example.com"}),
&unsampled_correlation(),
)
.await
.expect("durable acceptance");
// The checkpoint is the only statement that commits an attempt; it is refused here.
sqlx::raw_sql(sqlx::AssertSqlSafe(
"create function refuse_task_checkpoint() returns trigger language plpgsql as $$
begin
if new.current_attempt_id is not null and old.current_attempt_id is null then
raise exception 'durable checkpoint refused for test';
end if;
return new;
end $$;
create trigger refuse_task_checkpoint before update on tasks
for each row execute function refuse_task_checkpoint();",
))
.execute(registry.pool())
.await
.expect("checkpoint refusal trigger");
let worker = worker_for(
&registry,
&store,
&tasks,
&crypto,
"worker-guards-checkpoint",
);
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
tokio::time::sleep(Duration::from_secs(2)).await;
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
assert_eq!(
calls.load(Ordering::SeqCst),
0,
"an uncommitted durable checkpoint must never produce an outbound call"
);
let attempts: i64 =
sqlx::query_scalar("select count(*)::bigint from task_attempts where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap();
assert_eq!(
attempts, 0,
"no attempt history without a committed checkpoint"
);
assert_ne!(
state_of(&registry, acceptance.task_id).await.as_deref(),
Some("completed"),
"the refused attempt cannot complete the Task"
);
sqlx::raw_sql(sqlx::AssertSqlSafe(
"drop trigger refuse_task_checkpoint on tasks; drop function refuse_task_checkpoint();",
))
.execute(registry.pool())
.await
.expect("checkpoint refusal trigger removed");
server.abort();
}
/// AC6: an accepted absolute deadline that already elapsed fails the Task before any dispatch.
#[tokio::test]
async fn elapsed_deadline_fails_before_dispatch_without_upstream_call() {
let registry = common::test_registry().await;
let calls = Arc::new(AtomicUsize::new(0));
let upstream_app = Router::new().route(
"/crm/leads",
post({
let calls = Arc::clone(&calls);
move || {
let calls = Arc::clone(&calls);
async move {
calls.fetch_add(1, Ordering::SeqCst);
(StatusCode::OK, axum::Json(json!({"id": "lead-1"})))
}
}
}),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, upstream_app).await.unwrap() });
let upstream = format!("http://{address}");
let short_deadline = ToolExecutionPolicy {
mode: ToolExecutionMode::Background,
deadline_ms: 300,
no_tasks_fallback: crank_core::NoTasksFallback::Unsupported,
};
let (store, tasks, crypto, caller) = composition(
&registry,
&upstream,
short_deadline,
"attempt-guards-deadline",
"guards-deadline-key",
)
.await;
let acceptance = tasks
.accept_background(
&caller,
TOOL_NAME,
None,
"guards-deadline-key",
&json!({"email": "user@example.com"}),
&unsampled_correlation(),
)
.await
.expect("durable acceptance");
// The accepted deadline elapses before any worker claims the Task.
tokio::time::sleep(Duration::from_millis(500)).await;
let worker = worker_for(&registry, &store, &tasks, &crypto, "worker-guards-deadline");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
let mut state = None;
for _ in 0..100 {
tokio::time::sleep(Duration::from_millis(50)).await;
state = state_of(&registry, acceptance.task_id).await;
if state.as_deref() == Some("failed") {
break;
}
}
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
assert_eq!(state.as_deref(), Some("failed"));
assert_eq!(
calls.load(Ordering::SeqCst),
0,
"an elapsed deadline must fail before dispatch"
);
let attempts: i64 =
sqlx::query_scalar("select count(*)::bigint from task_attempts where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap();
assert_eq!(attempts, 0, "no attempt is checkpointed after the deadline");
let code: Option<String> =
sqlx::query_scalar("select error_envelope ->> 'code' from tasks where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap();
assert_eq!(code.as_deref(), Some("task_deadline_elapsed"));
assert_eq!(
sqlx::query_scalar::<_, String>("select state from tasks where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap(),
TaskState::Failed.as_str()
);
server.abort();
}
/// AC2: a write whose upstream idempotency is backed by server-owned evidence is retried with the
/// same stable key, in a bounded number of distinct durable attempts, and applies exactly one effect.
#[tokio::test]
async fn verified_idempotent_write_retries_with_stable_key_and_one_effect() {
use crank_community_mcp::VerifiedIdempotentWrite;
use crank_core::{IdempotencyMode, IdempotencyPolicy, RetryPolicy};
use std::collections::HashSet;
use std::sync::Mutex;
let registry = common::test_registry().await;
let requests = Arc::new(AtomicUsize::new(0));
let effects = Arc::new(AtomicUsize::new(0));
let keys = Arc::new(Mutex::new(Vec::<String>::new()));
let seen = Arc::new(Mutex::new(HashSet::<String>::new()));
let upstream_app = Router::new().route(
"/crm/leads",
post({
let requests = Arc::clone(&requests);
let effects = Arc::clone(&effects);
let keys = Arc::clone(&keys);
let seen = Arc::clone(&seen);
move |headers: axum::http::HeaderMap| {
let requests = Arc::clone(&requests);
let effects = Arc::clone(&effects);
let keys = Arc::clone(&keys);
let seen = Arc::clone(&seen);
async move {
let key = headers
.get("idempotency-key")
.and_then(|value| value.to_str().ok())
.unwrap_or_default()
.to_owned();
keys.lock().expect("keys").push(key.clone());
let first_application = seen.lock().expect("seen").insert(key);
if first_application {
effects.fetch_add(1, Ordering::SeqCst);
}
let attempt = requests.fetch_add(1, Ordering::SeqCst);
if attempt == 0 {
// The first application is stored but its acknowledgement is lost.
return (
StatusCode::INTERNAL_SERVER_ERROR,
axum::Json(json!({"error": "acknowledgement lost"})),
);
}
(
StatusCode::OK,
axum::Json(json!({"id": "lead-1", "deduped": !first_application})),
)
}
}
}),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, upstream_app).await.unwrap() });
let upstream = format!("http://{address}");
let mut operation = common::test_operation(&upstream, TOOL_NAME);
operation.execution_config.execution_policy = Some(background_policy());
operation.execution_config.retry_policy = Some(RetryPolicy { max_attempts: 2 });
operation.execution_config.idempotency = Some(IdempotencyPolicy {
mode: IdempotencyMode::Required,
ttl_ms: 60_000,
input_field: Some("idempotency_key".to_owned()),
header_name: Some("Idempotency-Key".to_owned()),
});
registry
.create_operation(&common::test_workspace_id(), &operation, Some("alice"))
.await
.unwrap();
publish(&registry, &operation.id, 1).await;
common::publish_agent_with_bindings(
&registry,
AGENT_SLUG,
vec![AgentOperationBinding {
operation_id: operation.id.clone(),
operation_version: 1,
tool_name: operation.name.clone(),
tool_title: operation.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
}],
)
.await;
let caller = principal(
&registry,
"attempt-guards-verified",
&[PlatformApiKeyScope::Read, PlatformApiKeyScope::Write],
)
.await;
let crypto = SecretCrypto::new(TEST_MASTER_KEY).unwrap();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &OffsetDateTime::now_utc(),
})
.await
.expect("test master-key identity registered");
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration"),
);
let tasks = TaskApplication::new(
registry.clone(),
Duration::ZERO,
Arc::new(InMemoryCoordinationStateStore::default()),
TaskService::with_secret_crypto(store.clone(), Arc::new(crypto.clone())),
);
let acceptance = tasks
.accept_background(
&caller,
TOOL_NAME,
None,
"guards-verified-key",
&json!({"email": "user@example.com", "idempotency_key": "k-1"}),
&unsampled_correlation(),
)
.await
.expect("durable acceptance");
let runtime = crank_runtime::community_with_outbound_policy(
crank_runtime::OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
)
.with_coordination_store(Arc::new(InMemoryCoordinationStateStore::default()))
.build();
let handler = Arc::new(
TaskWorkerHandler::from_application(&tasks, registry.clone(), runtime, crypto.clone())
.with_verified_writes(vec![VerifiedIdempotentWrite {
operation_id: operation.id.as_str().to_owned(),
operation_version: 1,
key_header: "Idempotency-Key".to_owned(),
validity_ms: 60_000,
}]),
);
let config = TaskWorkerConfig {
owner: "worker-guards-verified".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(50),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(5),
..TaskWorkerConfig::default()
};
let worker = TaskWorker::new(store.clone(), handler, config).expect("valid worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
let mut state = None;
for _ in 0..200 {
tokio::time::sleep(Duration::from_millis(50)).await;
state = state_of(&registry, acceptance.task_id).await;
if state.as_deref() == Some("completed") {
break;
}
}
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
server.abort();
assert_eq!(state.as_deref(), Some("completed"));
assert_eq!(
requests.load(Ordering::SeqCst),
2,
"exactly one retry of the write"
);
assert_eq!(
effects.load(Ordering::SeqCst),
1,
"the upstream applied the write exactly once"
);
assert_eq!(
keys.lock().expect("keys").as_slice(),
["k-1".to_owned(), "k-1".to_owned()],
"the same stable key is delivered on both attempts"
);
let attempts = store
.attempts(
&crank_core::TaskAccess {
workspace_id: common::test_workspace_id().to_string(),
owner_kind: crank_core::TaskOwnerKind::Agent,
owner_id: "agent_attempt_guards".to_owned(),
authorization_scope: String::new(),
},
acceptance.task_id,
)
.await;
// Owner scope of this principal is the agent of the published binding; read the history through
// the application entry instead of guessing the authorization scope string.
assert!(attempts.is_err(), "a foreign scope never reads the history");
let attempts = tasks
.attempts(&caller, TOOL_NAME, acceptance.task_id)
.await
.expect("owner-scoped attempt history");
assert_eq!(attempts.len(), 2, "two distinct durable attempts");
assert_eq!(attempts[0].ordinal, 1);
assert_eq!(attempts[1].ordinal, 2);
assert!(attempts.iter().all(|attempt| attempt.retry_allowed));
assert_eq!(
attempts[0].outcome,
Some(crank_core::AttemptOutcomeClass::RetryScheduled)
);
assert_eq!(
attempts[1].outcome,
Some(crank_core::AttemptOutcomeClass::Completed)
);
assert_eq!(
attempts[0].upstream_key_digest, attempts[1].upstream_key_digest,
"the same stable key digest is recorded on both attempts"
);
assert!(attempts[0].upstream_key_digest.is_some());
let record = tasks
.status(&caller, TOOL_NAME, acceptance.task_id)
.await
.expect("status");
assert_eq!(record.pins.retry.max_attempts, 2);
}
@@ -0,0 +1,432 @@
//! Story 4.3 ATDD: durable Attempt ID reaches the adapter attempt span (PF-1).
//!
//! End-to-end: a real durable Task worker dispatches a retrying GET through a real local
//! upstream, and the `adapter_attempt` spans carry the same durable Attempt ID that is
//! persisted in `tasks.current_attempt_id` on both attempts.
use super::super::common;
use axum::Router;
use axum::http::StatusCode;
use axum::routing::get;
use crank_community_mcp::TaskWorkerHandler;
use crank_core::{
AgentOperationBinding, CorrelationContext, RetryPolicy, Target, ToolExecutionPolicy,
};
use crank_runtime::{
OutboundHttpPolicy, SecretCrypto, TaskWorker, TaskWorkerConfig, community_with_outbound_policy,
};
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::{Arc, Mutex};
use time::OffsetDateTime;
use tracing::{Id, Subscriber, field::Visit};
use tracing_subscriber::{Layer, layer::SubscriberExt, registry::LookupSpan};
const AGENT_SLUG: &str = "reports";
const TOOL_NAME: &str = "attempt_trace";
const KEY: &str = "attempt-trace-key-1";
async fn publish(
registry: &crank_registry::PostgresRegistry,
operation_id: &crank_core::OperationId,
version: u32,
) {
registry
.publish_operation(crank_registry::PublishRequest {
workspace_id: &common::test_workspace_id(),
operation_id,
version,
published_at: &OffsetDateTime::parse(
"2026-03-26T10:00:00Z",
&time::format_description::well_known::Rfc3339,
)
.unwrap(),
published_by: Some("alice"),
})
.await
.unwrap();
}
async fn application_with_store(
registry: &crank_registry::PostgresRegistry,
) -> (
crank_community_mcp::TaskApplication,
std::sync::Arc<crank_registry::PostgresTaskStore>,
SecretCrypto,
) {
let crypto = SecretCrypto::new("test-master-key-00000000000000000000000000000000").unwrap();
let observed_at = OffsetDateTime::now_utc();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &observed_at,
})
.await
.expect("test master-key identity registered");
let store = std::sync::Arc::new(
crank_registry::PostgresTaskStore::new(
registry.pool().clone(),
crank_registry::AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration"),
);
let tasks = crank_community_mcp::TaskApplication::new(
registry.clone(),
std::time::Duration::ZERO,
std::sync::Arc::new(crank_runtime::InMemoryCoordinationStateStore::default()),
crank_runtime::TaskService::with_secret_crypto(
store.clone(),
std::sync::Arc::new(crypto.clone()),
),
);
(tasks, store, crypto)
}
async fn principal(
registry: &crank_registry::PostgresRegistry,
name: &str,
scopes: &[crank_core::PlatformApiKeyScope],
) -> crank_community_mcp::TaskPrincipal {
let token = common::create_platform_api_key(registry, AGENT_SLUG, name, scopes).await;
let secret_hash = common::hash_access_secret(&token);
let api_key = registry
.get_platform_api_key_by_secret_for_agent_slug(
common::test_workspace_slug(),
AGENT_SLUG,
&secret_hash,
)
.await
.unwrap()
.expect("an active platform key resolves");
registry
.touch_platform_api_key(
&api_key.api_key.workspace_id,
&api_key.api_key.id,
&OffsetDateTime::now_utc(),
)
.await
.unwrap();
crank_community_mcp::TaskPrincipal::Agent {
workspace_slug: common::test_workspace_slug().to_owned(),
agent_slug: AGENT_SLUG.to_owned(),
credential: crank_core::VerifiedMachineCredential {
machine_access_mode: crank_core::MachineAccessMode::StaticAgentKey,
max_security_level: crank_core::OperationSecurityLevel::Standard,
scopes: api_key.api_key.scopes.clone(),
platform_api_key_id: Some(api_key.api_key.id.clone()),
},
}
}
#[tokio::test]
async fn adapter_attempt_span_carries_durable_attempt_id() {
let registry = common::test_registry().await;
let calls = Arc::new(AtomicUsize::new(0));
let upstream_app = Router::new().route(
"/crm/leads",
get({
let calls = Arc::clone(&calls);
move || {
let calls = Arc::clone(&calls);
async move {
let n = calls.fetch_add(1, Ordering::SeqCst);
if n == 0 {
return (
StatusCode::INTERNAL_SERVER_ERROR,
axum::Json(serde_json::json!({"error":"boom"})),
);
}
(
StatusCode::OK,
axum::Json(serde_json::json!({"id":"lead-1"})),
)
}
}
}),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, upstream_app).await.unwrap() });
let upstream = format!("http://{address}");
let mut operation = common::test_operation(&upstream, TOOL_NAME);
let Target::Rest(target) = &mut operation.target;
target.method = crank_core::HttpMethod::Get;
operation.execution_config.execution_policy = Some(ToolExecutionPolicy {
mode: crank_core::ToolExecutionMode::Background,
deadline_ms: 60_000,
no_tasks_fallback: crank_core::NoTasksFallback::Unsupported,
});
operation.execution_config.retry_policy = Some(RetryPolicy { max_attempts: 2 });
registry
.create_operation(&common::test_workspace_id(), &operation, Some("alice"))
.await
.unwrap();
publish(&registry, &operation.id, 1).await;
common::publish_agent_with_bindings(
&registry,
AGENT_SLUG,
vec![AgentOperationBinding {
operation_id: operation.id.clone(),
operation_version: 1,
tool_name: operation.name.clone(),
tool_title: operation.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
}],
)
.await;
let caller = principal(
&registry,
"attempt-trace",
&[
crank_core::PlatformApiKeyScope::Read,
crank_core::PlatformApiKeyScope::Write,
],
)
.await;
let (tasks, store, crypto) = application_with_store(&registry).await;
let acceptance = tasks
.accept_background(
&caller,
TOOL_NAME,
None,
KEY,
&serde_json::json!({"email": "user@example.com"}),
&CorrelationContext::generate(),
)
.await
.expect("durable acceptance");
let runtime =
community_with_outbound_policy(OutboundHttpPolicy::allowing_hosts(["127.0.0.1"])).build();
let handler = Arc::new(TaskWorkerHandler::from_application(
&tasks,
registry.clone(),
runtime,
crypto,
));
let config = TaskWorkerConfig {
owner: "runtime-worker-attempt-trace".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: std::time::Duration::from_millis(50),
max_inflight: 1,
shutdown_timeout: std::time::Duration::from_secs(5),
..TaskWorkerConfig::default()
};
let worker = TaskWorker::new(store.clone(), handler, config).expect("valid worker");
let stop_flag = worker.stop_flag();
let capture = TraceCapture::default();
let subscriber = tracing_subscriber::registry().with(capture.clone());
let _guard = tracing::subscriber::set_default(subscriber);
let worker_task = tokio::spawn(async move { worker.run().await });
let mut state = None;
for _ in 0..200 {
tokio::time::sleep(std::time::Duration::from_millis(50)).await;
let current: Option<String> =
sqlx::query_scalar("select state from tasks where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap();
state = current;
if state.as_deref() == Some("completed") {
break;
}
}
stop_flag.store(true, Ordering::SeqCst);
tokio::time::timeout(std::time::Duration::from_secs(10), worker_task)
.await
.expect("worker stops inside the budget")
.expect("worker exits cleanly")
.expect("worker reports no error");
server.abort();
let durable_attempt_id: Option<String> = sqlx::query_scalar::<_, Option<String>>(
"select current_attempt_id from tasks where task_id = $1",
)
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap();
assert_eq!(state.as_deref(), Some("completed"));
// Story 4.4: every outbound send is its own durable attempt, so the two retries of one lease carry
// two distinct attempt identities, both recorded in the immutable attempt history.
let attempt_rows: Vec<(String, i32)> = sqlx::query_as(
"select attempt_id, ordinal from task_attempts where task_id = $1 order by ordinal",
)
.bind(acceptance.task_id.as_uuid())
.fetch_all(registry.pool())
.await
.unwrap();
assert_eq!(attempt_rows.len(), 2, "two durable attempts were committed");
assert_eq!(attempt_rows[0].1, 1);
assert_eq!(attempt_rows[1].1, 2);
assert_eq!(
Some(attempt_rows[1].0.as_str()),
durable_attempt_id.as_deref(),
"the Task points at its latest durable attempt"
);
let spans = capture.snapshot();
let attempt_spans: Vec<_> = spans
.iter()
.filter(|s| s.name == "adapter_attempt")
.collect();
assert_eq!(attempt_spans.len(), 2, "two adapter_attempt spans");
for span in &attempt_spans {
let aid = span.fields.get("attempt_id").cloned().unwrap_or_default();
assert!(!aid.is_empty(), "attempt_id must not be empty");
assert!(
!aid.contains("traceparent"),
"attempt_id must not contain traceparent"
);
assert!(
attempt_rows.iter().any(|(id, _)| id == &aid),
"every attempt span names an attempt that was durably committed"
);
}
let ids: Vec<String> = attempt_spans
.iter()
.map(|s| s.fields.get("attempt_id").cloned().unwrap_or_default())
.collect();
assert_ne!(ids[0], ids[1], "each retry has its own durable attempt id");
assert_eq!(
spans
.iter()
.filter(|span| span.name == "durable_task_attempt")
.count(),
1,
"one durable attempt span per dispatch"
);
let runtime_span = spans
.iter()
.find(|span| span.name == "runtime.execute")
.expect("runtime.execute span");
assert_eq!(
runtime_span.parent_name,
Some("durable_task_attempt"),
"the execution continues the durable attempt trace"
);
for span in &attempt_spans {
assert_eq!(
span.parent_name,
Some("runtime.execute"),
"each retry is a child attempt span under the durable execution"
);
}
let attempts: Vec<String> = attempt_spans
.iter()
.map(|s| s.fields.get("attempt").cloned().unwrap_or_default())
.collect();
assert!(attempts.contains(&"1".to_string()));
assert!(attempts.contains(&"2".to_string()));
for span in &attempt_spans {
assert_eq!(
span.fields.get("max_attempts").cloned().unwrap_or_default(),
"2"
);
}
}
// TraceCapture minimal port from crates/crank-runtime/tests/stages.rs
#[derive(Clone, Default)]
struct TraceCapture {
spans: Arc<Mutex<Vec<CapturedSpan>>>,
}
impl TraceCapture {
fn snapshot(&self) -> Vec<CapturedSpan> {
self.spans.lock().expect("span lock").clone()
}
}
#[derive(Clone, Debug)]
struct CapturedSpan {
name: &'static str,
parent_name: Option<&'static str>,
fields: std::collections::BTreeMap<String, String>,
}
impl<S> Layer<S> for TraceCapture
where
S: Subscriber + for<'lookup> LookupSpan<'lookup>,
{
fn on_new_span(
&self,
attributes: &tracing::span::Attributes<'_>,
id: &Id,
context: tracing_subscriber::layer::Context<'_, S>,
) {
let parent = attributes
.parent()
.and_then(|parent| context.span(parent))
.or_else(|| {
attributes
.is_contextual()
.then(|| context.lookup_current())
.flatten()
});
let mut visitor = FieldVisitor::default();
attributes.record(&mut visitor);
let mut spans = self.spans.lock().expect("span lock");
let index = spans.len();
spans.push(CapturedSpan {
name: attributes.metadata().name(),
parent_name: parent.map(|span| span.metadata().name()),
fields: visitor.fields,
});
context
.span(id)
.expect("span exists")
.extensions_mut()
.insert(index);
}
fn on_record(
&self,
id: &Id,
values: &tracing::span::Record<'_>,
context: tracing_subscriber::layer::Context<'_, S>,
) {
let mut visitor = FieldVisitor::default();
values.record(&mut visitor);
let span = context.span(id).expect("span exists");
let index = *span.extensions().get::<usize>().expect("capture index");
self.spans.lock().expect("span lock")[index]
.fields
.extend(visitor.fields);
}
}
#[derive(Default)]
struct FieldVisitor {
fields: std::collections::BTreeMap<String, String>,
}
impl Visit for FieldVisitor {
fn record_str(&mut self, field: &tracing::field::Field, value: &str) {
self.fields
.insert(field.name().to_owned(), value.to_owned());
}
fn record_debug(&mut self, field: &tracing::field::Field, value: &dyn std::fmt::Debug) {
self.fields
.insert(field.name().to_owned(), format!("{value:?}"));
}
}
@@ -0,0 +1,846 @@
//! ATDD Story 4.4 (Beam): cancel through TaskApplication -> PG -> worker -> upstream.
use super::common;
use super::unsampled_correlation;
use axum::Router;
use axum::http::StatusCode;
use axum::routing::post;
use crank_community_mcp::{TaskApplication, TaskWorkerHandler};
use crank_core::{
AgentOperationBinding, NoTasksFallback, PlatformApiKeyId, PlatformApiKeyScope, TaskState,
TaskStore, ToolExecutionMode, ToolExecutionPolicy, VerifiedMachineCredential,
};
use crank_registry::{AdmissionLimits, PostgresRegistry, PostgresTaskStore, PublishRequest};
use crank_runtime::{
InMemoryCoordinationStateStore, SecretCrypto, TaskService, TaskWorker, TaskWorkerConfig,
};
use serde_json::json;
use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::time::Duration;
use time::OffsetDateTime;
const AGENT_SLUG: &str = "reports";
const TOOL_NAME: &str = "report_gen";
const TEST_MASTER_KEY: &str = "test-master-key-00000000000000000000000000000000";
async fn state_of(registry: &PostgresRegistry, task_id: crank_core::TaskId) -> Option<String> {
sqlx::query_scalar("select state from tasks where task_id = $1")
.bind(task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap()
}
async fn seed_tool(registry: &PostgresRegistry, upstream: &str, policy: ToolExecutionPolicy) {
let mut operation = common::test_operation(upstream, TOOL_NAME);
operation.execution_config.execution_policy = Some(policy);
registry
.create_operation(&common::test_workspace_id(), &operation, Some("alice"))
.await
.unwrap();
registry
.publish_operation(PublishRequest {
workspace_id: &common::test_workspace_id(),
operation_id: &operation.id,
version: 1,
published_at: &OffsetDateTime::parse(
"2026-03-26T10:00:00Z",
&time::format_description::well_known::Rfc3339,
)
.unwrap(),
published_by: Some("alice"),
})
.await
.unwrap();
common::publish_agent_with_bindings(
registry,
AGENT_SLUG,
vec![AgentOperationBinding {
operation_id: operation.id.clone(),
operation_version: 1,
tool_name: operation.name.clone(),
tool_title: operation.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
}],
)
.await;
}
async fn principal(registry: &PostgresRegistry, name: &str) -> crank_community_mcp::TaskPrincipal {
let token = common::create_platform_api_key(
registry,
AGENT_SLUG,
name,
&[PlatformApiKeyScope::Read, PlatformApiKeyScope::Write],
)
.await;
let secret_hash = common::hash_access_secret(&token);
let api_key = registry
.get_platform_api_key_by_secret_for_agent_slug(
common::test_workspace_slug(),
AGENT_SLUG,
&secret_hash,
)
.await
.unwrap()
.expect("active key");
registry
.touch_platform_api_key(
&api_key.api_key.workspace_id,
&api_key.api_key.id,
&OffsetDateTime::now_utc(),
)
.await
.unwrap();
crank_community_mcp::TaskPrincipal::Agent {
workspace_slug: common::test_workspace_slug().to_owned(),
agent_slug: AGENT_SLUG.to_owned(),
credential: VerifiedMachineCredential {
machine_access_mode: crank_core::MachineAccessMode::StaticAgentKey,
max_security_level: crank_core::OperationSecurityLevel::Standard,
scopes: api_key.api_key.scopes.clone(),
platform_api_key_id: Some(api_key.api_key.id.clone()),
},
}
}
async fn application(registry: &PostgresRegistry) -> TaskApplication {
let crypto = SecretCrypto::new(TEST_MASTER_KEY).unwrap();
let observed_at = OffsetDateTime::now_utc();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &observed_at,
})
.await
.expect("master key");
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("store"),
);
TaskApplication::new(
registry.clone(),
Duration::ZERO,
Arc::new(InMemoryCoordinationStateStore::default()),
TaskService::with_secret_crypto(store, Arc::new(crypto)),
)
}
#[tokio::test]
async fn cancel_during_active_request_is_observed_and_leaves_one_unknown_outcome() {
let registry = common::test_registry().await;
let calls = Arc::new(AtomicUsize::new(0));
let (tx, rx) = tokio::sync::oneshot::channel::<()>();
let signal = Arc::new(std::sync::Mutex::new(Some(tx)));
let upstream_app = Router::new().route(
"/crm/leads",
post({
let calls = Arc::clone(&calls);
let signal = Arc::clone(&signal);
move || {
let calls = Arc::clone(&calls);
let signal = Arc::clone(&signal);
async move {
calls.fetch_add(1, Ordering::SeqCst);
if let Some(tx) = signal.lock().unwrap().take() {
let _ = tx.send(());
}
tokio::time::sleep(Duration::from_millis(1500)).await;
(StatusCode::OK, axum::Json(json!({"id":"lead"})))
}
}
}),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, upstream_app).await.unwrap() });
let upstream = format!("http://{address}");
let policy = ToolExecutionPolicy {
mode: ToolExecutionMode::Background,
deadline_ms: 30_000,
no_tasks_fallback: NoTasksFallback::Unsupported,
};
seed_tool(&registry, &upstream, policy).await;
let caller = principal(&registry, "cancel-s1").await;
let foreign = principal(&registry, "cancel-s1-foreign").await;
let app = application(&registry).await;
let acceptance = app
.accept_background(
&caller,
TOOL_NAME,
None,
"s1-key",
&json!({"email":"u"}),
&unsampled_correlation(),
)
.await
.expect("accept");
let crypto = SecretCrypto::new(TEST_MASTER_KEY).unwrap();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &OffsetDateTime::now_utc(),
})
.await
.expect("master key");
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("store"),
);
let runtime = crank_runtime::community_with_outbound_policy(
crank_runtime::OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
)
.build();
let handler = Arc::new(TaskWorkerHandler::from_application(
&app,
registry.clone(),
runtime,
crypto.clone(),
));
let worker = TaskWorker::new(
store,
handler,
TaskWorkerConfig {
owner: "worker-s1".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(50),
max_inflight: 1,
renewal_interval: Duration::from_millis(100),
shutdown_timeout: Duration::from_secs(5),
},
)
.expect("worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
let _ = tokio::time::timeout(Duration::from_secs(5), rx).await;
let cancel = app
.cancel(&caller, TOOL_NAME, acceptance.task_id, None)
.await
.expect("cancel");
match cancel {
crank_core::TaskCancelOutcome::OutcomeUnknown(record) => {
assert_eq!(record.state, TaskState::OutcomeUnknown);
}
other => panic!("expected OutcomeUnknown got {other:?}"),
}
let foreign_cancel = app
.cancel(&foreign, TOOL_NAME, acceptance.task_id, None)
.await;
assert!(matches!(
foreign_cancel,
Err(crank_community_mcp::TaskApplicationError::NotFound)
));
tokio::time::sleep(Duration::from_millis(1600)).await;
assert_eq!(calls.load(Ordering::SeqCst), 1);
assert_eq!(
state_of(&registry, acceptance.task_id).await.as_deref(),
Some("outcome_unknown")
);
let certainty: Option<String> = sqlx::query_scalar(
"select certainty from task_attempts where task_id = $1 order by ordinal desc limit 1",
)
.bind(acceptance.task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap();
assert_eq!(certainty.as_deref(), Some("unknown"));
let code: Option<String> =
sqlx::query_scalar("select error_envelope ->> 'code' from tasks where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap();
assert_eq!(code.as_deref(), Some("cancelled_after_dispatch"));
let attempts_count: i64 =
sqlx::query_scalar("select count(*)::bigint from task_attempts where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap();
assert_eq!(attempts_count, 1);
let cancel2 = app
.cancel(&caller, TOOL_NAME, acceptance.task_id, None)
.await
.expect("cancel2");
assert!(matches!(
cancel2,
crank_core::TaskCancelOutcome::OutcomeUnknown(_)
));
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
server.abort();
}
#[tokio::test]
async fn cancel_before_any_dispatch_is_cancelled_without_upstream_call() {
let registry = common::test_registry().await;
let calls = Arc::new(AtomicUsize::new(0));
let upstream_app = Router::new().route(
"/crm/leads",
post({
let calls = Arc::clone(&calls);
move || {
let calls = Arc::clone(&calls);
async move {
calls.fetch_add(1, Ordering::SeqCst);
(StatusCode::OK, axum::Json(json!({"id":"lead"})))
}
}
}),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, upstream_app).await.unwrap() });
let upstream = format!("http://{address}");
let policy = ToolExecutionPolicy {
mode: ToolExecutionMode::Background,
deadline_ms: 30_000,
no_tasks_fallback: NoTasksFallback::Unsupported,
};
seed_tool(&registry, &upstream, policy).await;
let caller = principal(&registry, "cancel-s3").await;
let app = application(&registry).await;
let acceptance = app
.accept_background(
&caller,
TOOL_NAME,
None,
"s3-key",
&json!({"email":"u"}),
&unsampled_correlation(),
)
.await
.expect("accept");
let cancel = app
.cancel(&caller, TOOL_NAME, acceptance.task_id, None)
.await
.expect("cancel");
match cancel {
crank_core::TaskCancelOutcome::Cancelled(record) => {
assert_eq!(record.state, TaskState::Cancelled)
}
other => panic!("expected Cancelled got {other:?}"),
}
assert_eq!(
state_of(&registry, acceptance.task_id).await.as_deref(),
Some("cancelled")
);
let attempts = app
.attempts(&caller, TOOL_NAME, acceptance.task_id)
.await
.expect("attempts");
assert!(attempts.is_empty());
assert_eq!(calls.load(Ordering::SeqCst), 0);
server.abort();
}
#[tokio::test]
async fn cancel_during_pinned_backoff_never_redispatches_and_is_outcome_unknown() {
use crank_community_mcp::VerifiedIdempotentWrite;
use crank_core::{IdempotencyMode, IdempotencyPolicy, RetryPolicy};
let registry = common::test_registry().await;
let requests = Arc::new(AtomicUsize::new(0));
let upstream_app = Router::new().route(
"/crm/leads",
post({
let requests = Arc::clone(&requests);
move |headers: axum::http::HeaderMap| {
let requests = Arc::clone(&requests);
async move {
let _ = headers.get("idempotency-key");
let attempt = requests.fetch_add(1, Ordering::SeqCst);
if attempt == 0 {
// Applied upstream, acknowledgement lost after the caller had time to observe that
// the request really left the worker.
tokio::time::sleep(Duration::from_millis(600)).await;
return (
StatusCode::INTERNAL_SERVER_ERROR,
axum::Json(json!({"error":"acknowledgement lost"})),
);
}
(StatusCode::OK, axum::Json(json!({"id":"lead-1"})))
}
}
}),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, upstream_app).await.unwrap() });
let upstream = format!("http://{address}");
let mut operation = common::test_operation(&upstream, TOOL_NAME);
operation.execution_config.execution_policy = Some(ToolExecutionPolicy {
mode: ToolExecutionMode::Background,
deadline_ms: 30_000,
no_tasks_fallback: NoTasksFallback::Unsupported,
});
operation.execution_config.retry_policy = Some(RetryPolicy { max_attempts: 2 });
operation.execution_config.idempotency = Some(IdempotencyPolicy {
mode: IdempotencyMode::Required,
ttl_ms: 60_000,
input_field: Some("idempotency_key".to_owned()),
header_name: Some("Idempotency-Key".to_owned()),
});
registry
.create_operation(&common::test_workspace_id(), &operation, Some("alice"))
.await
.unwrap();
registry
.publish_operation(PublishRequest {
workspace_id: &common::test_workspace_id(),
operation_id: &operation.id,
version: 1,
published_at: &OffsetDateTime::parse(
"2026-03-26T10:00:00Z",
&time::format_description::well_known::Rfc3339,
)
.unwrap(),
published_by: Some("alice"),
})
.await
.unwrap();
common::publish_agent_with_bindings(
&registry,
AGENT_SLUG,
vec![AgentOperationBinding {
operation_id: operation.id.clone(),
operation_version: 1,
tool_name: operation.name.clone(),
tool_title: operation.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
}],
)
.await;
let caller = principal(&registry, "cancel-s2").await;
let app = application(&registry).await;
let acceptance = app
.accept_background(
&caller,
TOOL_NAME,
None,
"s2-key",
&json!({"email":"u","idempotency_key":"k-2"}),
&unsampled_correlation(),
)
.await
.expect("accept");
let crypto = SecretCrypto::new(TEST_MASTER_KEY).unwrap();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &OffsetDateTime::now_utc(),
})
.await
.expect("master key");
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("store"),
);
let runtime = crank_runtime::community_with_outbound_policy(
crank_runtime::OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
)
.with_coordination_store(Arc::new(InMemoryCoordinationStateStore::default()))
.build();
let handler = Arc::new(
TaskWorkerHandler::from_application(&app, registry.clone(), runtime, crypto.clone())
.with_verified_writes(vec![VerifiedIdempotentWrite {
operation_id: operation.id.as_str().to_owned(),
operation_version: 1,
key_header: "Idempotency-Key".to_owned(),
validity_ms: 60_000,
}]),
);
// Reproduce deterministically the durable state the executor commits after an ambiguous attempt
// whose bounded retry is allowed: one attempt resolved with certainty=unknown, the retry armed and
// `current_attempt_id` cleared. The real retry window is a pinned millisecond backoff, so racing a
// running executor against the cancel would not be a repeatable oracle.
let lease = store
.claim_batch("worker-s2-prep", 1, 10_000)
.await
.expect("claim")
.pop()
.expect("one claimed lease");
let attempt = store
.checkpoint_dispatch(crank_core::BeginAttempt {
task_id: acceptance.task_id,
owner: lease.owner.clone(),
fence: lease.fence,
expected_version: lease.record.version,
attempt_id: format!("attempt-{}", uuid::Uuid::now_v7()),
input_fingerprint: lease.record.input_fingerprint.clone(),
upstream_key_digest: None,
retry_allowed: true,
})
.await
.expect("durable checkpoint");
let resolution = store
.resolve_attempt(
&attempt,
crank_core::AttemptOutcome::Uncertain {
error: crank_core::TaskEnvelope::try_redacted(
&json!({"code": "upstream_ack_lost", "phase": "dispatched"}),
)
.expect("bounded attempt error"),
repeatable: true,
phase: crank_core::AttemptPhase::Dispatched,
},
)
.await
.expect("attempt resolution");
assert!(
matches!(resolution, crank_core::AttemptResolution::RetryAt { .. }),
"an ambiguous but safely repeatable attempt arms the pinned retry"
);
let cancel = app
.cancel(&caller, TOOL_NAME, acceptance.task_id, None)
.await
.expect("cancel");
match cancel {
crank_core::TaskCancelOutcome::OutcomeUnknown(record) => {
assert_eq!(record.state, crank_core::TaskState::OutcomeUnknown)
}
other => panic!("cancel after an ambiguous attempt must stay uncertain, got {other:?}"),
}
// A real worker must never redispatch or terminate the Task afterwards: the uncertain state is not
// claimable and no second terminal write is possible.
let worker = TaskWorker::new(
store.clone(),
handler,
TaskWorkerConfig {
owner: "worker-s2".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(50),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(5),
..TaskWorkerConfig::default()
},
)
.expect("worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
tokio::time::sleep(Duration::from_secs(1)).await;
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
server.abort();
assert_eq!(
requests.load(Ordering::SeqCst),
0,
"no outbound send after the cancel"
);
assert_eq!(
state_of(&registry, acceptance.task_id).await.as_deref(),
Some("outcome_unknown")
);
let attempts_count: i64 =
sqlx::query_scalar("select count(*)::bigint from task_attempts where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap();
assert_eq!(attempts_count, 1, "no second durable attempt");
let certainty: Option<String> = sqlx::query_scalar(
"select certainty from task_attempts where task_id = $1 order by ordinal desc limit 1",
)
.bind(acceptance.task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap();
assert_eq!(certainty.as_deref(), Some("unknown"));
let code: Option<String> =
sqlx::query_scalar("select error_envelope ->> 'code' from tasks where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap();
assert_eq!(code.as_deref(), Some("cancelled_after_dispatch"));
}
#[tokio::test]
async fn refused_attempt_after_lost_ack_keeps_outcome_unknown() {
use crank_community_mcp::VerifiedIdempotentWrite;
use crank_core::{IdempotencyMode, IdempotencyPolicy, RetryPolicy};
let registry = common::test_registry().await;
let requests = Arc::new(AtomicUsize::new(0));
let upstream_app = Router::new().route(
"/crm/leads",
post({
let requests = Arc::clone(&requests);
move |headers: axum::http::HeaderMap| {
let requests = Arc::clone(&requests);
async move {
let _ = headers.get("idempotency-key");
let attempt = requests.fetch_add(1, Ordering::SeqCst);
if attempt == 0 {
// Applied upstream, acknowledgement lost after the caller had time to observe that
// the request really left the worker.
tokio::time::sleep(Duration::from_millis(600)).await;
return (
StatusCode::INTERNAL_SERVER_ERROR,
axum::Json(json!({"error":"acknowledgement lost"})),
);
}
(StatusCode::OK, axum::Json(json!({"id":"lead-1"})))
}
}
}),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, upstream_app).await.unwrap() });
let upstream = format!("http://{address}");
let mut operation = common::test_operation(&upstream, TOOL_NAME);
operation.execution_config.execution_policy = Some(ToolExecutionPolicy {
mode: ToolExecutionMode::Background,
deadline_ms: 30_000,
no_tasks_fallback: NoTasksFallback::Unsupported,
});
operation.execution_config.retry_policy = Some(RetryPolicy { max_attempts: 2 });
operation.execution_config.idempotency = Some(IdempotencyPolicy {
mode: IdempotencyMode::Required,
ttl_ms: 60_000,
input_field: Some("idempotency_key".to_owned()),
header_name: Some("Idempotency-Key".to_owned()),
});
registry
.create_operation(&common::test_workspace_id(), &operation, Some("alice"))
.await
.unwrap();
registry
.publish_operation(PublishRequest {
workspace_id: &common::test_workspace_id(),
operation_id: &operation.id,
version: 1,
published_at: &OffsetDateTime::parse(
"2026-03-26T10:00:00Z",
&time::format_description::well_known::Rfc3339,
)
.unwrap(),
published_by: Some("alice"),
})
.await
.unwrap();
common::publish_agent_with_bindings(
&registry,
AGENT_SLUG,
vec![AgentOperationBinding {
operation_id: operation.id.clone(),
operation_version: 1,
tool_name: operation.name.clone(),
tool_title: operation.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
}],
)
.await;
let caller = principal(&registry, "cancel-s4").await;
let app = application(&registry).await;
let acceptance = app
.accept_background(
&caller,
TOOL_NAME,
None,
"s4-key",
&json!({"email":"u","idempotency_key":"k-4"}),
&unsampled_correlation(),
)
.await
.expect("accept");
let crypto = SecretCrypto::new(TEST_MASTER_KEY).unwrap();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &OffsetDateTime::now_utc(),
})
.await
.expect("master key");
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("store"),
);
let runtime = crank_runtime::community_with_outbound_policy(
crank_runtime::OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
)
.with_coordination_store(Arc::new(InMemoryCoordinationStateStore::default()))
.build();
let handler = Arc::new(
TaskWorkerHandler::from_application(&app, registry.clone(), runtime, crypto.clone())
.with_verified_writes(vec![VerifiedIdempotentWrite {
operation_id: operation.id.as_str().to_owned(),
operation_version: 1,
key_header: "Idempotency-Key".to_owned(),
validity_ms: 60_000,
}]),
);
let worker = TaskWorker::new(
store.clone(),
handler,
TaskWorkerConfig {
owner: "worker-s4".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(50),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(5),
..TaskWorkerConfig::default()
},
)
.expect("worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
// Revoke the key the Task was accepted with while the first outbound write is provably in flight:
// that attempt still resolves as an uncertain write, and the next attempt is refused by the
// authority before its checkpoint.
for _ in 0..200 {
if requests.load(Ordering::SeqCst) >= 1 {
break;
}
tokio::time::sleep(Duration::from_millis(20)).await;
}
assert_eq!(
requests.load(Ordering::SeqCst),
1,
"the first attempt is dispatched before the refusal"
);
registry
.revoke_platform_api_key_for_agent(
&common::test_workspace_id(),
&common::test_agent_id(AGENT_SLUG),
&PlatformApiKeyId::new("pk_cancel-s4".to_owned()),
&OffsetDateTime::now_utc(),
)
.await
.expect("revoke");
let mut state = None;
for _ in 0..200 {
tokio::time::sleep(Duration::from_millis(50)).await;
state = state_of(&registry, acceptance.task_id).await;
if state.as_deref() == Some("outcome_unknown") {
break;
}
}
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
server.abort();
assert_eq!(
state.as_deref(),
Some("outcome_unknown"),
"a refused attempt after a lost acknowledgement must stay uncertain"
);
assert_eq!(
requests.load(Ordering::SeqCst),
1,
"the refused attempt is never dispatched"
);
let attempts_count: i64 =
sqlx::query_scalar("select count(*)::bigint from task_attempts where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_one(registry.pool())
.await
.unwrap();
assert_eq!(attempts_count, 1, "no second durable attempt");
let certainty: Option<String> = sqlx::query_scalar(
"select certainty from task_attempts where task_id = $1 order by ordinal desc limit 1",
)
.bind(acceptance.task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap();
assert_eq!(certainty.as_deref(), Some("unknown"));
let phase: Option<String> = sqlx::query_scalar(
"select phase from task_attempts where task_id = $1 order by ordinal desc limit 1",
)
.bind(acceptance.task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap();
assert_eq!(phase.as_deref(), Some("dispatched"));
let error: Option<(Option<String>, Option<String>)> = sqlx::query_as(
"select error_envelope ->> 'code', error_envelope ->> 'phase' from tasks where task_id = $1",
)
.bind(acceptance.task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap();
let (code, error_phase) = error.expect("the refused Task carries its bounded error");
assert_eq!(code.as_deref(), Some("task_authority_rejected"));
assert_eq!(
error_phase.as_deref(),
Some("dispatched"),
"the error keeps the phase of the attempt that may have reached the upstream"
);
}
@@ -0,0 +1,657 @@
//! ATDD Story 4.2 (FR-22/FR-23): the durable Task application entry over a real published catalog
//! and a real PostgreSQL schema.
//!
//! These oracles drive the production application entry (`TaskApplication`) that the mcp-server
//! composition root constructs: a verified machine credential is checked against the published Tool,
//! the Task is accepted durably, a freshly constructed application instance reads the same identity
//! and pins, replay/version conflicts never rebind, foreign callers receive no metadata, and a missing
//! Task store fails closed and reports readiness as not ready. The public `tools/call` wire stays
//! synchronous/unsupported: this suite never calls it.
use super::common;
use std::time::Duration;
use crank_community_mcp::{TaskApplication, TaskApplicationError, TaskPrincipal};
use crank_core::{
AgentOperationBinding, CorrelationContext, NoTasksFallback, PlatformApiKeyScope, RequestId,
TaskAcceptRequest, TaskEnvelope, TaskId, TaskOwnerKind, TaskPins, TaskState, TaskStore,
TaskSupportIds, ToolExecutionMode, ToolExecutionPolicy,
};
use crank_registry::{
AdmissionLimits, CreateVersionRequest, PostgresRegistry, PostgresTaskStore, PublishRequest,
};
use crank_runtime::{InMemoryCoordinationStateStore, SecretCrypto, TaskService};
use serde_json::json;
use sqlx::{PgPool, Row};
use time::{OffsetDateTime, format_description::well_known::Rfc3339};
const AGENT_SLUG: &str = "reports";
const TOOL_NAME: &str = "report_gen";
const KEY: &str = "retry-key-1";
fn operation(policy: ToolExecutionPolicy) -> crank_registry::RegistryOperation {
let mut operation = common::test_operation("https://api.example.com", TOOL_NAME);
operation.execution_config.execution_policy = Some(policy);
operation
}
fn background_policy() -> ToolExecutionPolicy {
ToolExecutionPolicy {
mode: ToolExecutionMode::Background,
deadline_ms: 60_000,
no_tasks_fallback: NoTasksFallback::Unsupported,
}
}
async fn seed_tool(registry: &PostgresRegistry, policy: ToolExecutionPolicy) {
let operation = operation(policy);
registry
.create_operation(&common::test_workspace_id(), &operation, Some("alice"))
.await
.unwrap();
publish(registry, &operation.id, 1).await;
// A second, explicitly synchronous Tool of the same Agent: the durable entry must refuse it.
let synchronous = common::test_operation("https://api.example.com", "report_sync");
registry
.create_operation(&common::test_workspace_id(), &synchronous, Some("alice"))
.await
.unwrap();
publish(registry, &synchronous.id, 1).await;
let binding = AgentOperationBinding {
operation_id: operation.id.clone(),
operation_version: 1,
tool_name: operation.name.clone(),
tool_title: operation.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
};
let synchronous_binding = AgentOperationBinding {
operation_id: synchronous.id.clone(),
operation_version: 1,
tool_name: synchronous.name.clone(),
tool_title: synchronous.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
};
common::publish_agent_with_bindings(registry, AGENT_SLUG, vec![binding, synchronous_binding])
.await;
}
async fn publish(
registry: &PostgresRegistry,
operation_id: &crank_core::OperationId,
version: u32,
) {
registry
.publish_operation(PublishRequest {
workspace_id: &common::test_workspace_id(),
operation_id,
version,
published_at: &OffsetDateTime::parse("2026-03-26T10:00:00Z", &Rfc3339).unwrap(),
published_by: Some("alice"),
})
.await
.unwrap();
}
async fn publish_tool_version(registry: &PostgresRegistry, version: u32) {
let mut next = operation(background_policy());
next.version = version;
next.status = crank_core::OperationStatus::Draft;
next.published_at = None;
registry
.create_version(CreateVersionRequest {
workspace_id: &common::test_workspace_id(),
snapshot: &next,
change_note: Some("next publication"),
created_by: Some("alice"),
})
.await
.unwrap();
publish(registry, &next.id, version).await;
// The agent catalog must bind the published Tool version for the new publication to be addressed.
let agent_id = common::test_agent_id(AGENT_SLUG);
let draft = crank_core::AgentVersion {
agent_id: agent_id.clone(),
version: 2,
status: crank_core::AgentStatus::Draft,
instructions: json!({}),
tool_selection_policy: Default::default(),
created_at: OffsetDateTime::parse("2026-03-26T10:01:00Z", &Rfc3339).unwrap(),
};
registry
.create_agent_draft_version(crank_registry::CreateAgentDraftVersionRequest {
workspace_id: &common::test_workspace_id(),
agent_id: &agent_id,
version: &draft,
bindings: &[AgentOperationBinding {
operation_id: next.id.clone(),
operation_version: version,
tool_name: next.name.clone(),
tool_title: next.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
}],
updated_at: &OffsetDateTime::parse("2026-03-26T10:01:00Z", &Rfc3339).unwrap(),
})
.await
.unwrap();
registry
.publish_agent(crank_registry::PublishAgentRequest {
workspace_id: &common::test_workspace_id(),
agent_id: &agent_id,
version: 2,
published_at: &OffsetDateTime::parse("2026-03-26T10:02:00Z", &Rfc3339).unwrap(),
published_by: Some("alice"),
expected_state: None,
})
.await
.unwrap();
}
const TEST_MASTER_KEY: &str = "test-master-key-00000000000000000000000000000000";
fn task_service(pool: PgPool, crypto: SecretCrypto) -> TaskService {
let store = PostgresTaskStore::new(
pool,
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration");
TaskService::with_secret_crypto(std::sync::Arc::new(store), std::sync::Arc::new(crypto))
}
/// The application entry under the same encrypted-input authority the production process uses.
async fn application(registry: &PostgresRegistry) -> TaskApplication {
let crypto = SecretCrypto::new(TEST_MASTER_KEY).expect("test master key");
let observed_at = OffsetDateTime::now_utc();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &observed_at,
})
.await
.expect("test master-key identity registered");
TaskApplication::new(
registry.clone(),
Duration::from_millis(0),
std::sync::Arc::new(InMemoryCoordinationStateStore::default()),
task_service(registry.pool().clone(), crypto),
)
}
async fn principal(
registry: &PostgresRegistry,
name: &str,
scopes: &[PlatformApiKeyScope],
) -> TaskPrincipal {
let token = common::create_platform_api_key(registry, AGENT_SLUG, name, scopes).await;
// The production machine-access path resolves the presented secret against the active platform
// key of the addressed Agent and derives the scope identity from that verified record.
let secret_hash = common::hash_access_secret(&token);
let api_key = registry
.get_platform_api_key_by_secret_for_agent_slug(
common::test_workspace_slug(),
AGENT_SLUG,
&secret_hash,
)
.await
.unwrap()
.expect("an active platform key resolves");
registry
.touch_platform_api_key(
&api_key.api_key.workspace_id,
&api_key.api_key.id,
&OffsetDateTime::now_utc(),
)
.await
.unwrap();
TaskPrincipal::Agent {
workspace_slug: common::test_workspace_slug().to_owned(),
agent_slug: AGENT_SLUG.to_owned(),
credential: crank_core::VerifiedMachineCredential {
machine_access_mode: crank_core::MachineAccessMode::StaticAgentKey,
max_security_level: crank_core::OperationSecurityLevel::Standard,
scopes: api_key.api_key.scopes.clone(),
platform_api_key_id: Some(api_key.api_key.id.clone()),
},
}
}
fn unsampled_correlation() -> CorrelationContext {
CorrelationContext::new(
crank_core::RequestId::generate(),
crank_core::TraceContext::parse("00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-00")
.unwrap(),
)
}
async fn counts(pool: &PgPool) -> (i64, i64) {
let row = sqlx::query(
"select
(select count(*)::bigint from tasks) as tasks,
(select count(*)::bigint from execution_admission_entries where task_id is not null)
as intents",
)
.fetch_one(pool)
.await
.unwrap();
(row.get("tasks"), row.get("intents"))
}
#[tokio::test]
async fn a_verified_caller_receives_a_durable_task_that_a_new_instance_reads() {
let registry = common::test_registry().await;
seed_tool(&registry, background_policy()).await;
let caller = principal(
&registry,
"task-accept",
&[PlatformApiKeyScope::Read, PlatformApiKeyScope::Write],
)
.await;
let correlation = unsampled_correlation();
let app = application(&registry).await;
let acceptance = app
.accept_background(
&caller,
TOOL_NAME,
None,
KEY,
&json!({"report": "q3"}),
&correlation,
)
.await
.expect("acceptance over the real publication");
assert_eq!(acceptance.state, TaskState::Queued);
assert!(!acceptance.replayed);
assert_eq!(acceptance.operation_version, 1);
assert_eq!(acceptance.support.request_id, *correlation.request_id());
assert_eq!(acceptance.support.trace_id, *correlation.trace_id());
assert!(!acceptance.support.trace_sampled);
assert_eq!(
uuid::Uuid::parse_str(&acceptance.task_id.to_string())
.unwrap()
.get_version_num(),
7,
"the durable Task id is UUIDv7"
);
// Task, dispatch intent and queue slot committed together.
assert_eq!(counts(registry.pool()).await, (1, 1));
// A freshly constructed application + service reads the same durable identity and pins.
let reader = application(&registry).await;
let record = reader
.status(&caller, TOOL_NAME, acceptance.task_id)
.await
.unwrap();
assert_eq!(record.state, TaskState::Queued);
assert_eq!(record.pins.operation_version, 1);
assert_eq!(record.pins.policy_mode, ToolExecutionMode::Background);
assert_eq!(record.support.request_id, acceptance.support.request_id);
assert_eq!(record.support.traceparent, acceptance.support.traceparent);
assert!(!record.support.trace_sampled);
let outcome = reader
.result(&caller, TOOL_NAME, acceptance.task_id)
.await
.unwrap();
assert!(outcome.result.is_none() && outcome.error.is_none());
// Replay of the same scoped key and input returns the accepted Task instead of a new one.
let replay = app
.accept_background(
&caller,
TOOL_NAME,
None,
KEY,
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.unwrap();
assert_eq!(replay.task_id, acceptance.task_id);
assert!(replay.replayed);
assert_eq!(counts(registry.pool()).await, (1, 1));
}
#[tokio::test]
async fn a_new_publication_never_rebinds_an_accepted_task() {
let registry = common::test_registry().await;
seed_tool(&registry, background_policy()).await;
let caller = principal(
&registry,
"task-rebind",
&[PlatformApiKeyScope::Read, PlatformApiKeyScope::Write],
)
.await;
let app = application(&registry).await;
let accepted = app
.accept_background(
&caller,
TOOL_NAME,
None,
KEY,
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.unwrap();
let before = app
.status(&caller, TOOL_NAME, accepted.task_id)
.await
.unwrap();
publish_tool_version(&registry, 2).await;
// The same retry keeps the accepted Task and its v1 pin; it never becomes a v2 Task.
let replay = app
.accept_background(
&caller,
TOOL_NAME,
None,
KEY,
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.unwrap();
assert_eq!(replay.task_id, accepted.task_id);
assert_eq!(replay.operation_version, 1);
let after = app
.status(&caller, TOOL_NAME, accepted.task_id)
.await
.unwrap();
assert_eq!(after.pins.operation_version, 1);
assert_eq!(after.support.request_id, before.support.request_id);
assert_eq!(counts(registry.pool()).await, (1, 1));
// An explicit other exact version is a typed conflict, not a rebind.
let explicit = app
.accept_background(
&caller,
TOOL_NAME,
Some(2),
KEY,
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.expect_err("a different explicit pin cannot rebind the accepted Task");
assert_eq!(explicit, TaskApplicationError::VersionUnavailable);
// The original explicit pin still resolves to the accepted Task.
let pinned = app
.accept_background(
&caller,
TOOL_NAME,
Some(1),
KEY,
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.unwrap();
assert_eq!(pinned.task_id, accepted.task_id);
assert_eq!(counts(registry.pool()).await, (1, 1));
}
#[tokio::test]
async fn foreign_and_insufficient_callers_receive_no_task_metadata() {
let registry = common::test_registry().await;
seed_tool(&registry, background_policy()).await;
let owner = principal(
&registry,
"task-owner",
&[PlatformApiKeyScope::Read, PlatformApiKeyScope::Write],
)
.await;
let app = application(&registry).await;
let accepted = app
.accept_background(
&owner,
TOOL_NAME,
None,
KEY,
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.unwrap();
// Positive control: the owner reads its Task.
assert!(
app.status(&owner, TOOL_NAME, accepted.task_id)
.await
.is_ok()
);
// A second verified key of the same Agent is another authorization scope: no payload, no metadata.
let foreign_scope =
principal(&registry, "task-other-scope", &[PlatformApiKeyScope::Read]).await;
match app
.status(&foreign_scope, TOOL_NAME, accepted.task_id)
.await
{
Err(TaskApplicationError::NotFound) => {}
other => panic!("foreign scope must be not-found without metadata, got {other:?}"),
}
assert!(matches!(
app.result(&foreign_scope, TOOL_NAME, accepted.task_id)
.await,
Err(TaskApplicationError::NotFound)
));
// A synchronous publication is not accepted as background work at all.
let synchronous_app = application(&registry).await;
let refused = synchronous_app
.accept_background(
&owner,
"report_sync",
None,
"sync-key",
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.expect_err("a synchronous policy has no durable acceptance");
assert_eq!(refused, TaskApplicationError::NotBackground);
assert_eq!(counts(registry.pool()).await, (1, 1));
}
#[tokio::test]
async fn admin_owned_tasks_are_scoped_to_that_admin_and_workspace() {
let registry = common::test_registry().await;
seed_tool(&registry, background_policy()).await;
let app = application(&registry).await;
let admin = TaskPrincipal::Admin {
workspace_slug: common::test_workspace_slug().to_owned(),
agent_slug: AGENT_SLUG.to_owned(),
user_id: "user_admin_1".to_owned(),
};
let accepted = app
.accept_background(
&admin,
TOOL_NAME,
None,
KEY,
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.expect("an authenticated administrator accepts durable work");
assert_eq!(accepted.state, TaskState::Queued);
// Positive control: the same administrator reads its own Task.
assert!(
app.status(&admin, TOOL_NAME, accepted.task_id)
.await
.is_ok()
);
// An administrator is never a wildcard owner: another admin user and the owning Agent of the
// same workspace both see nothing.
let other_admin = TaskPrincipal::Admin {
workspace_slug: common::test_workspace_slug().to_owned(),
agent_slug: AGENT_SLUG.to_owned(),
user_id: "user_admin_2".to_owned(),
};
assert_eq!(
app.status(&other_admin, TOOL_NAME, accepted.task_id)
.await
.unwrap_err(),
TaskApplicationError::NotFound
);
assert_eq!(
app.result(&other_admin, TOOL_NAME, accepted.task_id)
.await
.unwrap_err(),
TaskApplicationError::NotFound
);
let agent = principal(&registry, "task-admin-cross", &[PlatformApiKeyScope::Read]).await;
assert_eq!(
app.status(&agent, TOOL_NAME, accepted.task_id)
.await
.unwrap_err(),
TaskApplicationError::NotFound
);
// An admin identity addressed to another workspace cannot resolve the publication at all.
let foreign_workspace = TaskPrincipal::Admin {
workspace_slug: "other-workspace".to_owned(),
agent_slug: AGENT_SLUG.to_owned(),
user_id: "user_admin_1".to_owned(),
};
assert_eq!(
app.accept_background(
&foreign_workspace,
TOOL_NAME,
None,
"foreign-key",
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.unwrap_err(),
TaskApplicationError::NotFound
);
assert_eq!(counts(registry.pool()).await, (1, 1));
}
#[tokio::test]
async fn a_missing_task_store_fails_closed_and_readiness_reports_not_ready() {
let database_url = crank_test_support::postgres_schema_url("durable_tasks_outage").await;
let pool = PgPool::connect(&database_url).await.unwrap();
// No migration: the durable Task state does not exist.
let outage_registry = PostgresRegistry::connect(&database_url).await;
assert!(
outage_registry.is_err(),
"an unmigrated registry cannot be used"
);
crank_registry::MigrationAuthority::apply(&pool)
.await
.unwrap();
let registry = PostgresRegistry::connect(&database_url).await.unwrap();
seed_tool(&registry, background_policy()).await;
let caller = principal(
&registry,
"task-outage",
&[PlatformApiKeyScope::Read, PlatformApiKeyScope::Write],
)
.await;
let app = application(&registry).await;
assert!(app.health().await.is_ok());
// Fail closed: without the Task schema the entry accepts nothing and reports dependency failure.
sqlx::query("drop table tasks cascade")
.execute(&pool)
.await
.unwrap();
assert!(app.health().await.is_err());
let error = app
.accept_background(
&caller,
TOOL_NAME,
None,
KEY,
&json!({"report": "q3"}),
&unsampled_correlation(),
)
.await
.expect_err("no durable Task store, no acceptance");
assert!(matches!(
error,
TaskApplicationError::Task(crank_core::TaskError::Unavailable { .. })
));
// Readiness of the app that wires the same store reports the durable Task dependency as not ready.
let router = common::build_test_app(registry.clone(), Duration::from_millis(0), None);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move {
axum::serve(listener, router).await.unwrap();
});
let response = reqwest::get(format!("http://{address}/ready"))
.await
.unwrap();
assert_eq!(response.status(), reqwest::StatusCode::SERVICE_UNAVAILABLE);
let body: serde_json::Value = response.json().await.unwrap();
assert_eq!(body["checks"]["durable_tasks"], json!("not_ready"));
server.abort();
}
/// Commits and resolves exactly one durable attempt of the current lease through the real store.
///
/// Story 4.4 moved the dispatch checkpoint and the attempt resolution into the runtime seam, so a
/// worker-level test double persists exactly one attempt instead of inventing a second retry loop.
pub(crate) async fn commit_attempt(
context: &crank_runtime::TaskAttemptContext,
outcome: crank_core::AttemptOutcome,
) -> crank_runtime::TaskAttemptOutcome {
use crank_core::{AttemptResolution, BeginAttempt, TaskState};
let lease = context.lease();
let begin = BeginAttempt {
task_id: lease.record.task_id,
owner: lease.owner.clone(),
fence: lease.fence,
expected_version: lease.record.version,
attempt_id: format!("attempt-{}", uuid::Uuid::now_v7()),
input_fingerprint: lease.record.input_fingerprint.clone(),
upstream_key_digest: None,
retry_allowed: lease.record.pins.retry.max_attempts > 1,
};
match context.store().checkpoint_dispatch(begin).await {
Ok(attempt) => match context.store().resolve_attempt(&attempt, outcome).await {
Ok(AttemptResolution::Finished(task)) if task.state == TaskState::Completed => {
crank_runtime::TaskAttemptOutcome::Completed
}
Ok(AttemptResolution::Finished(_)) => crank_runtime::TaskAttemptOutcome::Failed {
code: "test_failure".to_owned(),
},
Ok(AttemptResolution::Unknown(_)) => crank_runtime::TaskAttemptOutcome::OutcomeUnknown,
Ok(AttemptResolution::RetryAt { .. }) | Err(_) => {
crank_runtime::TaskAttemptOutcome::Stopped
}
},
Err(_) => crank_runtime::TaskAttemptOutcome::Stopped,
}
}
/// Bounded success outcome used by the worker-level test doubles.
pub(crate) fn completed_outcome(value: serde_json::Value) -> crank_core::AttemptOutcome {
crank_core::AttemptOutcome::Completed(
crank_core::TaskEnvelope::try_redacted(&value).expect("test result is bounded"),
)
}
mod attempt_guards;
mod attempt_trace;
mod cancel_chain;
mod shutdown_window;
mod task_worker;
@@ -0,0 +1,220 @@
//! Story 4.3 QA caveat proof: the narrow post-checkpoint / pre-permit stop window.
//!
//! Split from `task_worker.rs` to keep both integration files inside the code-health limit.
use super::*;
/// A stop that lands after the durable dispatch checkpoint but before the sealed dispatch permit
/// exists finalizes the Task as failed(`shutdown_before_dispatch`) without any outbound call, and
/// does not consume work that is still queued.
///
/// The window is made deterministic instead of raced: a `BEFORE UPDATE` trigger delays only the
/// durable dispatch checkpoint, so the stop flag is set while that checkpoint statement is still in
/// flight (before `TaskDispatchPermit`/`TaskHandler::execute` can exist).
#[tokio::test]
async fn task_worker_stop_between_checkpoint_and_dispatch_never_reaches_upstream() {
use crank_core::{TaskAccess, TaskLease, TaskOwnerKind};
use crank_runtime::{
TaskAttemptOutcome, TaskHandler, TaskWorker, TaskWorkerConfig, TaskWorkerError,
};
use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};
let registry = common::test_registry().await;
// The test registry exposes a single pooled connection: delaying the checkpoint inside its own
// statement keeps the window deterministic without competing for a second connection.
sqlx::raw_sql(sqlx::AssertSqlSafe(
"create function delay_task_dispatch_checkpoint() returns trigger language plpgsql as $$
begin
if new.current_attempt_id is not null and old.current_attempt_id is null then
perform pg_sleep(1.5);
end if;
return new;
end $$;
create trigger delay_task_dispatch_checkpoint before update on tasks
for each row execute function delay_task_dispatch_checkpoint();",
))
.execute(registry.pool())
.await
.expect("checkpoint delay trigger");
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration"),
);
let mut accepted = Vec::new();
for (seed, scoped_key) in [("q", "a"), ("r", "c")] {
let request = TaskAcceptRequest {
task_id: TaskId::generate(),
workspace_id: common::test_workspace_id().to_string(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_stop_window".to_owned(),
platform_api_key_id: None,
operation_id: "op_report".to_owned(),
pins: TaskPins {
operation_version: 1,
policy_mode: ToolExecutionMode::Background,
execution_deadline_unix_ms: 1_900_000_000_000,
retain_until_unix_ms: 1_900_000_000_000 + 60_000,
approval_reference: None,
retry: crank_core::TaskRetryPolicy::single_attempt(),
},
scoped_key: scoped_key.repeat(64),
input_fingerprint: scoped_key.repeat(64),
input: TaskEnvelope::try_redacted(&json!({ "report": seed })).unwrap(),
support: TaskSupportIds::new(
RequestId::generate(),
&crank_core::TraceContext::parse(
"00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-00",
)
.unwrap(),
),
executable_input_ciphertext: None,
executable_input_key_version: None,
executable_input_master_key_epoch: None,
};
accepted.push(store.accept(request).await.expect("PG accepts"));
}
// The first revalidation waits for the test; the second Task is never claimed (batch_limit = 1).
let revalidate_entered = Arc::new(tokio::sync::Semaphore::new(0));
let revalidate_release = Arc::new(tokio::sync::Semaphore::new(0));
let executed = Arc::new(AtomicUsize::new(0));
struct GatedHandler {
entered: Arc<tokio::sync::Semaphore>,
release: Arc<tokio::sync::Semaphore>,
executed: Arc<AtomicUsize>,
}
#[async_trait::async_trait]
impl TaskHandler for GatedHandler {
async fn revalidate(&self, _lease: &TaskLease) -> Result<(), TaskWorkerError> {
self.entered.add_permits(1);
let _ = self.release.acquire().await;
Ok(())
}
async fn execute(&self, context: crank_runtime::TaskAttemptContext) -> TaskAttemptOutcome {
// The trigger delays the durable checkpoint; a stop that lands while it is in flight must
// finalize failed(shutdown_before_dispatch) without an outbound send.
let begin = crank_core::BeginAttempt {
task_id: context.lease().record.task_id,
owner: context.lease().owner.clone(),
fence: context.lease().fence,
expected_version: context.lease().record.version,
attempt_id: "attempt-stop-window".to_owned(),
input_fingerprint: context.lease().record.input_fingerprint.clone(),
upstream_key_digest: None,
retry_allowed: false,
};
let Ok(attempt) = context.store().checkpoint_dispatch(begin).await else {
return TaskAttemptOutcome::Stopped;
};
if context.is_stopping() {
let _ = context
.store()
.resolve_attempt(
&attempt,
crank_core::AttemptOutcome::Failed {
error: crank_core::TaskEnvelope::try_redacted(&json!({
"code": "shutdown_before_dispatch",
"phase": "before_dispatch",
}))
.expect("bounded error envelope"),
retryable: false,
phase: crank_core::AttemptPhase::BeforeDispatch,
},
)
.await;
return TaskAttemptOutcome::Failed {
code: "shutdown_before_dispatch".to_owned(),
};
}
self.executed.fetch_add(1, Ordering::SeqCst);
commit_attempt(&context, completed_outcome(json!({ "ok": true }))).await
}
}
let handler: Arc<dyn TaskHandler> = Arc::new(GatedHandler {
entered: revalidate_entered.clone(),
release: revalidate_release.clone(),
executed: executed.clone(),
});
let config = TaskWorkerConfig {
owner: "runtime-worker-stop-window".to_owned(),
batch_limit: 1,
lease_ms: 30_000,
poll_interval: Duration::from_millis(50),
renewal_interval: Duration::from_secs(20),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(20),
};
let worker = TaskWorker::new(store.clone(), handler, config).expect("valid worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
tokio::time::timeout(Duration::from_secs(5), revalidate_entered.acquire())
.await
.expect("worker must claim the first Task and enter revalidation")
.expect("revalidation gate is open")
.forget();
// Let revalidation continue; the checkpoint now runs while the trigger holds it for 1.5s.
revalidate_release.add_permits(1);
tokio::time::sleep(Duration::from_millis(300)).await;
// The stop now lands while the durable checkpoint statement is still in flight: the worker may
// finalize the Task, but no permit exists yet, so no outbound call can start.
let stop_requested = std::time::Instant::now();
stop_flag.store(true, Ordering::SeqCst);
tokio::time::timeout(Duration::from_secs(30), worker_task)
.await
.expect("worker stops inside the 30s budget")
.expect("worker exits cleanly")
.expect("worker reports no error");
assert!(
stop_requested.elapsed() >= Duration::from_millis(500),
"the checkpoint must have still been in flight when the stop arrived"
);
assert_eq!(
executed.load(Ordering::SeqCst),
0,
"a stop between checkpoint and permit must never reach the upstream adapter"
);
let access = TaskAccess {
workspace_id: common::test_workspace_id().to_string(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_stop_window".to_owned(),
};
let record = store.read(&access, accepted[0].task_id).await.unwrap();
assert_eq!(record.state, TaskState::Failed);
assert_eq!(record.attempts, 1, "the durable checkpoint did commit");
assert!(record.current_attempt_id.is_some());
assert_eq!(
record
.error
.as_ref()
.and_then(|error| error.value().get("code"))
.and_then(|code| code.as_str()),
Some("shutdown_before_dispatch")
);
// Work that was still queued is not consumed by the stopped worker.
let queued = store.read(&access, accepted[1].task_id).await.unwrap();
assert_eq!(queued.state, TaskState::Queued);
let recovered = store
.claim_batch("worker_after_stop", 4, 10_000)
.await
.expect("queued work stays claimable");
assert_eq!(recovered.len(), 1);
assert_eq!(recovered[0].record.task_id, accepted[1].task_id);
}
@@ -0,0 +1,821 @@
//! Story 4.3 ATDD: fenced durable Task worker leases, recovery and real upstream dispatch.
//!
//! Split from `durable_tasks.rs` to keep both integration files inside the code-health limit.
use super::*;
/// Store-level positive control only: PostgreSQL leases/fencing/recovery work without Valkey or a
/// task-worker process. This does not exercise TaskWorker startup, a bounded authoritative scan,
/// graceful shutdown, outbound dispatch, or runtime recovery timing, which remain unimplemented.
#[tokio::test]
async fn task_worker_store_lease_recovery_contract() {
let registry = common::test_registry().await;
let store = PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration");
// Positive control: no Valkey or worker process is involved; PostgreSQL alone accepts and
// leases durable work, preserving the same admission-backed intent.
let request = TaskAcceptRequest {
task_id: TaskId::generate(),
workspace_id: common::test_workspace_id().to_string(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_1".to_owned(),
platform_api_key_id: None,
operation_id: "op_report".to_owned(),
pins: TaskPins {
operation_version: 1,
policy_mode: ToolExecutionMode::Background,
execution_deadline_unix_ms: 1_900_000_000_000,
retain_until_unix_ms: 1_900_000_000_000 + 60_000,
approval_reference: None,
retry: crank_core::TaskRetryPolicy::single_attempt(),
},
scoped_key: "a".repeat(64),
input_fingerprint: "b".repeat(64),
input: TaskEnvelope::try_redacted(&json!({"report": "q3"})).unwrap(),
support: TaskSupportIds::new(
RequestId::generate(),
&crank_core::TraceContext::parse(
"00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-00",
)
.unwrap(),
),
// This store-level control exercises lease/fence behavior, not the encrypted acceptance path.
executable_input_ciphertext: None,
executable_input_key_version: None,
executable_input_master_key_epoch: None,
};
let accepted = store.accept(request).await.expect("PG accepts");
let leases = store
.claim_batch("worker_1", 4, 10_000)
.await
.expect("claim");
assert_eq!(leases.len(), 1);
assert_eq!(leases[0].record.task_id, accepted.task_id);
let lease = leases[0].clone();
// A stale or competing owner cannot win CAS with the issued fencing token.
assert!(
store
.renew_lease("worker_2", lease.record.task_id, lease.fence, 10_000)
.await
.is_err()
);
assert!(
store
.write_progress(
"worker_1",
lease.record.task_id,
lease.fence + 1,
TaskEnvelope::try_redacted(&json!({"progress": 1})).unwrap(),
)
.await
.is_err()
);
// A stale fencing token can never checkpoint dispatch, even under the issuing owner.
assert!(
store
.checkpoint_dispatch(crank_core::BeginAttempt {
task_id: lease.record.task_id,
owner: "worker_1".to_owned(),
fence: lease.fence + 1,
expected_version: lease.record.version,
attempt_id: "attempt-stale".to_owned(),
input_fingerprint: lease.record.input_fingerprint.clone(),
upstream_key_digest: None,
retry_allowed: false,
})
.await
.is_err()
);
// The valid owner still commits the dispatch checkpoint before any outbound work.
let committed = store
.checkpoint_dispatch(crank_core::BeginAttempt {
task_id: lease.record.task_id,
owner: "worker_1".to_owned(),
fence: lease.fence,
expected_version: lease.record.version,
attempt_id: "attempt-1".to_owned(),
input_fingerprint: lease.record.input_fingerprint.clone(),
upstream_key_digest: None,
retry_allowed: false,
})
.await
.unwrap();
assert_eq!(committed.attempt.attempt_id, "attempt-1");
assert_eq!(
committed.task.current_attempt_id.as_deref(),
Some("attempt-1")
);
}
#[tokio::test]
async fn task_worker_runtime_dispatch_and_recovery_contract() {
use async_trait::async_trait;
use crank_core::{TaskAccess, TaskLease, TaskOwnerKind};
use crank_runtime::{
TaskAttemptOutcome, TaskHandler, TaskWorker, TaskWorkerConfig, TaskWorkerError,
};
use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};
let registry = common::test_registry().await;
let pool = registry.pool().clone();
let store = Arc::new(
PostgresTaskStore::new(
pool.clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration"),
);
let observed = Arc::new(AtomicUsize::new(0));
// Positive control: PostgreSQL alone accepts the durable Task.
let request = TaskAcceptRequest {
task_id: TaskId::generate(),
workspace_id: common::test_workspace_id().to_string(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_runtime".to_owned(),
platform_api_key_id: None,
operation_id: "op_report".to_owned(),
pins: TaskPins {
operation_version: 1,
policy_mode: ToolExecutionMode::Background,
execution_deadline_unix_ms: 1_900_000_000_000,
retain_until_unix_ms: 1_900_000_000_000 + 60_000,
approval_reference: None,
retry: crank_core::TaskRetryPolicy::single_attempt(),
},
scoped_key: "r".repeat(64),
input_fingerprint: "b".repeat(64),
input: TaskEnvelope::try_redacted(&json!({"email": "q3"})).unwrap(),
support: TaskSupportIds::new(
RequestId::generate(),
&crank_core::TraceContext::parse(
"00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-00",
)
.unwrap(),
),
executable_input_ciphertext: None,
executable_input_key_version: None,
executable_input_master_key_epoch: None,
};
let accepted = store.accept(request).await.expect("PG accepts");
let task_id = accepted.task_id;
struct RuntimeHandler {
observed: Arc<AtomicUsize>,
}
#[async_trait]
impl TaskHandler for RuntimeHandler {
async fn revalidate(&self, _lease: &TaskLease) -> Result<(), TaskWorkerError> {
Ok(())
}
async fn execute(&self, context: crank_runtime::TaskAttemptContext) -> TaskAttemptOutcome {
self.observed.fetch_add(1, Ordering::SeqCst);
commit_attempt(&context, completed_outcome(serde_json::json!({"ok": true}))).await
}
}
let handler: Arc<dyn TaskHandler> = Arc::new(RuntimeHandler {
observed: observed.clone(),
});
// PG-only authoritative dispatch: no Valkey or upstream process is involved.
let config = TaskWorkerConfig {
owner: "runtime-worker-1".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(100),
renewal_interval: Duration::from_millis(1_000),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(5),
};
let worker = TaskWorker::new(store.clone(), handler, config).expect("valid worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
let access = TaskAccess {
workspace_id: common::test_workspace_id().to_string(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_runtime".to_owned(),
};
let mut state = None;
for _ in 0..100 {
tokio::time::sleep(Duration::from_millis(100)).await;
if let Ok(record) = store.read(&access, task_id).await {
state = Some(record.state);
if record.state == TaskState::Completed {
break;
}
}
}
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
assert_eq!(state, Some(TaskState::Completed));
assert_eq!(
observed.load(Ordering::SeqCst),
1,
"exactly one outbound attempt"
);
// Terminal state is not duplicated.
let record = store.read(&access, task_id).await.expect("terminal read");
assert_eq!(record.state, TaskState::Completed);
// Bounded shutdown with no fresh work returns promptly.
let idle_config = TaskWorkerConfig {
owner: "runtime-worker-idle".to_owned(),
poll_interval: Duration::from_millis(50),
shutdown_timeout: Duration::from_secs(2),
..TaskWorkerConfig::default()
};
let idle_worker = TaskWorker::new(
store.clone(),
handler_for_no_work(store.clone()),
idle_config,
)
.expect("valid idle worker");
idle_worker.stop();
let started = std::time::Instant::now();
idle_worker.run().await.expect("idle worker stops");
assert!(started.elapsed() <= Duration::from_secs(30));
}
/// Story 4.3 ATDD: failed pre-dispatch revalidation finishes the attempt without any outbound call.
#[tokio::test]
async fn task_worker_revalidation_denial_never_dispatches() {
use crank_core::{TaskAccess, TaskError, TaskLease, TaskOwnerKind};
use crank_runtime::{
TaskAttemptOutcome, TaskHandler, TaskWorker, TaskWorkerConfig, TaskWorkerError,
};
use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};
let registry = common::test_registry().await;
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration"),
);
let request = TaskAcceptRequest {
task_id: TaskId::generate(),
workspace_id: common::test_workspace_id().to_string(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_denied".to_owned(),
platform_api_key_id: None,
operation_id: "op_report".to_owned(),
pins: TaskPins {
operation_version: 1,
policy_mode: ToolExecutionMode::Background,
execution_deadline_unix_ms: 1_900_000_000_000,
retain_until_unix_ms: 1_900_000_000_000 + 60_000,
approval_reference: None,
retry: crank_core::TaskRetryPolicy::single_attempt(),
},
scoped_key: "d".repeat(64),
input_fingerprint: "b".repeat(64),
input: TaskEnvelope::try_redacted(&json!({"report": "q3"})).unwrap(),
support: TaskSupportIds::new(
RequestId::generate(),
&crank_core::TraceContext::parse(
"00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-01",
)
.unwrap(),
),
executable_input_ciphertext: None,
executable_input_key_version: None,
executable_input_master_key_epoch: None,
};
let accepted = store.accept(request).await.expect("PG accepts");
struct DeniedHandler {
observed: Arc<AtomicUsize>,
}
#[async_trait::async_trait]
impl TaskHandler for DeniedHandler {
async fn revalidate(&self, _lease: &TaskLease) -> Result<(), TaskWorkerError> {
Err(TaskWorkerError::Store(TaskError::invalid(
"publication or credential revoked",
)))
}
async fn execute(&self, context: crank_runtime::TaskAttemptContext) -> TaskAttemptOutcome {
self.observed.fetch_add(1, Ordering::SeqCst);
commit_attempt(&context, completed_outcome(json!({"ok": true}))).await
}
}
let observed = Arc::new(AtomicUsize::new(0));
let handler: Arc<dyn TaskHandler> = Arc::new(DeniedHandler {
observed: observed.clone(),
});
let config = TaskWorkerConfig {
owner: "runtime-worker-denied".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(50),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(5),
..TaskWorkerConfig::default()
};
let worker = TaskWorker::new(store.clone(), handler, config).expect("valid worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
let access = TaskAccess {
workspace_id: common::test_workspace_id().to_string(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_denied".to_owned(),
};
let mut state = None;
for _ in 0..100 {
tokio::time::sleep(Duration::from_millis(50)).await;
if let Ok(record) = store.read(&access, accepted.task_id).await {
state = Some(record.state);
if record.state == TaskState::Failed {
break;
}
}
}
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
assert_eq!(state, Some(TaskState::Failed));
assert_eq!(
observed.load(Ordering::SeqCst),
0,
"invalidated work never reaches the upstream"
);
}
fn handler_for_no_work(
_store: std::sync::Arc<PostgresTaskStore>,
) -> std::sync::Arc<dyn crank_runtime::TaskHandler> {
struct NoWorkHandler;
#[async_trait::async_trait]
impl crank_runtime::TaskHandler for NoWorkHandler {
async fn revalidate(
&self,
_lease: &crank_core::TaskLease,
) -> Result<(), crank_runtime::TaskWorkerError> {
Ok(())
}
async fn execute(
&self,
context: crank_runtime::TaskAttemptContext,
) -> crank_runtime::TaskAttemptOutcome {
commit_attempt(&context, completed_outcome(serde_json::json!({"ok": true}))).await
}
}
std::sync::Arc::new(NoWorkHandler)
}
/// Story 4.3 end-to-end ATDD: the production durable-Task handler, the real runtime executor and a
/// real local HTTP upstream observe exactly one dispatch after the fenced checkpoint, while the
/// accepted parent trace is continued on the wire.
#[tokio::test]
async fn task_worker_dispatches_one_real_upstream_attempt_with_parent_trace() {
use axum::Router;
use axum::http::{HeaderMap, StatusCode};
use axum::routing::post;
use crank_community_mcp::TaskWorkerHandler;
use crank_core::{AgentOperationBinding, PlatformApiKeyScope};
use crank_runtime::{SecretCrypto, TaskWorker, TaskWorkerConfig};
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::{Arc, Mutex};
let registry = common::test_registry().await;
let calls = Arc::new(AtomicUsize::new(0));
let seen_traceparent = Arc::new(Mutex::new(None::<String>));
let app = Router::new().route(
"/crm/leads",
post({
let calls = Arc::clone(&calls);
let seen = Arc::clone(&seen_traceparent);
move |headers: HeaderMap| {
let calls = Arc::clone(&calls);
let seen = Arc::clone(&seen);
async move {
calls.fetch_add(1, Ordering::SeqCst);
if let Some(value) = headers
.get("traceparent")
.and_then(|value| value.to_str().ok())
{
*seen.lock().unwrap() = Some(value.to_owned());
}
(StatusCode::OK, axum::Json(json!({"id": "lead-1"})))
}
}
}),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
let upstream = format!("http://{address}");
let mut operation = common::test_operation(&upstream, TOOL_NAME);
operation.execution_config.execution_policy = Some(background_policy());
registry
.create_operation(&common::test_workspace_id(), &operation, Some("alice"))
.await
.unwrap();
publish(&registry, &operation.id, 1).await;
common::publish_agent_with_bindings(
&registry,
AGENT_SLUG,
vec![AgentOperationBinding {
operation_id: operation.id.clone(),
operation_version: 1,
tool_name: operation.name.clone(),
tool_title: operation.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
}],
)
.await;
let caller = principal(
&registry,
"task-worker-e2e",
&[PlatformApiKeyScope::Read, PlatformApiKeyScope::Write],
)
.await;
let correlation = unsampled_correlation();
let crypto = SecretCrypto::new("test-master-key-00000000000000000000000000000000").unwrap();
let observed_at = OffsetDateTime::now_utc();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &observed_at,
})
.await
.expect("test master-key identity registered");
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration"),
);
let tasks = TaskApplication::new(
registry.clone(),
Duration::ZERO,
Arc::new(InMemoryCoordinationStateStore::default()),
TaskService::with_secret_crypto(store.clone(), Arc::new(crypto.clone())),
);
let acceptance = tasks
.accept_background(
&caller,
TOOL_NAME,
None,
"worker-e2e-key",
&json!({"email": "user@example.com"}),
&correlation,
)
.await
.expect("durable acceptance with executable input");
let runtime = crank_runtime::community_with_outbound_policy(
crank_runtime::OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
)
.build();
let handler = Arc::new(TaskWorkerHandler::from_application(
&tasks,
registry.clone(),
runtime,
crypto,
));
let config = TaskWorkerConfig {
owner: "runtime-worker-e2e".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(50),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(5),
..TaskWorkerConfig::default()
};
let worker = TaskWorker::new(store.clone(), handler, config).expect("valid worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
let mut state = None;
for _ in 0..200 {
tokio::time::sleep(Duration::from_millis(50)).await;
let current: Option<String> =
sqlx::query_scalar("select state from tasks where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap();
state = current;
if state.as_deref() == Some("completed") {
break;
}
}
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
server.abort();
assert_eq!(state.as_deref(), Some("completed"));
assert_eq!(calls.load(Ordering::SeqCst), 1, "exactly one upstream call");
assert_eq!(
seen_traceparent.lock().unwrap().as_deref(),
Some(correlation.trace_context().traceparent()),
"the durable attempt continues the accepted parent trace"
);
}
/// Story 4.3 end-to-end ATDD: an ambiguous post-checkpoint upstream failure is persisted as
/// `outcome_unknown` (not a retryable failure) and is not redispatched automatically.
#[tokio::test]
async fn task_worker_ambiguous_upstream_error_becomes_outcome_unknown() {
use axum::Router;
use axum::http::StatusCode;
use axum::routing::post;
use crank_community_mcp::TaskWorkerHandler;
use crank_core::{AgentOperationBinding, PlatformApiKeyScope};
use crank_runtime::{SecretCrypto, TaskWorker, TaskWorkerConfig};
use std::sync::Arc;
use std::sync::atomic::Ordering;
let registry = common::test_registry().await;
let upstream_app = Router::new().route(
"/crm/leads",
post(|| async { (StatusCode::SERVICE_UNAVAILABLE, "upstream unavailable") }),
);
let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
let address = listener.local_addr().unwrap();
let server = tokio::spawn(async move { axum::serve(listener, upstream_app).await.unwrap() });
let upstream = format!("http://{address}");
let mut operation = common::test_operation(&upstream, TOOL_NAME);
operation.execution_config.execution_policy = Some(background_policy());
registry
.create_operation(&common::test_workspace_id(), &operation, Some("alice"))
.await
.unwrap();
publish(&registry, &operation.id, 1).await;
common::publish_agent_with_bindings(
&registry,
AGENT_SLUG,
vec![AgentOperationBinding {
operation_id: operation.id.clone(),
operation_version: 1,
tool_name: operation.name.clone(),
tool_title: operation.tool_description.title.clone(),
tool_description_override: None,
enabled: true,
}],
)
.await;
let caller = principal(
&registry,
"task-worker-uncertain",
&[PlatformApiKeyScope::Read, PlatformApiKeyScope::Write],
)
.await;
let correlation = unsampled_correlation();
let crypto = SecretCrypto::new("test-master-key-00000000000000000000000000000000").unwrap();
let observed_at = OffsetDateTime::now_utc();
registry
.verify_or_register_master_key_identity(crank_registry::MasterKeyIdentityCandidate {
epoch: crypto.master_key_epoch(),
fingerprint: crypto.master_key_fingerprint(),
cipher_contract: crank_registry::MASTER_KEY_CIPHER_CONTRACT,
observed_at: &observed_at,
})
.await
.expect("test master-key identity registered");
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration"),
);
let tasks = TaskApplication::new(
registry.clone(),
Duration::ZERO,
Arc::new(InMemoryCoordinationStateStore::default()),
TaskService::with_secret_crypto(store.clone(), Arc::new(crypto.clone())),
);
let acceptance = tasks
.accept_background(
&caller,
TOOL_NAME,
None,
"worker-uncertain-key",
&json!({"email": "user@example.com"}),
&correlation,
)
.await
.expect("durable acceptance with executable input");
let runtime = crank_runtime::community_with_outbound_policy(
crank_runtime::OutboundHttpPolicy::allowing_hosts(["127.0.0.1"]),
)
.build();
let handler = Arc::new(TaskWorkerHandler::from_application(
&tasks,
registry.clone(),
runtime,
crypto,
));
let config = TaskWorkerConfig {
owner: "runtime-worker-uncertain".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(50),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(5),
..TaskWorkerConfig::default()
};
let worker = TaskWorker::new(store.clone(), handler, config).expect("valid worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
let mut state = None;
for _ in 0..200 {
tokio::time::sleep(Duration::from_millis(50)).await;
let current: Option<String> =
sqlx::query_scalar("select state from tasks where task_id = $1")
.bind(acceptance.task_id.as_uuid())
.fetch_optional(registry.pool())
.await
.unwrap();
state = current;
if state.as_deref() == Some("outcome_unknown") {
break;
}
}
stop_flag.store(true, Ordering::SeqCst);
let _ = tokio::time::timeout(Duration::from_secs(10), worker_task).await;
server.abort();
assert_eq!(state.as_deref(), Some("outcome_unknown"));
// No automatic retry: another authoritative scan never redispaches the uncertain attempt.
assert!(
store
.claim_batch("worker_after", 4, 10_000)
.await
.unwrap()
.is_empty()
);
}
/// Story 4.3 ATDD: graceful stop stops new claims but lets an active attempt reach its fenced
/// terminal commit within the 30s budget.
#[tokio::test]
async fn task_worker_graceful_stop_drains_active_attempt() {
use crank_core::{TaskAccess, TaskLease, TaskOwnerKind};
use crank_runtime::{
TaskAttemptOutcome, TaskHandler, TaskWorker, TaskWorkerConfig, TaskWorkerError,
};
use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};
let registry = common::test_registry().await;
let store = Arc::new(
PostgresTaskStore::new(
registry.pool().clone(),
AdmissionLimits {
global_limit: 4,
agent_limit: 4,
tool_limit: 4,
queue_limit: 8,
},
Default::default(),
)
.expect("valid task store configuration"),
);
let request = TaskAcceptRequest {
task_id: TaskId::generate(),
workspace_id: common::test_workspace_id().to_string(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_drain".to_owned(),
platform_api_key_id: None,
operation_id: "op_report".to_owned(),
pins: TaskPins {
operation_version: 1,
policy_mode: ToolExecutionMode::Background,
execution_deadline_unix_ms: 1_900_000_000_000,
retain_until_unix_ms: 1_900_000_000_000 + 60_000,
approval_reference: None,
retry: crank_core::TaskRetryPolicy::single_attempt(),
},
scoped_key: "s".repeat(64),
input_fingerprint: "d".repeat(64),
input: TaskEnvelope::try_redacted(&json!({"report": "q3"})).unwrap(),
support: TaskSupportIds::new(
RequestId::generate(),
&crank_core::TraceContext::parse(
"00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-00",
)
.unwrap(),
),
executable_input_ciphertext: None,
executable_input_key_version: None,
executable_input_master_key_epoch: None,
};
let accepted = store.accept(request).await.expect("PG accepts");
struct SlowHandler {
started: Arc<AtomicUsize>,
}
#[async_trait::async_trait]
impl TaskHandler for SlowHandler {
async fn revalidate(&self, _lease: &TaskLease) -> Result<(), TaskWorkerError> {
Ok(())
}
async fn execute(&self, context: crank_runtime::TaskAttemptContext) -> TaskAttemptOutcome {
self.started.fetch_add(1, Ordering::SeqCst);
tokio::time::sleep(Duration::from_millis(500)).await;
commit_attempt(&context, completed_outcome(json!({"ok": true}))).await
}
}
let started = Arc::new(AtomicUsize::new(0));
let handler: Arc<dyn TaskHandler> = Arc::new(SlowHandler {
started: started.clone(),
});
let config = TaskWorkerConfig {
owner: "runtime-worker-drain".to_owned(),
batch_limit: 1,
lease_ms: 10_000,
poll_interval: Duration::from_millis(50),
max_inflight: 1,
shutdown_timeout: Duration::from_secs(20),
..TaskWorkerConfig::default()
};
let worker = TaskWorker::new(store.clone(), handler, config).expect("valid worker");
let stop_flag = worker.stop_flag();
let worker_task = tokio::spawn(async move { worker.run().await });
for _ in 0..100 {
if started.load(Ordering::SeqCst) > 0 {
break;
}
tokio::time::sleep(Duration::from_millis(20)).await;
}
assert_eq!(started.load(Ordering::SeqCst), 1, "attempt started");
let stop_requested = std::time::Instant::now();
stop_flag.store(true, Ordering::SeqCst);
tokio::time::timeout(Duration::from_secs(30), worker_task)
.await
.expect("worker stops inside the 30s budget")
.expect("worker exits cleanly")
.expect("worker reports no error");
assert!(stop_requested.elapsed() <= Duration::from_secs(30));
let access = TaskAccess {
workspace_id: common::test_workspace_id().to_string(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_drain".to_owned(),
};
let record = store.read(&access, accepted.task_id).await.unwrap();
assert_eq!(record.state, TaskState::Completed);
}
+5
View File
@@ -1,5 +1,10 @@
FROM node:22-alpine AS build
# The CI host uplink drops TLS 1.3 ClientHello carrying the hybrid
# X25519MLKEM768 key share; pin the npm stage to TLS 1.2 until that network
# defect is fixed. Certificates, SRI and the lockfile are unaffected.
ENV NODE_OPTIONS=--tls-max-v1.2
WORKDIR /app
COPY apps/ui/package.json apps/ui/package-lock.json ./
+2
View File
@@ -1036,6 +1036,7 @@ var TRANSLATIONS = {
'wizard.error.secret_name_required': 'Secret name is required.',
'wizard.error.secret_value_required': 'Secret value is required.',
'wizard.error.save_upstream_first': 'Save the API host after creating a new authorization profile before continuing.',
'wizard.error.execution_policy_invalid': 'Execution policy is invalid: mode, deadline from 1 to 3600000 ms and no-Tasks fallback are required.',
'wizard.toast.auth_profile_created_title': 'Authorization profile created',
'wizard.toast.auth_profile_created_body': 'New authorization profile is ready and selected.',
'wizard.toast.auth_profile_error_title': 'Authorization profile creation failed',
@@ -2353,6 +2354,7 @@ var TRANSLATIONS = {
'wizard.error.secret_name_required': 'Нужно указать имя секрета.',
'wizard.error.secret_value_required': 'Нужно указать значение секрета.',
'wizard.error.save_upstream_first': 'Сначала сохраните API-хост после создания нового профиля авторизации.',
'wizard.error.execution_policy_invalid': 'Политика выполнения некорректна: нужны режим, deadline от 1 до 3600000 мс и поведение без Tasks.',
'wizard.toast.auth_profile_created_title': 'Профиль авторизации создан',
'wizard.toast.auth_profile_created_body': 'Новый профиль авторизации готов и уже выбран.',
'wizard.toast.auth_profile_error_title': 'Не удалось создать профиль авторизации',
+30
View File
@@ -214,10 +214,33 @@ function parseExecutionConfig(text) {
auth_profile_ref: authProfileRef,
headers: headers,
};
var executionPolicy = parseExecutionPolicy(value.execution_policy);
if (executionPolicy) {
config.execution_policy = executionPolicy;
}
return config;
}
// Story 4.1: the published execution policy must survive a wizard save. The editor is a free-form
// document, so unknown/legacy shapes stay untouched: only a fully valid policy is projected, and an
// explicit invalid value is reported instead of being silently replaced by defaults.
function parseExecutionPolicy(raw) {
if (raw === undefined || raw === null) return null;
if (typeof raw !== 'object' || Array.isArray(raw)) {
throw new Error(tKey('wizard.error.execution_policy_invalid'));
}
var mode = raw.mode === undefined ? 'synchronous' : raw.mode;
var fallback = raw.no_tasks_fallback === undefined ? 'synchronous' : raw.no_tasks_fallback;
var deadline = Number(raw.deadline_ms);
if (['synchronous', 'background', 'adaptive'].indexOf(mode) < 0
|| ['synchronous', 'unsupported'].indexOf(fallback) < 0
|| !Number.isInteger(deadline) || deadline < 1 || deadline > 3600000) {
throw new Error(tKey('wizard.error.execution_policy_invalid'));
}
return { mode: mode, deadline_ms: deadline, no_tasks_fallback: fallback };
}
function currentUpstream() {
var selected = upstreams.find(function(item) { return item.id === selectedUpstreamId; });
if (selected) return selected;
@@ -415,6 +438,13 @@ function executionConfigToEditorValue(config) {
var value = {
timeout_ms: config.timeout_ms,
};
if (config.execution_policy) {
value.execution_policy = {
mode: config.execution_policy.mode,
deadline_ms: config.execution_policy.deadline_ms,
no_tasks_fallback: config.execution_policy.no_tasks_fallback,
};
}
if (config.retry_policy) {
value.retry = { max_attempts: config.retry_policy.max_attempts };
}
@@ -0,0 +1,139 @@
const { test, expect } = require('@playwright/test');
const { getCurrentWorkspace, login } = require('./helpers');
// Story 4.1 (FR-21): the published execution policy is part of the Tool version and must survive a
// wizard save. The editor is a free-form document, so a no-op edit must not erase it and an explicit
// invalid value must be reported instead of being silently replaced by defaults.
function draftSnapshot(operationId) {
return {
id: operationId,
name: operationId,
display_name: 'Policy rates',
category: 'general',
protocol: 'rest',
target: {
kind: 'rest',
base_url: 'https://api.example.test',
method: 'GET',
path_template: '/rates',
static_headers: {},
},
input_schema: {
type: 'object',
description: null,
required: false,
nullable: false,
default_value: null,
fields: {},
items: null,
enum_values: [],
variants: [],
},
output_schema: {
type: 'object',
description: null,
required: false,
nullable: false,
default_value: null,
fields: {},
items: null,
enum_values: [],
variants: [],
},
input_mapping: { rules: [] },
output_mapping: { rules: [{ source: '$.response.body.rate', target: '$.output.rate', required: true }] },
execution_config: {
timeout_ms: 10000,
retry_policy: null,
auth_profile_ref: null,
headers: {},
approval_policy: null,
execution_policy: {
mode: 'adaptive',
deadline_ms: 45000,
no_tasks_fallback: 'unsupported',
},
},
tool_description: { title: 'Policy rates', description: 'Get rates.', tags: [], examples: [] },
wizard_state: { input_sample: {}, output_sample: {}, test_input: {}, import_findings: [] },
};
}
test('wizard round-trips the published execution policy and rejects an invalid edit', async ({ page }) => {
await login(page);
const workspace = await getCurrentWorkspace(page);
const operationId = 'policy_rates_41';
const operationPath = `/api/admin/workspaces/${encodeURIComponent(workspace.id)}/operations/${operationId}`;
const savedPayloads = [];
await page.route(operationPath, async (route) => {
if (route.request().method() === 'PATCH') {
savedPayloads.push(route.request().postDataJSON());
await route.fulfill({
contentType: 'application/json',
body: JSON.stringify({ operation_id: operationId, workspace_id: workspace.id, version: 2, status: 'draft' }),
});
return;
}
await route.fulfill({
contentType: 'application/json',
body: JSON.stringify({
id: operationId,
workspace_id: workspace.id,
name: operationId,
display_name: 'Policy rates',
category: 'general',
protocol: 'rest',
status: 'draft',
current_draft_version: 1,
draft_version_ref: { version: 1, status: 'draft' },
published_version_ref: null,
agent_refs: [],
}),
});
});
await page.route(`${operationPath}/versions/*`, async (route) => {
await route.fulfill({
contentType: 'application/json',
body: JSON.stringify({
operation_id: operationId,
workspace_id: workspace.id,
version: 1,
status: 'draft',
snapshot: draftSnapshot(operationId),
}),
});
});
await page.route('**/operations/analyze-quality', async (route) => {
await route.fulfill({ contentType: 'application/json', body: JSON.stringify({ blocking: false, findings: [] }) });
});
page.on('dialog', (dialog) => dialog.accept());
await page.goto(`/wizard/?mode=edit&operationId=${operationId}`);
await expect(page.locator('#tool-display-name')).toHaveValue('Policy rates');
const editorValue = await page.locator('#tool-exec-config').inputValue();
expect(editorValue).toContain('execution_policy');
expect(editorValue).toContain('adaptive');
expect(editorValue).toContain('45000');
expect(editorValue).toContain('unsupported');
// A real edit still saves the pinned policy unchanged.
await page.evaluate(() => window.CrankWizardShell.doGoToStep(4));
await page.locator('#tool-display-name').fill('Policy rates edited');
await page.locator('.btn-save-draft').click();
await expect.poll(() => savedPayloads.length).toBe(1);
expect(savedPayloads[0].execution_config.execution_policy).toEqual({
mode: 'adaptive',
deadline_ms: 45000,
no_tasks_fallback: 'unsupported',
});
// An explicit invalid policy is refused instead of being silently rewritten to defaults.
await page.evaluate(() => window.CrankWizardShell.doGoToStep(5));
await expect(page.locator('#tool-exec-config')).toBeVisible();
await page.locator('#tool-exec-config').fill('timeout_ms: 10000\nexecution_policy:\n mode: background\n deadline_ms: 0\n');
await page.locator('.btn-save-draft').click();
await expect(page.locator('#tool-exec-config')).toHaveValue(/deadline_ms: 0/);
expect(savedPayloads.length).toBe(1);
});
+4 -1
View File
@@ -272,7 +272,10 @@ fn record_unknown_approval_access_denial(reason: &'static str) {
);
}
fn allows_scope(scopes: &[PlatformApiKeyScope], required_scope: PlatformApiKeyScope) -> bool {
pub(super) fn allows_scope(
scopes: &[PlatformApiKeyScope],
required_scope: PlatformApiKeyScope,
) -> bool {
match required_scope {
PlatformApiKeyScope::Read => scopes.iter().any(|scope| {
matches!(
+105 -16
View File
@@ -51,13 +51,15 @@ mod auth_resolution;
mod dispatch;
mod endpoint;
mod legacy_dispatch;
mod task_worker;
mod tasks;
use self::approvals::{
approve_request, deny_request, get_approval_request, list_pending_approvals,
};
pub(super) use self::auth_resolution::resolve_operation_auth;
use self::endpoint::{mcp_delete, mcp_get, mcp_post};
pub use self::task_worker::{TaskWorkerHandler, VerifiedIdempotentWrite};
pub use self::tasks::{TaskApplication, TaskApplicationError, TaskPrincipal};
// Test-only import surface: the BUILD-internal `app/tests.rs` module reaches these two
// through `super`, and the plain lib build does not use them (same pattern as
// `#[cfg(test)] use invocation_history::observe_invocation_history_outcome;` above).
@@ -92,6 +94,9 @@ pub(super) struct AppState {
pub(super) signal_streams: Arc<crate::protocol::StreamRegistry>,
/// Story 3.4: the bounded table of the `tools/call` requests a legacy client can still cancel.
pub(super) in_flight: Arc<self::lifecycle::InFlightCalls>,
/// Story 4.2: durable background Tasks over the same registry pool. The public wire stays
/// synchronous/unsupported; readiness depends on this mandatory store.
tasks: Arc<TaskApplication>,
}
#[derive(Clone)]
@@ -137,6 +142,7 @@ pub fn build_app(
DEFAULT_MAX_CONCURRENT_SESSIONS,
false,
None,
None,
)
}
@@ -192,6 +198,7 @@ pub fn build_app_with_background_workers_and_limits(
max_concurrent_sessions,
true,
None,
None,
)
}
@@ -227,6 +234,41 @@ pub fn build_app_with_artifacts(
max_concurrent_sessions,
true,
Some(artifact_store),
None,
)
}
/// Production launcher entry point with the durable Task application entry constructed by the
/// process composition root from the same registry pool and admission limits as the runtime.
#[allow(clippy::too_many_arguments)]
pub fn build_app_with_artifacts_and_tasks(
registry: PostgresRegistry,
refresh_interval: Duration,
public_base_url: Option<String>,
secret_crypto: SecretCrypto,
runtime: RuntimeExecutor,
api_rate_limiter: RequestRateLimiter,
coordination_store: Arc<dyn CoordinationStateStore>,
sessions: SharedSessionStore,
credential_verifier: SharedMachineCredentialVerifier,
max_concurrent_sessions: usize,
artifact_store: crank_artifacts::ArtifactStore,
tasks: Arc<TaskApplication>,
) -> Router {
build_app_inner(
registry,
refresh_interval,
public_base_url,
secret_crypto,
runtime,
api_rate_limiter,
coordination_store,
sessions,
credential_verifier,
max_concurrent_sessions,
true,
Some(artifact_store),
Some(tasks),
)
}
@@ -244,8 +286,23 @@ fn build_app_inner(
max_concurrent_sessions: usize,
start_background_workers: bool,
artifact_store: Option<crank_artifacts::ArtifactStore>,
tasks: Option<Arc<TaskApplication>>,
) -> Router {
let session_metrics = ActiveSessionMetrics::start(Arc::clone(&sessions));
let tasks = tasks.unwrap_or_else(|| {
let store = crank_registry::PostgresTaskStore::new(
registry.pool().clone(),
crank_registry::AdmissionLimits::default(),
crank_core::AdmissionSchedulerConfig::default(),
)
.expect("default admission limits and scheduler weights are valid");
Arc::new(TaskApplication::new(
registry.clone(),
refresh_interval,
Arc::clone(&coordination_store),
crank_runtime::TaskService::new(Arc::new(store)),
))
});
let state = Arc::new(AppState {
registry: registry.clone(),
catalog: PublishedToolCatalog::new(registry, refresh_interval, coordination_store),
@@ -263,6 +320,7 @@ fn build_app_inner(
in_flight: Arc::new(lifecycle::InFlightCalls::new(
lifecycle::MAX_IN_FLIGHT_CALLS,
)),
tasks,
});
if start_background_workers {
spawn_approval_recovery(Arc::clone(&state));
@@ -311,24 +369,55 @@ async fn health() -> Json<Value> {
}
async fn readiness(State(state): State<Arc<AppState>>) -> Response {
match state.registry.ping().await {
Ok(()) => Json(json!({
let postgres_ready = state.registry.ping().await.is_ok();
// Story 4.1: admission state is a mandatory dependency of every execution, so readiness must
// reflect it rather than reporting ready on a bare `select 1`.
let admission_ready = state.registry.execution_admission_health().await.is_ok();
// Story 4.2: the durable Task store is mandatory for a durable acceptance acknowledgement.
let durable_tasks_ready = state.tasks.health().await.is_ok();
if postgres_ready && admission_ready && durable_tasks_ready {
return Json(json!({
"service": "mcp-server",
"status": "ready",
"checks": { "postgres": "ready" }
"checks": {
"postgres": "ready",
"execution_admission": "ready",
"durable_tasks": "ready"
}
}))
.into_response(),
Err(_) => (
StatusCode::SERVICE_UNAVAILABLE,
Json(json!({
"service": "mcp-server",
"status": "not_ready",
"checks": { "postgres": "not_ready" },
"error": "database is unavailable"
})),
)
.into_response(),
.into_response();
}
let mut checks = serde_json::Map::new();
checks.insert(
"postgres".to_owned(),
json!(if postgres_ready { "ready" } else { "not_ready" }),
);
checks.insert(
"execution_admission".to_owned(),
json!(if admission_ready {
"ready"
} else {
"not_ready"
}),
);
checks.insert(
"durable_tasks".to_owned(),
json!(if durable_tasks_ready {
"ready"
} else {
"not_ready"
}),
);
(
StatusCode::SERVICE_UNAVAILABLE,
Json(json!({
"service": "mcp-server",
"status": "not_ready",
"checks": checks,
"error": "a mandatory dependency is unavailable"
})),
)
.into_response()
}
#[cfg(test)]
@@ -19,18 +19,31 @@ pub(crate) async fn resolve_operation_auth(
workspace_id: &crank_core::WorkspaceId,
execution_config: &crank_core::ExecutionConfig,
) -> Result<Option<ResolvedAuth>, RuntimeError> {
if execution_config.auth_profile_ref.is_none() {
return Ok(None);
}
let span = Stage::AuthResolve.span();
let result = resolve_runtime_auth_for_task(
resolve_operation_auth_with(
&state.registry,
&state.secret_crypto,
workspace_id,
execution_config,
)
.instrument(span.clone())
.await;
.await
}
/// Same resolution for callers that hold the authoritative registry/secret-crypto pair without the
/// HTTP application state (the durable Task worker).
pub(crate) async fn resolve_operation_auth_with(
registry: &PostgresRegistry,
secret_crypto: &SecretCrypto,
workspace_id: &crank_core::WorkspaceId,
execution_config: &crank_core::ExecutionConfig,
) -> Result<Option<ResolvedAuth>, RuntimeError> {
if execution_config.auth_profile_ref.is_none() {
return Ok(None);
}
let span = Stage::AuthResolve.span();
let result =
resolve_runtime_auth_for_task(registry, secret_crypto, workspace_id, execution_config)
.instrument(span.clone())
.await;
match &result {
Ok(_) => StageOutcome::Success.record(&span),
Err(_) => {
@@ -0,0 +1,330 @@
//! Application-level durable Task handler and attempt authority (Story 4.3/4.4).
//!
//! The handler owns everything that touches product authority: the exact published pin, the typed
//! Platform API key, the approval policy, the encrypted executable input and the resolved upstream
//! auth. The runtime owns the single bounded retry loop and the durable attempt checkpoint/resolution;
//! this module only revalidates and prepares one attempt at a time.
use std::time::{Duration, Instant};
use async_trait::async_trait;
use crank_core::{
AgentId, ExecutionOrigin, MachineAccessMode, OperationSecurityLevel, PlatformApiKeyScope,
TaskError, TaskLease, TaskOwnerKind, TraceContext, VerifiedMachineCredential, WorkspaceId,
current_unix_ms,
};
use crank_registry::PostgresRegistry;
use crank_registry::PublishedAgentTool;
use crank_runtime::{
ExecutionAuthorization, RuntimeExecutor, RuntimeOperation, SecretCrypto, TaskAttemptAuthority,
TaskAttemptContext, TaskAttemptOutcome, TaskAttemptPlan, TaskAttemptRefusal, TaskHandler,
TaskService, TaskWorkerError,
};
use sha2::{Digest, Sha256};
use super::TaskApplication;
use super::auth_resolution::resolve_operation_auth_with;
use super::tool_resolution::runtime_operation;
use crate::access::{allows_scope, credential_allows_security_level};
use crate::catalog::PublishedToolCatalog;
/// Server-owned evidence that one published operation deduplicates repeated sends of the same stable
/// idempotency key.
///
/// The descriptor is constructed by the application for an integration whose upstream contract was
/// actually qualified (key delivery, dedup scope, canonical input binding and a dedup window that
/// covers the pinned retry budget). It is never deserialized from Task input, so a caller key or
/// header alone can never make a write retryable. Production wiring passes an empty set until such a
/// contract exists.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct VerifiedIdempotentWrite {
pub operation_id: String,
pub operation_version: u32,
/// Header through which the stable key is delivered to this exact upstream.
pub key_header: String,
/// Dedup window of the upstream contract, which must cover the pinned retry budget.
pub validity_ms: i64,
}
pub struct TaskWorkerHandler {
registry: PostgresRegistry,
catalog: PublishedToolCatalog,
runtime: RuntimeExecutor,
secret_crypto: SecretCrypto,
service: TaskService,
verified_writes: Vec<VerifiedIdempotentWrite>,
}
impl TaskWorkerHandler {
/// Production construction from the same application entry, registry and runtime the MCP
/// process serves requests with. The worker needs no HTTP state, only these authoritative
/// dependencies.
pub fn from_application(
tasks: &TaskApplication,
registry: PostgresRegistry,
runtime: RuntimeExecutor,
secret_crypto: SecretCrypto,
) -> Self {
Self {
registry,
catalog: tasks.catalog(),
runtime,
secret_crypto,
service: tasks.task_service(),
verified_writes: Vec::new(),
}
}
/// Registers server-owned verified-idempotency evidence of qualified integrations.
pub fn with_verified_writes(mut self, verified_writes: Vec<VerifiedIdempotentWrite>) -> Self {
self.verified_writes = verified_writes;
self
}
/// Authoritative pre-dispatch authorization of one leased Task (Story 4.3).
///
/// Resolves the exact pinned Published Version in the current publication, repeats the approval
/// policy and revalidates the typed persisted Platform API key identity against the registry:
/// the key must still be active, unexpired and bound to this workspace/agent, and its exact
/// scopes/security level must still authorize the operation. Any denial is a typed error, so the
/// worker finishes the attempt without dispatch (fail closed before any network use).
///
/// An administrator Task has no authoritative Task revalidation API in the registry, so it stays
/// fail-closed here instead of inventing authority.
async fn authorize(&self, lease: &TaskLease) -> Result<PublishedAgentTool, TaskError> {
if lease.record.owner_kind != TaskOwnerKind::Agent {
return Err(TaskError::invalid(
"admin task authority revalidation is not implemented",
));
}
let Some(key_id) = lease.record.platform_api_key_id.as_ref() else {
return Err(TaskError::invalid("task authority reference is missing"));
};
let workspace_id = WorkspaceId::new(lease.record.workspace_id.as_str());
let agent_id = AgentId::new(lease.record.owner_id.as_str());
let workspace = self
.registry
.get_workspace(&workspace_id)
.await
.map_err(|_| TaskError::unavailable("published catalog unavailable"))?
.ok_or_else(|| TaskError::invalid("task workspace no longer exists"))?;
let agents = self
.registry
.list_agents(&workspace_id)
.await
.map_err(|_| TaskError::unavailable("published catalog unavailable"))?;
let agent = agents
.into_iter()
.find(|agent| agent.id.as_str() == lease.record.owner_id.as_str())
.ok_or_else(|| TaskError::invalid("task agent no longer exists"))?;
let catalog = self
.catalog
.get_catalog(&workspace.workspace.slug, &agent.slug)
.await
.map_err(|_| TaskError::unavailable("published catalog unavailable"))?;
let tool = catalog
.tools
.into_iter()
.find(|tool| {
tool.operation.id.as_str() == lease.record.operation_id.as_str()
&& tool.operation.version == lease.record.pins.operation_version
})
.ok_or_else(|| TaskError::invalid("task published operation pin no longer matches"))?;
if tool.workspace_id.as_str() != lease.record.workspace_id.as_str() {
return Err(TaskError::invalid("task workspace binding mismatch"));
}
if tool.agent_id.as_str() != lease.record.owner_id.as_str() {
return Err(TaskError::invalid("task agent binding mismatch"));
}
if tool
.operation
.execution_config
.approval_policy
.as_ref()
.is_some_and(|policy| policy.required)
{
return Err(TaskError::invalid(
"missing approval proof for task operation",
));
}
// Authoritative revalidation of the exact immutable key identity. A revoked, expired,
// foreign-workspace or foreign-agent key is absent here and never reaches the network.
let key = self
.registry
.revalidate_platform_api_key(&workspace_id, &agent_id, key_id)
.await
.map_err(|_| TaskError::unavailable("task authority store unavailable"))?
.ok_or_else(|| TaskError::invalid("task authority is no longer active"))?;
let credential = VerifiedMachineCredential {
machine_access_mode: MachineAccessMode::StaticAgentKey,
max_security_level: OperationSecurityLevel::Standard,
scopes: key.api_key.scopes.clone(),
platform_api_key_id: Some(key.api_key.id.clone()),
};
if !allows_scope(&credential.scopes, PlatformApiKeyScope::Write) {
return Err(TaskError::invalid("task authority scope is insufficient"));
}
if !credential_allows_security_level(&credential, tool.operation.security_level) {
return Err(TaskError::invalid(
"task authority security level is insufficient",
));
}
Ok(tool)
}
/// Safe-retry classification of one attempt and the stable upstream key digest.
///
/// A read is always safe to repeat. A write is only safe to repeat when this process holds
/// server-owned evidence for the exact published operation and that evidence's dedup window still
/// covers the accepted retry budget; the presence of a client key or an idempotency header alone
/// never qualifies.
fn retry_classification(
&self,
operation: &RuntimeOperation,
input: &serde_json::Value,
retry_window_ms: i64,
) -> (bool, Option<String>) {
if matches!(
&operation.target,
crank_core::Target::Rest(target) if target.method == crank_core::HttpMethod::Get
) {
return (true, None);
}
let Some(policy) = operation.execution_config.idempotency.as_ref() else {
return (false, None);
};
let Some(field) = policy
.input_field
.as_deref()
.filter(|field| !field.trim().is_empty())
else {
return (false, None);
};
let Some(header) = policy
.header_name
.as_deref()
.filter(|value| !value.is_empty())
else {
return (false, None);
};
let key = match input.get(field) {
Some(serde_json::Value::String(value)) if !value.trim().is_empty() => value.clone(),
Some(serde_json::Value::Number(value)) => value.to_string(),
_ => return (false, None),
};
let evidence = self.verified_writes.iter().find(|evidence| {
evidence.operation_id == operation.operation_id.as_str()
&& evidence.operation_version == operation.operation_version
&& evidence.key_header.eq_ignore_ascii_case(header)
&& evidence.validity_ms >= retry_window_ms
});
let Some(evidence) = evidence else {
return (false, None);
};
// Only a digest of the stable key is persisted; the key itself stays in the encrypted input.
let mut digest = Sha256::new();
digest.update(evidence.key_header.as_bytes());
digest.update([0u8]);
digest.update(key.as_bytes());
(true, Some(format!("{:x}", digest.finalize())))
}
/// Maps a preparation failure onto the attempt refusal vocabulary: a transient authority-store
/// outage leaves the Task leased for recovery, everything else is a proven pre-dispatch refusal.
fn refusal(error: TaskError) -> TaskAttemptRefusal {
match error {
TaskError::Unavailable { .. } => TaskAttemptRefusal::Stopped,
_ => TaskAttemptRefusal::Failed {
code: "task_authority_rejected".to_owned(),
},
}
}
}
#[async_trait]
impl TaskHandler for TaskWorkerHandler {
async fn revalidate(&self, lease: &TaskLease) -> Result<(), TaskWorkerError> {
self.authorize(lease)
.await
.map_err(TaskWorkerError::Store)?;
Ok(())
}
async fn execute(&self, context: TaskAttemptContext) -> TaskAttemptOutcome {
self.runtime.execute_task_outcome(&context, self).await
}
}
#[async_trait]
impl TaskAttemptAuthority for TaskWorkerHandler {
async fn prepare_attempt(
&self,
lease: &TaskLease,
) -> Result<TaskAttemptPlan, TaskAttemptRefusal> {
let tool = self.authorize(lease).await.map_err(Self::refusal)?;
let input = self
.service
.executable_input(lease)
.await
.map_err(Self::refusal)?;
let operation = runtime_operation(&tool);
let now_ms = current_unix_ms();
let deadline_ms = lease.record.pins.execution_deadline_unix_ms;
if deadline_ms <= now_ms {
// The accepted absolute deadline already elapsed: no further attempt may be dispatched,
// and any previous uncertainty was already recorded as such.
return Err(TaskAttemptRefusal::Failed {
code: "task_deadline_elapsed".to_owned(),
});
}
let deadline = Instant::now() + Duration::from_millis((deadline_ms - now_ms) as u64);
let workspace_id = WorkspaceId::new(lease.record.workspace_id.as_str());
let agent_id = AgentId::new(lease.record.owner_id.as_str());
let resolved_auth = resolve_operation_auth_with(
&self.registry,
&self.secret_crypto,
&workspace_id,
&operation.execution_config,
)
.await
.map_err(|_| TaskAttemptRefusal::Failed {
code: "task_auth_resolution_failed".to_owned(),
})?;
// Continue the parent trace persisted with the accepted Task; every attempt adds its own
// child `adapter_attempt` span carrying the committed durable attempt ID.
let trace_context =
TraceContext::parse(&lease.record.support.traceparent).map_err(|_| {
TaskAttemptRefusal::Failed {
code: "task_trace_parent_invalid".to_owned(),
}
})?;
let request_context = crank_runtime::RuntimeRequestContext::new(
lease.record.support.request_id.clone(),
trace_context,
)
.with_metering_context(
workspace_id.clone(),
Some(agent_id.clone()),
crank_core::InvocationSource::AgentToolCall,
);
let (retry_allowed, upstream_key_digest) =
self.retry_classification(&operation, &input, deadline_ms - now_ms);
Ok(TaskAttemptPlan {
workspace_id,
origin: ExecutionOrigin::AgentSnapshot,
agent_id: Some(agent_id),
operation,
input,
authorization: ExecutionAuthorization::Authorized,
resolved_auth,
request_context,
deadline,
retry_allowed,
upstream_key_digest,
})
}
}
+377
View File
@@ -0,0 +1,377 @@
//! Application entry of durable background Tasks (Story 4.2, FR-22/FR-23).
//!
//! The entry takes an already authenticated principal (a verified machine credential, or an
//! administrator identity verified by the calling admin surface), resolves the exact published Tool,
//! repeats the caller's authorization, pins the Published Version/policy/deadline and delegates the
//! durable acceptance to the runtime Task service. The repository owns the transaction, so the
//! returned Task ID exists only after Task + dispatch intent + queue slot committed.
//!
//! Deliberately not wired: the public MCP `tools/call`/Tasks wire (Story 4.7) and the worker/leases
//! (Story 4.3). The existing synchronous/unsupported `tools/call` path is unchanged, and this entry
//! never spawns an ephemeral execution future.
use std::sync::Arc;
use std::time::{Duration, Instant};
use crank_core::{
CorrelationContext, PlatformApiKeyId, TaskAcceptance, TaskAccess, TaskError, TaskId,
TaskOwnerKind, TaskRecord, ToolExecutionMode,
};
use crank_registry::{PostgresRegistry, PublishedAgentTool};
use crank_runtime::{TaskAcceptDraft, TaskOutcome, TaskService};
use serde_json::Value;
use crate::access::credential_allows_security_level;
use crate::catalog::PublishedToolCatalog;
/// Authenticated principal of one Task operation. The authorization scope identity is derived here
/// from the verified credential/admin identity, never from a caller-supplied request string.
#[derive(Clone, Debug)]
pub enum TaskPrincipal {
Agent {
workspace_slug: String,
agent_slug: String,
credential: crank_core::VerifiedMachineCredential,
},
/// Administrator identity that the admin surface already authenticated. It is never a wildcard:
/// owner, workspace and authorization scope all participate in the stored predicate.
Admin {
workspace_slug: String,
agent_slug: String,
user_id: String,
},
}
impl TaskPrincipal {
fn workspace_slug(&self) -> &str {
match self {
Self::Agent { workspace_slug, .. } | Self::Admin { workspace_slug, .. } => {
workspace_slug
}
}
}
fn agent_slug(&self) -> &str {
match self {
Self::Agent { agent_slug, .. } | Self::Admin { agent_slug, .. } => agent_slug,
}
}
fn owner(&self, tool: &PublishedAgentTool) -> (TaskOwnerKind, String) {
match self {
Self::Agent { .. } => (TaskOwnerKind::Agent, tool.agent_id.as_str().to_owned()),
Self::Admin { user_id, .. } => (TaskOwnerKind::Admin, user_id.clone()),
}
}
/// Typed immutable Platform API key identity of this principal. A verified agent credential
/// without a key ID cannot be revalidated by a worker, so it is refused instead of accepted;
/// an administrator has no machine key and is stored as `None` (the worker fails closed).
fn platform_api_key_id(&self) -> Result<Option<PlatformApiKeyId>, TaskApplicationError> {
match self {
Self::Agent { credential, .. } => credential
.platform_api_key_id
.clone()
.map(Some)
.ok_or(TaskApplicationError::Denied),
Self::Admin { .. } => Ok(None),
}
}
/// Bounded server-derived authorization scope identity of this principal.
fn authorization_scope(&self) -> String {
let identity = match self {
Self::Agent { credential, .. } => {
let mut scopes = credential
.scopes
.iter()
.map(|scope| format!("{scope:?}"))
.collect::<Vec<_>>();
scopes.sort();
format!(
"key:{}|mode:{:?}|level:{:?}|scopes:{}",
credential
.platform_api_key_id
.as_ref()
.map(|id| id.as_str().to_owned())
.unwrap_or_else(|| "-".to_owned()),
credential.machine_access_mode,
credential.max_security_level,
scopes.join(",")
)
}
Self::Admin { user_id, .. } => format!("admin:{user_id}"),
};
identity.chars().take(160).collect()
}
}
/// Typed refusal of the application entry. `NotFound`/`Denied` never reveal whether a foreign Task
/// exists, and no variant carries Task payload or fingerprint.
#[derive(Clone, Debug, PartialEq, Eq, thiserror::Error)]
pub enum TaskApplicationError {
#[error("task request was not found")]
NotFound,
#[error("task request is not authorized")]
Denied,
#[error("the published Tool policy does not accept durable background work")]
NotBackground,
#[error("the requested published Tool version is not available")]
VersionUnavailable,
#[error("this Tool requires a verified approval before acceptance")]
ApprovalRequired,
#[error(transparent)]
Task(#[from] TaskError),
}
#[derive(Clone)]
pub struct TaskApplication {
catalog: PublishedToolCatalog,
service: TaskService,
}
impl TaskApplication {
/// Production construction of the application entry over the process registry, the published
/// catalog and the durable Task service.
pub fn new(
registry: PostgresRegistry,
refresh_interval: Duration,
coordination_store: Arc<dyn crank_core::CoordinationStateStore>,
service: TaskService,
) -> Self {
Self {
catalog: PublishedToolCatalog::new(registry, refresh_interval, coordination_store),
service,
}
}
/// Runtime task service carrying encrypted executable input.
pub fn task_service(&self) -> TaskService {
self.service.clone()
}
/// Published catalog of the same application entry, for the durable Task worker's pre-dispatch
/// revalidation.
pub(super) fn catalog(&self) -> PublishedToolCatalog {
self.catalog.clone()
}
/// Readiness of the durable Task store this entry depends on.
pub async fn health(&self) -> Result<(), TaskError> {
self.service.health().await
}
/// Accepts one background request for an already authenticated principal.
///
/// `expected_version` is the caller's explicit Published Version pin. Without it the current
/// publication is pinned for new work and a retry of the same scoped key returns the stored Task
/// even after a newer publication. An explicit pin may only retry the Task accepted with that
/// exact version: a different pin is a typed conflict, never a rebind.
pub async fn accept_background(
&self,
principal: &TaskPrincipal,
tool_name: &str,
expected_version: Option<u32>,
scoped_key: &str,
input: &Value,
correlation: &CorrelationContext,
) -> Result<TaskAcceptance, TaskApplicationError> {
let tool = self.resolve_authorized_tool(principal, tool_name).await?;
let current_version = tool.operation.version;
let Some(policy) = tool.operation.execution_config.execution_policy else {
return Err(TaskApplicationError::NotBackground);
};
if policy.mode == ToolExecutionMode::Synchronous {
return Err(TaskApplicationError::NotBackground);
}
if tool
.operation
.execution_config
.approval_policy
.as_ref()
.is_some_and(|approval| approval.required)
{
// Without a verified, non-consuming approval reference this entry fails closed instead
// of accepting work a worker could never safely execute.
return Err(TaskApplicationError::ApprovalRequired);
}
let platform_api_key_id = principal.platform_api_key_id()?;
let draft = |operation_version: u32| {
let (owner_kind, owner_id) = principal.owner(&tool);
TaskAcceptDraft {
workspace_id: tool.workspace_id.as_str().to_owned(),
owner_kind,
owner_id,
authorization_scope: principal.authorization_scope(),
platform_api_key_id: platform_api_key_id.clone(),
operation_id: tool.operation.id.as_str().to_owned(),
operation_version,
policy_mode: policy.mode,
scoped_key: scoped_key.to_owned(),
input: input.clone(),
correlation: correlation.clone(),
deadline: Instant::now() + Duration::from_millis(policy.deadline_ms.max(1)),
retention_ms: TaskService::default_retention_ms(),
approval_reference: None,
retry: task_retry_policy(&tool),
}
};
let Some(requested) = expected_version else {
return Ok(self.service.accept(draft(current_version)).await?);
};
let draft = draft(requested);
match self.service.lookup(&draft).await? {
Some(record) => {
if record.input_fingerprint != TaskService::input_fingerprint(input)? {
return Err(TaskApplicationError::Task(TaskError::FingerprintConflict));
}
if record.pins.operation_version != requested {
return Err(TaskApplicationError::VersionUnavailable);
}
Ok(TaskAcceptance {
task_id: record.task_id,
state: record.state,
operation_version: record.pins.operation_version,
support: record.support,
created_at_unix_ms: record.created_at_unix_ms,
retain_until_unix_ms: record.pins.retain_until_unix_ms,
replayed: true,
})
}
None if requested == current_version => Ok(self.service.accept(draft).await?),
// A new Task can only be accepted against the currently published version.
None => Err(TaskApplicationError::VersionUnavailable),
}
}
/// Owner-scoped status read of one Task of the addressed published Tool.
pub async fn status(
&self,
principal: &TaskPrincipal,
tool_name: &str,
task_id: TaskId,
) -> Result<TaskRecord, TaskApplicationError> {
let access = self.task_access(principal, tool_name).await?;
match self.service.status(&access, task_id).await {
Ok(record) => Ok(record),
Err(TaskError::NotFound) => Err(TaskApplicationError::NotFound),
Err(error) => Err(TaskApplicationError::Task(error)),
}
}
/// Owner-scoped result projection of one Task of the addressed published Tool.
pub async fn result(
&self,
principal: &TaskPrincipal,
tool_name: &str,
task_id: TaskId,
) -> Result<TaskOutcome, TaskApplicationError> {
let access = self.task_access(principal, tool_name).await?;
match self.service.result(&access, task_id).await {
Ok(outcome) => Ok(outcome),
Err(TaskError::NotFound) => Err(TaskApplicationError::NotFound),
Err(error) => Err(TaskApplicationError::Task(error)),
}
}
/// Owner-scoped cancellation of one Task of the addressed published Tool (Story 4.4).
///
/// Cancellation is decided by the durable store under a row lock: before the dispatch checkpoint
/// it is an honest cancellation with zero outbound calls, after it the Task becomes the
/// non-terminal uncertain outcome instead of a plain cancel or a retry.
pub async fn cancel(
&self,
principal: &TaskPrincipal,
tool_name: &str,
task_id: TaskId,
expected_version: Option<u32>,
) -> Result<crank_core::TaskCancelOutcome, TaskApplicationError> {
let access = self.task_access(principal, tool_name).await?;
match self
.service
.cancel(&access, task_id, expected_version, "user_cancelled")
.await
{
Ok(outcome) => Ok(outcome),
Err(TaskError::NotFound) => Err(TaskApplicationError::NotFound),
Err(error) => Err(TaskApplicationError::Task(error)),
}
}
/// Owner-scoped attempt history of one Task of the addressed published Tool.
pub async fn attempts(
&self,
principal: &TaskPrincipal,
tool_name: &str,
task_id: TaskId,
) -> Result<Vec<crank_core::TaskAttempt>, TaskApplicationError> {
let access = self.task_access(principal, tool_name).await?;
match self.service.attempts(&access, task_id).await {
Ok(attempts) => Ok(attempts),
Err(TaskError::NotFound) => Err(TaskApplicationError::NotFound),
Err(error) => Err(TaskApplicationError::Task(error)),
}
}
async fn task_access(
&self,
principal: &TaskPrincipal,
tool_name: &str,
) -> Result<TaskAccess, TaskApplicationError> {
let tool = self.resolve_authorized_tool(principal, tool_name).await?;
let (owner_kind, owner_id) = principal.owner(&tool);
Ok(TaskAccess {
workspace_id: tool.workspace_id.as_str().to_owned(),
owner_kind,
owner_id,
authorization_scope: principal.authorization_scope(),
})
}
/// Resolves the addressed Tool of the principal's published catalog and repeats its authorization
/// check. A foreign workspace, an unknown Tool and an insufficient security level are typed
/// refusals without metadata.
async fn resolve_authorized_tool(
&self,
principal: &TaskPrincipal,
tool_name: &str,
) -> Result<PublishedAgentTool, TaskApplicationError> {
let catalog = self
.catalog
.get_catalog(principal.workspace_slug(), principal.agent_slug())
.await
.map_err(|_| TaskApplicationError::NotFound)?;
let tool = catalog
.tools
.into_iter()
.find(|tool| tool.tool_name == tool_name)
.ok_or(TaskApplicationError::NotFound)?;
if let TaskPrincipal::Admin { workspace_slug, .. } = principal
&& tool.workspace_slug != *workspace_slug
{
return Err(TaskApplicationError::Denied);
}
if let TaskPrincipal::Agent { credential, .. } = principal
&& !credential_allows_security_level(credential, tool.operation.security_level)
{
return Err(TaskApplicationError::Denied);
}
Ok(tool)
}
}
/// Bounded retry snapshot pinned from the exact published operation at acceptance. A later
/// publication, a restart or a new binary never changes the numbers an accepted Task was admitted
/// with; a missing or oversized policy is clamped to the closed story bound.
fn task_retry_policy(tool: &PublishedAgentTool) -> crank_core::TaskRetryPolicy {
let max_attempts = tool
.operation
.execution_config
.retry_policy
.as_ref()
.map_or(1, |policy| {
policy.max_attempts.clamp(1, crank_core::TASK_MAX_ATTEMPTS)
});
crank_core::TaskRetryPolicy::new(max_attempts, 1)
}
@@ -550,6 +550,7 @@ pub(super) fn test_operation(base_url: &str) -> Operation<Schema, MappingSet> {
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
},
+3 -1
View File
@@ -16,6 +16,8 @@ mod tool_search;
mod transport;
pub use app::{
build_app, build_app_with_artifacts, build_app_with_background_workers,
TaskApplication, TaskApplicationError, TaskPrincipal, TaskWorkerHandler,
VerifiedIdempotentWrite, build_app, build_app_with_artifacts,
build_app_with_artifacts_and_tasks, build_app_with_background_workers,
build_app_with_background_workers_and_limits,
};
@@ -136,6 +136,7 @@ fn published_tool(name: &str, title: &str, description: &str, version: u32) -> P
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
},
@@ -284,6 +284,7 @@ fn operation() -> RegistryOperation {
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
},
+531
View File
@@ -0,0 +1,531 @@
//! Typed execution-admission seam (Story 4.1, FR-21/FR-24).
//!
//! Admission is the single gate in front of every real Tool execution: it decides whether the
//! caller receives a scoped permit or a durable enqueue decision, and it owns the deployment-wide
//! global/per-Agent/per-Tool capacity and the bounded waiting queue. The authoritative state lives
//! in PostgreSQL (`crank-registry`), so several processes share one budget.
//!
//! Deliberately absent from this contract: credentials, connection or protocol revision (they must
//! never widen a quota), request bodies and prepared headers (payload ownership is Story 4.2).
use async_trait::async_trait;
use serde::{Deserialize, Serialize};
use crate::{NoTasksFallback, ToolExecutionMode};
pub const ADMISSION_FINGERPRINT_MAX_BYTES: usize = 128;
pub const ADMISSION_OWNER_TOKEN_MAX_BYTES: usize = 64;
pub const MAX_RETRY_AFTER_MS: u64 = 60_000;
/// Bounded retry guidance for an admission refusal (capacity or queue wait exhausted). One value for
/// the store and every runtime so the client contract stays stable.
pub const ADMISSION_OVERLOAD_RETRY_AFTER_MS: u64 = 250;
/// Current wall clock in unix milliseconds (absolute deadlines are persisted as unix ms).
pub fn current_unix_ms() -> i64 {
i64::try_from(
std::time::SystemTime::now()
.duration_since(std::time::UNIX_EPOCH)
.map(|elapsed| elapsed.as_millis())
.unwrap_or_default(),
)
.unwrap_or(i64::MAX)
}
/// Bucket used for an execution that has no Agent (admin draft/test-run): an explicit admin origin
/// bucket that is still bounded by the global limit.
pub const ADMIN_ORIGIN_KEY: &str = "admin";
/// Exact execution identity pinned at admission time.
#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
pub struct AdmissionScope {
pub workspace_id: String,
/// Agent id, or [`ADMIN_ORIGIN_KEY`] for an Agent-less administrative execution.
pub agent_key: String,
pub operation_id: String,
/// Quota keys exclude the version; the version is pinned for execution semantics only.
pub operation_version: u32,
}
impl AdmissionScope {
/// Quota key of the Agent bucket: `(workspace, agent)`.
pub fn agent_bucket(&self) -> String {
format!("{}:{}", self.workspace_id, self.agent_key)
}
/// Quota key of the Tool bucket: `(workspace, operation)` without version.
pub fn tool_bucket(&self) -> String {
format!("{}:{}", self.workspace_id, self.operation_id)
}
}
#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum AdmissionMode {
/// The caller waits for a scoped permit and dispatches inside it.
Synchronous,
/// The store commits a durable enqueue decision and reserves queue capacity.
Background,
}
impl AdmissionMode {
pub const fn as_str(self) -> &'static str {
match self {
Self::Synchronous => "synchronous",
Self::Background => "background",
}
}
}
/// Non-consuming approval evidence gathered by the runtime before admission.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum ApprovalEvidence {
/// The operation does not require confirmation.
NotRequired,
/// A verified approval exists for this exact scope and input.
Granted,
/// Confirmation is required and no verified approval exists yet.
NotGranted,
}
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct AdmissionRequest {
pub scope: AdmissionScope,
pub mode: AdmissionMode,
/// Published policy mode that produced this request.
pub policy_mode: ToolExecutionMode,
pub no_tasks_fallback: NoTasksFallback,
pub approval: ApprovalEvidence,
/// Bounded request identity used only to correlate one admission entry.
pub fingerprint: String,
/// Absolute deadline (unix milliseconds), already including queue waiting time.
pub expires_at_unix_ms: i64,
}
impl AdmissionRequest {
pub fn validate(&self) -> Result<(), AdmissionStoreError> {
if self.fingerprint.is_empty()
|| self.fingerprint.len() > ADMISSION_FINGERPRINT_MAX_BYTES
|| self
.fingerprint
.bytes()
.any(|byte| byte.is_ascii_control() || byte == b' ')
{
return Err(AdmissionStoreError::invalid(
"admission fingerprint is outside its allowed bounds",
));
}
if self.scope.workspace_id.is_empty() || self.scope.workspace_id.len() > 160 {
return Err(AdmissionStoreError::invalid(
"admission workspace is invalid",
));
}
if self.scope.agent_key.is_empty() || self.scope.agent_key.len() > 160 {
return Err(AdmissionStoreError::invalid("admission agent is invalid"));
}
if self.scope.operation_id.is_empty() || self.scope.operation_id.len() > 160 {
return Err(AdmissionStoreError::invalid(
"admission operation is invalid",
));
}
if self.scope.operation_version == 0 {
return Err(AdmissionStoreError::invalid(
"admission operation version must be positive",
));
}
if self.expires_at_unix_ms <= 0 {
return Err(AdmissionStoreError::invalid(
"admission deadline is invalid",
));
}
Ok(())
}
}
/// Scoped permit. Held for the whole local execution future (including retries) and released
/// explicitly; a dropped permit is cleaned up by a bounded owner-token CAS, never reused by TTL.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct AdmissionPermit {
pub ticket_id: u64,
pub owner_token: String,
pub scope: AdmissionScope,
pub mode: AdmissionMode,
pub policy_mode: ToolExecutionMode,
}
/// Durable enqueue decision: a reserved queue slot with exact Published Version semantics.
/// It is not a public Task id and it never authorises a worker dispatch (Story 4.2).
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct AdmissionDecision {
pub ticket_id: u64,
pub scope: AdmissionScope,
pub expires_at_unix_ms: i64,
pub policy_mode: ToolExecutionMode,
}
#[derive(Clone, Debug, PartialEq, Eq)]
pub enum AdmissionOutcome {
Permitted(AdmissionPermit),
Enqueued(AdmissionDecision),
/// Capacity or queue bounds were exhausted before any unbounded allocation.
Overloaded {
retry_after_ms: u64,
},
/// The queue wait reached the absolute deadline; nothing was dispatched.
DeadlineElapsed,
/// Confirmation is mandatory for this operation and no verified approval exists.
ApprovalRequired,
/// The published policy does not advertise durable Tasks and refuses the fallback.
Unsupported,
}
#[derive(Clone, Debug, Default, PartialEq, Eq)]
pub struct AdmissionQueueSnapshot {
pub waiting: u64,
pub executing: u64,
/// Age of the oldest waiting request in the authoritative queue, in milliseconds.
pub oldest_waiting_ms: Option<u64>,
}
#[derive(Clone, Debug, PartialEq, Eq, thiserror::Error)]
pub enum AdmissionStoreError {
/// The mandatory safety/capacity store cannot confirm state; every dependent path fails closed.
#[error("execution admission store is unavailable: {message}")]
Unavailable { message: String },
/// The store rejected a malformed admission request.
#[error("execution admission request is invalid: {message}")]
Invalid { message: String },
}
impl AdmissionStoreError {
pub fn unavailable(message: impl Into<String>) -> Self {
Self::Unavailable {
message: message.into(),
}
}
pub fn invalid(message: impl Into<String>) -> Self {
Self::Invalid {
message: message.into(),
}
}
}
#[async_trait]
pub trait ExecutionAdmission: Send + Sync {
/// Proves the mandatory admission state is usable (schema and operations reachable).
async fn health(&self) -> Result<(), AdmissionStoreError>;
/// Waits (bounded by `expires_at_unix_ms`) for a scoped permit, or commits a durable enqueue
/// decision for a background request, or returns a typed refusal.
async fn admit(
&self,
request: AdmissionRequest,
) -> Result<AdmissionOutcome, AdmissionStoreError>;
/// Releases the caller's own permit. Idempotent: a repeated release returns `false`.
async fn release(&self, permit: &AdmissionPermit) -> Result<bool, AdmissionStoreError>;
/// Removes the caller's own waiting ticket. Idempotent and never touches a foreign ticket.
async fn cancel(&self, ticket_id: u64, owner_token: &str) -> Result<bool, AdmissionStoreError>;
/// Bounded authoritative queue readings for low-cardinality metrics.
async fn snapshot(&self) -> Result<AdmissionQueueSnapshot, AdmissionStoreError>;
}
/// One waiting entry as seen by the fairness scheduler.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct AdmissionQueueEntry {
pub ticket_id: u64,
/// Authoritative insertion order inside `(Agent, Tool)`; never rewritten.
pub sequence: u64,
pub agent_key: String,
pub operation_id: String,
}
/// Weighted round-robin state persisted with the authoritative queue.
#[derive(Clone, Debug, Default, PartialEq, Eq)]
pub struct AdmissionCursor {
pub agent: Option<String>,
pub agent_credit: u32,
pub tool: Option<String>,
pub tool_credit: u32,
}
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct AdmissionSchedulerConfig {
pub agent_weight: u32,
pub tool_weight: u32,
}
impl Default for AdmissionSchedulerConfig {
fn default() -> Self {
Self {
agent_weight: 1,
tool_weight: 1,
}
}
}
impl AdmissionSchedulerConfig {
pub fn validate(&self) -> Result<(), AdmissionStoreError> {
if !(1..=64).contains(&self.agent_weight) || !(1..=64).contains(&self.tool_weight) {
return Err(AdmissionStoreError::invalid(
"admission weights must be between 1 and 64",
));
}
Ok(())
}
}
/// Result of one fair-selection step: the queue index to grant and the cursor to persist when the
/// grant actually happens. The cursor is not mutated on a blocked head, so a blocked head keeps its
/// turn instead of being skipped by newer arrivals.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct AdmissionSelection {
pub index: usize,
pub cursor: AdmissionCursor,
}
/// Hierarchical weighted round-robin: Agents first, then the Tools of the chosen Agent, FIFO inside
/// `(Agent, Tool)`. A new waiter never resets the cursor and never inherits idle credit.
pub fn select_next_grant(
queue: &[AdmissionQueueEntry],
cursor: &AdmissionCursor,
config: AdmissionSchedulerConfig,
) -> Option<AdmissionSelection> {
if queue.is_empty() {
return None;
}
let agent_weight = config.agent_weight.max(1);
let tool_weight = config.tool_weight.max(1);
let mut agents: Vec<&str> = queue.iter().map(|entry| entry.agent_key.as_str()).collect();
agents.sort_unstable();
agents.dedup();
let mut next = cursor.clone();
let agent = match cursor.agent.as_deref() {
Some(current) if cursor.agent_credit > 0 && agents.contains(&current) => current.to_owned(),
_ => {
let chosen = match cursor.agent.as_deref() {
None => agents[0].to_owned(),
Some(current) => agents
.iter()
.copied()
.find(|candidate| *candidate > current)
.unwrap_or(agents[0])
.to_owned(),
};
if cursor.agent.as_deref() != Some(chosen.as_str()) {
next.tool = None;
next.tool_credit = 0;
}
next.agent = Some(chosen.clone());
next.agent_credit = agent_weight;
chosen
}
};
let mut tools: Vec<&str> = queue
.iter()
.filter(|entry| entry.agent_key == agent)
.map(|entry| entry.operation_id.as_str())
.collect();
tools.sort_unstable();
tools.dedup();
let tool = match cursor.tool.as_deref() {
Some(current) if cursor.tool_credit > 0 && tools.contains(&current) => current.to_owned(),
_ => {
let chosen = match cursor.tool.as_deref() {
None => tools[0].to_owned(),
Some(current) => tools
.iter()
.copied()
.find(|candidate| *candidate > current)
.unwrap_or(tools[0])
.to_owned(),
};
next.tool = Some(chosen.clone());
next.tool_credit = tool_weight;
chosen
}
};
let index = queue
.iter()
.position(|entry| entry.agent_key == agent && entry.operation_id == tool)?;
next.agent_credit = next.agent_credit.saturating_sub(1);
next.tool_credit = next.tool_credit.saturating_sub(1);
Some(AdmissionSelection {
index,
cursor: next,
})
}
#[cfg(test)]
mod tests {
use super::{
ADMIN_ORIGIN_KEY, AdmissionCursor, AdmissionMode, AdmissionRequest,
AdmissionSchedulerConfig, AdmissionScope, AdmissionStoreError, ApprovalEvidence,
select_next_grant,
};
use crate::{NoTasksFallback, ToolExecutionMode};
fn entry(ticket_id: u64, agent: &str, tool: &str) -> super::AdmissionQueueEntry {
super::AdmissionQueueEntry {
ticket_id,
sequence: ticket_id,
agent_key: agent.to_owned(),
operation_id: tool.to_owned(),
}
}
fn request() -> AdmissionRequest {
AdmissionRequest {
scope: AdmissionScope {
workspace_id: "ws_default".to_owned(),
agent_key: ADMIN_ORIGIN_KEY.to_owned(),
operation_id: "op_leads".to_owned(),
operation_version: 3,
},
mode: AdmissionMode::Synchronous,
policy_mode: ToolExecutionMode::Synchronous,
no_tasks_fallback: NoTasksFallback::Synchronous,
approval: ApprovalEvidence::NotRequired,
fingerprint: "fp_0123456789".to_owned(),
expires_at_unix_ms: 1_790_000_000_000,
}
}
#[test]
fn quota_keys_exclude_version_and_keep_agent_bucket() {
let scope = request().scope;
assert_eq!(scope.agent_bucket(), "ws_default:admin");
assert_eq!(scope.tool_bucket(), "ws_default:op_leads");
}
#[test]
fn request_validation_rejects_zero_deadline_and_oversized_fingerprint() {
let mut invalid = request();
invalid.expires_at_unix_ms = 0;
assert!(matches!(
invalid.validate(),
Err(AdmissionStoreError::Invalid { .. })
));
let mut invalid = request();
invalid.fingerprint = "x".repeat(super::ADMISSION_FINGERPRINT_MAX_BYTES + 1);
assert!(invalid.validate().is_err());
let mut invalid = request();
invalid.scope.operation_version = 0;
assert!(invalid.validate().is_err());
assert!(request().validate().is_ok());
}
#[test]
fn weights_are_bounded_positive_configuration() {
assert!(
AdmissionSchedulerConfig {
agent_weight: 0,
tool_weight: 1
}
.validate()
.is_err()
);
assert!(
AdmissionSchedulerConfig {
agent_weight: 1,
tool_weight: 65
}
.validate()
.is_err()
);
assert!(AdmissionSchedulerConfig::default().validate().is_ok());
}
#[test]
fn weighted_round_robin_serves_each_agent_its_quantum() {
let mut queue = vec![
entry(1, "agent_a", "op_1"),
entry(2, "agent_a", "op_1"),
entry(3, "agent_a", "op_1"),
entry(4, "agent_b", "op_1"),
entry(5, "agent_b", "op_1"),
];
let config = AdmissionSchedulerConfig {
agent_weight: 2,
tool_weight: 1,
};
let mut cursor = AdmissionCursor::default();
let mut order = Vec::new();
while !queue.is_empty() {
let selection = select_next_grant(&queue, &cursor, config).unwrap();
order.push(queue.remove(selection.index).agent_key.clone());
cursor = selection.cursor;
}
assert_eq!(
order,
vec!["agent_a", "agent_a", "agent_b", "agent_b", "agent_a"]
);
}
#[test]
fn fifo_inside_one_tool_and_round_robin_between_tools() {
let mut queue = vec![
entry(1, "agent_a", "op_1"),
entry(2, "agent_a", "op_1"),
entry(3, "agent_a", "op_2"),
];
let config = AdmissionSchedulerConfig::default();
let mut cursor = AdmissionCursor::default();
let mut order = Vec::new();
while !queue.is_empty() {
let selection = select_next_grant(&queue, &cursor, config).unwrap();
// A granted ticket leaves the waiting queue: granted entries are never re-granted.
order.push(queue.remove(selection.index).ticket_id);
cursor = selection.cursor;
}
// One grant per Tool per round; FIFO is preserved inside each `(Agent, Tool)` queue.
assert_eq!(order, vec![1, 3, 2]);
}
#[test]
fn new_arrival_does_not_reset_cursor_or_receive_idle_credit() {
let queue = vec![entry(1, "agent_a", "op_1")];
let config = AdmissionSchedulerConfig {
agent_weight: 2,
tool_weight: 1,
};
let cursor = AdmissionCursor::default();
let first = select_next_grant(&queue, &cursor, config).unwrap();
assert_eq!(first.cursor.agent.as_deref(), Some("agent_a"));
assert_eq!(first.cursor.agent_credit, 1);
// A later Agent joins; the cursor keeps its remaining quantum for `agent_a`.
let queue = vec![entry(1, "agent_a", "op_1"), entry(2, "agent_b", "op_9")];
let second = select_next_grant(&queue, &first.cursor, config).unwrap();
assert_eq!(queue[second.index].agent_key, "agent_a");
assert_eq!(second.cursor.agent_credit, 0);
let third = select_next_grant(&queue, &second.cursor, config).unwrap();
assert_eq!(queue[third.index].agent_key, "agent_b");
assert_eq!(third.cursor.agent_credit, 1);
}
#[test]
fn empty_queue_keeps_cursor_untouched() {
let cursor = AdmissionCursor {
agent: Some("agent_b".to_owned()),
agent_credit: 1,
tool: Some("op_1".to_owned()),
tool_credit: 0,
};
assert_eq!(
select_next_grant(&[], &cursor, AdmissionSchedulerConfig::default()),
None
);
assert_eq!(cursor.agent.as_deref(), Some("agent_b"));
}
}
+44 -9
View File
@@ -1,4 +1,5 @@
pub mod access;
pub mod admission;
pub mod agent;
pub mod approval;
pub mod auth;
@@ -17,6 +18,7 @@ pub mod prompt_composition;
pub mod protocol;
pub mod resource;
pub mod secret;
pub mod task;
pub mod tool_catalog;
pub mod tool_quality;
pub mod tool_search;
@@ -60,10 +62,11 @@ pub mod domain {
pub use crate::onboarding::{OnboardingProjection, OnboardingStep, OnboardingStepId};
pub use crate::operation::{
ConfigExport, ConfirmationPolicy, ExecutionConfig, GeneratedDraft, GeneratedDraftStatus,
IdempotencyMode, IdempotencyPolicy, Operation, OperationApprovalMode,
IdempotencyMode, IdempotencyPolicy, NoTasksFallback, Operation, OperationApprovalMode,
OperationApprovalPayloadPreviewMode, OperationApprovalPolicy, OperationApprovalRiskLevel,
OperationSafetyClass, OperationSafetyPolicy, OperationStatus, ResponseCachePolicy,
RestTarget, RetryPolicy, Samples, Target, ToolDescription, ToolExample, WizardState,
RestTarget, RetryPolicy, Samples, Target, ToolDescription, ToolExample, ToolExecutionMode,
ToolExecutionPolicy, WizardState,
};
pub use crate::product_event::{
OnboardingMilestone, PRODUCT_EVENT_IDEMPOTENCY_KEY_MAX_BYTES, PRODUCT_EVENT_SCHEMA_VERSION,
@@ -71,6 +74,18 @@ pub mod domain {
};
pub use crate::protocol::{AuthKind, ExportMode, HttpMethod, Protocol};
pub use crate::secret::{Secret, SecretKind, SecretStatus, SecretVersion};
pub use crate::task::{
AttemptCertainty, AttemptLease, AttemptOutcome, AttemptOutcomeClass, AttemptPhase,
AttemptResolution, BeginAttempt, TASK_APPROVAL_REFERENCE_MAX_BYTES,
TASK_CANCEL_REASON_MAX_BYTES, TASK_DEFAULT_RETENTION_MS, TASK_ENVELOPE_MAX_BYTES,
TASK_ENVELOPE_MAX_DEPTH, TASK_ENVELOPE_MAX_NODES,
TASK_EXECUTABLE_INPUT_CIPHERTEXT_MAX_BYTES, TASK_EXECUTABLE_INPUT_MAX_BYTES,
TASK_IDENTITY_MAX_BYTES, TASK_MAX_ATTEMPTS, TASK_MAX_BACKOFF_STEP_MS,
TASK_MAX_RETENTION_MS, TASK_PURGE_BATCH_LIMIT, TASK_SCOPED_KEY_MAX_BYTES,
TaskAcceptRequest, TaskAcceptance, TaskAccess, TaskAttempt, TaskCancelOutcome,
TaskCancelRequest, TaskEnvelope, TaskError, TaskFinishRequest, TaskId, TaskOwnerKind,
TaskPins, TaskPurgeReport, TaskRecord, TaskRetryPolicy, TaskState, TaskSupportIds,
};
pub use crate::tool_catalog::{ToolCatalogAnalysis, ToolCatalogBudget};
pub use crate::tool_quality::{
ToolQualityCatalogTool, ToolQualityFinding, ToolQualityMappingRule, ToolQualityMappingSet,
@@ -108,12 +123,20 @@ pub mod ports {
PreparedRequest, ProtocolAdapter, ProtocolAdapterError, ResponseCacheScope,
RuntimeRequestContext, SharedProtocolAdapter,
};
pub use crate::task::TaskStore;
}
pub use access::{
InvitationStatus, InvitationToken, Membership, MembershipRole, PlatformApiKey,
PlatformApiKeyKind, PlatformApiKeyScope, PlatformApiKeyStatus, User, UserStatus,
};
pub use admission::{
ADMIN_ORIGIN_KEY, ADMISSION_FINGERPRINT_MAX_BYTES, ADMISSION_OVERLOAD_RETRY_AFTER_MS,
AdmissionCursor, AdmissionDecision, AdmissionMode, AdmissionOutcome, AdmissionPermit,
AdmissionQueueEntry, AdmissionQueueSnapshot, AdmissionRequest, AdmissionSchedulerConfig,
AdmissionScope, AdmissionSelection, AdmissionStoreError, ApprovalEvidence, ExecutionAdmission,
MAX_RETRY_AFTER_MS, current_unix_ms, select_next_grant,
};
pub use agent::{
Agent, AgentOperationBinding, AgentStatus, AgentVersion, ToolAccessMode, ToolGroup,
ToolSearchSettings, ToolSelectionPolicy, ToolSelectionPolicyError,
@@ -168,13 +191,14 @@ pub use observability::{
pub use onboarding::{OnboardingProjection, OnboardingStep, OnboardingStepId};
pub use operation::{
ConfigExport, ConfirmationPolicy, ExecutionConfig, GeneratedDraft, GeneratedDraftStatus,
IdempotencyMode, IdempotencyPolicy, Operation, OperationApprovalMode,
OperationApprovalPayloadPreviewMode, OperationApprovalPolicy, OperationApprovalRiskLevel,
OperationAvailability, OperationLifecycle, OperationLifecycleAction, OperationLifecycleError,
OperationSafetyClass, OperationSafetyPolicy, OperationStatus, OperationVersionState,
ResponseCachePolicy, RestRequestContract, RestRequestContractVersion, RestRequestDescriptor,
RestResponseDescriptor, RestTarget, RetryPolicy, Samples, Target, ToolDescription, ToolExample,
WizardState,
IdempotencyMode, IdempotencyPolicy, MAX_TOOL_DEADLINE_MS, NoTasksFallback, Operation,
OperationApprovalMode, OperationApprovalPayloadPreviewMode, OperationApprovalPolicy,
OperationApprovalRiskLevel, OperationAvailability, OperationLifecycle,
OperationLifecycleAction, OperationLifecycleError, OperationSafetyClass, OperationSafetyPolicy,
OperationStatus, OperationVersionState, ResponseCachePolicy, RestRequestContract,
RestRequestContractVersion, RestRequestDescriptor, RestResponseDescriptor, RestTarget,
RetryPolicy, Samples, Target, ToolDescription, ToolExample, ToolExecutionMode,
ToolExecutionPolicy, ToolExecutionPolicyError, WizardState,
};
pub use product_event::{
OnboardingMilestone, PRODUCT_EVENT_IDEMPOTENCY_KEY_MAX_BYTES, PRODUCT_EVENT_SCHEMA_VERSION,
@@ -182,6 +206,17 @@ pub use product_event::{
};
pub use protocol::{AuthKind, ExportMode, HttpMethod, Protocol};
pub use secret::{Secret, SecretKind, SecretStatus, SecretVersion};
pub use task::{
AttemptCertainty, AttemptLease, AttemptOutcome, AttemptOutcomeClass, AttemptPhase,
AttemptResolution, BeginAttempt, TASK_DEFAULT_RETENTION_MS, TASK_ENVELOPE_MAX_BYTES,
TASK_EXECUTABLE_INPUT_CIPHERTEXT_MAX_BYTES, TASK_EXECUTABLE_INPUT_MAX_BYTES, TASK_MAX_ATTEMPTS,
TASK_MAX_BACKOFF_STEP_MS, TASK_MAX_RETENTION_MS, TASK_PURGE_BATCH_LIMIT,
TASK_SCOPED_KEY_MAX_BYTES, TaskAcceptRequest, TaskAcceptance, TaskAccess, TaskAttempt,
TaskCancelOutcome, TaskCancelRequest, TaskEnvelope, TaskError, TaskFinishRequest, TaskId,
TaskLease, TaskOwnerKind, TaskPins, TaskPurgeReport, TaskRecord, TaskRetryPolicy, TaskState,
TaskStore, TaskSupportIds, contains_sensitive_task_keys, redact_task_envelope,
sensitive_task_key, validate_task_envelope,
};
pub use tool_catalog::{
RECOMMENDED_TOOL_CATALOG_CONTEXT_TOKENS, ToolCatalogAnalysis, ToolCatalogAnalysisError,
ToolCatalogBudget, analyze_tool_catalog,
+144 -2
View File
@@ -270,6 +270,72 @@ pub struct OperationApprovalPolicy {
pub elicitation_message: Option<String>,
}
/// Published execution mode of one Tool (Story 4.1, FR-21).
///
/// This is a scheduling policy of the Published Tool version. It is deliberately separate from
/// `ext::protocol::ExecutionMode`, which describes the protocol adapter (streaming vs unary).
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum ToolExecutionMode {
#[default]
Synchronous,
Background,
Adaptive,
}
/// Behaviour promised to a client that does not negotiate durable Tasks (Story 4.1).
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum NoTasksFallback {
/// Execute through the same synchronous admission path.
#[default]
Synchronous,
/// Reject with a typed unsupported-execution-mode failure; never a Task ID.
Unsupported,
}
pub const MAX_TOOL_DEADLINE_MS: u64 = 3_600_000;
/// Execution policy pinned by a Published Tool version.
///
/// Absence of the policy keeps the pre-4.1 behaviour: synchronous execution bounded by
/// `ExecutionConfig::timeout_ms`, with no background path.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct ToolExecutionPolicy {
#[serde(default)]
pub mode: ToolExecutionMode,
/// Absolute budget of one execution, counted from the start of admission.
pub deadline_ms: u64,
#[serde(default)]
pub no_tasks_fallback: NoTasksFallback,
}
impl Default for ToolExecutionPolicy {
fn default() -> Self {
Self {
mode: ToolExecutionMode::Synchronous,
deadline_ms: 30_000,
no_tasks_fallback: NoTasksFallback::Synchronous,
}
}
}
impl ToolExecutionPolicy {
pub fn validate(&self) -> Result<(), ToolExecutionPolicyError> {
if !(1..=MAX_TOOL_DEADLINE_MS).contains(&self.deadline_ms) {
return Err(ToolExecutionPolicyError::DeadlineOutOfRange);
}
Ok(())
}
}
#[derive(Clone, Copy, Debug, PartialEq, Eq, thiserror::Error)]
pub enum ToolExecutionPolicyError {
#[error("tool execution deadline must be between 1 and {MAX_TOOL_DEADLINE_MS} ms")]
DeadlineOutOfRange,
}
#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct ExecutionConfig {
@@ -285,6 +351,8 @@ pub struct ExecutionConfig {
#[serde(skip_serializing_if = "Option::is_none")]
pub approval_policy: Option<OperationApprovalPolicy>,
#[serde(skip_serializing_if = "Option::is_none")]
pub execution_policy: Option<ToolExecutionPolicy>,
#[serde(skip_serializing_if = "Option::is_none")]
pub auth_profile_ref: Option<AuthProfileId>,
#[serde(default, skip_serializing_if = "BTreeMap::is_empty")]
pub headers: BTreeMap<String, String>,
@@ -444,9 +512,10 @@ mod tests {
edition::OperationSecurityLevel,
ids::{AuthProfileId, OperationId},
operation::{
ConfigExport, ExecutionConfig, Operation, OperationAvailability, OperationLifecycle,
OperationLifecycleAction, OperationLifecycleError, OperationStatus,
ConfigExport, ExecutionConfig, NoTasksFallback, Operation, OperationAvailability,
OperationLifecycle, OperationLifecycleAction, OperationLifecycleError, OperationStatus,
OperationVersionState, RestTarget, Samples, Target, ToolDescription, ToolExample,
ToolExecutionMode, ToolExecutionPolicy,
},
protocol::{AuthKind, ExportMode, HttpMethod, Protocol},
};
@@ -664,6 +733,7 @@ updated_at: 2026-03-25T08:10:00Z
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: Some(AuthProfileId::new("auth_01")),
headers: BTreeMap::new(),
},
@@ -684,4 +754,76 @@ updated_at: 2026-03-25T08:10:00Z
published_at: None,
}
}
#[test]
fn execution_policy_round_trips_all_modes_and_fallbacks() {
for (mode, fallback) in [
(ToolExecutionMode::Synchronous, NoTasksFallback::Synchronous),
(ToolExecutionMode::Background, NoTasksFallback::Synchronous),
(ToolExecutionMode::Background, NoTasksFallback::Unsupported),
(ToolExecutionMode::Adaptive, NoTasksFallback::Unsupported),
] {
let config = ExecutionConfig {
execution_policy: Some(ToolExecutionPolicy {
mode,
deadline_ms: 45_000,
no_tasks_fallback: fallback,
}),
..ExecutionConfig::default()
};
let encoded = serde_json::to_value(&config).unwrap();
assert_eq!(encoded["execution_policy"]["deadline_ms"], json!(45_000));
let decoded: ExecutionConfig = serde_json::from_value(encoded).unwrap();
assert_eq!(decoded, config);
assert_eq!(
decoded.execution_policy.unwrap().mode,
mode,
"the published mode must survive the JSON round-trip"
);
}
}
#[test]
fn legacy_execution_config_without_policy_stays_synchronous() {
let legacy: ExecutionConfig =
serde_json::from_value(json!({ "timeout_ms": 12_000 })).unwrap();
assert_eq!(legacy.execution_policy, None);
assert_eq!(legacy.timeout_ms, 12_000);
}
#[test]
fn execution_policy_rejects_explicit_invalid_deadline() {
assert!(
ToolExecutionPolicy {
mode: ToolExecutionMode::Background,
deadline_ms: 0,
no_tasks_fallback: NoTasksFallback::Synchronous,
}
.validate()
.is_err()
);
assert!(
ToolExecutionPolicy {
mode: ToolExecutionMode::Adaptive,
deadline_ms: crate::operation::MAX_TOOL_DEADLINE_MS + 1,
no_tasks_fallback: NoTasksFallback::Unsupported,
}
.validate()
.is_err()
);
assert!(ToolExecutionPolicy::default().validate().is_ok());
}
#[test]
fn execution_policy_rejects_unknown_fields() {
assert!(
serde_json::from_value::<ToolExecutionPolicy>(json!({
"mode": "background",
"deadline_ms": 1000,
"no_tasks_fallback": "synchronous",
"task_queue": "default"
}))
.is_err()
);
}
}
+285
View File
@@ -0,0 +1,285 @@
//! Durable attempt, retry-snapshot and cancellation vocabulary of a Task (Story 4.4).
use serde::{Deserialize, Serialize};
use super::{TASK_IDENTITY_MAX_BYTES, TaskAccess, TaskEnvelope, TaskError, TaskId, TaskRecord};
/// Upper bound of the retry budget a Task may pin at acceptance (Story 4.4).
pub const TASK_MAX_ATTEMPTS: u32 = 5;
/// Upper bound of the pinned backoff step between two durable attempts (Story 4.4).
pub const TASK_MAX_BACKOFF_STEP_MS: i64 = 1_000;
/// Bound of a stored cancellation reason.
pub const TASK_CANCEL_REASON_MAX_BYTES: usize = 160;
/// Immutable retry snapshot of one accepted Task.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct TaskRetryPolicy {
/// Total durable outbound attempts (1 = a single attempt, never retried).
pub max_attempts: u32,
/// Pinned base backoff between two attempts; the delay grows by this step per attempt.
pub backoff_step_ms: i64,
}
impl TaskRetryPolicy {
/// A Task that may be dispatched exactly once (legacy Tasks and unverified writes).
pub const fn single_attempt() -> Self {
Self {
max_attempts: 1,
backoff_step_ms: 0,
}
}
pub fn new(max_attempts: u32, backoff_step_ms: i64) -> Self {
Self {
max_attempts,
backoff_step_ms,
}
}
pub fn validate(&self) -> Result<(), TaskError> {
if !(1..=TASK_MAX_ATTEMPTS).contains(&self.max_attempts) {
return Err(TaskError::invalid("task retry budget is invalid"));
}
if !(0..=TASK_MAX_BACKOFF_STEP_MS).contains(&self.backoff_step_ms) {
return Err(TaskError::invalid("task retry backoff is invalid"));
}
Ok(())
}
/// Delay before attempt number `next_ordinal` (2..=max_attempts), capped so a pinned snapshot can
/// never stretch the accepted execution deadline on its own.
pub fn backoff_before(&self, next_ordinal: u32) -> i64 {
let steps = i64::from(next_ordinal.saturating_sub(1)).max(1);
self.backoff_step_ms
.saturating_mul(steps)
.min(self.backoff_step_ms.saturating_mul(4))
}
}
/// Certainty of one durable attempt outcome.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum AttemptCertainty {
/// The outcome of the outbound operation is proven (success or a definite failure).
Known,
/// The operation may have been applied upstream; the effect cannot be proven locally.
Unknown,
}
impl AttemptCertainty {
pub const fn as_str(self) -> &'static str {
match self {
Self::Known => "known",
Self::Unknown => "unknown",
}
}
pub fn parse(value: &str) -> Result<Self, TaskError> {
match value {
"known" => Ok(Self::Known),
"unknown" => Ok(Self::Unknown),
_ => Err(TaskError::invalid("unknown attempt certainty")),
}
}
}
/// Phase of one durable attempt at the moment it was resolved.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum AttemptPhase {
/// The durable checkpoint exists but no outbound send was observed.
BeforeDispatch,
/// The request was handed to the transport; its effect may exist upstream.
Dispatched,
}
impl AttemptPhase {
pub const fn as_str(self) -> &'static str {
match self {
Self::BeforeDispatch => "before_dispatch",
Self::Dispatched => "dispatched",
}
}
pub fn parse(value: &str) -> Result<Self, TaskError> {
match value {
"before_dispatch" => Ok(Self::BeforeDispatch),
"dispatched" => Ok(Self::Dispatched),
_ => Err(TaskError::invalid("unknown attempt phase")),
}
}
}
/// Closed outcome vocabulary of one resolved attempt.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum AttemptOutcomeClass {
Completed,
Failed,
/// A bounded further attempt was scheduled instead of a terminal outcome.
RetryScheduled,
OutcomeUnknown,
}
impl AttemptOutcomeClass {
pub const fn as_str(self) -> &'static str {
match self {
Self::Completed => "completed",
Self::Failed => "failed",
Self::RetryScheduled => "retry_scheduled",
Self::OutcomeUnknown => "outcome_unknown",
}
}
pub fn parse(value: &str) -> Result<Self, TaskError> {
match value {
"completed" => Ok(Self::Completed),
"failed" => Ok(Self::Failed),
"retry_scheduled" => Ok(Self::RetryScheduled),
"outcome_unknown" => Ok(Self::OutcomeUnknown),
_ => Err(TaskError::invalid("unknown attempt outcome")),
}
}
}
/// One durable outbound attempt. The row is created by the dispatch checkpoint before any network
/// use and becomes immutable once resolved, so a stale worker can never append to the history.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskAttempt {
pub attempt_id: String,
pub task_id: TaskId,
pub ordinal: u32,
pub owner: String,
pub fence: u64,
/// Canonical fingerprint of the sealed executable input this attempt was admitted for.
pub input_fingerprint: String,
/// Digest of the stable upstream idempotency key (never the raw key or headers).
pub upstream_key_digest: Option<String>,
/// Whether a retry of this attempt is safe (read or server-verified idempotent write).
pub retry_allowed: bool,
pub checkpoint_at_unix_ms: i64,
pub resolved_at_unix_ms: Option<i64>,
pub outcome: Option<AttemptOutcomeClass>,
pub certainty: Option<AttemptCertainty>,
pub phase: Option<AttemptPhase>,
}
/// Committed dispatch intent of one durable attempt (Story 4.4). The store owns the transaction: the
/// attempt row, the incremented budget and the new Task version commit together before the network is
/// touched, so an uncommitted checkpoint can never produce an outbound call.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct BeginAttempt {
pub task_id: TaskId,
pub owner: String,
pub fence: u64,
/// Version the worker read from the lease; a concurrent change rejects the checkpoint.
pub expected_version: u32,
pub attempt_id: String,
pub input_fingerprint: String,
pub upstream_key_digest: Option<String>,
/// The runtime's decision that a retry of this attempt is safe; the store still enforces the
/// pinned budget and the absolute deadline.
pub retry_allowed: bool,
}
impl BeginAttempt {
pub fn validate(&self) -> Result<(), TaskError> {
if self.owner.is_empty() || self.owner.len() > TASK_IDENTITY_MAX_BYTES {
return Err(TaskError::invalid("task attempt owner is invalid"));
}
if self.attempt_id.is_empty() || self.attempt_id.len() > TASK_IDENTITY_MAX_BYTES {
return Err(TaskError::invalid("task attempt identity is invalid"));
}
if self.input_fingerprint.len() != 64 {
return Err(TaskError::invalid("task attempt fingerprint is invalid"));
}
if let Some(digest) = self.upstream_key_digest.as_deref()
&& (digest.is_empty() || digest.len() > TASK_IDENTITY_MAX_BYTES)
{
return Err(TaskError::invalid(
"task attempt upstream key digest is invalid",
));
}
Ok(())
}
}
/// Leased, committed view of one in-flight attempt and the Task version it was admitted under.
#[derive(Clone, Debug, PartialEq)]
pub struct AttemptLease {
pub attempt: TaskAttempt,
pub task: TaskRecord,
}
/// Typed outcome of one attempt, handed to the store for a single atomic resolution.
#[derive(Clone, Debug, PartialEq, Eq)]
pub enum AttemptOutcome {
/// Proven success; the Task becomes `completed` with this result.
Completed(TaskEnvelope),
/// Proven failure. `retryable` only schedules another attempt when the store's pinned budget,
/// the attempt's safe-retry decision and the absolute deadline all still allow it; otherwise the
/// Task fails with this error.
Failed {
error: TaskEnvelope,
retryable: bool,
phase: AttemptPhase,
},
/// The outbound effect cannot be proven. `repeatable` only schedules another bounded attempt when
/// repeating it is safe and the pinned budget/deadline still allow it; otherwise the Task becomes
/// the non-terminal `outcome_unknown`, never a plain failure.
Uncertain {
error: TaskEnvelope,
repeatable: bool,
phase: AttemptPhase,
},
}
/// Single-winner result of resolving one attempt.
#[derive(Clone, Debug, PartialEq)]
pub enum AttemptResolution {
/// A terminal transition committed (completed or failed).
Finished(TaskRecord),
/// The attempt failed in a known way and another bounded attempt was armed.
RetryAt {
task: TaskRecord,
next_attempt_at_unix_ms: i64,
},
/// The attempt resolved to the non-terminal uncertain state.
Unknown(TaskRecord),
}
/// Owner-side cancellation request. `expected_version` is optional so a repeated cancel of the same
/// accepted Task stays idempotent while a stale version is reported as a conflict.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskCancelRequest {
pub access: TaskAccess,
pub task_id: TaskId,
pub expected_version: Option<u32>,
pub reason: String,
}
impl TaskCancelRequest {
pub fn validate(&self) -> Result<(), TaskError> {
self.access.validate()?;
if self.reason.is_empty()
|| self.reason.len() > TASK_CANCEL_REASON_MAX_BYTES
|| self.reason.bytes().any(|byte| byte.is_ascii_control())
{
return Err(TaskError::invalid("task cancellation reason is invalid"));
}
Ok(())
}
}
/// Typed result of an owner-scoped cancellation.
#[derive(Clone, Debug, PartialEq)]
pub enum TaskCancelOutcome {
/// Cancelled before the dispatch checkpoint: no outbound call can have happened.
Cancelled(TaskRecord),
/// The Task already had a checkpointed attempt: non-terminal uncertain, never a plain cancel.
OutcomeUnknown(TaskRecord),
/// The Task was already in a terminal state; its accepted result is unchanged.
AlreadyTerminal(TaskRecord),
/// The caller's expected version is stale; `current` is the authorized current state.
Conflict(TaskRecord),
}
+956
View File
@@ -0,0 +1,956 @@
//! Durable background Task seam (Story 4.2, FR-22/FR-23).
//!
//! A Task is the internal, protocol-independent record of one accepted background request. This
//! module owns the domain vocabulary only: the lifecycle states, the accepted identity/pins, the
//! bounded persisted envelopes and the `TaskStore` port. The protocol adapter (Story 4.7) and the
//! worker (Story 4.3) do not define these semantics; the authoritative store lives in PostgreSQL
//! (`crank-registry`), so an accepted Task survives a crash or restart.
//!
//! Deliberately absent: credentials, prepared headers, approval bearer tokens and any claim that a
//! Task was executed. A terminal state is written by a conditional transition only.
mod attempt;
use async_trait::async_trait;
pub use attempt::*;
use serde::{Deserialize, Serialize};
use serde_json::Value;
use uuid::Uuid;
use crate::correlation::{RequestId, TraceContext, TraceId};
use crate::ids::PlatformApiKeyId;
use crate::operation::ToolExecutionMode;
/// Byte budget of every persisted Task envelope (input, progress, result, error, schema metadata).
pub const TASK_ENVELOPE_MAX_BYTES: usize = 32_768;
/// Maximum byte budget of the decrypted executable Task input.
pub const TASK_EXECUTABLE_INPUT_MAX_BYTES: usize = TASK_ENVELOPE_MAX_BYTES;
/// Maximum byte budget of the stored encrypted executable Task input ciphertext.
pub const TASK_EXECUTABLE_INPUT_CIPHERTEXT_MAX_BYTES: usize = 98_304;
/// Maximum nesting depth of a persisted Task envelope.
pub const TASK_ENVELOPE_MAX_DEPTH: usize = 8;
/// Maximum number of JSON members/nodes of a persisted Task envelope.
pub const TASK_ENVELOPE_MAX_NODES: usize = 256;
/// Bound of every scope/owner/operation identity column.
pub const TASK_IDENTITY_MAX_BYTES: usize = 160;
/// Bound of the caller-supplied idempotency key before it becomes a scoped digest.
pub const TASK_SCOPED_KEY_MAX_BYTES: usize = 128;
/// Bound of a saved approval reference (an identifier, never a bearer token).
pub const TASK_APPROVAL_REFERENCE_MAX_BYTES: usize = 160;
/// Default retention of a terminal Task after its execution deadline (internal bounded default).
pub const TASK_DEFAULT_RETENTION_MS: i64 = 86_400_000;
/// Upper bound on the retention an accepted Task may pin.
pub const TASK_MAX_RETENTION_MS: i64 = 30 * 86_400_000;
/// Upper bound of one retention/housekeeping batch.
pub const TASK_PURGE_BATCH_LIMIT: u32 = 128;
/// Stable identity of one accepted Task. Generated as UUIDv7 and returned only after the accepting
/// transaction committed, so the ID never exists without its durable record.
#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
#[serde(transparent)]
pub struct TaskId(Uuid);
impl TaskId {
pub fn generate() -> Self {
Self(Uuid::now_v7())
}
pub fn parse(value: &str) -> Result<Self, TaskError> {
Uuid::parse_str(value)
.map(Self)
.map_err(|_| TaskError::invalid("task identity is not a UUID"))
}
pub fn as_uuid(self) -> Uuid {
self.0
}
}
impl std::fmt::Display for TaskId {
fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
self.0.fmt(formatter)
}
}
/// Closed Task lifecycle vocabulary. `outcome_unknown` is a non-terminal state: a value commit that
/// failed to confirm its dispatch is never collapsed into a plain failure.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum TaskState {
Queued,
Working,
InputRequired,
OutcomeUnknown,
Completed,
Failed,
Cancelled,
}
impl TaskState {
pub const fn as_str(self) -> &'static str {
match self {
Self::Queued => "queued",
Self::Working => "working",
Self::InputRequired => "input_required",
Self::OutcomeUnknown => "outcome_unknown",
Self::Completed => "completed",
Self::Failed => "failed",
Self::Cancelled => "cancelled",
}
}
pub fn parse(value: &str) -> Result<Self, TaskError> {
match value {
"queued" => Ok(Self::Queued),
"working" => Ok(Self::Working),
"input_required" => Ok(Self::InputRequired),
"outcome_unknown" => Ok(Self::OutcomeUnknown),
"completed" => Ok(Self::Completed),
"failed" => Ok(Self::Failed),
"cancelled" => Ok(Self::Cancelled),
_ => Err(TaskError::invalid("unknown task state")),
}
}
pub const fn is_terminal(self) -> bool {
matches!(self, Self::Completed | Self::Failed | Self::Cancelled)
}
/// One conditional transition may only move along this table. A terminal state has no successor,
/// so a repeated terminal write can never rewrite an accepted result.
pub const fn allows_transition_to(self, next: Self) -> bool {
match self {
Self::Queued => matches!(next, Self::Working | Self::Failed | Self::Cancelled),
Self::Working => matches!(
next,
Self::InputRequired
| Self::OutcomeUnknown
| Self::Completed
| Self::Failed
| Self::Cancelled
),
Self::InputRequired => matches!(next, Self::Working | Self::Failed | Self::Cancelled),
// Story 4.4: an uncertain outbound outcome is never resolved by a generic transition.
// Reconciliation is Story 4.6, so no ordinary finish/cancel/retry may leave this state.
Self::OutcomeUnknown => false,
Self::Completed | Self::Failed | Self::Cancelled => false,
}
}
}
/// Who owns the accepted Task. An administrator is never a wildcard owner of every Task: the owner
/// identity and the authorization scope are separate, both-pinned columns.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum TaskOwnerKind {
Agent,
Admin,
}
impl TaskOwnerKind {
pub const fn as_str(self) -> &'static str {
match self {
Self::Agent => "agent",
Self::Admin => "admin",
}
}
pub fn parse(value: &str) -> Result<Self, TaskError> {
match value {
"agent" => Ok(Self::Agent),
"admin" => Ok(Self::Admin),
_ => Err(TaskError::invalid("unknown task owner kind")),
}
}
}
/// Support identities captured at acceptance. They are stored and returned independently of
/// sampling, and a replay never rewrites them.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskSupportIds {
pub request_id: RequestId,
pub trace_id: TraceId,
pub traceparent: String,
pub trace_sampled: bool,
}
impl TaskSupportIds {
pub fn new(request_id: RequestId, trace_context: &TraceContext) -> Self {
Self {
request_id,
trace_id: trace_context.trace_id().clone(),
traceparent: trace_context.traceparent().to_owned(),
trace_sampled: trace_context.is_sampled(),
}
}
/// Re-validates both the trace identity and the parent context: a Task must never persist an
/// inconsistent or malformed support pair, and an unsampled context is still a valid parent.
pub fn validate(&self) -> Result<(), TaskError> {
if self.trace_id.as_str().is_empty() {
return Err(TaskError::invalid("task trace identity is missing"));
}
let context = TraceContext::parse(&self.traceparent)
.map_err(|_| TaskError::invalid("task trace parent is invalid"))?;
if context.trace_id() != &self.trace_id || context.is_sampled() != self.trace_sampled {
return Err(TaskError::invalid("task trace parent is inconsistent"));
}
Ok(())
}
}
/// Execution semantics pinned at acceptance: a later publication never rewrites them.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskPins {
pub operation_version: u32,
pub policy_mode: ToolExecutionMode,
pub execution_deadline_unix_ms: i64,
/// Retention deadline of the record; never the same thing as the execution deadline.
pub retain_until_unix_ms: i64,
/// Verified approval evidence reference (identifier only, never a bearer token).
pub approval_reference: Option<String>,
/// Bounded retry budget pinned at acceptance (Story 4.4). A later publication, a restart or a
/// new binary never changes the numbers an accepted Task was admitted with.
pub retry: TaskRetryPolicy,
}
impl TaskPins {
pub fn validate(&self) -> Result<(), TaskError> {
if self.operation_version == 0 {
return Err(TaskError::invalid(
"task operation version must be positive",
));
}
if self.execution_deadline_unix_ms <= 0 {
return Err(TaskError::invalid("task execution deadline is invalid"));
}
if self.retain_until_unix_ms < self.execution_deadline_unix_ms {
return Err(TaskError::invalid(
"task retention deadline precedes execution",
));
}
if self.retain_until_unix_ms - self.execution_deadline_unix_ms > TASK_MAX_RETENTION_MS {
return Err(TaskError::invalid("task retention exceeds its bound"));
}
if let Some(reference) = self.approval_reference.as_deref()
&& (reference.is_empty()
|| reference.len() > TASK_APPROVAL_REFERENCE_MAX_BYTES
|| reference.bytes().any(|byte| byte.is_ascii_control()))
{
return Err(TaskError::invalid("task approval reference is invalid"));
}
self.retry.validate()?;
Ok(())
}
}
/// Bounded, classified payload of a Task. Construction validates the byte/depth/node budget and
/// redacts secret-bearing fields before anything is hashed or persisted.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskEnvelope {
value: Value,
}
impl TaskEnvelope {
/// Validates the caller-provided value, then stores a redacted copy of it. The caller hashes the
/// original value separately, so redaction never collapses two different requests into one.
pub fn try_redacted(input: &Value) -> Result<Self, TaskError> {
validate_task_envelope(input)?;
Ok(Self {
value: redact_task_envelope(input),
})
}
/// Accepts a value that was already validated and redacted at a trust boundary.
pub fn from_persisted(value: Value) -> Result<Self, TaskError> {
validate_task_envelope(&value)?;
Ok(Self { value })
}
pub fn value(&self) -> &Value {
&self.value
}
pub fn into_value(self) -> Value {
self.value
}
}
/// Validates the size, nesting depth and node count of one Task envelope. Every byte, collection and
/// node is counted before the value is cloned, hashed or persisted.
pub fn validate_task_envelope(value: &Value) -> Result<(), TaskError> {
let encoded = serde_json::to_vec(value)
.map_err(|_| TaskError::invalid("task envelope is not serializable"))?;
if encoded.len() > TASK_ENVELOPE_MAX_BYTES {
return Err(TaskError::invalid("task envelope exceeds its byte budget"));
}
let mut nodes = 0usize;
check_budget(value, 1, &mut nodes)?;
Ok(())
}
fn check_budget(value: &Value, depth: usize, nodes: &mut usize) -> Result<(), TaskError> {
if depth > TASK_ENVELOPE_MAX_DEPTH {
return Err(TaskError::invalid("task envelope exceeds its depth budget"));
}
*nodes += 1;
if *nodes > TASK_ENVELOPE_MAX_NODES {
return Err(TaskError::invalid("task envelope exceeds its node budget"));
}
match value {
Value::Object(object) => {
for (key, child) in object {
check_budget_name(key)?;
check_budget(child, depth + 1, nodes)?;
}
}
Value::Array(values) => {
for child in values {
check_budget(child, depth + 1, nodes)?;
}
}
_ => {}
}
Ok(())
}
fn check_budget_name(key: &str) -> Result<(), TaskError> {
if key.len() > TASK_IDENTITY_MAX_BYTES {
return Err(TaskError::invalid("task envelope key exceeds its budget"));
}
Ok(())
}
/// Field-name classification of secret-bearing values. A matching field keeps its position but never
/// its plaintext in the persisted envelope.
pub fn sensitive_task_key(name: &str) -> bool {
let lowered = name.to_ascii_lowercase();
[
"password",
"passwd",
"secret",
"token",
"api_key",
"apikey",
"authorization",
"credential",
"private_key",
"cookie",
"signature",
]
.iter()
.any(|marker| lowered.contains(marker))
}
/// Redacted projection of an already validated envelope.
pub fn redact_task_envelope(value: &Value) -> Value {
match value {
Value::Object(object) => Value::Object(
object
.iter()
.map(|(key, child)| {
let value = if sensitive_task_key(key) {
Value::String("[redacted]".to_owned())
} else {
redact_task_envelope(child)
};
(key.clone(), value)
})
.collect(),
),
Value::Array(values) => Value::Array(values.iter().map(redact_task_envelope).collect()),
other => other.clone(),
}
}
/// Recursive check used to fail closed on sensitive fields before encryption.
pub fn contains_sensitive_task_keys(value: &Value) -> bool {
match value {
Value::Object(object) => object
.iter()
.any(|(key, child)| sensitive_task_key(key) || contains_sensitive_task_keys(child)),
Value::Array(values) => values.iter().any(contains_sensitive_task_keys),
_ => false,
}
}
/// Write request of a new durable Task. The store owns the transaction: Task, dispatch intent and
/// queue capacity commit together, and no acknowledgement exists before that commit.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskAcceptRequest {
pub task_id: TaskId,
pub workspace_id: String,
pub owner_kind: TaskOwnerKind,
pub owner_id: String,
pub authorization_scope: String,
/// Typed identity of the Platform API key that authorized the acceptance (Story 4.3). Only the
/// immutable key ID is stored, never the bearer secret; a worker revalidates it authoritatively.
/// `None` means the accepting principal has no machine key (administrator) and fail-closes.
pub platform_api_key_id: Option<PlatformApiKeyId>,
pub operation_id: String,
pub pins: TaskPins,
/// Digest of the caller key over the full owner/authorization/Tool scope.
pub scoped_key: String,
/// Digest of the canonical validated executable input.
pub input_fingerprint: String,
pub input: TaskEnvelope,
pub support: TaskSupportIds,
/// Bounded encrypted executable input, persisted alongside the redacted input projection.
pub executable_input_ciphertext: Option<String>,
pub executable_input_key_version: Option<String>,
pub executable_input_master_key_epoch: Option<i64>,
}
impl TaskAcceptRequest {
pub fn validate(&self) -> Result<(), TaskError> {
for (field, value) in [
("workspace", self.workspace_id.as_str()),
("owner", self.owner_id.as_str()),
("authorization scope", self.authorization_scope.as_str()),
("operation", self.operation_id.as_str()),
] {
if value.is_empty() || value.len() > TASK_IDENTITY_MAX_BYTES {
return Err(TaskError::invalid(format!("task {field} is invalid")));
}
}
if let Some(key_id) = self.platform_api_key_id.as_ref()
&& (key_id.as_str().is_empty() || key_id.as_str().len() > TASK_IDENTITY_MAX_BYTES)
{
return Err(TaskError::invalid(
"task platform api key reference is invalid",
));
}
if self.scoped_key.len() < 16 || self.scoped_key.len() > TASK_SCOPED_KEY_MAX_BYTES {
return Err(TaskError::invalid("task scoped key digest is invalid"));
}
if self.input_fingerprint.len() != 64
|| !self
.input_fingerprint
.bytes()
.all(|byte| byte.is_ascii_hexdigit() && !byte.is_ascii_uppercase())
{
return Err(TaskError::invalid("task input fingerprint is invalid"));
}
self.pins.validate()?;
self.support.validate()?;
let has_ciphertext = self.executable_input_ciphertext.is_some();
let has_key = self.executable_input_key_version.is_some();
let has_epoch = self.executable_input_master_key_epoch.is_some();
match (has_ciphertext, has_key, has_epoch) {
(false, false, false) => {}
(true, true, true) => {
let ciphertext = self.executable_input_ciphertext.as_deref().unwrap();
if ciphertext.is_empty()
|| ciphertext.len() > TASK_EXECUTABLE_INPUT_CIPHERTEXT_MAX_BYTES
{
return Err(TaskError::invalid(
"task executable input ciphertext is invalid",
));
}
let key_version = self.executable_input_key_version.as_deref().unwrap();
if key_version.is_empty() || key_version.len() > 64 {
return Err(TaskError::invalid(
"task executable input ciphertext is invalid",
));
}
if self.executable_input_master_key_epoch.unwrap_or_default() < 1 {
return Err(TaskError::invalid(
"task executable input ciphertext is invalid",
));
}
}
_ => {
return Err(TaskError::invalid(
"task executable input ciphertext is invalid",
));
}
}
Ok(())
}
}
/// Successful acceptance, returned only after the accepting transaction committed. A replay returns
/// the stored identity and support IDs of the original acceptance.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskAcceptance {
pub task_id: TaskId,
pub state: TaskState,
/// Exact Published Version pinned by this acceptance; a replay returns the stored pin.
pub operation_version: u32,
pub support: TaskSupportIds,
pub created_at_unix_ms: i64,
pub retain_until_unix_ms: i64,
pub replayed: bool,
}
/// Owner-side access proof of a read or a terminal write. It is built by the application layer from
/// the verified principal, never from a caller-supplied marker, and every store query carries the
/// whole predicate so a foreign row is never fetched.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskAccess {
pub workspace_id: String,
pub owner_kind: TaskOwnerKind,
pub owner_id: String,
pub authorization_scope: String,
}
impl TaskAccess {
pub fn validate(&self) -> Result<(), TaskError> {
for (field, value) in [
("workspace", self.workspace_id.as_str()),
("owner", self.owner_id.as_str()),
("authorization scope", self.authorization_scope.as_str()),
] {
if value.is_empty() || value.len() > TASK_IDENTITY_MAX_BYTES {
return Err(TaskError::invalid(format!("task {field} is invalid")));
}
}
Ok(())
}
}
/// Read model of one Task. Lifecycle fields, pins and support IDs stay readable together; the
/// protocol adapter only maps them.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskRecord {
pub task_id: TaskId,
pub state: TaskState,
pub workspace_id: String,
pub owner_kind: TaskOwnerKind,
pub owner_id: String,
pub authorization_scope: String,
/// Typed immutable identity of the Platform API key that authorized the acceptance; the worker
/// revalidates this exact key against the registry before any outbound dispatch.
pub platform_api_key_id: Option<PlatformApiKeyId>,
pub operation_id: String,
/// Canonical digest of the accepted executable input (compare-and-conflict identity).
pub input_fingerprint: String,
pub pins: TaskPins,
pub support: TaskSupportIds,
pub input: TaskEnvelope,
pub progress: Option<TaskEnvelope>,
pub result: Option<TaskEnvelope>,
pub error: Option<TaskEnvelope>,
pub attempts: u32,
/// Monotonic lease/fencing token. It never resets and protects Task state mutations.
pub fence: u64,
/// Random process-instance lease owner, if a worker currently owns the Task.
pub lease_owner: Option<String>,
/// Lease expiry epoch milliseconds, if the Task is claimed.
pub lease_until_unix_ms: Option<i64>,
/// Current dispatch attempt identity, committed before any outbound adapter call.
pub current_attempt_id: Option<String>,
/// Dispatch checkpoint epoch milliseconds, if an outbound attempt was committed.
pub dispatch_started_at_unix_ms: Option<i64>,
/// Earliest moment the next durable attempt may be dispatched (pinned backoff).
pub next_attempt_at_unix_ms: Option<i64>,
/// Durable cancellation request that arrived while an attempt was already checkpointed.
pub cancel_requested_at_unix_ms: Option<i64>,
/// Bounded reason of the cancellation request.
pub cancel_reason: Option<String>,
/// Monotonic lifecycle version used by the terminal CAS.
pub version: u32,
pub created_at_unix_ms: i64,
pub updated_at_unix_ms: i64,
}
/// One conditional terminal transition. Exactly one concurrent writer wins; the losers observe the
/// accepted state and never rewrite the result.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct TaskFinishRequest {
pub access: TaskAccess,
pub task_id: TaskId,
pub expected_state: TaskState,
pub expected_version: u32,
pub target_state: TaskState,
pub result: Option<TaskEnvelope>,
pub error: Option<TaskEnvelope>,
}
impl TaskFinishRequest {
pub fn validate(&self) -> Result<(), TaskError> {
self.access.validate()?;
if !self.target_state.is_terminal() {
return Err(TaskError::invalid("task finish target is not terminal"));
}
if self.expected_state.is_terminal() {
return Err(TaskError::invalid("task finish source is already terminal"));
}
if !self.expected_state.allows_transition_to(self.target_state) {
return Err(TaskError::invalid("task state transition is not allowed"));
}
if self.target_state == TaskState::Completed && self.result.is_none() {
return Err(TaskError::invalid("a completed task requires a result"));
}
Ok(())
}
}
/// Bounded retention outcome.
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
pub struct TaskPurgeReport {
/// Queued Tasks that passed their execution deadline and were failed with their intent removed.
pub expired_queued: u64,
/// Terminal Tasks that passed their retention deadline.
pub expired_terminal: u64,
}
/// Typed Task failure. A denial and a missing Task share one outward shape, so a foreign caller can
/// never distinguish "not yours" from "does not exist".
#[derive(Clone, Debug, PartialEq, Eq, thiserror::Error)]
pub enum TaskError {
#[error("task request is invalid: {message}")]
Invalid { message: String },
#[error("task store is unavailable: {message}")]
Unavailable { message: String },
#[error("task was not found")]
NotFound,
#[error("task request conflicts with the accepted request")]
FingerprintConflict,
#[error("task lifecycle transition was rejected")]
StateConflict,
#[error("background task capacity is exhausted")]
Overloaded { retry_after_ms: u64 },
}
impl TaskError {
pub fn invalid(message: impl Into<String>) -> Self {
Self::Invalid {
message: message.into(),
}
}
pub fn unavailable(message: impl Into<String>) -> Self {
Self::Unavailable {
message: message.into(),
}
}
/// Whether the caller may tell this failure apart from an absent Task.
pub fn is_not_found(&self) -> bool {
matches!(self, Self::NotFound)
}
}
/// Leased view of a claimed Task: its durable record plus lease fencing identity.
#[derive(Clone, Debug, PartialEq)]
pub struct TaskLease {
pub record: TaskRecord,
pub owner: String,
pub fence: u64,
pub lease_until: i64,
}
/// Authoritative durable Task port.
#[async_trait]
pub trait TaskStore: Send + Sync {
/// Proves the durable Task state is usable (schema and operations reachable).
async fn health(&self) -> Result<(), TaskError>;
/// Atomically accepts a Task: Task record, dispatch intent and one queue slot commit together.
async fn accept(&self, request: TaskAcceptRequest) -> Result<TaskAcceptance, TaskError>;
/// Reads one Task inside the caller's owner scope.
async fn read(&self, access: &TaskAccess, task_id: TaskId) -> Result<TaskRecord, TaskError>;
/// Scoped lookup of an already accepted Task by its full-scope idempotency digest. Used to keep a
/// retry of an explicitly pinned version from becoming new work; `None` means "not accepted yet".
async fn find(
&self,
access: &TaskAccess,
operation_id: &str,
scoped_key: &str,
) -> Result<Option<TaskRecord>, TaskError>;
/// Applies exactly one conditional terminal transition and releases its intent atomically.
async fn finish(&self, request: TaskFinishRequest) -> Result<TaskRecord, TaskError>;
/// Reads encrypted executable input only under a valid, unexpired current lease.
async fn read_executable_input(
&self,
owner: &str,
task_id: TaskId,
fence: u64,
) -> Result<(String, String, i64), TaskError>;
/// Bounded retention...
async fn purge_expired(&self, batch: u32) -> Result<TaskPurgeReport, TaskError>;
/// Claims a bounded batch of queued or safely recoverable Tasks under leases.
async fn claim_batch(
&self,
owner: &str,
batch_limit: u32,
lease_ms: i64,
) -> Result<Vec<TaskLease>, TaskError>;
/// Renews a lease owned by `owner` when fence and expiry still match.
async fn renew_lease(
&self,
owner: &str,
task_id: TaskId,
fence: u64,
lease_ms: i64,
) -> Result<TaskLease, TaskError>;
/// Persists redacted progress under a valid lease.
async fn write_progress(
&self,
owner: &str,
task_id: TaskId,
fence: u64,
progress: TaskEnvelope,
) -> Result<TaskRecord, TaskError>;
/// Commits one durable outbound attempt before network use: attempt row, incremented budget and
/// the new Task version commit in one transaction. A returned lease is the only authority to
/// dispatch that attempt; an unknown commit means zero outbound calls.
async fn checkpoint_dispatch(&self, request: BeginAttempt) -> Result<AttemptLease, TaskError>;
/// Resolves exactly one committed attempt under its lease/fence/version. Only the current attempt
/// of the current owner may be resolved, and the attempt becomes immutable afterwards.
async fn resolve_attempt(
&self,
lease: &AttemptLease,
outcome: AttemptOutcome,
) -> Result<AttemptResolution, TaskError>;
/// Owner-scoped atomic cancellation. It serializes with claim and checkpoint on the same Task.
async fn cancel(&self, request: TaskCancelRequest) -> Result<TaskCancelOutcome, TaskError>;
/// Owner-scoped bounded history of the durable attempts of one Task.
async fn attempts(
&self,
access: &TaskAccess,
task_id: TaskId,
) -> Result<Vec<TaskAttempt>, TaskError>;
/// Applies a terminal transition only under a valid lease and releases intent atomically.
async fn finish_leased(
&self,
request: TaskFinishRequest,
owner: &str,
fence: u64,
) -> Result<TaskRecord, TaskError>;
/// Resolves a leased Task whose next attempt was refused before its dispatch checkpoint (elapsed
/// deadline, rejected authority). A Task that still carries a proven ambiguity is never downgraded
/// to a plain failure: it ends non-terminal `outcome_unknown` with `code`; otherwise `failed`.
async fn refuse_attempt(&self, lease: &TaskLease, code: &str) -> Result<TaskRecord, TaskError>;
}
#[cfg(test)]
mod tests {
use super::*;
use serde_json::json;
fn request() -> TaskAcceptRequest {
TaskAcceptRequest {
task_id: TaskId::generate(),
workspace_id: "ws_default".to_owned(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_1".to_owned(),
platform_api_key_id: Some(PlatformApiKeyId::new("pk_1")),
operation_id: "op_report".to_owned(),
pins: TaskPins {
operation_version: 3,
policy_mode: ToolExecutionMode::Background,
execution_deadline_unix_ms: 1_790_000_000_000,
retain_until_unix_ms: 1_790_000_000_000 + 1_000,
approval_reference: Some("appr_1".to_owned()),
retry: TaskRetryPolicy::new(3, 1),
},
scoped_key: "a".repeat(64),
input_fingerprint: "b".repeat(64),
input: TaskEnvelope::try_redacted(&json!({"report": "q3"})).unwrap(),
support: TaskSupportIds::new(
RequestId::generate(),
&TraceContext::parse("00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-00")
.unwrap(),
),
executable_input_ciphertext: None,
executable_input_key_version: None,
executable_input_master_key_epoch: None,
}
}
#[test]
fn uuid_v7_task_ids_are_unique_and_round_trip() {
let first = TaskId::generate();
let second = TaskId::generate();
assert_ne!(first, second);
assert_eq!(TaskId::parse(&first.to_string()).unwrap(), first);
}
#[test]
fn terminal_states_have_no_successor() {
for terminal in [
TaskState::Completed,
TaskState::Failed,
TaskState::Cancelled,
] {
assert!(terminal.is_terminal());
for next in [
TaskState::Queued,
TaskState::Working,
TaskState::Completed,
TaskState::Failed,
] {
assert!(!terminal.allows_transition_to(next));
}
}
assert!(TaskState::Queued.allows_transition_to(TaskState::Working));
assert!(TaskState::Working.allows_transition_to(TaskState::OutcomeUnknown));
assert!(!TaskState::Queued.allows_transition_to(TaskState::Completed));
assert!(!TaskState::InputRequired.allows_transition_to(TaskState::Completed));
}
#[test]
fn envelope_bounds_reject_bytes_depth_and_nodes() {
assert!(TaskEnvelope::try_redacted(&json!({"a": 1})).is_ok());
assert!(
TaskEnvelope::try_redacted(&json!({"a": "x".repeat(TASK_ENVELOPE_MAX_BYTES)})).is_err()
);
let mut deep = json!(1);
for _ in 0..=TASK_ENVELOPE_MAX_DEPTH {
deep = json!({"n": deep});
}
assert!(TaskEnvelope::try_redacted(&deep).is_err());
let wide = Value::Array((0..=TASK_ENVELOPE_MAX_NODES).map(|_| json!(1)).collect());
assert!(TaskEnvelope::try_redacted(&wide).is_err());
}
#[test]
fn secret_bearing_fields_are_redacted_and_keep_their_shape() {
let envelope = TaskEnvelope::try_redacted(&json!({
"query": "leads",
"auth": {"api_key": "sk_live_1", "nested": [{"password": "p"}]}
}))
.unwrap();
let value = envelope.value();
assert_eq!(value["query"], json!("leads"));
assert_eq!(value["auth"]["api_key"], json!("[redacted]"));
assert_eq!(value["auth"]["nested"][0]["password"], json!("[redacted]"));
}
#[test]
fn unsampled_support_ids_are_valid_and_consistent() {
let support = TaskSupportIds::new(
RequestId::generate(),
&TraceContext::parse("00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-00")
.unwrap(),
);
assert!(!support.trace_sampled);
assert!(support.traceparent.ends_with("-00"));
support.validate().unwrap();
let mut inconsistent = support.clone();
inconsistent.traceparent = inconsistent.traceparent.replace("-00", "-01");
assert!(inconsistent.validate().is_err());
}
#[test]
fn accept_request_validation_holds_the_identity_and_pin_bounds() {
assert!(request().validate().is_ok());
let mut oversized = request();
oversized.workspace_id = "w".repeat(TASK_IDENTITY_MAX_BYTES + 1);
assert!(oversized.validate().is_err());
let mut bad_fingerprint = request();
bad_fingerprint.input_fingerprint = "XYZ".to_owned();
assert!(bad_fingerprint.validate().is_err());
let mut bad_retention = request();
bad_retention.pins.retain_until_unix_ms = bad_retention.pins.execution_deadline_unix_ms - 1;
assert!(bad_retention.validate().is_err());
// The persisted authority reference is bounded and never empty.
let mut empty_key_reference = request();
empty_key_reference.platform_api_key_id = Some(PlatformApiKeyId::new(""));
assert!(empty_key_reference.validate().is_err());
}
#[test]
fn outcome_unknown_has_no_generic_successor() {
// Story 4.4: reconciliation of an uncertain outcome is Story 4.6. Until then no ordinary
// finish/cancel/retry may leave the non-terminal uncertain state.
for next in [
TaskState::Queued,
TaskState::Working,
TaskState::InputRequired,
TaskState::OutcomeUnknown,
TaskState::Completed,
TaskState::Failed,
TaskState::Cancelled,
] {
assert!(
!TaskState::OutcomeUnknown.allows_transition_to(next),
"outcome_unknown must not transition to {next:?}"
);
}
assert!(!TaskState::OutcomeUnknown.is_terminal());
}
#[test]
fn retry_policy_bounds_are_pinned_and_bounded() {
assert!(TaskRetryPolicy::single_attempt().validate().is_ok());
assert!(
TaskRetryPolicy::new(TASK_MAX_ATTEMPTS, TASK_MAX_BACKOFF_STEP_MS)
.validate()
.is_ok()
);
assert!(TaskRetryPolicy::new(0, 0).validate().is_err());
assert!(
TaskRetryPolicy::new(TASK_MAX_ATTEMPTS + 1, 0)
.validate()
.is_err()
);
assert!(
TaskRetryPolicy::new(1, TASK_MAX_BACKOFF_STEP_MS + 1)
.validate()
.is_err()
);
// The pinned step grows the delay but never beyond four steps.
assert_eq!(TaskRetryPolicy::new(5, 2).backoff_before(2), 2);
assert_eq!(TaskRetryPolicy::new(5, 2).backoff_before(5), 8);
assert_eq!(TaskRetryPolicy::single_attempt().backoff_before(2), 0);
// A pin whose retry snapshot is invalid is rejected at acceptance.
let mut invalid = request();
invalid.pins.retry = TaskRetryPolicy::new(0, 0);
assert!(invalid.validate().is_err());
}
#[test]
fn finish_requires_a_terminal_reachable_transition() {
let access = TaskAccess {
workspace_id: "ws_default".to_owned(),
owner_kind: TaskOwnerKind::Agent,
owner_id: "agent_a".to_owned(),
authorization_scope: "credential:pk_1".to_owned(),
};
let base = TaskFinishRequest {
access,
task_id: TaskId::generate(),
expected_state: TaskState::Working,
expected_version: 1,
target_state: TaskState::Completed,
result: Some(TaskEnvelope::try_redacted(&json!({"ok": true})).unwrap()),
error: None,
};
assert!(base.validate().is_ok());
let mut non_terminal = base.clone();
non_terminal.target_state = TaskState::Working;
assert!(non_terminal.validate().is_err());
let mut terminal_source = base.clone();
terminal_source.expected_state = TaskState::Completed;
assert!(terminal_source.validate().is_err());
let mut completed_without_result = base;
completed_without_result.result = None;
assert!(completed_without_result.validate().is_err());
}
}
@@ -231,6 +231,7 @@ fn default_execution_config() -> ExecutionConfig {
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
}
+1
View File
@@ -564,6 +564,7 @@ fn default_execution_config() -> ExecutionConfig {
idempotency: None,
safety: None,
approval_policy: None,
execution_policy: None,
auth_profile_ref: None,
headers: BTreeMap::new(),
}
+28
View File
@@ -217,6 +217,34 @@ string_enum! {
Concurrency => "concurrency",
RateLimit => "rate_limit",
McpStream => "mcp_stream",
Admission => "admission",
}
}
string_enum! {
pub enum RejectionReason {
Limit => "limit",
Capacity => "capacity",
QueueFull => "queue_full",
Deadline => "deadline",
Unsupported => "unsupported",
DependencyUnavailable => "dependency_unavailable",
}
}
string_enum! {
pub enum ScopeKind {
Process => "process",
Global => "global",
Agent => "agent",
Tool => "tool",
}
}
string_enum! {
pub enum AdmissionModeLabel {
Synchronous => "synchronous",
Background => "background",
}
}
+6 -5
View File
@@ -16,14 +16,15 @@ mod schema;
pub use exemplar::{ExemplarObservation, ExemplarTraceId, exemplar_snapshot};
pub use labels::{
CacheOutcome, ConfirmationOutcome, DbPoolState, Exporter, HttpMethod, HttpRoute,
HttpStatusClass, IdempotencyOutcome, InvocationSource, LimitStage, McpMethod, McpOutcome,
McpResponseMode, SignalType, ToolErrorKind, ToolOutcome, UpstreamOperationKind,
UpstreamOutcome,
AdmissionModeLabel, CacheOutcome, ConfirmationOutcome, DbPoolState, Exporter, HttpMethod,
HttpRoute, HttpStatusClass, IdempotencyOutcome, InvocationSource, LimitStage, McpMethod,
McpOutcome, McpResponseMode, RejectionReason, ScopeKind, SignalType, ToolErrorKind,
ToolOutcome, UpstreamOperationKind, UpstreamOutcome,
};
pub use record::{
InFlightGuard, ToolInvocationMetrics, UpstreamRequestMetrics, bind_metric_service,
bound_metric_service, initialize_gauges, normalize_http_route, record_cache_outcome,
bound_metric_service, initialize_gauges, normalize_http_route, record_admission_oldest_waiting,
record_admission_queue_wait, record_admission_rejection, record_cache_outcome,
record_confirmation_outcome, record_export_failure, record_http_request,
record_http_request_with_exemplar, record_idempotency_outcome, record_invocation_history_lost,
record_limit_rejection, record_mcp_request, record_mcp_request_with_exemplar,
+41 -5
View File
@@ -5,10 +5,10 @@ use metrics::Gauge;
use crate::schema::{MetricDefinition, metric_schema};
use crate::{
CacheOutcome, ConfirmationOutcome, DbPoolState, ExemplarTraceId, Exporter, HttpMethod,
HttpRoute, HttpStatusClass, IdempotencyOutcome, InvocationSource, LimitStage, McpMethod,
McpOutcome, McpResponseMode, MetricBindingError, MetricService, SignalType, ToolErrorKind,
ToolOutcome, UpstreamOperationKind, UpstreamOutcome,
AdmissionModeLabel, CacheOutcome, ConfirmationOutcome, DbPoolState, ExemplarTraceId, Exporter,
HttpMethod, HttpRoute, HttpStatusClass, IdempotencyOutcome, InvocationSource, LimitStage,
McpMethod, McpOutcome, McpResponseMode, MetricBindingError, MetricService, RejectionReason,
ScopeKind, SignalType, ToolErrorKind, ToolOutcome, UpstreamOperationKind, UpstreamOutcome,
};
static BINDING: OnceLock<MetricService> = OnceLock::new();
@@ -352,11 +352,47 @@ pub fn record_limit_rejection(stage: LimitStage) {
}
metrics::counter!(
"crank_runtime_limit_rejections_total",
"stage" => stage.as_str()
"stage" => stage.as_str(),
"reason" => RejectionReason::Limit.as_str(),
"scope_kind" => ScopeKind::Process.as_str()
)
.increment(1);
}
/// Typed admission refusal with finite, low-cardinality labels only: no workspace, Agent, Tool,
/// request id, fingerprint or arbitrary error text ever becomes a label.
pub fn record_admission_rejection(reason: RejectionReason, scope_kind: ScopeKind) {
if bound_family("crank_runtime_limit_rejections_total").is_none() {
return;
}
metrics::counter!(
"crank_runtime_limit_rejections_total",
"stage" => LimitStage::Admission.as_str(),
"reason" => reason.as_str(),
"scope_kind" => scope_kind.as_str()
)
.increment(1);
}
pub fn record_admission_queue_wait(mode: AdmissionModeLabel, waited_ms: u64) {
if bound_family("crank_runtime_admission_queue_wait_seconds").is_none() {
return;
}
metrics::gauge!(
"crank_runtime_admission_queue_wait_seconds",
"mode" => mode.as_str()
)
.set(waited_ms as f64 / 1000.0);
}
pub fn record_admission_oldest_waiting(oldest_waiting_ms: Option<u64>) {
if bound_family("crank_runtime_admission_oldest_waiting_seconds").is_none() {
return;
}
let seconds = oldest_waiting_ms.unwrap_or_default() as f64 / 1000.0;
metrics::gauge!("crank_runtime_admission_oldest_waiting_seconds").set(seconds);
}
pub fn record_cache_outcome(outcome: CacheOutcome) {
if bound_family("crank_runtime_cache_total").is_none() {
return;
+38 -8
View File
@@ -3,16 +3,17 @@
use super::routes::{BOTH, HTTP_ROUTE_DOMAIN, MCP};
use crate::labels::{
CacheOutcome, ConfirmationOutcome, DbPoolState, Exporter, HttpMethod, HttpStatusClass,
IdempotencyOutcome, InvocationSource, LimitStage, McpMethod, McpOutcome, McpResponseMode,
SignalType, ToolErrorKind, ToolOutcome, UpstreamOperationKind, UpstreamOutcome,
AdmissionModeLabel, CacheOutcome, ConfirmationOutcome, DbPoolState, Exporter, HttpMethod,
HttpStatusClass, IdempotencyOutcome, InvocationSource, LimitStage, McpMethod, McpOutcome,
McpResponseMode, RejectionReason, ScopeKind, SignalType, ToolErrorKind, ToolOutcome,
UpstreamOperationKind, UpstreamOutcome,
};
pub const METRIC_SCHEMA_VERSION: u32 = 2;
pub const MAX_METRIC_FAMILIES: usize = 128;
pub const MAX_PRODUCT_LABELS_PER_FAMILY: usize = 4;
pub const MAX_LOGICAL_SERIES_PER_PROCESS: usize = 6_285;
pub const MAX_RENDERED_SERIES_PER_PROCESS: usize = 19_515;
pub const MAX_LOGICAL_SERIES_PER_PROCESS: usize = 6_381;
pub const MAX_RENDERED_SERIES_PER_PROCESS: usize = 19_611;
pub const MAX_EXEMPLAR_SLOTS_PER_PROCESS: usize = 12_348;
pub const MAX_EXPOSITION_BYTES: usize = 8 * 1024 * 1024;
@@ -179,6 +180,9 @@ const TOOL_ERRORS: &[&str] = ToolErrorKind::VALUES;
const OPERATION_KINDS: &[&str] = UpstreamOperationKind::VALUES;
const UPSTREAM_OUTCOMES: &[&str] = UpstreamOutcome::VALUES;
const LIMIT_STAGES: &[&str] = LimitStage::VALUES;
const REJECTION_REASONS: &[&str] = RejectionReason::VALUES;
const SCOPE_KINDS: &[&str] = ScopeKind::VALUES;
const ADMISSION_MODES: &[&str] = AdmissionModeLabel::VALUES;
const CACHE_OUTCOMES: &[&str] = CacheOutcome::VALUES;
const IDEMPOTENCY_OUTCOMES: &[&str] = IdempotencyOutcome::VALUES;
const CONFIRMATION_OUTCOMES: &[&str] = ConfirmationOutcome::VALUES;
@@ -403,13 +407,39 @@ pub(crate) const METRIC_SCHEMA: &[MetricDefinition] = &[
"crank_runtime_limit_rejections_total",
MetricKind::Counter,
MetricUnit::Count,
&["stage"],
&[domain("stage", LIMIT_STAGES)],
&["stage", "reason", "scope_kind"],
&[
domain("stage", LIMIT_STAGES),
domain("reason", REJECTION_REASONS),
domain("scope_kind", SCOPE_KINDS),
],
&[],
BOTH,
3,
96,
"Runtime executions rejected by a bounded limit.",
),
metric(
"crank_runtime_admission_queue_wait_seconds",
MetricKind::Gauge,
MetricUnit::Seconds,
&["mode"],
&[domain("mode", ADMISSION_MODES)],
&[],
BOTH,
2,
"Wait of the most recently admitted execution before it received a scoped permit.",
),
metric(
"crank_runtime_admission_oldest_waiting_seconds",
MetricKind::Gauge,
MetricUnit::Seconds,
&[],
&[],
&[],
BOTH,
1,
"Age of the oldest waiting admission request.",
),
metric(
"crank_runtime_cache_total",
MetricKind::Counter,
+13 -6
View File
@@ -1,12 +1,13 @@
use std::{collections::BTreeSet, time::Duration};
use crank_metrics::{
CacheOutcome, ConfirmationOutcome, DbPoolState, Exporter, HttpMethod, HttpRoute,
HttpStatusClass, IdempotencyOutcome, InFlightGuard, InvocationSource, LimitStage, McpMethod,
McpOutcome, McpResponseMode, MetricService, SignalType, ToolErrorKind, ToolInvocationMetrics,
ToolOutcome, UpstreamOperationKind, UpstreamOutcome, UpstreamRequestMetrics,
bind_metric_service, initialize_gauges, metric_schema, record_cache_outcome,
record_confirmation_outcome, record_export_failure, record_http_request,
AdmissionModeLabel, CacheOutcome, ConfirmationOutcome, DbPoolState, Exporter, HttpMethod,
HttpRoute, HttpStatusClass, IdempotencyOutcome, InFlightGuard, InvocationSource, LimitStage,
McpMethod, McpOutcome, McpResponseMode, MetricService, RejectionReason, ScopeKind, SignalType,
ToolErrorKind, ToolInvocationMetrics, ToolOutcome, UpstreamOperationKind, UpstreamOutcome,
UpstreamRequestMetrics, bind_metric_service, initialize_gauges, metric_schema,
record_admission_oldest_waiting, record_admission_queue_wait, record_admission_rejection,
record_cache_outcome, record_confirmation_outcome, record_export_failure, record_http_request,
record_idempotency_outcome, record_invocation_history_lost, record_limit_rejection,
record_mcp_request, record_tool_invocation, record_upstream_request, set_catalog,
set_db_pool_connections, set_mcp_active_sessions, set_mcp_session_metrics_fresh,
@@ -71,6 +72,9 @@ fn schema_names_are_unique_and_labels_are_closed() {
"state",
"signal_type",
"exporter",
"reason",
"scope_kind",
"mode",
];
for metric in schema {
assert!(
@@ -265,6 +269,9 @@ fn every_recording_api_matches_the_declared_schema() {
Duration::from_millis(1),
);
record_limit_rejection(LimitStage::Concurrency);
record_admission_rejection(RejectionReason::Capacity, ScopeKind::Global);
record_admission_queue_wait(AdmissionModeLabel::Synchronous, 5);
record_admission_oldest_waiting(Some(5));
record_cache_outcome(CacheOutcome::Hit);
record_idempotency_outcome(IdempotencyOutcome::Replay);
record_confirmation_outcome(ConfirmationOutcome::Required);
+8 -8
View File
@@ -48,10 +48,10 @@ fn schema_v1_is_complete_deterministic_and_inside_the_frozen_budget() {
#[test]
fn frozen_budget_edges_are_exact() {
assert!(validate_budget(6_284, 19_514).is_ok());
assert!(validate_budget(6_285, 19_515).is_ok());
assert!(validate_budget(6_286, 16_000).is_err());
assert!(validate_budget(6_000, 19_516).is_err());
assert!(validate_budget(6_380, 19_610).is_ok());
assert!(validate_budget(6_381, 19_611).is_ok());
assert!(validate_budget(6_382, 19_000).is_err());
assert!(validate_budget(6_000, 19_612).is_err());
assert_eq!(
crank_metrics::max_exemplar_slots(MetricService::AdminApi),
12_348
@@ -79,15 +79,15 @@ fn canonical_schema_validation_reports_its_category() {
assert_eq!(
schema_budget(MetricService::AdminApi),
Ok(crank_metrics::SchemaBudget {
logical_series: 6_285,
rendered_series: 19_515
logical_series: 6_381,
rendered_series: 19_611
})
);
assert_eq!(
schema_budget(MetricService::McpServer),
Ok(crank_metrics::SchemaBudget {
logical_series: 1_083,
rendered_series: 3_558
logical_series: 1_179,
rendered_series: 3_654
})
);
}
+11 -11
View File
@@ -543,22 +543,22 @@ fn root_budgets_are_per_process_only() {
"root обязан объявлять ровно один ключ budgets"
);
assert!(rendered.ends_with(
",\"budgets\":[{\"process\":\"admin_api\",\"families\":15,\"logical_series\":6285,\
\"rendered_series\":19515,\"exemplar_slots\":12348},{\"process\":\"mcp_server\",\
\"families\":23,\"logical_series\":1083,\"rendered_series\":3558,\"exemplar_slots\":2310}]}\n"
",\"budgets\":[{\"process\":\"admin_api\",\"families\":17,\"logical_series\":6381,\
\"rendered_series\":19611,\"exemplar_slots\":12348},{\"process\":\"mcp_server\",\
\"families\":25,\"logical_series\":1179,\"rendered_series\":3654,\"exemplar_slots\":2310}]}\n"
));
assert_eq!(
schema_budget(MetricService::AdminApi),
Ok(crank_metrics::SchemaBudget {
logical_series: 6_285,
rendered_series: 19_515
logical_series: 6_381,
rendered_series: 19_611
})
);
assert_eq!(
schema_budget(MetricService::McpServer),
Ok(crank_metrics::SchemaBudget {
logical_series: 1_083,
rendered_series: 3_558
logical_series: 1_179,
rendered_series: 3_654
})
);
}
@@ -566,11 +566,11 @@ fn root_budgets_are_per_process_only() {
/// TD-M3 — frozen caps и границы `validate_exemplar_budget` точны.
#[test]
fn frozen_caps_and_edges_are_exact() {
assert_eq!(MAX_LOGICAL_SERIES_PER_PROCESS, 6_285);
assert_eq!(MAX_RENDERED_SERIES_PER_PROCESS, 19_515);
assert_eq!(MAX_LOGICAL_SERIES_PER_PROCESS, 6_381);
assert_eq!(MAX_RENDERED_SERIES_PER_PROCESS, 19_611);
assert_eq!(MAX_EXEMPLAR_SLOTS_PER_PROCESS, 12_348);
assert!(validate_budget(6_285, 19_515).is_ok());
assert!(validate_budget(6_286, 19_515).is_err());
assert!(validate_budget(6_381, 19_611).is_ok());
assert!(validate_budget(6_382, 19_611).is_err());
assert!(validate_exemplar_budget(12_347).is_ok());
assert!(validate_exemplar_budget(12_348).is_ok());
assert!(validate_exemplar_budget(12_349).is_err());
+1 -1
View File
@@ -13,7 +13,7 @@ mod schema;
pub use config::{ObservabilityConfig, ObservabilityConfigError, ServiceIdentity};
pub use crank_metrics::{
DURATION_BUCKETS_SECONDS, MetricDefinition, MetricKind, MetricService, MetricUnit,
bind_metric_service, bound_metric_service, metric_schema,
bind_metric_service, bound_metric_service, metric_schema, record_admission_oldest_waiting,
};
pub use error_reporting::{
CriticalErrorCategory, SentryConfig, SentryConfigError, capture_critical_error,
+1
View File
@@ -7,6 +7,7 @@ publish.workspace = true
version.workspace = true
[dependencies]
async-trait = "0.1"
crank-artifacts = { path = "../crank-artifacts" }
crank-core = { path = "../crank-core" }
crank-mapping = { path = "../crank-mapping" }
+3 -2
View File
@@ -116,6 +116,7 @@ pub use model::{
YamlImportJob, YamlImportJobCompletion, YamlImportJobId, YamlImportJobStatus,
};
pub use postgres::{
AuthorizedResourceRead, BoundResourceAdmission, PostgresPoolConfig, PostgresPoolConfigError,
PostgresRegistry,
AdmissionLimits, AuthorizedResourceRead, BoundResourceAdmission, PostgresExecutionAdmission,
PostgresPoolConfig, PostgresPoolConfigError, PostgresRegistry, PostgresTaskStore,
admission_health, admission_snapshot, tasks_health,
};
+12
View File
@@ -6,6 +6,12 @@ mod artifact_cleanup_indexes_v13;
mod artifact_metadata_v12;
mod authority;
mod baseline_v1;
mod durable_tasks_v25;
mod durable_tasks_v26;
mod durable_tasks_v27;
mod durable_tasks_v28;
mod durable_tasks_v29;
mod execution_admission_v24;
mod execution_outcome_v5;
mod import_plan_v14;
mod master_key_identity_v7;
@@ -35,6 +41,12 @@ mod schema_guard_v19;
mod schema_guard_v20;
mod schema_guard_v22;
mod schema_guard_v23;
mod schema_guard_v24;
mod schema_guard_v25;
mod schema_guard_v26;
mod schema_guard_v27;
mod schema_guard_v28;
mod schema_guard_v29;
mod schema_guard_v7;
mod schema_guard_v8;
mod schema_guard_v9;
@@ -4,6 +4,12 @@ use super::agent_resource_bindings_v18;
use super::approval_side_effects_v10;
use super::artifact_cleanup_indexes_v13;
use super::artifact_metadata_v12;
use super::durable_tasks_v25;
use super::durable_tasks_v26;
use super::durable_tasks_v27;
use super::durable_tasks_v28;
use super::durable_tasks_v29;
use super::execution_admission_v24;
use super::execution_outcome_v5;
use super::import_plan_v14;
use super::master_key_identity_v7;
@@ -39,9 +45,10 @@ use preflight::*;
use contract::baseline_source_digest;
use contract::validate_descriptors;
const MIGRATION_LOCK_ID: i64 = 0x4352_414E_4B4D_4947;
const CURRENT_VERSION: i64 = 23;
const CURRENT_VERSION: i64 = 29;
const IMPLEMENTED_VERSIONS: &[i64] = &[
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23,
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26,
27, 28, 29,
];
const BASELINE_SOURCE_SHA256: &str =
"eb1656fc5b4b5be9ee390d237d1d58e4b2274ae5ba9b7ba06a2f3f860dfda675";
@@ -330,6 +337,42 @@ impl MigrationAuthority {
prompt_dependencies_v23::SOURCE_SHA256,
22,
),
expand_descriptor(
24,
execution_admission_v24::NAME,
execution_admission_v24::SOURCE_SHA256,
23,
),
expand_descriptor(
25,
durable_tasks_v25::NAME,
durable_tasks_v25::SOURCE_SHA256,
24,
),
expand_descriptor(
26,
durable_tasks_v26::NAME,
durable_tasks_v26::SOURCE_SHA256,
25,
),
expand_descriptor(
27,
durable_tasks_v27::NAME,
durable_tasks_v27::SOURCE_SHA256,
26,
),
expand_descriptor(
28,
durable_tasks_v28::NAME,
durable_tasks_v28::SOURCE_SHA256,
27,
),
expand_descriptor(
29,
durable_tasks_v29::NAME,
durable_tasks_v29::SOURCE_SHA256,
28,
),
]
}
pub fn validate_sequence() -> Result<(), MigrationError> {
@@ -489,6 +532,24 @@ impl MigrationAuthority {
if from < 23 {
prompt_dependencies_v23::apply(&mut transaction, &Self::sequence()[22]).await?;
}
if from < 24 {
execution_admission_v24::apply(&mut transaction, &Self::sequence()[23]).await?;
}
if from < 25 {
durable_tasks_v25::apply(&mut transaction, &Self::sequence()[24]).await?;
}
if from < 26 {
durable_tasks_v26::apply(&mut transaction, &Self::sequence()[25]).await?;
}
if from < 27 {
durable_tasks_v27::apply(&mut transaction, &Self::sequence()[26]).await?;
}
if from < 28 {
durable_tasks_v28::apply(&mut transaction, &Self::sequence()[27]).await?;
}
if from < 29 {
durable_tasks_v29::apply(&mut transaction, &Self::sequence()[28]).await?;
}
match inspect(&mut transaction).await? {
MigrationPreflight::Current {
version: CURRENT_VERSION,
@@ -140,6 +140,8 @@ pub(super) fn validate_descriptors(
!= agent_resource_bindings_v18::SOURCE_SHA256
|| sha256_hex(resource_security_revocations_v19::SOURCE.as_bytes())
!= resource_security_revocations_v19::SOURCE_SHA256
|| sha256_hex(execution_admission_v24::SOURCE.as_bytes())
!= execution_admission_v24::SOURCE_SHA256
{
return Err(MigrationError::new(
"invalid_contract",
@@ -148,6 +148,7 @@ pub(super) async fn inspect(
owned_relations::PROMPTS,
owned_relations::PROMPT_PUBLICATIONS,
owned_relations::PROMPT_DEPENDENCIES,
owned_relations::EXECUTION_ADMISSION,
] {
validate_required_relations(connection, relations, CURRENT_VERSION).await?;
}
@@ -0,0 +1,45 @@
use sqlx::{Postgres, Transaction, query};
use super::authority::{MigrationDescriptor, MigrationError};
pub(super) const NAME: &str = "durable-tasks-v25";
pub(super) const SOURCE: &str = include_str!("durable_tasks_v25.sql");
pub(super) const SOURCE_SHA256: &str =
"ada48bfb7d7130c84ab38217795608cd368ccb28af66faafbdfc4fc7609f67cd";
pub(super) async fn apply(
transaction: &mut Transaction<'_, Postgres>,
descriptor: &MigrationDescriptor,
) -> Result<(), MigrationError> {
sqlx::raw_sql(SOURCE)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.durable_tasks_v25",
Some(25),
"restore_known_good_backup",
)
})?;
query(
"insert into __crank_migrations (version, name, checksum, phase, compatibility)
values ($1, $2, $3, $4, $5)",
)
.bind(descriptor.version)
.bind(descriptor.name)
.bind(&descriptor.checksum)
.bind(descriptor.phase)
.bind(descriptor.compatibility)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.canonical_ledger",
Some(25),
"restore_known_good_backup",
)
})?;
Ok(())
}
@@ -0,0 +1,90 @@
-- V25: durable background Tasks (Story 4.2, FR-22, FR-23).
--
-- Одна строка `tasks` — принятая durable работа. Идентификатор Task (UUIDv7), owner scope,
-- authorization scope, запрошенный Tool и точная published-версия, canonical fingerprint входа,
-- support IDs (Request/Trace/parent TraceContext независимо от sampling) и bounded envelopes
-- закрепляются в той же PostgreSQL transaction, что и dispatch intent
-- (`execution_admission_entries.task_id`) и его queue slot. До commit наружу не выдаётся ни ID, ни ack.
--
-- `scoped_key` — digest caller-ключа по полному (workspace, owner kind/id, authorization scope,
-- operation) scope; UNIQUE по этому набору и есть защита идемпотентности. Повтор того же ключа с тем
-- же canonical fingerprint возвращает сохранённую Task, другой fingerprint — typed conflict.
--
-- `retain_until` — срок хранения записи, а не execution deadline: пока запись в retention, replay
-- находит тот же Task. Удаление terminal-записей возможно только после retain_until; queued/working/
-- input_required/outcome_unknown не удаляются никогда ради лимита.
--
-- Строка не хранит credentials, bearer/prepared headers и confirmation tokens: секретные поля входа
-- редактируются до записи, а approval хранится безопасной ссылкой.
create table tasks (
task_id uuid not null,
workspace_id text not null,
owner_kind text not null,
owner_id text not null,
authorization_scope text not null,
operation_id text not null,
operation_version integer not null,
scoped_key text not null,
input_fingerprint text not null,
request_id text not null,
trace_id text not null,
traceparent text not null,
trace_sampled boolean not null,
policy_mode text not null,
approval_reference text,
state text not null,
version integer not null default 1,
attempts integer not null default 0,
input_envelope jsonb not null,
progress_envelope jsonb,
result_envelope jsonb,
error_envelope jsonb,
execution_deadline timestamptz not null,
retain_until timestamptz not null,
created_at timestamptz not null default now(),
updated_at timestamptz not null default now(),
constraint tasks_pkey primary key (task_id),
constraint tasks_workspace_check check (octet_length(workspace_id) between 1 and 160),
constraint tasks_owner_kind_check check (owner_kind in ('agent', 'admin')),
constraint tasks_owner_check check (octet_length(owner_id) between 1 and 160),
constraint tasks_authorization_scope_check check (
octet_length(authorization_scope) between 1 and 160
),
constraint tasks_operation_check check (octet_length(operation_id) between 1 and 160),
constraint tasks_operation_version_check check (operation_version >= 1),
constraint tasks_scoped_key_check check (octet_length(scoped_key) between 16 and 128),
constraint tasks_fingerprint_check check (octet_length(input_fingerprint) = 64),
constraint tasks_request_check check (octet_length(request_id) between 1 and 128),
constraint tasks_trace_check check (octet_length(trace_id) = 32),
constraint tasks_traceparent_check check (octet_length(traceparent) = 55),
constraint tasks_policy_mode_check check (
policy_mode in ('synchronous', 'background', 'adaptive')
),
constraint tasks_state_check check (
state in ('queued', 'working', 'input_required', 'outcome_unknown',
'completed', 'failed', 'cancelled')
),
constraint tasks_lifecycle_check check (version >= 1 and attempts >= 0),
constraint tasks_retention_check check (retain_until >= execution_deadline),
constraint tasks_approval_check check (
approval_reference is null or octet_length(approval_reference) between 1 and 160
)
);
create unique index tasks_scoped_key_uidx
on tasks (workspace_id, owner_kind, owner_id, authorization_scope, operation_id, scoped_key);
create index tasks_retention_idx on tasks (state, retain_until, task_id);
create index tasks_deadline_idx on tasks (state, execution_deadline, task_id);
-- Dispatch intent: прежняя reserved-строка admission становится точкой связи Task и capacity.
-- FK без cascade: удаление принятой Task не маскируется housekeeping-удалением intent.
alter table execution_admission_entries
add column task_id uuid,
add constraint execution_admission_entries_task_fkey
foreign key (task_id) references tasks (task_id);
create unique index execution_admission_entries_task_uidx
on execution_admission_entries (task_id) where task_id is not null;
@@ -0,0 +1,45 @@
use sqlx::{Postgres, Transaction, query};
use super::authority::{MigrationDescriptor, MigrationError};
pub(super) const NAME: &str = "durable-tasks-v26";
pub(super) const SOURCE: &str = include_str!("durable_tasks_v26.sql");
pub(super) const SOURCE_SHA256: &str =
"6dcbf9990deabb3984a333acbfc3cb490cfc886c2ec30093e3b46d9113616089";
pub(super) async fn apply(
transaction: &mut Transaction<'_, Postgres>,
descriptor: &MigrationDescriptor,
) -> Result<(), MigrationError> {
sqlx::raw_sql(SOURCE)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.durable_tasks_v26",
Some(26),
"restore_known_good_backup",
)
})?;
query(
"insert into __crank_migrations (version, name, checksum, phase, compatibility)
values ($1, $2, $3, $4, $5)",
)
.bind(descriptor.version)
.bind(descriptor.name)
.bind(&descriptor.checksum)
.bind(descriptor.phase)
.bind(descriptor.compatibility)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.durable_tasks_v26",
Some(26),
"restore_known_good_backup",
)
})?;
Ok(())
}
@@ -0,0 +1,6 @@
-- V26: PostgreSQL lease/fencing/CAS additions for durable background Tasks (Story 4.3).
alter table tasks add column if not exists fence bigint not null default 0;
alter table tasks add column if not exists lease_owner text;
alter table tasks add column if not exists lease_until timestamptz;
alter table tasks add column if not exists current_attempt_id text;
alter table tasks add column if not exists dispatch_started_at timestamptz;
@@ -0,0 +1,45 @@
use sqlx::{Postgres, Transaction, query};
use super::authority::{MigrationDescriptor, MigrationError};
pub(super) const NAME: &str = "durable-tasks-v27";
pub(super) const SOURCE: &str = include_str!("durable_tasks_v27.sql");
pub(super) const SOURCE_SHA256: &str =
"a3c5335cbf81c244359faaee993487ae132d8b3be6b9c63fa322788021273dc8";
pub(super) async fn apply(
transaction: &mut Transaction<'_, Postgres>,
descriptor: &MigrationDescriptor,
) -> Result<(), MigrationError> {
sqlx::raw_sql(SOURCE)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.durable_tasks_v27",
Some(27),
"restore_known_good_backup",
)
})?;
query(
"insert into __crank_migrations (version, name, checksum, phase, compatibility)
values ($1, $2, $3, $4, $5)",
)
.bind(descriptor.version)
.bind(descriptor.name)
.bind(&descriptor.checksum)
.bind(descriptor.phase)
.bind(descriptor.compatibility)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.durable_tasks_v27",
Some(27),
"restore_known_good_backup",
)
})?;
Ok(())
}
@@ -0,0 +1,19 @@
-- V27: encrypted executable Task input (Story 4.3).
alter table tasks add column if not exists executable_input_ciphertext text;
alter table tasks add column if not exists executable_input_key_version text;
alter table tasks add column if not exists executable_input_master_key_epoch bigint;
alter table tasks add constraint tasks_executable_input_key_version_check check (
executable_input_ciphertext is null
or (executable_input_key_version is not null and octet_length(executable_input_key_version) between 1 and 64)
);
alter table tasks add column if not exists target_executable_input_ciphertext text;
alter table tasks add column if not exists target_executable_input_key_version text;
alter table tasks add column if not exists target_executable_input_master_key_epoch bigint;
alter table tasks add constraint tasks_executable_input_epoch_check check (
executable_input_ciphertext is null
or (executable_input_master_key_epoch is not null
and executable_input_master_key_epoch >= 1)
);
@@ -0,0 +1,45 @@
use sqlx::{Postgres, Transaction, query};
use super::authority::{MigrationDescriptor, MigrationError};
pub(super) const NAME: &str = "durable-tasks-v28";
pub(super) const SOURCE: &str = include_str!("durable_tasks_v28.sql");
pub(super) const SOURCE_SHA256: &str =
"4dc77739c2ea1e32a4b00abd70df0f00fd82b7cf7933372f869f2c1e997a3676";
pub(super) async fn apply(
transaction: &mut Transaction<'_, Postgres>,
descriptor: &MigrationDescriptor,
) -> Result<(), MigrationError> {
sqlx::raw_sql(SOURCE)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.durable_tasks_v28",
Some(28),
"restore_known_good_backup",
)
})?;
query(
"insert into __crank_migrations (version, name, checksum, phase, compatibility)
values ($1, $2, $3, $4, $5)",
)
.bind(descriptor.version)
.bind(descriptor.name)
.bind(&descriptor.checksum)
.bind(descriptor.phase)
.bind(descriptor.compatibility)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.durable_tasks_v28",
Some(28),
"restore_known_good_backup",
)
})?;
Ok(())
}
@@ -0,0 +1,8 @@
-- V28: typed Platform API key authority reference of a durable Task (Story 4.3).
-- Only the immutable key identifier is stored, never the bearer secret.
alter table tasks add column if not exists platform_api_key_id text;
alter table tasks add constraint tasks_platform_api_key_id_check check (
platform_api_key_id is null
or octet_length(platform_api_key_id) between 1 and 160
);
@@ -0,0 +1,45 @@
use sqlx::{Postgres, Transaction, query};
use super::authority::{MigrationDescriptor, MigrationError};
pub(super) const NAME: &str = "durable-tasks-v29";
pub(super) const SOURCE: &str = include_str!("durable_tasks_v29.sql");
pub(super) const SOURCE_SHA256: &str =
"0666c7d2aa440e6aae2170ceaf2fce2549118612e1505ef78817be60436a75dd";
pub(super) async fn apply(
transaction: &mut Transaction<'_, Postgres>,
descriptor: &MigrationDescriptor,
) -> Result<(), MigrationError> {
sqlx::raw_sql(SOURCE)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.durable_tasks_v29",
Some(29),
"restore_known_good_backup",
)
})?;
query(
"insert into __crank_migrations (version, name, checksum, phase, compatibility)
values ($1, $2, $3, $4, $5)",
)
.bind(descriptor.version)
.bind(descriptor.name)
.bind(&descriptor.checksum)
.bind(descriptor.phase)
.bind(descriptor.compatibility)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.durable_tasks_v29",
Some(29),
"restore_known_good_backup",
)
})?;
Ok(())
}
@@ -0,0 +1,52 @@
-- V29: durable per-outbound attempt history, pinned retry policy and cancellation (Story 4.4).
-- An accepted Task keeps the retry snapshot it was admitted with; legacy Tasks get a single attempt.
alter table tasks add column if not exists max_attempts integer not null default 1;
alter table tasks add column if not exists retry_backoff_step_ms bigint not null default 0;
alter table tasks add column if not exists next_attempt_at timestamptz;
alter table tasks add column if not exists cancel_requested_at timestamptz;
alter table tasks add column if not exists cancel_reason text;
alter table tasks add constraint tasks_max_attempts_check check (max_attempts between 1 and 5);
alter table tasks add constraint tasks_retry_backoff_step_check
check (retry_backoff_step_ms between 0 and 1000);
alter table tasks add constraint tasks_cancel_reason_check check (
cancel_reason is null or octet_length(cancel_reason) between 1 and 160
);
-- One row per committed outbound attempt. The authoritative outcome of a Task stays on `tasks`;
-- this history only records what was attempted and how each attempt resolved.
create table if not exists task_attempts (
attempt_id text primary key,
task_id uuid not null references tasks (task_id) on delete cascade,
ordinal integer not null,
owner text not null,
fence bigint not null default 0,
input_fingerprint text not null,
upstream_key_digest text,
retry_allowed boolean not null default false,
checkpoint_at timestamptz not null default now(),
resolved_at timestamptz,
outcome text,
certainty text,
phase text,
constraint task_attempts_task_ordinal_unique unique (task_id, ordinal),
constraint task_attempts_ordinal_check check (ordinal between 1 and 5),
constraint task_attempts_outcome_check check (
outcome is null
or outcome in ('completed', 'failed', 'retry_scheduled', 'outcome_unknown')
),
constraint task_attempts_certainty_check check (
certainty is null or certainty in ('known', 'unknown')
),
constraint task_attempts_phase_check check (
phase is null or phase in ('before_dispatch', 'dispatched')
),
-- A resolved attempt is complete: an unresolved row carries no resolution at all, so a stale
-- writer can never append a second outcome to the same attempt.
constraint task_attempts_resolution_check check (
(resolved_at is null and outcome is null and certainty is null and phase is null)
or (resolved_at is not null and outcome is not null and certainty is not null)
)
);
create index if not exists task_attempts_task_id_idx on task_attempts (task_id, ordinal);
@@ -0,0 +1,45 @@
use sqlx::{Postgres, Transaction, query};
use super::authority::{MigrationDescriptor, MigrationError};
pub(super) const NAME: &str = "execution-admission-v24";
pub(super) const SOURCE: &str = include_str!("execution_admission_v24.sql");
pub(super) const SOURCE_SHA256: &str =
"efe021ebcd68bbed64edaa1e90c3a851f7b167b94ba65d88f525d58160f02681";
pub(super) async fn apply(
transaction: &mut Transaction<'_, Postgres>,
descriptor: &MigrationDescriptor,
) -> Result<(), MigrationError> {
sqlx::raw_sql(SOURCE)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.execution_admission_v24",
Some(24),
"restore_known_good_backup",
)
})?;
query(
"insert into __crank_migrations (version, name, checksum, phase, compatibility)
values ($1, $2, $3, $4, $5)",
)
.bind(descriptor.version)
.bind(descriptor.name)
.bind(&descriptor.checksum)
.bind(descriptor.phase)
.bind(descriptor.compatibility)
.execute(&mut **transaction)
.await
.map_err(|_| {
MigrationError::new(
"apply_failed",
"apply.canonical_ledger",
Some(24),
"restore_known_good_backup",
)
})?;
Ok(())
}
@@ -0,0 +1,92 @@
-- V24: authoritative execution admission state (Story 4.1, FR-21, FR-24).
--
-- Один deployment-wide admission: PostgreSQL хранит cursor/fairness state, ожидающие тикеты,
-- выданные permits и durable enqueue decisions, поэтому несколько процессов делят общий бюджет.
--
-- `execution_admission_entries` — bounded набор: waiting (ожидает capacity), executing (scoped
-- permit выдан, держится до явного release) и reserved (durable enqueue decision с зарезервированным
-- queue slot; не исполняется и не рекламируется как Task до Story 4.2).
--
-- Строка не хранит ни credentials, ни connection/session id, ни prepared headers, ни тело запроса:
-- quota key — (workspace, agent) и (workspace, operation) без version. `operation_version` пишется
-- только как pin исполняемых semantics и не участвует в квотах.
--
-- `expires_at` — абсолютный deadline. Queue decisions без dispatch можно expire по deadline; активный
-- permit НЕ освобождается по TTL (иначе paused/crashed owner обойдёт global limit): его снимает только
-- явный release/cancel владельца.
create table execution_admission_controllers (
workspace_id text not null,
cursor_agent text,
cursor_agent_credit integer not null default 0,
cursor_tool text,
cursor_tool_credit integer not null default 0,
updated_at timestamptz not null default now(),
constraint execution_admission_controllers_pkey primary key (workspace_id),
constraint execution_admission_controllers_workspace_check check (
octet_length(workspace_id) between 1 and 160
),
constraint execution_admission_controllers_agent_credit_check check (
cursor_agent_credit between 0 and 64
),
constraint execution_admission_controllers_tool_credit_check check (
cursor_tool_credit between 0 and 64
)
);
create table execution_admission_entries (
ticket_id bigint generated always as identity,
workspace_id text not null,
agent_key text not null,
operation_id text not null,
operation_version integer not null,
mode text not null,
policy_mode text not null,
state text not null,
owner_token text not null,
fingerprint text not null,
created_at timestamptz not null default now(),
expires_at timestamptz not null,
granted_at timestamptz,
constraint execution_admission_entries_pkey primary key (ticket_id),
constraint execution_admission_entries_workspace_check check (
octet_length(workspace_id) between 1 and 160
),
constraint execution_admission_entries_agent_check check (
octet_length(agent_key) between 1 and 160
),
constraint execution_admission_entries_operation_check check (
octet_length(operation_id) between 1 and 160
),
constraint execution_admission_entries_version_check check (operation_version >= 1),
constraint execution_admission_entries_mode_check check (
mode in ('synchronous', 'background')
),
constraint execution_admission_entries_policy_mode_check check (
policy_mode in ('synchronous', 'background', 'adaptive')
),
constraint execution_admission_entries_state_check check (
state in ('waiting', 'executing', 'reserved')
),
constraint execution_admission_entries_owner_check check (
octet_length(owner_token) between 1 and 64
),
constraint execution_admission_entries_fingerprint_check check (
octet_length(fingerprint) between 1 and 128
),
constraint execution_admission_entries_reserved_mode_check check (
state <> 'reserved' or mode = 'background'
)
);
create unique index execution_admission_entries_owner_uidx
on execution_admission_entries (owner_token);
create index execution_admission_entries_state_idx
on execution_admission_entries (state, workspace_id, ticket_id);
create index execution_admission_entries_agent_idx
on execution_admission_entries (state, workspace_id, agent_key);
create index execution_admission_entries_operation_idx
on execution_admission_entries (state, workspace_id, operation_id);
@@ -55,6 +55,11 @@ pub(super) const PROMPT_PUBLICATIONS: &[&str] = &["prompt_publications", "agent_
pub(super) const PROMPT_DEPENDENCIES: &[&str] =
&["prompt_dependencies", "prompt_message_resource_refs"];
pub(super) const EXECUTION_ADMISSION: &[&str] = &[
"execution_admission_controllers",
"execution_admission_entries",
];
pub(super) const ARTIFACT_METADATA: &[&str] = &["artifact_blobs", "artifact_sources"];
pub(super) const IMPORT_PLAN: &[&str] = &["import_plans", "import_plan_applications"];
@@ -468,6 +468,71 @@ pub(super) async fn validate_schema_fingerprint(
} else {
super::schema_guard_v23::validate_v23_prompt_dependencies(connection).await?;
}
if current_version < 24 {
if !super::schema_guard_v24::validate_v24_absent(connection).await? {
return Err(schema_error(current_version));
}
} else {
super::schema_guard_v24::validate_v24_execution_admission(connection, current_version)
.await?;
}
if current_version < 25 {
if !super::schema_guard_v25::validate_v25_absent(connection).await? {
return Err(schema_error(current_version));
}
if !super::schema_guard_v26::validate_v26_absent(connection).await? {
return Err(schema_error(current_version));
}
if !super::schema_guard_v27::validate_v27_absent(connection).await? {
return Err(schema_error(current_version));
}
if !super::schema_guard_v28::validate_v28_absent(connection).await? {
return Err(schema_error(current_version));
}
if !super::schema_guard_v29::validate_v29_absent(connection).await? {
return Err(schema_error(current_version));
}
} else if current_version < 26 {
super::schema_guard_v25::validate_v25_durable_tasks(connection).await?;
if !super::schema_guard_v26::validate_v26_absent(connection).await? {
return Err(schema_error(current_version));
}
if !super::schema_guard_v28::validate_v28_absent(connection).await? {
return Err(schema_error(current_version));
}
if !super::schema_guard_v29::validate_v29_absent(connection).await? {
return Err(schema_error(current_version));
}
} else if current_version < 27 {
super::schema_guard_v26::validate_v26_durable_tasks(connection).await?;
if !super::schema_guard_v27::validate_v27_absent(connection).await? {
return Err(schema_error(current_version));
}
if !super::schema_guard_v28::validate_v28_absent(connection).await? {
return Err(schema_error(current_version));
}
if !super::schema_guard_v29::validate_v29_absent(connection).await? {
return Err(schema_error(current_version));
}
} else if current_version < 28 {
super::schema_guard_v26::validate_v26_base_columns(connection).await?;
super::schema_guard_v27::validate_v27_durable_tasks(connection).await?;
if !super::schema_guard_v28::validate_v28_absent(connection).await? {
return Err(schema_error(current_version));
}
if !super::schema_guard_v29::validate_v29_absent(connection).await? {
return Err(schema_error(current_version));
}
} else if current_version < 29 {
super::schema_guard_v26::validate_v26_base_columns(connection).await?;
super::schema_guard_v28::validate_v28_durable_tasks(connection).await?;
if !super::schema_guard_v29::validate_v29_absent(connection).await? {
return Err(schema_error(current_version));
}
} else {
super::schema_guard_v26::validate_v26_base_columns(connection).await?;
super::schema_guard_v29::validate_v29_durable_attempts(connection).await?;
}
Ok(())
}
@@ -51,6 +51,8 @@ pub(crate) const OWNED_RELATIONS: &[&str] = &[
"prompt_messages",
"prompt_dependencies",
"prompt_message_resource_refs",
"execution_admission_controllers",
"execution_admission_entries",
];
pub(crate) const REQUIRED_COLUMNS: &[(&str, &[&str])] = &[
@@ -0,0 +1,277 @@
//! V24 schema guard: authoritative execution admission state (Story 4.1, FR-21/FR-24).
//!
//! Схема V24 либо полностью отсутствует (миграция ещё не применена), либо присутствует ровно в
//! ожидаемой форме: controller-row с weighted round-robin cursor и bounded entries с закрытым
//! словарём mode/policy_mode/state, уникальным owner token и индексами квот.
use sqlx::{PgConnection, Row, query};
use super::authority::MigrationError;
use super::schema_guard::relation_exists;
const CONTROLLERS: &str = "execution_admission_controllers";
const ENTRIES: &str = "execution_admission_entries";
const OWNED_RELATIONS: &[&str] = &[CONTROLLERS, ENTRIES];
const CONTROLLER_COLUMNS: &[(&str, &str, bool)] = &[
("workspace_id", "text", false),
("cursor_agent", "text", true),
("cursor_agent_credit", "integer", false),
("cursor_tool", "text", true),
("cursor_tool_credit", "integer", false),
("updated_at", "timestamp with time zone", false),
];
const ENTRY_COLUMNS: &[(&str, &str, bool)] = &[
("ticket_id", "bigint", false),
("workspace_id", "text", false),
("agent_key", "text", false),
("operation_id", "text", false),
("operation_version", "integer", false),
("mode", "text", false),
("policy_mode", "text", false),
("state", "text", false),
("owner_token", "text", false),
("fingerprint", "text", false),
("created_at", "timestamp with time zone", false),
("expires_at", "timestamp with time zone", false),
("granted_at", "timestamp with time zone", true),
];
/// V25 adds the nullable Task dispatch-intent link to the same reserved row. At version 24 the
/// column must not exist yet, so the guard validates the shape of the version it describes.
fn entry_columns(current_version: i64) -> Vec<(&'static str, &'static str, bool)> {
let mut columns = ENTRY_COLUMNS.to_vec();
if current_version >= 25 {
columns.push(("task_id", "uuid", true));
}
columns
}
fn constraint_count(table: &str, current_version: i64) -> Option<usize> {
match table {
CONTROLLERS => Some(4),
ENTRIES if current_version >= 25 => Some(12),
ENTRIES => Some(11),
_ => None,
}
}
const UNIQUE_INDEXES: &[&str] = &["execution_admission_entries_owner_uidx"];
fn v24_error(stage: &'static str) -> MigrationError {
MigrationError::new(
"partial_sequence",
stage,
Some(24),
"restore_known_good_backup",
)
}
/// Схема V24 отсутствует целиком (для current_version < 24).
pub(super) async fn validate_v24_absent(
connection: &mut PgConnection,
) -> Result<bool, MigrationError> {
for relation in OWNED_RELATIONS {
if relation_exists(connection, relation).await? {
return Ok(false);
}
}
for index in UNIQUE_INDEXES {
if index_exists(connection, index).await? {
return Ok(false);
}
}
Ok(true)
}
/// Схема V24 присутствует в ожидаемой форме (для current_version >= 24).
pub(super) async fn validate_v24_execution_admission(
connection: &mut PgConnection,
current_version: i64,
) -> Result<(), MigrationError> {
for relation in OWNED_RELATIONS {
if !relation_exists(connection, relation).await? {
return Err(v24_error("v24.relations"));
}
}
let tables = [
(CONTROLLERS, CONTROLLER_COLUMNS.to_vec()),
(ENTRIES, entry_columns(current_version)),
];
for (table, columns) in tables {
validate_columns(connection, table, &columns).await?;
}
for table in [CONTROLLERS, ENTRIES] {
let expected = constraint_count(table, current_version).unwrap_or_default();
let count: i64 = query(
"select count(*)::bigint as count
from pg_catalog.pg_constraint c
join pg_catalog.pg_class t on t.oid = c.conrelid
join pg_catalog.pg_namespace n on n.oid = t.relnamespace
where n.nspname = current_schema() and t.relname = $1",
)
.bind(table)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get("count")
.map_err(|_| MigrationError::storage("preflight.schema"))?;
if count != i64::try_from(expected).unwrap_or_default() {
return Err(v24_error("v24.constraints"));
}
}
for (table, constraint) in [
(CONTROLLERS, "execution_admission_controllers_pkey"),
(ENTRIES, "execution_admission_entries_pkey"),
(ENTRIES, "execution_admission_entries_state_check"),
] {
if !constraint_exists(connection, table, constraint).await? {
return Err(v24_error("v24.constraint_missing"));
}
}
if !normalized_expression(
connection,
ENTRIES,
"execution_admission_entries_state_check",
)
.await?
.is_some_and(|expression| {
super::schema_guard::enum_constraint_matches(
&expression,
"state",
&["waiting", "executing", "reserved"],
)
}) {
return Err(v24_error("v24.state_vocabulary"));
}
for index in UNIQUE_INDEXES {
let unique = index_unique(connection, index).await?;
if unique != Some(true) {
return Err(v24_error("v24.indexes"));
}
}
Ok(())
}
async fn validate_columns(
connection: &mut PgConnection,
table: &str,
columns: &[(&str, &str, bool)],
) -> Result<(), MigrationError> {
let count: i64 = query(
"select count(*)::bigint as count from information_schema.columns
where table_schema = current_schema() and table_name = $1",
)
.bind(table)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get("count")
.map_err(|_| MigrationError::storage("preflight.schema"))?;
if count != i64::try_from(columns.len()).unwrap_or_default() {
return Err(v24_error("v24.columns"));
}
for (column, data_type, nullable) in columns {
let row = query(
"select data_type, is_nullable from information_schema.columns
where table_schema = current_schema() and table_name = $1 and column_name = $2",
)
.bind(table)
.bind(column)
.fetch_optional(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?;
let valid = row.is_some_and(|row| {
row.try_get::<String, _>("data_type").ok().as_deref() == Some(*data_type)
&& row.try_get::<String, _>("is_nullable").ok().as_deref()
== Some(if *nullable { "YES" } else { "NO" })
});
if !valid {
return Err(v24_error("v24.columns"));
}
}
Ok(())
}
async fn constraint_exists(
connection: &mut PgConnection,
table: &str,
constraint: &str,
) -> Result<bool, MigrationError> {
query(
"select exists (
select 1 from pg_catalog.pg_constraint c
join pg_catalog.pg_class t on t.oid = c.conrelid
join pg_catalog.pg_namespace n on n.oid = t.relnamespace
where n.nspname = current_schema() and t.relname = $1 and c.conname = $2
) as present",
)
.bind(table)
.bind(constraint)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get::<bool, _>("present")
.map_err(|_| MigrationError::storage("preflight.schema"))
}
async fn normalized_expression(
connection: &mut PgConnection,
table: &str,
constraint: &str,
) -> Result<Option<String>, MigrationError> {
let row = query(
"select pg_get_expr(c.conbin, c.conrelid) as expression
from pg_catalog.pg_constraint c
join pg_catalog.pg_class t on t.oid = c.conrelid
join pg_catalog.pg_namespace n on n.oid = t.relnamespace
where n.nspname = current_schema() and t.relname = $1 and c.conname = $2
and c.contype = 'c'",
)
.bind(table)
.bind(constraint)
.fetch_optional(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?;
Ok(row.and_then(|row| row.try_get::<String, _>("expression").ok()))
}
async fn index_exists(connection: &mut PgConnection, index: &str) -> Result<bool, MigrationError> {
query(
"select exists (
select 1 from pg_catalog.pg_class idx
join pg_catalog.pg_namespace n on n.oid = idx.relnamespace
where n.nspname = current_schema() and idx.relname = $1
) as present",
)
.bind(index)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get::<bool, _>("present")
.map_err(|_| MigrationError::storage("preflight.schema"))
}
async fn index_unique(
connection: &mut PgConnection,
index: &str,
) -> Result<Option<bool>, MigrationError> {
let row = query(
"select i.indisunique, i.indisvalid
from pg_catalog.pg_index i
join pg_catalog.pg_class idx on idx.oid = i.indexrelid
join pg_catalog.pg_namespace n on n.oid = idx.relnamespace
where n.nspname = current_schema() and idx.relname = $1",
)
.bind(index)
.fetch_optional(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?;
Ok(row.and_then(|row| {
let unique = row.try_get::<bool, _>("indisunique").ok()?;
let valid = row.try_get::<bool, _>("indisvalid").ok()?;
Some(unique && valid)
}))
}
@@ -0,0 +1,247 @@
//! V25 schema guard: durable background Tasks (Story 4.2, FR-22/FR-23).
//!
//! Схема V25 либо полностью отсутствует (миграция ещё не применена), либо присутствует ровно в
//! ожидаемой форме: таблица `tasks` с закрытым словарём state/policy_mode и scoped uniqueness, плюс
//! nullable `task_id` на существующей reserved-строке admission с FK и UNIQUE-индексом.
use sqlx::{PgConnection, Row, query};
use super::authority::MigrationError;
use super::schema_guard::relation_exists;
const TASKS: &str = "tasks";
const ENTRIES: &str = "execution_admission_entries";
const SCOPED_INDEX: &str = "tasks_scoped_key_uidx";
const TASK_INDEX: &str = "execution_admission_entries_task_uidx";
const TASK_FK: &str = "execution_admission_entries_task_fkey";
const TASK_COLUMNS: &[(&str, &str, bool)] = &[
("task_id", "uuid", false),
("workspace_id", "text", false),
("owner_kind", "text", false),
("owner_id", "text", false),
("authorization_scope", "text", false),
("operation_id", "text", false),
("operation_version", "integer", false),
("scoped_key", "text", false),
("input_fingerprint", "text", false),
("request_id", "text", false),
("trace_id", "text", false),
("traceparent", "text", false),
("trace_sampled", "boolean", false),
("policy_mode", "text", false),
("approval_reference", "text", true),
("state", "text", false),
("version", "integer", false),
("attempts", "integer", false),
("input_envelope", "jsonb", false),
("progress_envelope", "jsonb", true),
("result_envelope", "jsonb", true),
("error_envelope", "jsonb", true),
("execution_deadline", "timestamp with time zone", false),
("retain_until", "timestamp with time zone", false),
("created_at", "timestamp with time zone", false),
("updated_at", "timestamp with time zone", false),
];
fn v25_error(stage: &'static str) -> MigrationError {
MigrationError::new(
"partial_sequence",
stage,
Some(25),
"restore_known_good_backup",
)
}
/// Схема V25 отсутствует целиком (для current_version < 25).
pub(super) async fn validate_v25_absent(
connection: &mut PgConnection,
) -> Result<bool, MigrationError> {
if relation_exists(connection, TASKS).await? {
return Ok(false);
}
if super::schema_guard::column_exists(connection, ENTRIES, "task_id").await? {
return Ok(false);
}
for index in [SCOPED_INDEX, TASK_INDEX] {
if index_exists(connection, index).await? {
return Ok(false);
}
}
Ok(true)
}
/// Схема V25 присутствует в ожидаемой форме (для current_version >= 25).
pub(super) async fn validate_v25_durable_tasks(
connection: &mut PgConnection,
) -> Result<(), MigrationError> {
if !relation_exists(connection, TASKS).await? {
return Err(v25_error("v25.relations"));
}
let count: i64 = query(
"select count(*)::bigint as count from information_schema.columns
where table_schema = current_schema() and table_name = $1",
)
.bind(TASKS)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get("count")
.map_err(|_| MigrationError::storage("preflight.schema"))?;
if count != i64::try_from(TASK_COLUMNS.len()).unwrap_or_default() {
return Err(v25_error("v25.columns"));
}
for (column, data_type, nullable) in TASK_COLUMNS {
let row = query(
"select data_type, is_nullable from information_schema.columns
where table_schema = current_schema() and table_name = $1 and column_name = $2",
)
.bind(TASKS)
.bind(column)
.fetch_optional(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?;
let valid = row.is_some_and(|row| {
row.try_get::<String, _>("data_type").ok().as_deref() == Some(*data_type)
&& row.try_get::<String, _>("is_nullable").ok().as_deref()
== Some(if *nullable { "YES" } else { "NO" })
});
if !valid {
return Err(v25_error("v25.columns"));
}
}
if !super::schema_guard::column_exists(connection, ENTRIES, "task_id").await? {
return Err(v25_error("v25.intent_column"));
}
for (constraint, vocabulary) in [
(
"tasks_state_check",
[
"queued",
"working",
"input_required",
"outcome_unknown",
"completed",
"failed",
"cancelled",
]
.as_slice(),
),
(
"tasks_policy_mode_check",
["synchronous", "background", "adaptive"].as_slice(),
),
] {
let expression = normalized_expression(connection, TASKS, constraint).await?;
if !expression.is_some_and(|expression| {
super::schema_guard::enum_constraint_matches(
&expression,
column_of(constraint),
vocabulary,
)
}) {
return Err(v25_error("v25.vocabulary"));
}
}
for (index, table) in [(SCOPED_INDEX, TASKS), (TASK_INDEX, ENTRIES)] {
if index_unique(connection, table, index).await? != Some(true) {
return Err(v25_error("v25.indexes"));
}
}
if !constraint_exists(connection, ENTRIES, TASK_FK).await? {
return Err(v25_error("v25.intent_fk"));
}
Ok(())
}
fn column_of(constraint: &str) -> &'static str {
if constraint.contains("policy_mode") {
"policy_mode"
} else {
"state"
}
}
async fn constraint_exists(
connection: &mut PgConnection,
table: &str,
constraint: &str,
) -> Result<bool, MigrationError> {
query(
"select exists (
select 1 from pg_catalog.pg_constraint c
join pg_catalog.pg_class t on t.oid = c.conrelid
join pg_catalog.pg_namespace n on n.oid = t.relnamespace
where n.nspname = current_schema() and t.relname = $1 and c.conname = $2
) as present",
)
.bind(table)
.bind(constraint)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get::<bool, _>("present")
.map_err(|_| MigrationError::storage("preflight.schema"))
}
async fn normalized_expression(
connection: &mut PgConnection,
table: &str,
constraint: &str,
) -> Result<Option<String>, MigrationError> {
let row = query(
"select pg_get_expr(c.conbin, c.conrelid) as expression
from pg_catalog.pg_constraint c
join pg_catalog.pg_class t on t.oid = c.conrelid
join pg_catalog.pg_namespace n on n.oid = t.relnamespace
where n.nspname = current_schema() and t.relname = $1 and c.conname = $2
and c.contype = 'c'",
)
.bind(table)
.bind(constraint)
.fetch_optional(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?;
Ok(row.and_then(|row| row.try_get::<String, _>("expression").ok()))
}
async fn index_exists(connection: &mut PgConnection, index: &str) -> Result<bool, MigrationError> {
query(
"select exists (
select 1 from pg_catalog.pg_class idx
join pg_catalog.pg_namespace n on n.oid = idx.relnamespace
where n.nspname = current_schema() and idx.relname = $1
) as present",
)
.bind(index)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get::<bool, _>("present")
.map_err(|_| MigrationError::storage("preflight.schema"))
}
async fn index_unique(
connection: &mut PgConnection,
table: &str,
index: &str,
) -> Result<Option<bool>, MigrationError> {
let row = query(
"select i.indisunique, i.indisvalid
from pg_catalog.pg_index i
join pg_catalog.pg_class idx on idx.oid = i.indexrelid
join pg_catalog.pg_class t on t.oid = i.indrelid
join pg_catalog.pg_namespace n on n.oid = idx.relnamespace
where n.nspname = current_schema() and idx.relname = $1 and t.relname = $2",
)
.bind(index)
.bind(table)
.fetch_optional(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?;
Ok(row.and_then(|row| {
let unique = row.try_get::<bool, _>("indisunique").ok()?;
let valid = row.try_get::<bool, _>("indisvalid").ok()?;
Some(unique && valid)
}))
}
@@ -0,0 +1,120 @@
//! V26 schema guard: lease/fencing columns for durable background Tasks (Story 4.3).
use sqlx::{PgConnection, Row, query};
use super::authority::MigrationError;
use super::schema_guard::relation_exists;
const TASKS: &str = "tasks";
const TASK_COLUMNS: &[(&str, &str, bool)] = &[
("task_id", "uuid", false),
("workspace_id", "text", false),
("owner_kind", "text", false),
("owner_id", "text", false),
("authorization_scope", "text", false),
("operation_id", "text", false),
("operation_version", "integer", false),
("scoped_key", "text", false),
("input_fingerprint", "text", false),
("request_id", "text", false),
("trace_id", "text", false),
("traceparent", "text", false),
("trace_sampled", "boolean", false),
("policy_mode", "text", false),
("approval_reference", "text", true),
("state", "text", false),
("version", "integer", false),
("attempts", "integer", false),
("input_envelope", "jsonb", false),
("progress_envelope", "jsonb", true),
("result_envelope", "jsonb", true),
("error_envelope", "jsonb", true),
("execution_deadline", "timestamp with time zone", false),
("retain_until", "timestamp with time zone", false),
("created_at", "timestamp with time zone", false),
("updated_at", "timestamp with time zone", false),
("fence", "bigint", false),
("lease_owner", "text", true),
("lease_until", "timestamp with time zone", true),
("current_attempt_id", "text", true),
("dispatch_started_at", "timestamp with time zone", true),
];
fn v26_error(stage: &'static str) -> MigrationError {
MigrationError::new(
"partial_sequence",
stage,
Some(26),
"restore_known_good_backup",
)
}
pub(super) async fn validate_v26_absent(
connection: &mut PgConnection,
) -> Result<bool, MigrationError> {
if relation_exists(connection, TASKS).await? {
// delegate to v25 absent for relation mostly; V26 columns must not exist.
for column in [
"fence",
"lease_owner",
"lease_until",
"current_attempt_id",
"dispatch_started_at",
] {
if super::schema_guard::column_exists(connection, TASKS, column).await? {
return Ok(false);
}
}
}
Ok(true)
}
pub(super) async fn validate_v26_base_columns(
connection: &mut PgConnection,
) -> Result<(), MigrationError> {
if !relation_exists(connection, TASKS).await? {
return Err(v26_error("v26.relations"));
}
for (column, data_type, nullable) in TASK_COLUMNS {
let row = query(
"select data_type, is_nullable from information_schema.columns
where table_schema = current_schema() and table_name = $1 and column_name = $2",
)
.bind(TASKS)
.bind(column)
.fetch_optional(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?;
let valid = row.is_some_and(|row| {
row.try_get::<String, _>("data_type").ok().as_deref() == Some(*data_type)
&& row.try_get::<String, _>("is_nullable").ok().as_deref()
== Some(if *nullable { "YES" } else { "NO" })
});
if !valid {
return Err(v26_error("v26.columns"));
}
}
Ok(())
}
pub(super) async fn validate_v26_durable_tasks(
connection: &mut PgConnection,
) -> Result<(), MigrationError> {
validate_v26_base_columns(connection).await?;
let count: i64 = query(
"select count(*)::bigint as count from information_schema.columns
where table_schema = current_schema() and table_name = $1",
)
.bind(TASKS)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get("count")
.map_err(|_| MigrationError::storage("preflight.schema"))?;
if count != i64::try_from(TASK_COLUMNS.len()).unwrap_or_default() {
return Err(v26_error("v26.columns"));
}
// Constraints/index checks inherited conceptually from V25; V26 adds columns only.
Ok(())
}
@@ -0,0 +1,85 @@
//! V27 schema guard: encrypted executable Task input columns (Story 4.3).
use sqlx::{PgConnection, Row, query};
use super::authority::MigrationError;
use super::schema_guard::relation_exists;
use super::schema_guard_v26::validate_v26_base_columns;
const TASKS: &str = "tasks";
pub(super) async fn validate_v27_absent(
connection: &mut PgConnection,
) -> Result<bool, MigrationError> {
if relation_exists(connection, TASKS).await? {
for column in [
"executable_input_ciphertext",
"executable_input_key_version",
"executable_input_master_key_epoch",
] {
if super::schema_guard::column_exists(connection, TASKS, column).await? {
return Ok(false);
}
}
}
Ok(true)
}
pub(super) async fn validate_v27_durable_tasks(
connection: &mut PgConnection,
) -> Result<(), MigrationError> {
validate_v26_base_columns(connection).await?;
// Additional V27 columns.
let extra = [
("executable_input_ciphertext", "text", true),
("executable_input_key_version", "text", true),
("executable_input_master_key_epoch", "bigint", true),
("target_executable_input_ciphertext", "text", true),
("target_executable_input_key_version", "text", true),
("target_executable_input_master_key_epoch", "bigint", true),
];
for (column, data_type, nullable) in extra {
let row = query(
"select data_type, is_nullable from information_schema.columns
where table_schema = current_schema() and table_name = $1 and column_name = $2",
)
.bind(TASKS)
.bind(column)
.fetch_optional(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?;
let valid = row.is_some_and(|row| {
row.try_get::<String, _>("data_type").ok().as_deref() == Some(data_type)
&& row.try_get::<String, _>("is_nullable").ok().as_deref()
== Some(if nullable { "YES" } else { "NO" })
});
if !valid {
return Err(MigrationError::new(
"partial_sequence",
"v27.columns",
Some(27),
"restore_known_good_backup",
));
}
}
// Exact total column count for version 27: 43 V26 columns + 3 V27 columns.
let count: i64 = query(
"select count(*)::bigint as count from information_schema.columns
where table_schema = current_schema() and table_name = $1",
)
.bind(TASKS)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get("count")
.map_err(|_| MigrationError::storage("preflight.schema"))?;
if count != 37_i64 {
return Err(MigrationError::new(
"partial_sequence",
"v27.columns",
Some(27),
"restore_known_good_backup",
));
}
Ok(())
}
@@ -0,0 +1,89 @@
//! V28 schema guard: typed Platform API key authority reference of a durable Task (Story 4.3).
use sqlx::{PgConnection, Row, query};
use super::authority::MigrationError;
use super::schema_guard::relation_exists;
const TASKS: &str = "tasks";
pub(super) async fn validate_v28_absent(
connection: &mut PgConnection,
) -> Result<bool, MigrationError> {
if relation_exists(connection, TASKS).await?
&& super::schema_guard::column_exists(connection, TASKS, "platform_api_key_id").await?
{
return Ok(false);
}
Ok(true)
}
pub(super) async fn validate_v28_durable_tasks(
connection: &mut PgConnection,
) -> Result<(), MigrationError> {
// The V27 column set is still required and unchanged; V28 only appends one column.
validate_v27_base_columns(connection).await?;
let row = query(
"select data_type, is_nullable from information_schema.columns
where table_schema = current_schema() and table_name = $1 and column_name = $2",
)
.bind(TASKS)
.bind("platform_api_key_id")
.fetch_optional(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?;
let valid = row.is_some_and(|row| {
row.try_get::<String, _>("data_type").ok().as_deref() == Some("text")
&& row.try_get::<String, _>("is_nullable").ok().as_deref() == Some("YES")
});
if !valid {
return Err(MigrationError::new(
"partial_sequence",
"v28.columns",
Some(28),
"restore_known_good_backup",
));
}
// Exact total column count for version 28: the 37 V27 columns + 1 V28 column.
let count: i64 = query(
"select count(*)::bigint as count from information_schema.columns
where table_schema = current_schema() and table_name = $1",
)
.bind(TASKS)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get("count")
.map_err(|_| MigrationError::storage("preflight.schema"))?;
if count != 38_i64 {
return Err(MigrationError::new(
"partial_sequence",
"v28.columns",
Some(28),
"restore_known_good_backup",
));
}
Ok(())
}
/// V27 columns must still be exactly present before V28 is considered applied.
async fn validate_v27_base_columns(connection: &mut PgConnection) -> Result<(), MigrationError> {
for column in [
"executable_input_ciphertext",
"executable_input_key_version",
"executable_input_master_key_epoch",
"target_executable_input_ciphertext",
"target_executable_input_key_version",
"target_executable_input_master_key_epoch",
] {
if !super::schema_guard::column_exists(connection, TASKS, column).await? {
return Err(MigrationError::new(
"partial_sequence",
"v28.columns",
Some(28),
"restore_known_good_backup",
));
}
}
Ok(())
}
@@ -0,0 +1,87 @@
//! V29 schema guard: durable per-outbound attempt history, pinned retry policy and cancellation
//! (Story 4.4).
use sqlx::{PgConnection, Row, query};
use super::authority::MigrationError;
use super::schema_guard::{column_exists, relation_exists};
const TASKS: &str = "tasks";
const ATTEMPTS: &str = "task_attempts";
pub(super) async fn validate_v29_absent(
connection: &mut PgConnection,
) -> Result<bool, MigrationError> {
if relation_exists(connection, TASKS).await? {
if column_exists(connection, TASKS, "max_attempts").await? {
return Ok(false);
}
if relation_exists(connection, ATTEMPTS).await? {
return Ok(false);
}
}
Ok(true)
}
pub(super) async fn validate_v29_durable_attempts(
connection: &mut PgConnection,
) -> Result<(), MigrationError> {
// The V28 column set is still required and unchanged; V29 only appends columns and one table.
for column in [
"platform_api_key_id",
"max_attempts",
"retry_backoff_step_ms",
"next_attempt_at",
"cancel_requested_at",
"cancel_reason",
] {
if !column_exists(connection, TASKS, column).await? {
return Err(invalid(29, "v29.columns"));
}
}
// Exact total column count for version 29: the 38 V28 columns + 5 V29 columns.
let count: i64 = query(
"select count(*)::bigint as count from information_schema.columns
where table_schema = current_schema() and table_name = $1",
)
.bind(TASKS)
.fetch_one(&mut *connection)
.await
.map_err(|_| MigrationError::storage("preflight.schema"))?
.try_get("count")
.map_err(|_| MigrationError::storage("preflight.schema"))?;
if count != 43_i64 {
return Err(invalid(29, "v29.columns"));
}
if !relation_exists(connection, ATTEMPTS).await? {
return Err(invalid(29, "v29.attempts"));
}
for column in [
"attempt_id",
"task_id",
"ordinal",
"owner",
"fence",
"input_fingerprint",
"upstream_key_digest",
"retry_allowed",
"checkpoint_at",
"resolved_at",
"outcome",
"certainty",
"phase",
] {
if !column_exists(connection, ATTEMPTS, column).await? {
return Err(invalid(29, "v29.attempts"));
}
}
Ok(())
}
fn invalid(version: i64, stage: &'static str) -> MigrationError {
MigrationError::new(
"partial_sequence",
stage,
Some(version),
"restore_known_good_backup",
)
}
@@ -0,0 +1,819 @@
//! PostgreSQL implementation of the mandatory execution-admission store (Story 4.1, FR-21/FR-24).
//!
//! Source of truth for capacity and durable enqueue decisions of every process: a short transaction
//! locks one deployment controller row and one workspace controller row in a fixed order, then reads
//! bounded counters, inserts/transitions at most one entry and persists the weighted round-robin
//! cursor. No network call happens under the transaction and no connection is held while a waiter is
//! idle between attempts.
use std::collections::BTreeMap;
use std::time::Duration;
use async_trait::async_trait;
use crank_core::{
AdmissionCursor, AdmissionDecision, AdmissionMode, AdmissionOutcome, AdmissionPermit,
AdmissionQueueEntry, AdmissionQueueSnapshot, AdmissionRequest, AdmissionSchedulerConfig,
AdmissionStoreError, ApprovalEvidence, ExecutionAdmission, NoTasksFallback, TaskError,
select_next_grant,
};
use sqlx::{PgPool, Postgres, Row, Transaction};
use uuid::Uuid;
const DEPLOYMENT_CONTROLLER: &str = "*";
const OVERLOAD_RETRY_AFTER_MS: u64 = crank_core::ADMISSION_OVERLOAD_RETRY_AFTER_MS;
const MAX_WAIT_BACKOFF_MS: u64 = 250;
const INITIAL_WAIT_BACKOFF_MS: u64 = 5;
const EXPIRED_CLEANUP_BATCH: i64 = 128;
const QUEUE_READ_LIMIT: i64 = 1024;
/// Deployment-wide bounds of one admission store. They come from validated deployment
/// configuration, never from a request.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct AdmissionLimits {
pub global_limit: usize,
pub agent_limit: usize,
pub tool_limit: usize,
pub queue_limit: usize,
}
impl Default for AdmissionLimits {
fn default() -> Self {
Self {
global_limit: 64,
agent_limit: 16,
tool_limit: 8,
queue_limit: 32,
}
}
}
impl AdmissionLimits {
pub fn validate(&self) -> Result<(), AdmissionStoreError> {
for (field, value) in [
("global_limit", self.global_limit),
("agent_limit", self.agent_limit),
("tool_limit", self.tool_limit),
] {
if !(1..=65_535).contains(&value) {
return Err(AdmissionStoreError::invalid(format!(
"admission {field} is outside its allowed bounds"
)));
}
}
if !(1..=1_024).contains(&self.queue_limit) {
return Err(AdmissionStoreError::invalid(
"admission queue_limit is outside its allowed bounds",
));
}
Ok(())
}
}
#[derive(Clone)]
pub struct PostgresExecutionAdmission {
pool: PgPool,
limits: AdmissionLimits,
scheduler: AdmissionSchedulerConfig,
}
impl PostgresExecutionAdmission {
pub fn new(
pool: PgPool,
limits: AdmissionLimits,
scheduler: AdmissionSchedulerConfig,
) -> Result<Self, AdmissionStoreError> {
limits.validate()?;
scheduler.validate()?;
Ok(Self {
pool,
limits,
scheduler,
})
}
fn unavailable(error: impl std::fmt::Display) -> AdmissionStoreError {
AdmissionStoreError::unavailable(format!(
"execution admission store is unavailable: {error}"
))
}
}
struct AttemptContext<'a> {
request: &'a AdmissionRequest,
owner_token: &'a str,
}
enum Attempt {
Granted,
Enqueued(u64),
Waiting,
Overloaded,
Expired,
}
#[async_trait]
impl ExecutionAdmission for PostgresExecutionAdmission {
async fn health(&self) -> Result<(), AdmissionStoreError> {
admission_health(&self.pool).await
}
async fn admit(
&self,
request: AdmissionRequest,
) -> Result<AdmissionOutcome, AdmissionStoreError> {
request.validate()?;
if request.approval == ApprovalEvidence::NotGranted {
return Ok(AdmissionOutcome::ApprovalRequired);
}
if request.mode == AdmissionMode::Background
&& request.no_tasks_fallback == NoTasksFallback::Unsupported
{
return Ok(AdmissionOutcome::Unsupported);
}
let owner_token = format!("adm_{}", Uuid::now_v7().simple());
let mut ticket: Option<u64> = None;
let mut backoff = INITIAL_WAIT_BACKOFF_MS;
loop {
if crank_core::current_unix_ms() >= request.expires_at_unix_ms {
if let Some(ticket_id) = ticket {
self.cancel(ticket_id, &owner_token).await?;
}
return Ok(AdmissionOutcome::DeadlineElapsed);
}
let attempt = self
.attempt(
AttemptContext {
request: &request,
owner_token: &owner_token,
},
&mut ticket,
)
.await?;
match attempt {
Attempt::Granted => {
return Ok(AdmissionOutcome::Permitted(AdmissionPermit {
ticket_id: ticket.unwrap_or_default(),
owner_token,
scope: request.scope.clone(),
mode: request.mode,
policy_mode: request.policy_mode,
}));
}
Attempt::Enqueued(ticket_id) => {
return Ok(AdmissionOutcome::Enqueued(AdmissionDecision {
ticket_id,
scope: request.scope.clone(),
expires_at_unix_ms: request.expires_at_unix_ms,
policy_mode: request.policy_mode,
}));
}
Attempt::Expired => {
return Ok(AdmissionOutcome::DeadlineElapsed);
}
Attempt::Overloaded => {
return Ok(AdmissionOutcome::Overloaded {
retry_after_ms: OVERLOAD_RETRY_AFTER_MS,
});
}
Attempt::Waiting => {
let remaining = request
.expires_at_unix_ms
.saturating_sub(crank_core::current_unix_ms())
.max(1);
let sleep_ms = i64::try_from(backoff).unwrap_or(remaining).min(remaining);
tokio::time::sleep(Duration::from_millis(
u64::try_from(sleep_ms).unwrap_or(backoff),
))
.await;
backoff = (backoff * 2).min(MAX_WAIT_BACKOFF_MS);
}
}
}
}
async fn release(&self, permit: &AdmissionPermit) -> Result<bool, AdmissionStoreError> {
let result = sqlx::query(
"delete from execution_admission_entries
where ticket_id = $1 and owner_token = $2 and state = 'executing'
and task_id is null",
)
.bind(i64::try_from(permit.ticket_id).unwrap_or(i64::MAX))
.bind(&permit.owner_token)
.execute(&self.pool)
.await
.map_err(Self::unavailable)?;
Ok(result.rows_affected() > 0)
}
async fn cancel(&self, ticket_id: u64, owner_token: &str) -> Result<bool, AdmissionStoreError> {
let result = sqlx::query(
"delete from execution_admission_entries
where ticket_id = $1 and owner_token = $2 and state in ('waiting', 'reserved')
and task_id is null",
)
.bind(i64::try_from(ticket_id).unwrap_or(i64::MAX))
.bind(owner_token)
.execute(&self.pool)
.await
.map_err(Self::unavailable)?;
Ok(result.rows_affected() > 0)
}
async fn snapshot(&self) -> Result<AdmissionQueueSnapshot, AdmissionStoreError> {
admission_snapshot(&self.pool).await
}
}
/// Readiness probe of the mandatory admission state: both tables must be reachable and readable.
pub async fn admission_health(pool: &PgPool) -> Result<(), AdmissionStoreError> {
sqlx::query(
"select count(*)::bigint from execution_admission_controllers where workspace_id = $1",
)
.bind(DEPLOYMENT_CONTROLLER)
.fetch_one(pool)
.await
.map_err(PostgresExecutionAdmission::unavailable)?;
sqlx::query("select count(*)::bigint from execution_admission_entries where false")
.fetch_one(pool)
.await
.map_err(PostgresExecutionAdmission::unavailable)?;
Ok(())
}
/// Authoritative queue readings of the deployment-wide admission state.
pub async fn admission_snapshot(
pool: &PgPool,
) -> Result<AdmissionQueueSnapshot, AdmissionStoreError> {
let row = sqlx::query(
"select
count(*) filter (where state in ('waiting', 'reserved')) as queued,
count(*) filter (where state = 'executing') as executing,
coalesce(
(extract(epoch from (now() - min(created_at) filter (where state in ('waiting', 'reserved')))) * 1000)::bigint,
-1
) as oldest_waiting_ms
from execution_admission_entries",
)
.fetch_one(pool)
.await
.map_err(PostgresExecutionAdmission::unavailable)?;
let queued: i64 = row
.try_get("queued")
.map_err(PostgresExecutionAdmission::unavailable)?;
let executing: i64 = row
.try_get("executing")
.map_err(PostgresExecutionAdmission::unavailable)?;
let oldest: i64 = row
.try_get("oldest_waiting_ms")
.map_err(PostgresExecutionAdmission::unavailable)?;
Ok(AdmissionQueueSnapshot {
waiting: u64::try_from(queued).unwrap_or_default(),
executing: u64::try_from(executing).unwrap_or_default(),
oldest_waiting_ms: (oldest >= 0).then(|| u64::try_from(oldest).unwrap_or_default()),
})
}
impl PostgresExecutionAdmission {
async fn attempt(
&self,
context: AttemptContext<'_>,
ticket: &mut Option<u64>,
) -> Result<Attempt, AdmissionStoreError> {
let mut transaction = self.pool.begin().await.map_err(Self::unavailable)?;
sqlx::query("set local lock_timeout = '3000ms'")
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
sqlx::query("set local statement_timeout = '5000ms'")
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
match self.attempt_locked(&mut transaction, context, ticket).await {
Ok(attempt) => {
transaction.commit().await.map_err(Self::unavailable)?;
Ok(attempt)
}
Err(AttemptError::Contended) => {
transaction.rollback().await.ok();
Ok(Attempt::Waiting)
}
Err(AttemptError::Overloaded) => {
transaction.rollback().await.ok();
Ok(Attempt::Overloaded)
}
Err(AttemptError::Failure(error)) => {
transaction.rollback().await.ok();
Err(AdmissionStoreError::unavailable(error.to_string()))
}
Err(AttemptError::AdmissionFailure(error)) => {
transaction.rollback().await.ok();
Err(error)
}
Err(AttemptError::UniqueViolation) => {
transaction.rollback().await.ok();
Err(AdmissionStoreError::unavailable(
"execution admission hit an unexpected unique conflict",
))
}
}
}
async fn attempt_locked(
&self,
transaction: &mut Transaction<'_, Postgres>,
context: AttemptContext<'_>,
ticket: &mut Option<u64>,
) -> Result<Attempt, AttemptError> {
let workspace_id = context.request.scope.workspace_id.as_str();
let cursor = lock_admission_in_tx(transaction, workspace_id).await?;
// Expired queue decisions and waiting tickets leave the bounded queue; held permits and
// Task-linked dispatch intents never do (Story 4.2 owns the durable Task lifetime).
cleanup_expired_unlinked(transaction).await?;
if ticket.is_none() {
let queued = bounded_queue_len(transaction).await?;
if queued >= u64::try_from(self.limits.queue_limit).unwrap_or(u64::MAX) {
return Err(AttemptError::Overloaded);
}
let state = match context.request.mode {
AdmissionMode::Synchronous => "waiting",
AdmissionMode::Background => "reserved",
};
let ticket_id = insert_admission_entry(
transaction,
context.request,
context.owner_token,
state,
None,
)
.await?;
*ticket = Some(ticket_id);
if context.request.mode == AdmissionMode::Background {
return Ok(Attempt::Enqueued(ticket_id));
}
}
self.grant_next(transaction, workspace_id, &cursor).await?;
let Some(ticket_id) = *ticket else {
return Ok(Attempt::Waiting);
};
let state: Option<String> =
sqlx::query("select state from execution_admission_entries where ticket_id = $1")
.bind(i64::try_from(ticket_id).unwrap_or(i64::MAX))
.fetch_optional(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?
.map(|row| row.try_get("state"))
.transpose()
.map_err(AttemptError::from_sql)?;
match state.as_deref() {
Some("executing") => Ok(Attempt::Granted),
Some(_) => Ok(Attempt::Waiting),
// The ticket expired inside the store (or was removed by another cleanup).
None => Ok(Attempt::Expired),
}
}
/// Grants at most one waiting ticket per transaction, following the persisted weighted
/// round-robin cursor. A blocked head keeps its turn: nothing later is granted in its place.
async fn grant_next(
&self,
transaction: &mut Transaction<'_, Postgres>,
workspace_id: &str,
cursor: &AdmissionCursor,
) -> Result<(), AttemptError> {
let rows = sqlx::query(
"select ticket_id, agent_key, operation_id from execution_admission_entries
where workspace_id = $1 and state = 'waiting'
order by ticket_id
limit $2",
)
.bind(workspace_id)
.bind(QUEUE_READ_LIMIT)
.fetch_all(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
if rows.is_empty() {
return Ok(());
}
let queue = rows
.iter()
.map(|row| {
Ok(AdmissionQueueEntry {
ticket_id: u64::try_from(
row.try_get::<i64, _>("ticket_id")
.map_err(AttemptError::from_sql)?,
)
.unwrap_or_default(),
sequence: 0,
agent_key: row.try_get("agent_key").map_err(AttemptError::from_sql)?,
operation_id: row
.try_get("operation_id")
.map_err(AttemptError::from_sql)?,
})
})
.collect::<Result<Vec<_>, AttemptError>>()?;
let Some(selection) = select_next_grant(&queue, cursor, self.scheduler) else {
return Ok(());
};
let head = &queue[selection.index];
let global_executing: i64 = sqlx::query(
"select count(*)::bigint as count from execution_admission_entries
where state = 'executing'",
)
.fetch_one(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?
.try_get("count")
.map_err(AttemptError::from_sql)?;
let agent_counts = scoped_counts(transaction, workspace_id, "agent_key").await?;
let tool_counts = scoped_counts(transaction, workspace_id, "operation_id").await?;
if u64::try_from(global_executing).unwrap_or(u64::MAX)
>= u64::try_from(self.limits.global_limit).unwrap_or(u64::MAX)
{
return Ok(());
}
let agent_executing = agent_counts
.get(head.agent_key.as_str())
.copied()
.unwrap_or_default();
if agent_executing >= u64::try_from(self.limits.agent_limit).unwrap_or(u64::MAX) {
return Ok(());
}
let tool_executing = tool_counts
.get(head.operation_id.as_str())
.copied()
.unwrap_or_default();
if tool_executing >= u64::try_from(self.limits.tool_limit).unwrap_or(u64::MAX) {
return Ok(());
}
sqlx::query(
"update execution_admission_entries
set state = 'executing', granted_at = now()
where ticket_id = $1 and state = 'waiting'",
)
.bind(i64::try_from(head.ticket_id).unwrap_or(i64::MAX))
.execute(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
sqlx::query(
"update execution_admission_controllers
set cursor_agent = $2,
cursor_agent_credit = $3,
cursor_tool = $4,
cursor_tool_credit = $5,
updated_at = now()
where workspace_id = $1",
)
.bind(workspace_id)
.bind(selection.cursor.agent.as_deref())
.bind(i32::try_from(selection.cursor.agent_credit).unwrap_or_default())
.bind(selection.cursor.tool.as_deref())
.bind(i32::try_from(selection.cursor.tool_credit).unwrap_or_default())
.execute(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
Ok(())
}
}
/// Takes both controller locks in the fixed `'*'` -> workspace order. Shared by the synchronous
/// admission attempt and the durable Task repository, so both serialize on the same controller rows
/// and one task/queue invariant holds. Callers own the transaction; nothing here commits or sleeps.
pub(super) async fn lock_admission_in_tx(
transaction: &mut Transaction<'_, Postgres>,
workspace_id: &str,
) -> Result<AdmissionCursor, AttemptError> {
sqlx::query(
"insert into execution_admission_controllers (workspace_id)
values ($1), ($2)
on conflict (workspace_id) do nothing",
)
.bind(DEPLOYMENT_CONTROLLER)
.bind(workspace_id)
.execute(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
lock_controller(transaction, DEPLOYMENT_CONTROLLER).await?;
lock_controller(transaction, workspace_id).await
}
/// Bounded cleanup of expired waiting/reserved rows that are not a durable Task dispatch intent.
pub(super) async fn cleanup_expired_unlinked(
transaction: &mut Transaction<'_, Postgres>,
) -> Result<(), AttemptError> {
sqlx::query(
"delete from execution_admission_entries
where ticket_id in (
select ticket_id from execution_admission_entries
where state in ('waiting', 'reserved') and expires_at <= now()
and task_id is null
order by ticket_id
limit $1
)",
)
.bind(EXPIRED_CLEANUP_BATCH)
.execute(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
Ok(())
}
/// Authoritative bounded queue length: waiting and reserved rows share one budget.
pub(super) async fn bounded_queue_len(
transaction: &mut Transaction<'_, Postgres>,
) -> Result<u64, AttemptError> {
let queued: i64 = sqlx::query(
"select count(*)::bigint as count from execution_admission_entries
where state in ('waiting', 'reserved')",
)
.fetch_one(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?
.try_get("count")
.map_err(AttemptError::from_sql)?;
Ok(u64::try_from(queued).unwrap_or(u64::MAX))
}
/// Inserts one bounded admission row. A Task-linked row is the durable dispatch intent of the Task
/// accepted in the same transaction, so its capacity slot can never be silently reclaimed.
pub(super) async fn insert_admission_entry(
transaction: &mut Transaction<'_, Postgres>,
request: &AdmissionRequest,
owner_token: &str,
state: &str,
task_id: Option<Uuid>,
) -> Result<u64, AttemptError> {
let row = sqlx::query(
"insert into execution_admission_entries (
workspace_id, agent_key, operation_id, operation_version, mode, policy_mode,
state, owner_token, fingerprint, expires_at, task_id
) values ($1, $2, $3, $4, $5, $6, $7, $8, $9,
to_timestamp($10::double precision / 1000.0), $11)
returning ticket_id",
)
.bind(&request.scope.workspace_id)
.bind(&request.scope.agent_key)
.bind(&request.scope.operation_id)
.bind(i32::try_from(request.scope.operation_version).unwrap_or(i32::MAX))
.bind(request.mode.as_str())
.bind(policy_mode_str(request))
.bind(state)
.bind(owner_token)
.bind(&request.fingerprint)
.bind(request.expires_at_unix_ms)
.bind(task_id)
.fetch_one(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
let ticket_id: i64 = row.try_get("ticket_id").map_err(AttemptError::from_sql)?;
Ok(u64::try_from(ticket_id).unwrap_or_default())
}
/// Shared grant step used by the durable Task store: grants one reserved Task intent if capacity allows,
/// marks it `executing`, persists the weighted round-robin cursor, and returns its Task id. It reuses
/// the same selector/capacity logic as synchronous admission so no second policy exists.
pub(super) async fn grant_next_task(
transaction: &mut Transaction<'_, Postgres>,
workspace_id: &str,
cursor: &AdmissionCursor,
limits: &AdmissionLimits,
scheduler: AdmissionSchedulerConfig,
) -> Result<(Option<Uuid>, AdmissionCursor), AttemptError> {
let rows = sqlx::query(
"select ticket_id, agent_key, operation_id, task_id
from execution_admission_entries
where state = 'reserved' and task_id is not null
order by ticket_id
limit $1
for update skip locked",
)
.bind(QUEUE_READ_LIMIT)
.fetch_all(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
if rows.is_empty() {
return Ok((None, cursor.clone()));
}
struct Candidate {
entry: AdmissionQueueEntry,
task_id: Uuid,
}
let mut candidates = Vec::with_capacity(rows.len());
for row in rows {
let ticket_id: i64 = row.try_get("ticket_id").map_err(AttemptError::from_sql)?;
let task_id: Uuid = row.try_get("task_id").map_err(AttemptError::from_sql)?;
candidates.push(Candidate {
entry: AdmissionQueueEntry {
ticket_id: u64::try_from(ticket_id).unwrap_or_default(),
sequence: 0,
agent_key: row.try_get("agent_key").map_err(AttemptError::from_sql)?,
operation_id: row
.try_get("operation_id")
.map_err(AttemptError::from_sql)?,
},
task_id,
});
}
let queue: Vec<AdmissionQueueEntry> = candidates.iter().map(|c| c.entry.clone()).collect();
let Some(selection) = select_next_grant(&queue, cursor, scheduler) else {
return Ok((None, cursor.clone()));
};
let head_index = selection.index;
let head = &candidates[head_index];
let global_executing: i64 = sqlx::query(
"select count(*)::bigint as count from execution_admission_entries where state = 'executing'",
)
.fetch_one(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?
.try_get("count")
.map_err(AttemptError::from_sql)?;
let agent_counts = sqlx::query(
"select agent_key as key, count(*)::bigint as count
from execution_admission_entries where state = 'executing' group by agent_key",
)
.fetch_all(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
let mut agent_map = std::collections::BTreeMap::new();
for row in agent_counts {
let key: String = row.try_get("key").map_err(AttemptError::from_sql)?;
let count: i64 = row.try_get("count").map_err(AttemptError::from_sql)?;
agent_map.insert(key, u64::try_from(count).unwrap_or_default());
}
let tool_counts = sqlx::query(
"select operation_id as key, count(*)::bigint as count
from execution_admission_entries where state = 'executing' group by operation_id",
)
.fetch_all(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
let mut tool_map = std::collections::BTreeMap::new();
for row in tool_counts {
let key: String = row.try_get("key").map_err(AttemptError::from_sql)?;
let count: i64 = row.try_get("count").map_err(AttemptError::from_sql)?;
tool_map.insert(key, u64::try_from(count).unwrap_or_default());
}
if u64::try_from(global_executing).unwrap_or(u64::MAX)
>= u64::try_from(limits.global_limit).unwrap_or(u64::MAX)
{
return Ok((None, cursor.clone()));
}
let agent_executing = agent_map
.get(head.entry.agent_key.as_str())
.copied()
.unwrap_or_default();
if agent_executing >= u64::try_from(limits.agent_limit).unwrap_or(u64::MAX) {
return Ok((None, cursor.clone()));
}
let tool_executing = tool_map
.get(head.entry.operation_id.as_str())
.copied()
.unwrap_or_default();
if tool_executing >= u64::try_from(limits.tool_limit).unwrap_or(u64::MAX) {
return Ok((None, cursor.clone()));
}
sqlx::query(
"update execution_admission_entries
set state = 'executing', granted_at = clock_timestamp()
where ticket_id = $1 and state = 'reserved'",
)
.bind(i64::try_from(head.entry.ticket_id).unwrap_or(i64::MAX))
.execute(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
sqlx::query(
"update execution_admission_controllers
set cursor_agent = $2,
cursor_agent_credit = $3,
cursor_tool = $4,
cursor_tool_credit = $5,
updated_at = now()
where workspace_id = $1",
)
.bind(workspace_id)
.bind(selection.cursor.agent.as_deref())
.bind(i32::try_from(selection.cursor.agent_credit).unwrap_or_default())
.bind(selection.cursor.tool.as_deref())
.bind(i32::try_from(selection.cursor.tool_credit).unwrap_or_default())
.execute(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
Ok((Some(head.task_id), selection.cursor))
}
#[derive(Debug)]
pub(super) enum AttemptError {
/// Lock or statement timeout: the transaction made no decision and the waiter retries.
Contended,
/// Bounded queue capacity is exhausted: a typed overload outcome, never an unbounded wait.
Overloaded,
/// A foreign unique constraint aborted the transaction; the caller rolls back and rereads.
UniqueViolation,
/// Durable-Task side failure (the Task repository maps it to its own typed error).
Failure(TaskError),
/// Admission side failure.
AdmissionFailure(AdmissionStoreError),
}
impl AttemptError {
pub(super) fn from_sql(error: sqlx::Error) -> Self {
if let Some(database) = error.as_database_error()
&& matches!(database.code().as_deref(), Some("55P03" | "57014"))
{
return Self::Contended;
}
Self::AdmissionFailure(PostgresExecutionAdmission::unavailable(error))
}
}
async fn lock_controller(
transaction: &mut Transaction<'_, Postgres>,
workspace_id: &str,
) -> Result<AdmissionCursor, AttemptError> {
sqlx::query(
"update execution_admission_controllers set updated_at = updated_at
where workspace_id = $1",
)
.bind(workspace_id)
.execute(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
let row = sqlx::query(
"select cursor_agent, cursor_agent_credit, cursor_tool, cursor_tool_credit
from execution_admission_controllers where workspace_id = $1",
)
.bind(workspace_id)
.fetch_one(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
Ok(AdmissionCursor {
agent: row
.try_get("cursor_agent")
.map_err(AttemptError::from_sql)?,
agent_credit: u32::try_from(
row.try_get::<i32, _>("cursor_agent_credit")
.map_err(AttemptError::from_sql)?
.max(0),
)
.unwrap_or_default(),
tool: row.try_get("cursor_tool").map_err(AttemptError::from_sql)?,
tool_credit: u32::try_from(
row.try_get::<i32, _>("cursor_tool_credit")
.map_err(AttemptError::from_sql)?
.max(0),
)
.unwrap_or_default(),
})
}
async fn scoped_counts(
transaction: &mut Transaction<'_, Postgres>,
workspace_id: &str,
column: &str,
) -> Result<BTreeMap<String, u64>, AttemptError> {
let statement = if column == "agent_key" {
"select agent_key as key, count(*)::bigint as count
from execution_admission_entries
where workspace_id = $1 and state = 'executing'
group by agent_key"
} else {
"select operation_id as key, count(*)::bigint as count
from execution_admission_entries
where workspace_id = $1 and state = 'executing'
group by operation_id"
};
let rows = sqlx::query(statement)
.bind(workspace_id)
.fetch_all(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
let mut counts = BTreeMap::new();
for row in rows {
let key: String = row.try_get("key").map_err(AttemptError::from_sql)?;
let count: i64 = row.try_get("count").map_err(AttemptError::from_sql)?;
counts.insert(key, u64::try_from(count).unwrap_or_default());
}
Ok(counts)
}
fn policy_mode_str(request: &AdmissionRequest) -> &'static str {
match request.policy_mode {
crank_core::ToolExecutionMode::Synchronous => "synchronous",
crank_core::ToolExecutionMode::Background => "background",
crank_core::ToolExecutionMode::Adaptive => "adaptive",
}
}
@@ -170,6 +170,69 @@ impl PostgresRegistry {
.transpose()
}
/// Authoritative revalidation of the Platform API key that authorized a durable Task (Story 4.3).
///
/// The key is returned only when the exact immutable ID is still active, unexpired and bound to
/// the same workspace and agent; otherwise `None`, so a worker fails closed. Only the immutable
/// identity is supplied here — the bearer secret is never stored or re-read.
pub async fn revalidate_platform_api_key(
&self,
workspace_id: &WorkspaceId,
agent_id: &AgentId,
key_id: &PlatformApiKeyId,
) -> Result<Option<PlatformApiKeyRecord>, RegistryError> {
let row = sqlx::query(
"select
k.id,
k.workspace_id,
k.agent_id,
k.key_kind,
k.name,
k.prefix,
k.scopes_json,
k.status,
k.created_at,
k.last_used_at,
k.expires_at,
k.allowed_origins_json
from platform_api_keys k
where k.id = $1
and k.workspace_id = $2
and k.agent_id = $3
and k.key_kind = 'mcp_client'
and k.status = 'active'
and (k.expires_at is null or k.expires_at > now())
limit 1",
)
.bind(key_id.as_str())
.bind(workspace_id.as_str())
.bind(agent_id.as_str())
.fetch_optional(&self.pool)
.await?;
row.map(|row| {
Ok(PlatformApiKeyRecord {
api_key: PlatformApiKey {
id: PlatformApiKeyId::new(row.get::<String, _>("id")),
workspace_id: WorkspaceId::new(row.get::<String, _>("workspace_id")),
agent_id: row.get::<Option<String>, _>("agent_id").map(AgentId::new),
key_kind: deserialize_enum_text(&row.get::<String, _>("key_kind"), "key_kind")?,
name: row.get("name"),
prefix: row.get("prefix"),
scopes: deserialize_json_value(row.get::<Value, _>("scopes_json"))?,
status: deserialize_enum_text(&row.get::<String, _>("status"), "status")?,
created_at: row.get("created_at"),
last_used_at: row.get("last_used_at"),
expires_at: row.get("expires_at"),
allowed_origins: deserialize_json_value(
row.get::<Value, _>("allowed_origins_json"),
)?,
},
})
})
.transpose()
}
pub async fn get_approval_api_key_by_secret_for_agent_slug(
&self,
workspace_slug: &str,
@@ -46,6 +46,24 @@ impl PostgresRegistry {
Ok(())
}
/// Readiness probe of the mandatory execution-admission state (Story 4.1). A missing or
/// unreadable admission schema is a dependency failure, never a degraded-but-ready state.
pub async fn execution_admission_health(&self) -> Result<(), crank_core::AdmissionStoreError> {
super::admission::admission_health(&self.pool).await
}
/// Readiness probe of the durable Task state (Story 4.2).
pub async fn durable_tasks_health(&self) -> Result<(), crank_core::TaskError> {
super::tasks::tasks_health(&self.pool).await
}
/// Authoritative admission queue readings for low-cardinality metrics.
pub async fn execution_admission_snapshot(
&self,
) -> Result<crank_core::AdmissionQueueSnapshot, crank_core::AdmissionStoreError> {
super::admission::admission_snapshot(&self.pool).await
}
async fn connect_in_schema(
database_url: &str,
schema: Option<&str>,
@@ -1,6 +1,6 @@
use super::*;
const MASTER_KEY_PAGE_LIMIT_MAX: i64 = 10_000;
pub(super) const MASTER_KEY_PAGE_LIMIT_MAX: i64 = 10_000;
impl PostgresRegistry {
pub async fn verify_or_register_master_key_identity(
@@ -356,12 +356,19 @@ impl PostgresRegistry {
return Err(RegistryError::MasterKeyRotationConflict);
}
let total = sqlx::query_scalar::<_, i64>(
let secret_total = sqlx::query_scalar::<_, i64>(
"select count(*) from secret_versions where master_key_epoch = $1",
)
.bind(source_epoch)
.fetch_one(&mut *tx)
.await?;
let task_total = sqlx::query_scalar::<_, i64>(
"select count(*) from tasks where executable_input_master_key_epoch = $1",
)
.bind(source_epoch)
.fetch_one(&mut *tx)
.await?;
let total = secret_total + task_total;
let rotation_id = format!("master-key-e{source_epoch}-to-e{target_epoch}");
let aborted = sqlx::query(
"select
@@ -403,6 +410,16 @@ impl PostgresRegistry {
.bind(target_epoch)
.execute(&mut *tx)
.await?;
sqlx::query(
"update tasks
set target_executable_input_ciphertext = null,
target_executable_input_key_version = null,
target_executable_input_master_key_epoch = null
where target_executable_input_master_key_epoch = $1",
)
.bind(target_epoch)
.execute(&mut *tx)
.await?;
sqlx::query(
"update master_key_rotations
set state = 'running',
@@ -662,9 +679,24 @@ impl PostgresRegistry {
.bind(rotation.target_epoch)
.execute(&mut *tx)
.await?;
sqlx::query(
"update tasks
set executable_input_ciphertext = target_executable_input_ciphertext,
executable_input_key_version = target_executable_input_key_version,
executable_input_master_key_epoch = target_executable_input_master_key_epoch,
target_executable_input_ciphertext = null,
target_executable_input_key_version = null,
target_executable_input_master_key_epoch = null
where executable_input_master_key_epoch = $1
and target_executable_input_master_key_epoch = $2",
)
.bind(rotation.source_epoch)
.bind(rotation.target_epoch)
.execute(&mut *tx)
.await?;
sqlx::query(
"update master_key_rotations
set state = 'promoted',
set state = 'promoted',
updated_at = $2::timestamptz
where id = $1",
)
@@ -697,6 +729,16 @@ impl PostgresRegistry {
.bind(rotation.target_epoch)
.execute(&mut *tx)
.await?;
sqlx::query(
"update tasks
set target_executable_input_ciphertext = null,
target_executable_input_key_version = null,
target_executable_input_master_key_epoch = null
where target_executable_input_master_key_epoch = $1",
)
.bind(rotation.target_epoch)
.execute(&mut *tx)
.await?;
sqlx::query(
"update master_key_rotations
set state = 'aborted',
@@ -883,7 +925,7 @@ async fn select_active_master_key_identity_for_update(
Ok(row.map(map_master_key_identity))
}
async fn select_master_key_rotation_for_update(
pub(super) async fn select_master_key_rotation_for_update(
transaction: &mut sqlx::Transaction<'_, sqlx::Postgres>,
rotation_id: &str,
) -> Result<MasterKeyRotationRecord, RegistryError> {
@@ -915,16 +957,19 @@ async fn select_master_key_rotation_for_update(
})
}
async fn count_staged_rotation_targets(
pub(super) async fn count_staged_rotation_targets(
transaction: &mut sqlx::Transaction<'_, sqlx::Postgres>,
source_epoch: i64,
target_epoch: i64,
) -> Result<i64, RegistryError> {
let count = sqlx::query_scalar::<_, i64>(
"select count(*)
from secret_versions
where master_key_epoch = $1
and target_master_key_epoch = $2",
"select
(select count(*) from secret_versions
where master_key_epoch = $1
and target_master_key_epoch = $2) +
(select count(*) from tasks
where executable_input_master_key_epoch = $1
and target_executable_input_master_key_epoch = $2)",
)
.bind(source_epoch)
.bind(target_epoch)
@@ -0,0 +1,144 @@
//! Task ciphertext paging/staging of the master-key rotation (Story 4.3).
//!
//! Extracted from `master_key.rs` without logic changes so that module stays within the
//! code-health line limit; the operations and their SQL are unchanged.
use super::master_key::{
MASTER_KEY_PAGE_LIMIT_MAX, count_staged_rotation_targets, select_master_key_rotation_for_update,
};
use super::*;
fn map_master_key_task_version(
row: sqlx::postgres::PgRow,
) -> Result<MasterKeyTaskVersion, RegistryError> {
Ok(MasterKeyTaskVersion {
task_id: row.get::<String, _>("task_id"),
ciphertext: row
.get::<Option<String>, _>("executable_input_ciphertext")
.unwrap_or_default(),
key_version: row
.get::<Option<String>, _>("executable_input_key_version")
.unwrap_or_default(),
epoch: row
.get::<Option<i64>, _>("executable_input_master_key_epoch")
.unwrap_or_default(),
target_ciphertext: row.get::<Option<String>, _>("target_executable_input_ciphertext"),
target_key_version: row.get::<Option<String>, _>("target_executable_input_key_version"),
target_epoch: row.get::<Option<i64>, _>("target_executable_input_master_key_epoch"),
})
}
impl PostgresRegistry {
pub async fn list_tasks_for_master_key_epoch_page(
&self,
epoch: i64,
after_task_id: Option<&str>,
limit: i64,
) -> Result<Vec<MasterKeyTaskVersion>, RegistryError> {
if epoch < 1 || !(1..=MASTER_KEY_PAGE_LIMIT_MAX).contains(&limit) {
return Err(RegistryError::InvalidMasterKeyIdentity);
}
let rows = sqlx::query(
"select task_id::text as task_id, executable_input_ciphertext, executable_input_key_version, executable_input_master_key_epoch, target_executable_input_ciphertext, target_executable_input_key_version, target_executable_input_master_key_epoch from tasks where executable_input_master_key_epoch = $1 and ($2::text is null or task_id::text > $2::text) order by task_id asc limit $3",
)
.bind(epoch)
.bind(after_task_id)
.bind(limit)
.fetch_all(&self.pool)
.await?;
rows.into_iter().map(map_master_key_task_version).collect()
}
pub async fn list_target_tasks_for_master_key_rotation_page(
&self,
target_epoch: i64,
after_task_id: Option<&str>,
limit: i64,
) -> Result<Vec<MasterKeyTaskVersion>, RegistryError> {
if target_epoch < 1 || !(1..=MASTER_KEY_PAGE_LIMIT_MAX).contains(&limit) {
return Err(RegistryError::InvalidMasterKeyIdentity);
}
let rows = sqlx::query(
"select task_id::text as task_id, executable_input_ciphertext, executable_input_key_version, executable_input_master_key_epoch, target_executable_input_ciphertext, target_executable_input_key_version, target_executable_input_master_key_epoch from tasks where target_executable_input_master_key_epoch = $1 and ($2::text is null or task_id::text > $2::text) order by task_id asc limit $3",
)
.bind(target_epoch)
.bind(after_task_id)
.bind(limit)
.fetch_all(&self.pool)
.await?;
rows.into_iter().map(map_master_key_task_version).collect()
}
#[allow(clippy::too_many_arguments)]
pub async fn stage_master_key_rotation_task_ciphertext(
&self,
rotation_id: &str,
task_id: &str,
source_epoch: i64,
target_ciphertext: &str,
target_key_version: &str,
target_epoch: i64,
now: &OffsetDateTime,
) -> Result<MasterKeyRotationRecord, RegistryError> {
let mut tx = self.pool.begin().await?;
let rotation = select_master_key_rotation_for_update(&mut tx, rotation_id).await?;
if rotation.state != "running"
|| rotation.source_epoch != source_epoch
|| rotation.target_epoch != target_epoch
{
return Err(RegistryError::MasterKeyRotationConflict);
}
let result = sqlx::query(
"update tasks
set target_executable_input_ciphertext = $3,
target_executable_input_key_version = $4,
target_executable_input_master_key_epoch = $5
where task_id::text = $1
and executable_input_master_key_epoch = $2
and (
target_executable_input_master_key_epoch is null
or target_executable_input_master_key_epoch = $5
)
and (
target_executable_input_ciphertext is null
or target_executable_input_ciphertext = $3
)",
)
.bind(task_id)
.bind(source_epoch)
.bind(target_ciphertext)
.bind(target_key_version)
.bind(target_epoch)
.execute(&mut *tx)
.await?;
if result.rows_affected() != 1 {
return Err(RegistryError::MasterKeyRotationConflict);
}
let processed = count_staged_rotation_targets(&mut tx, source_epoch, target_epoch).await?;
sqlx::query(
"update master_key_rotations
set checkpoint_secret_id = checkpoint_secret_id,
processed_secret_versions = $2,
updated_at = $3::timestamptz
where id = $1",
)
.bind(rotation_id)
.bind(processed)
.bind(now)
.execute(&mut *tx)
.await?;
let updated = select_master_key_rotation_for_update(&mut tx, rotation_id).await?;
tx.commit().await?;
Ok(updated)
}
}
pub struct MasterKeyTaskVersion {
pub task_id: String,
pub ciphertext: String,
pub key_version: String,
pub epoch: i64,
pub target_ciphertext: Option<String>,
pub target_key_version: Option<String>,
pub target_epoch: Option<i64>,
}
@@ -1,3 +1,4 @@
mod admission;
mod agent;
mod agent_catalog;
mod agent_prompt_bindings;
@@ -9,6 +10,7 @@ mod auth;
mod connection;
mod import_job;
mod master_key;
mod master_key_tasks;
mod observability;
mod onboarding;
mod onboarding_import;
@@ -20,6 +22,7 @@ mod product_event;
mod prompt;
mod resource;
mod secret;
mod tasks;
mod upstream;
mod workspace;
mod yaml_import;
@@ -36,8 +39,13 @@ use serde_json::Value;
use sqlx::{PgPool, Postgres, Row, Transaction, postgres::PgRow, types::Json};
use time::OffsetDateTime;
pub use admission::{
AdmissionLimits, PostgresExecutionAdmission, admission_health, admission_snapshot,
};
pub use pool_config::{PostgresPoolConfig, PostgresPoolConfigError};
pub use resource::{AuthorizedResourceRead, BoundResourceAdmission};
pub use tasks::PostgresTaskStore;
pub use tasks::tasks_health;
use crate::{
error::RegistryError,
@@ -0,0 +1,195 @@
//! Acceptance half of the PostgreSQL Task store (Story 4.2/4.3): the one-transaction acceptance
//! path, its bounded unique-conflict reread and the typed miss classification.
use crank_core::{
AdmissionMode, AdmissionRequest, AdmissionScope, ApprovalEvidence, NoTasksFallback,
PlatformApiKeyId, TaskAcceptRequest, TaskAcceptance, TaskError, TaskFinishRequest, TaskState,
};
use sqlx::{Postgres, Row, Transaction};
use super::super::admission::{
AttemptError, bounded_queue_len, cleanup_expired_unlinked, insert_admission_entry,
lock_admission_in_tx,
};
use super::{
Accepted, PostgresTaskStore, TASK_PROJECTION, acceptance_from_record, policy_mode_str,
record_from_row, task_sql_error,
};
impl PostgresTaskStore {
pub(super) async fn accept_locked(
&self,
transaction: &mut Transaction<'_, Postgres>,
request: &TaskAcceptRequest,
) -> Result<Accepted, AttemptError> {
// Replay first: an already accepted Task is returned without capacity mutation, so it works
// with a full queue and after the original execution deadline while the record is retained.
if let Some((acceptance, fingerprint)) =
self.stored_acceptance(transaction, request).await?
{
return Ok(if fingerprint == request.input_fingerprint {
Accepted::Replayed(acceptance)
} else {
Accepted::Conflict
});
}
// Serialize with synchronous admission and any concurrent acceptance of the same scope.
lock_admission_in_tx(transaction, &request.workspace_id).await?;
// The competitor may have committed while this transaction waited for the controller lock.
if let Some((acceptance, fingerprint)) =
self.stored_acceptance(transaction, request).await?
{
return Ok(if fingerprint == request.input_fingerprint {
Accepted::Replayed(acceptance)
} else {
Accepted::Conflict
});
}
cleanup_expired_unlinked(transaction).await?;
let queued = bounded_queue_len(transaction).await?;
if queued >= u64::try_from(self.limits.queue_limit).unwrap_or(u64::MAX) {
return Err(AttemptError::Overloaded);
}
let created_ms: i64 = sqlx::query(
"insert into tasks (
task_id, workspace_id, owner_kind, owner_id, authorization_scope, operation_id,
operation_version, scoped_key, input_fingerprint, request_id, trace_id, traceparent,
trace_sampled, policy_mode, approval_reference, state, input_envelope,
execution_deadline, retain_until, executable_input_ciphertext, executable_input_key_version,
executable_input_master_key_epoch, platform_api_key_id, max_attempts, retry_backoff_step_ms
) values (
$1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14, $15, 'queued', $16,
to_timestamp($17::double precision / 1000.0),
to_timestamp($18::double precision / 1000.0), $19, $20, $21, $22, $23, $24
)
returning (extract(epoch from created_at) * 1000)::bigint as created_ms",
)
.bind(request.task_id.as_uuid())
.bind(&request.workspace_id)
.bind(request.owner_kind.as_str())
.bind(&request.owner_id)
.bind(&request.authorization_scope)
.bind(&request.operation_id)
.bind(i32::try_from(request.pins.operation_version).unwrap_or(i32::MAX))
.bind(&request.scoped_key)
.bind(&request.input_fingerprint)
.bind(request.support.request_id.as_str())
.bind(request.support.trace_id.as_str())
.bind(&request.support.traceparent)
.bind(request.support.trace_sampled)
.bind(policy_mode_str(request.pins.policy_mode))
.bind(request.pins.approval_reference.as_deref())
.bind(request.input.value().clone())
.bind(request.pins.execution_deadline_unix_ms)
.bind(request.pins.retain_until_unix_ms)
.bind(request.executable_input_ciphertext.as_deref())
.bind(request.executable_input_key_version.as_deref())
.bind(request.executable_input_master_key_epoch)
.bind(
request
.platform_api_key_id
.as_ref()
.map(PlatformApiKeyId::as_str),
)
.bind(i32::try_from(request.pins.retry.max_attempts).unwrap_or(1))
.bind(request.pins.retry.backoff_step_ms)
.fetch_one(&mut **transaction)
.await
.map_err(task_sql_error)?
.try_get("created_ms")
.map_err(AttemptError::from_sql)?;
let admission_request = AdmissionRequest {
scope: AdmissionScope {
workspace_id: request.workspace_id.clone(),
agent_key: request.owner_id.clone(),
operation_id: request.operation_id.clone(),
operation_version: request.pins.operation_version,
},
mode: AdmissionMode::Background,
policy_mode: request.pins.policy_mode,
no_tasks_fallback: NoTasksFallback::Unsupported,
// Never reached: the intent row is inserted by this transaction, not by `admit`.
approval: ApprovalEvidence::NotRequired,
fingerprint: request.input_fingerprint.clone(),
expires_at_unix_ms: request.pins.execution_deadline_unix_ms,
};
let owner_token = format!("task_{}", request.task_id);
insert_admission_entry(
transaction,
&admission_request,
&owner_token,
"reserved",
Some(request.task_id.as_uuid()),
)
.await?;
Ok(Accepted::Created(TaskAcceptance {
task_id: request.task_id,
state: TaskState::Queued,
operation_version: request.pins.operation_version,
support: request.support.clone(),
created_at_unix_ms: created_ms,
retain_until_unix_ms: request.pins.retain_until_unix_ms,
replayed: false,
}))
}
/// Bounded scoped reread after a foreign unique conflict aborted the transaction.
pub(super) async fn reread_after_unique(
&self,
request: &TaskAcceptRequest,
) -> Result<TaskAcceptance, TaskError> {
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"select {TASK_PROJECTION} from tasks
where workspace_id = $1 and owner_kind = $2 and owner_id = $3
and authorization_scope = $4 and operation_id = $5 and scoped_key = $6"
)))
.bind(&request.workspace_id)
.bind(request.owner_kind.as_str())
.bind(&request.owner_id)
.bind(&request.authorization_scope)
.bind(&request.operation_id)
.bind(&request.scoped_key)
.fetch_optional(&self.pool)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
return Err(TaskError::unavailable(
"durable task acceptance ended with an unresolved unique conflict",
));
};
let record = record_from_row(&row)?;
if record.input_fingerprint != request.input_fingerprint {
return Err(TaskError::FingerprintConflict);
}
Ok(acceptance_from_record(&record))
}
/// A rejected conditional update is either a lost CAS race or a foreign/absent Task. Both are
/// reported without payload, and a foreign Task is indistinguishable from an absent one.
pub(super) async fn classify_missed_finish(&self, request: &TaskFinishRequest) -> TaskError {
let access = &request.access;
match sqlx::query(
"select state, version from tasks
where task_id = $1 and workspace_id = $2 and owner_kind = $3
and owner_id = $4 and authorization_scope = $5",
)
.bind(request.task_id.as_uuid())
.bind(&access.workspace_id)
.bind(access.owner_kind.as_str())
.bind(&access.owner_id)
.bind(&access.authorization_scope)
.fetch_optional(&self.pool)
.await
{
Ok(Some(_)) => TaskError::StateConflict,
Ok(None) => TaskError::NotFound,
Err(error) => Self::unavailable(error),
}
}
}
@@ -0,0 +1,583 @@
//! Durable per-outbound attempt persistence of a Task (Story 4.4): the dispatch checkpoint, the
//! single-winner attempt resolution and the owner-scoped cancellation live here so the acceptance
//! and retention half of the store stays readable.
use crank_core::{
AttemptCertainty, AttemptLease, AttemptOutcome, AttemptOutcomeClass, AttemptPhase,
AttemptResolution, BeginAttempt, TaskAccess, TaskAttempt, TaskCancelOutcome, TaskCancelRequest,
TaskEnvelope, TaskError, TaskId, TaskRecord, TaskState,
};
use sqlx::Row;
use uuid::Uuid;
use super::{PostgresTaskStore, TASK_PROJECTION, record_from_row, unavailable_row};
const ATTEMPT_PROJECTION: &str = "attempt_id, task_id, ordinal, owner, fence, input_fingerprint,
upstream_key_digest, retry_allowed,
(extract(epoch from checkpoint_at) * 1000)::bigint as checkpoint_at_ms,
(extract(epoch from resolved_at) * 1000)::bigint as resolved_at_ms,
outcome, certainty, phase";
/// Bounded attempt-resolution helper: `retry` may only be armed when the attempt itself is marked
/// safe-to-retry, the pinned budget still has room and the accepted absolute deadline still does.
impl PostgresTaskStore {
fn retry_delay(
&self,
record: &TaskRecord,
lease: &AttemptLease,
safe: bool,
now: i64,
) -> Option<i64> {
let delay = record.pins.retry.backoff_before(lease.attempt.ordinal + 1);
(safe
&& lease.attempt.retry_allowed
&& lease.attempt.ordinal < record.pins.retry.max_attempts
&& now.saturating_add(delay) < record.pins.execution_deadline_unix_ms)
.then_some(delay)
}
}
impl PostgresTaskStore {
/// Authoritative word on whether an earlier resolved attempt really reached upstream despite
/// the current one not having dispatched. Returns Some(phase) only when the latest resolved
/// attempt that actually dispatched (phase = 'dispatched') is still of unknown certainty.
pub(super) async fn unproven_dispatch_phase(
&self,
task_id: TaskId,
transaction: &mut sqlx::Transaction<'_, sqlx::Postgres>,
) -> Result<Option<String>, TaskError> {
let row = sqlx::query(
"select certainty, phase from task_attempts where task_id = $1 and resolved_at is not null and phase = 'dispatched' order by ordinal desc limit 1",
)
.bind(task_id.as_uuid())
.fetch_optional(&mut **transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
return Ok(None);
};
let certainty: Option<String> = row.try_get("certainty").map_err(Self::unavailable)?;
if certainty.as_deref() != Some(AttemptCertainty::as_str(AttemptCertainty::Unknown)) {
return Ok(None);
}
let phase: Option<String> = row.try_get("phase").map_err(Self::unavailable)?;
Ok(Some(phase.unwrap_or_else(|| "dispatched".to_owned())))
}
}
fn attempt_from_row(row: &sqlx::postgres::PgRow) -> Result<TaskAttempt, TaskError> {
let attempt_id: String = row.try_get("attempt_id").map_err(unavailable_row)?;
let task_id: Uuid = row.try_get("task_id").map_err(unavailable_row)?;
let outcome: Option<String> = row.try_get("outcome").map_err(unavailable_row)?;
let certainty: Option<String> = row.try_get("certainty").map_err(unavailable_row)?;
let phase: Option<String> = row.try_get("phase").map_err(unavailable_row)?;
Ok(TaskAttempt {
attempt_id,
task_id: TaskId::parse(&task_id.to_string())?,
ordinal: u32::try_from(row.try_get::<i32, _>("ordinal").map_err(unavailable_row)?)
.unwrap_or_default(),
owner: row.try_get("owner").map_err(unavailable_row)?,
fence: u64::try_from(row.try_get::<i64, _>("fence").map_err(unavailable_row)?)
.unwrap_or_default(),
input_fingerprint: row.try_get("input_fingerprint").map_err(unavailable_row)?,
upstream_key_digest: row
.try_get("upstream_key_digest")
.map_err(unavailable_row)?,
retry_allowed: row.try_get("retry_allowed").map_err(unavailable_row)?,
checkpoint_at_unix_ms: row.try_get("checkpoint_at_ms").map_err(unavailable_row)?,
resolved_at_unix_ms: row.try_get("resolved_at_ms").map_err(unavailable_row)?,
outcome: outcome
.as_deref()
.map(AttemptOutcomeClass::parse)
.transpose()?,
certainty: certainty
.as_deref()
.map(AttemptCertainty::parse)
.transpose()?,
phase: phase.as_deref().map(AttemptPhase::parse).transpose()?,
})
}
impl PostgresTaskStore {
pub(super) async fn checkpoint_dispatch_attempt(
&self,
request: BeginAttempt,
) -> Result<AttemptLease, TaskError> {
request.validate()?;
let mut transaction = self.start().await?;
// One row lock serializes this checkpoint with claim, cancellation and the resolution of the
// same Task, so a cancel that wins first can never be followed by an outbound call.
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"select {TASK_PROJECTION} from tasks where task_id = $1 for update"
)))
.bind(request.task_id.as_uuid())
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::NotFound);
};
let record = record_from_row(&row)?;
if record.input_fingerprint != request.input_fingerprint {
transaction.rollback().await.ok();
return Err(TaskError::FingerprintConflict);
}
let now = crank_core::current_unix_ms();
let eligible = record.state == TaskState::Working
&& record.lease_owner.as_deref() == Some(request.owner.as_str())
&& record.fence == request.fence
&& record.version == request.expected_version
&& record.current_attempt_id.is_none()
&& record.cancel_requested_at_unix_ms.is_none()
&& record.lease_until_unix_ms.is_some_and(|until| until > now)
&& record.attempts < record.pins.retry.max_attempts
&& record.pins.execution_deadline_unix_ms > now
&& record.next_attempt_at_unix_ms.is_none_or(|at| at <= now);
if !eligible {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
}
let ordinal = record.attempts.saturating_add(1);
sqlx::query(
"insert into task_attempts (
attempt_id, task_id, ordinal, owner, fence, input_fingerprint,
upstream_key_digest, retry_allowed
) values ($1, $2, $3, $4, $5, $6, $7, $8)",
)
.bind(&request.attempt_id)
.bind(request.task_id.as_uuid())
.bind(i32::try_from(ordinal).unwrap_or(i32::MAX))
.bind(&request.owner)
.bind(i64::try_from(request.fence).unwrap_or_default())
.bind(&request.input_fingerprint)
.bind(request.upstream_key_digest.as_deref())
.bind(request.retry_allowed)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"update tasks
set current_attempt_id = $1,
dispatch_started_at = now(),
attempts = attempts + 1,
next_attempt_at = null,
version = version + 1,
updated_at = now()
where task_id = $2 and current_attempt_id is null and version = $3
returning {TASK_PROJECTION}"
)))
.bind(&request.attempt_id)
.bind(request.task_id.as_uuid())
.bind(i64::from(record.version))
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
};
let task = record_from_row(&row)?;
transaction.commit().await.map_err(Self::unavailable)?;
let attempt = TaskAttempt {
attempt_id: request.attempt_id,
task_id: request.task_id,
ordinal,
owner: request.owner,
fence: request.fence,
input_fingerprint: request.input_fingerprint,
upstream_key_digest: request.upstream_key_digest,
retry_allowed: request.retry_allowed,
checkpoint_at_unix_ms: task.dispatch_started_at_unix_ms.unwrap_or(now),
resolved_at_unix_ms: None,
outcome: None,
certainty: None,
phase: None,
};
Ok(AttemptLease { attempt, task })
}
pub(super) async fn resolve_attempt_outcome(
&self,
lease: &AttemptLease,
outcome: AttemptOutcome,
) -> Result<AttemptResolution, TaskError> {
let mut transaction = self.start().await?;
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"select {TASK_PROJECTION} from tasks where task_id = $1 for update"
)))
.bind(lease.attempt.task_id.as_uuid())
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::NotFound);
};
let record = record_from_row(&row)?;
let now = crank_core::current_unix_ms();
let prior_uncertain_phase = self
.unproven_dispatch_phase(lease.attempt.task_id, &mut transaction)
.await?;
// Only the current attempt of the current owner, under the version it was admitted with, may
// be resolved. A stale worker can never append a second outcome to the history or the Task.
let current = record.state == TaskState::Working
&& record.lease_owner.as_deref() == Some(lease.attempt.owner.as_str())
&& record.fence == lease.attempt.fence
&& record.version == lease.task.version
&& record.current_attempt_id.as_deref() == Some(lease.attempt.attempt_id.as_str())
&& record.lease_until_unix_ms.is_some_and(|until| until > now);
if !current {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
}
let (class, certainty, phase) = match &outcome {
AttemptOutcome::Completed(_) => (
AttemptOutcomeClass::Completed,
AttemptCertainty::Known,
AttemptPhase::Dispatched,
),
AttemptOutcome::Failed { phase, .. } => {
(AttemptOutcomeClass::Failed, AttemptCertainty::Known, *phase)
}
AttemptOutcome::Uncertain { phase, .. } => (
AttemptOutcomeClass::OutcomeUnknown,
AttemptCertainty::Unknown,
*phase,
),
};
// A failed or uncertain attempt that may be safely repeated keeps the lease and arms the
// pinned backoff. Exhausted budget or deadline never turns uncertainty into a plain failure.
let retry = match &outcome {
AttemptOutcome::Failed {
retryable,
phase: _,
error: _,
} => self.retry_delay(&record, lease, *retryable, now),
AttemptOutcome::Uncertain {
repeatable,
phase: _,
error: _,
} => self.retry_delay(&record, lease, *repeatable, now),
AttemptOutcome::Completed(_) => None,
};
let attempt_class = if retry.is_some() {
AttemptOutcomeClass::RetryScheduled
} else {
class
};
let closed = sqlx::query(
"update task_attempts
set resolved_at = now(), outcome = $1, certainty = $2, phase = $3
where attempt_id = $4 and resolved_at is null",
)
.bind(attempt_class.as_str())
.bind(certainty.as_str())
.bind(phase.as_str())
.bind(&lease.attempt.attempt_id)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
if closed.rows_affected() != 1 {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
}
let (state, result, error, keep_lease, next_attempt_at) = match outcome {
AttemptOutcome::Completed(result) => {
(TaskState::Completed, Some(result), None, false, None)
}
AttemptOutcome::Failed {
error,
retryable: _,
phase,
} => match retry {
Some(delay) => (
TaskState::Working,
None,
None,
true,
Some(now.saturating_add(delay)),
),
None => {
if phase == AttemptPhase::BeforeDispatch {
if let Some(prior) = prior_uncertain_phase.as_ref() {
let mut value = error.value().clone();
if let serde_json::Value::Object(map) = &mut value {
map.insert(
"phase".to_string(),
serde_json::Value::String(prior.clone()),
);
}
let replaced = TaskEnvelope::try_redacted(&value)?;
(TaskState::OutcomeUnknown, None, Some(replaced), false, None)
} else {
(TaskState::Failed, None, Some(error), false, None)
}
} else {
(TaskState::Failed, None, Some(error), false, None)
}
}
},
AttemptOutcome::Uncertain {
error,
repeatable: _,
phase: _,
} => match retry {
Some(delay) => (
TaskState::Working,
None,
None,
true,
Some(now.saturating_add(delay)),
),
None => (TaskState::OutcomeUnknown, None, Some(error), false, None),
},
};
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"update tasks
set state = $1,
result_envelope = $2,
error_envelope = $3,
lease_owner = case when $4 then lease_owner else null end,
lease_until = case when $4 then lease_until else null end,
-- A retry-scheduled attempt is already resolved, so it no longer holds the Task:
-- the next durable attempt needs its own checkpoint under the same lease.
current_attempt_id = case when $8 then null else current_attempt_id end,
next_attempt_at = case when $5::bigint is null then null
else to_timestamp($5::double precision / 1000.0) end,
version = version + 1,
updated_at = now()
where task_id = $6 and version = $7
returning {TASK_PROJECTION}"
)))
.bind(state.as_str())
.bind(result.as_ref().map(|value| value.value().clone()))
.bind(error.as_ref().map(|value| value.value().clone()))
.bind(keep_lease)
.bind(next_attempt_at)
.bind(lease.attempt.task_id.as_uuid())
.bind(i64::from(record.version))
.bind(retry.is_some())
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
};
let task = record_from_row(&row)?;
if !keep_lease {
sqlx::query("delete from execution_admission_entries where task_id = $1")
.bind(lease.attempt.task_id.as_uuid())
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
}
transaction.commit().await.map_err(Self::unavailable)?;
Ok(match retry {
Some(_) => AttemptResolution::RetryAt {
next_attempt_at_unix_ms: next_attempt_at.unwrap_or(now),
task,
},
None if state == TaskState::OutcomeUnknown => AttemptResolution::Unknown(task),
None => AttemptResolution::Finished(task),
})
}
pub(super) async fn cancel_task(
&self,
request: TaskCancelRequest,
) -> Result<TaskCancelOutcome, TaskError> {
request.validate()?;
let access = &request.access;
let mut transaction = self.start().await?;
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"select {TASK_PROJECTION} from tasks
where task_id = $1 and workspace_id = $2 and owner_kind = $3
and owner_id = $4 and authorization_scope = $5
for update"
)))
.bind(request.task_id.as_uuid())
.bind(&access.workspace_id)
.bind(access.owner_kind.as_str())
.bind(&access.owner_id)
.bind(&access.authorization_scope)
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::NotFound);
};
let record = record_from_row(&row)?;
if let Some(expected) = request.expected_version
&& record.version != expected
{
transaction.rollback().await.ok();
return Ok(TaskCancelOutcome::Conflict(record));
}
if record.state.is_terminal() {
transaction.rollback().await.ok();
return Ok(TaskCancelOutcome::AlreadyTerminal(record));
}
if record.state == TaskState::OutcomeUnknown {
transaction.rollback().await.ok();
return Ok(TaskCancelOutcome::OutcomeUnknown(record));
}
// A checkpointed attempt may already have reached the upstream: the honest answer is the
// non-terminal uncertain state, never a plain cancellation and never a retry.
let checkpointed = record.current_attempt_id.clone();
// Preserve an uncertainty recorded by a previously resolved attempt even after RetryAt
// cleared current_attempt_id: the request already reached the upstream in that case.
let prior_uncertain = self
.unproven_dispatch_phase(request.task_id, &mut transaction)
.await?
.is_some();
let (state, code) = match (&checkpointed, prior_uncertain) {
(Some(_), _) => (TaskState::OutcomeUnknown, "cancelled_after_dispatch"),
(None, true) => (TaskState::OutcomeUnknown, "cancelled_after_dispatch"),
(None, false) => (TaskState::Cancelled, "cancelled_before_dispatch"),
};
if let Some(attempt_id) = checkpointed {
let closed = sqlx::query(
"update task_attempts
set resolved_at = now(), outcome = 'outcome_unknown', certainty = 'unknown',
phase = 'dispatched'
where task_id = $1 and attempt_id = $2 and resolved_at is null",
)
.bind(request.task_id.as_uuid())
.bind(&attempt_id)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let _ = closed;
}
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"update tasks
set state = $1,
error_envelope = jsonb_build_object('code', $2),
cancel_requested_at = now(),
cancel_reason = $3,
lease_owner = null,
lease_until = null,
version = version + 1,
updated_at = now()
where task_id = $4 and version = $5
returning {TASK_PROJECTION}"
)))
.bind(state.as_str())
.bind(code)
.bind(&request.reason)
.bind(request.task_id.as_uuid())
.bind(i64::from(record.version))
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
};
let task = record_from_row(&row)?;
sqlx::query("delete from execution_admission_entries where task_id = $1")
.bind(request.task_id.as_uuid())
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
transaction.commit().await.map_err(Self::unavailable)?;
Ok(match state {
TaskState::Cancelled => TaskCancelOutcome::Cancelled(task),
_ => TaskCancelOutcome::OutcomeUnknown(task),
})
}
pub(super) async fn refuse_attempt_resolution(
&self,
lease: &crank_core::TaskLease,
code: &str,
) -> Result<TaskRecord, TaskError> {
let mut transaction = self.start().await?;
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"select {TASK_PROJECTION} from tasks\n where task_id = $1\n for update"
)))
.bind(lease.record.task_id.as_uuid())
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::NotFound);
};
let record = record_from_row(&row).map_err(Self::unavailable)?;
let now = crank_core::current_unix_ms();
let eligible = matches!(record.state, TaskState::Queued | TaskState::Working)
&& record.lease_owner.as_deref() == Some(lease.owner.as_str())
&& record.fence == lease.fence
&& record.version == lease.record.version
&& record.current_attempt_id.is_none()
&& record.lease_until_unix_ms.is_some_and(|until| until > now);
if !eligible {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
}
let prior_phase = self
.unproven_dispatch_phase(lease.record.task_id, &mut transaction)
.await?;
let uncertain = prior_phase.is_some();
let state = if uncertain {
TaskState::OutcomeUnknown
} else {
TaskState::Failed
};
let phase = prior_phase.unwrap_or_else(|| "before_dispatch".to_owned());
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"update tasks\n set state = $1,\n error_envelope = jsonb_build_object('code', $2, 'phase', $3),\n lease_owner = null,\n lease_until = null,\n next_attempt_at = null,\n version = version + 1,\n updated_at = now()\n where task_id = $4 and version = $5\n returning {TASK_PROJECTION}"
)))
.bind(state.as_str())
.bind(code)
.bind(&phase)
.bind(lease.record.task_id.as_uuid())
.bind(i64::from(record.version))
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
};
let task = record_from_row(&row).map_err(Self::unavailable)?;
sqlx::query("delete from execution_admission_entries where task_id = $1")
.bind(lease.record.task_id.as_uuid())
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
transaction.commit().await.map_err(Self::unavailable)?;
Ok(task)
}
pub(super) async fn attempt_history(
&self,
access: &TaskAccess,
task_id: TaskId,
) -> Result<Vec<TaskAttempt>, TaskError> {
access.validate()?;
let rows = sqlx::query(sqlx::AssertSqlSafe(format!(
"select {ATTEMPT_PROJECTION} from task_attempts
where task_id = $1 and exists (
select 1 from tasks
where tasks.task_id = task_attempts.task_id and tasks.workspace_id = $2
and tasks.owner_kind = $3 and tasks.owner_id = $4
and tasks.authorization_scope = $5
)
order by ordinal"
)))
.bind(task_id.as_uuid())
.bind(&access.workspace_id)
.bind(access.owner_kind.as_str())
.bind(&access.owner_id)
.bind(&access.authorization_scope)
.fetch_all(&self.pool)
.await
.map_err(Self::unavailable)?;
rows.iter().map(attempt_from_row).collect()
}
}
@@ -0,0 +1,906 @@
//! PostgreSQL implementation of the durable Task store (Story 4.2, FR-22/FR-23).
//!
//! One PostgreSQL transaction owns the whole acceptance: Task record, dispatch intent
//! (`execution_admission_entries.task_id`) and its queue slot commit together, so no Task ID and no
//! acknowledgement can exist without the durable record. The transaction reuses the transaction-aware
//! primitives of the execution-admission store (`lock_admission_in_tx`, `bounded_queue_len`,
//! `insert_admission_entry`) instead of copying capacity SQL or starting a nested transaction.
//!
//! Replay is resolved before any capacity mutation and never touches the accepted row. An ambiguous
//! COMMIT is reported as an unavailable store: it is never treated as a proven rollback, so no
//! compensating delete and no synchronous fallback can happen after an uncertain acceptance.
use async_trait::async_trait;
use crank_core::{
AttemptLease, AttemptOutcome, AttemptResolution, BeginAttempt, PlatformApiKeyId,
TaskAcceptRequest, TaskAcceptance, TaskAccess, TaskAttempt, TaskCancelOutcome,
TaskCancelRequest, TaskEnvelope, TaskError, TaskFinishRequest, TaskId, TaskLease,
TaskOwnerKind, TaskPins, TaskPurgeReport, TaskRecord, TaskRetryPolicy, TaskState, TaskStore,
TaskSupportIds, ToolExecutionMode,
};
use serde_json::Value;
use sqlx::{PgPool, Postgres, Row, Transaction};
use uuid::Uuid;
use super::admission::{AdmissionLimits, AttemptError};
mod acceptance;
mod attempts;
const TASK_PROJECTION: &str = "task_id, state, workspace_id, owner_kind, owner_id,
authorization_scope, platform_api_key_id, operation_id, input_fingerprint, operation_version,
policy_mode,
approval_reference,
request_id, trace_id, traceparent, trace_sampled, input_envelope, progress_envelope,
result_envelope, error_envelope, attempts, fence, lease_owner, lease_until,
(extract(epoch from lease_until) * 1000)::bigint as lease_until_ms,
current_attempt_id, dispatch_started_at,
(extract(epoch from dispatch_started_at) * 1000)::bigint as dispatch_started_at_ms, version,
max_attempts, retry_backoff_step_ms,
(extract(epoch from next_attempt_at) * 1000)::bigint as next_attempt_at_ms,
(extract(epoch from cancel_requested_at) * 1000)::bigint as cancel_requested_at_ms,
cancel_reason,
(extract(epoch from execution_deadline) * 1000)::bigint as execution_deadline_ms,
(extract(epoch from retain_until) * 1000)::bigint as retain_until_ms,
(extract(epoch from created_at) * 1000)::bigint as created_ms,
(extract(epoch from updated_at) * 1000)::bigint as updated_ms";
#[derive(Clone)]
pub struct PostgresTaskStore {
pool: PgPool,
limits: AdmissionLimits,
scheduler: crank_core::AdmissionSchedulerConfig,
}
impl PostgresTaskStore {
pub fn new(
pool: PgPool,
limits: AdmissionLimits,
scheduler: crank_core::AdmissionSchedulerConfig,
) -> Result<Self, TaskError> {
limits
.validate()
.map_err(|error| TaskError::invalid(error.to_string()))?;
scheduler
.validate()
.map_err(|error| TaskError::invalid(error.to_string()))?;
Ok(Self {
pool,
limits,
scheduler,
})
}
fn unavailable(error: impl std::fmt::Display) -> TaskError {
TaskError::unavailable(format!("durable task store is unavailable: {error}"))
}
async fn start(&self) -> Result<Transaction<'_, Postgres>, TaskError> {
let mut transaction = self.pool.begin().await.map_err(Self::unavailable)?;
sqlx::query("set local lock_timeout = '3000ms'")
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
sqlx::query("set local statement_timeout = '5000ms'")
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
Ok(transaction)
}
/// Scoped lookup of an accepted Task. The full owner/authorization/Tool predicate is part of the
/// query, so a foreign row is never fetched.
async fn stored_acceptance(
&self,
transaction: &mut Transaction<'_, Postgres>,
request: &TaskAcceptRequest,
) -> Result<Option<(TaskAcceptance, String)>, AttemptError> {
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"select {TASK_PROJECTION} from tasks
where workspace_id = $1 and owner_kind = $2 and owner_id = $3
and authorization_scope = $4 and operation_id = $5 and scoped_key = $6"
)))
.bind(&request.workspace_id)
.bind(request.owner_kind.as_str())
.bind(&request.owner_id)
.bind(&request.authorization_scope)
.bind(&request.operation_id)
.bind(&request.scoped_key)
.fetch_optional(&mut **transaction)
.await
.map_err(AttemptError::from_sql)?;
let Some(row) = row else {
return Ok(None);
};
let record = record_from_row(&row).map_err(AttemptError::Failure)?;
Ok(Some((
acceptance_from_record(&record),
record.input_fingerprint,
)))
}
}
fn acceptance_from_record(record: &TaskRecord) -> TaskAcceptance {
TaskAcceptance {
task_id: record.task_id,
state: record.state,
operation_version: record.pins.operation_version,
support: record.support.clone(),
created_at_unix_ms: record.created_at_unix_ms,
retain_until_unix_ms: record.pins.retain_until_unix_ms,
replayed: true,
}
}
/// Readiness probe of the durable Task state: the table must be reachable and readable. A missing or
/// unreadable schema is a dependency failure, never a degraded-but-ready state.
pub async fn tasks_health(pool: &PgPool) -> Result<(), TaskError> {
sqlx::query("select count(*)::bigint from tasks where false")
.fetch_one(pool)
.await
.map_err(PostgresTaskStore::unavailable)?;
Ok(())
}
/// One transaction's decision of the acceptance path.
enum Accepted {
Created(TaskAcceptance),
Replayed(TaskAcceptance),
Conflict,
}
#[async_trait]
impl TaskStore for PostgresTaskStore {
async fn health(&self) -> Result<(), TaskError> {
sqlx::query("select count(*)::bigint from tasks where false")
.fetch_one(&self.pool)
.await
.map_err(Self::unavailable)?;
super::admission::admission_health(&self.pool)
.await
.map_err(|error| TaskError::unavailable(error.to_string()))?;
Ok(())
}
async fn accept(&self, request: TaskAcceptRequest) -> Result<TaskAcceptance, TaskError> {
request.validate()?;
let mut transaction = self.start().await?;
super::master_key::ensure_no_active_master_key_rotation(&mut transaction)
.await
.map_err(|error| match error {
crate::RegistryError::Storage(_) => TaskError::unavailable(error.to_string()),
_ => TaskError::StateConflict,
})?;
if let Some(epoch) = request.executable_input_master_key_epoch {
super::master_key::ensure_active_master_key_epoch(&mut transaction, epoch)
.await
.map_err(|error| match error {
crate::RegistryError::Storage(_) => TaskError::unavailable(error.to_string()),
_ => TaskError::StateConflict,
})?;
}
let outcome = self.accept_locked(&mut transaction, &request).await;
match outcome {
Ok(Accepted::Created(acceptance) | Accepted::Replayed(acceptance)) => {
// The acknowledgement of a created Task is only possible after a real COMMIT.
transaction.commit().await.map_err(Self::unavailable)?;
Ok(TaskAcceptance {
replayed: acceptance.replayed,
..acceptance
})
}
Ok(Accepted::Conflict) => {
transaction.rollback().await.ok();
Err(TaskError::FingerprintConflict)
}
Err(AttemptError::Overloaded) => {
transaction.rollback().await.ok();
Err(TaskError::Overloaded {
retry_after_ms: crank_core::ADMISSION_OVERLOAD_RETRY_AFTER_MS,
})
}
Err(AttemptError::Contended) => {
transaction.rollback().await.ok();
Err(TaskError::unavailable(
"durable task acceptance could not take the admission locks",
))
}
Err(AttemptError::UniqueViolation) => {
// Roll back the aborted transaction, then reread the scoped key once outside it.
transaction.rollback().await.ok();
self.reread_after_unique(&request).await
}
Err(AttemptError::Failure(error)) => {
transaction.rollback().await.ok();
Err(error)
}
Err(AttemptError::AdmissionFailure(error)) => {
transaction.rollback().await.ok();
Err(TaskError::unavailable(error.to_string()))
}
}
}
async fn read(&self, access: &TaskAccess, task_id: TaskId) -> Result<TaskRecord, TaskError> {
access.validate()?;
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"select {TASK_PROJECTION} from tasks
where task_id = $1 and workspace_id = $2 and owner_kind = $3
and owner_id = $4 and authorization_scope = $5"
)))
.bind(task_id.as_uuid())
.bind(&access.workspace_id)
.bind(access.owner_kind.as_str())
.bind(&access.owner_id)
.bind(&access.authorization_scope)
.fetch_optional(&self.pool)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
return Err(TaskError::NotFound);
};
record_from_row(&row).map_err(Self::unavailable)
}
async fn find(
&self,
access: &TaskAccess,
operation_id: &str,
scoped_key: &str,
) -> Result<Option<TaskRecord>, TaskError> {
access.validate()?;
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"select {TASK_PROJECTION} from tasks
where workspace_id = $1 and owner_kind = $2 and owner_id = $3
and authorization_scope = $4 and operation_id = $5 and scoped_key = $6"
)))
.bind(&access.workspace_id)
.bind(access.owner_kind.as_str())
.bind(&access.owner_id)
.bind(&access.authorization_scope)
.bind(operation_id)
.bind(scoped_key)
.fetch_optional(&self.pool)
.await
.map_err(Self::unavailable)?;
row.map(|row| record_from_row(&row)).transpose()
}
async fn finish(&self, request: TaskFinishRequest) -> Result<TaskRecord, TaskError> {
request.validate()?;
let mut transaction = self.start().await?;
let access = &request.access;
let update = sqlx::query(sqlx::AssertSqlSafe(format!(
"update tasks
set state = $1,
result_envelope = $2,
error_envelope = $3,
version = version + 1,
updated_at = now()
where task_id = $4 and workspace_id = $5 and owner_kind = $6
and owner_id = $7 and authorization_scope = $8
and state = $9 and version = $10
and (lease_owner is null or fence = 0)
and current_attempt_id is null
returning {TASK_PROJECTION}"
)))
.bind(request.target_state.as_str())
.bind(request.result.as_ref().map(|value| value.value().clone()))
.bind(request.error.as_ref().map(|value| value.value().clone()))
.bind(request.task_id.as_uuid())
.bind(&access.workspace_id)
.bind(access.owner_kind.as_str())
.bind(&access.owner_id)
.bind(&access.authorization_scope)
.bind(request.expected_state.as_str())
.bind(i64::from(request.expected_version))
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = update else {
transaction.rollback().await.ok();
return Err(self.classify_missed_finish(&request).await);
};
let record = record_from_row(&row).map_err(Self::unavailable)?;
// Terminal Task and its dispatch intent leave together: a half-done release is rolled back.
sqlx::query("delete from execution_admission_entries where task_id = $1")
.bind(request.task_id.as_uuid())
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
transaction.commit().await.map_err(Self::unavailable)?;
Ok(record)
}
async fn purge_expired(&self, batch: u32) -> Result<TaskPurgeReport, TaskError> {
let batch = i64::from(batch.clamp(1, crank_core::TASK_PURGE_BATCH_LIMIT));
let mut transaction = self.start().await?;
super::master_key::ensure_no_active_master_key_rotation(&mut transaction)
.await
.map_err(|error| match error {
crate::RegistryError::Storage(_) => TaskError::unavailable(error.to_string()),
_ => TaskError::StateConflict,
})?;
let expired = sqlx::query(
"with expired as (
update tasks
set state = 'failed',
error_envelope = jsonb_build_object('reason', 'execution_deadline_elapsed'),
version = version + 1,
updated_at = now()
where task_id in (
select task_id from tasks
where state = 'queued' and execution_deadline <= now()
and lease_owner is null
order by task_id
limit $1
)
returning task_id
), released as (
delete from execution_admission_entries
where task_id in (select task_id from expired)
returning task_id
)
select (select count(*) from expired)::bigint as expired_queued",
)
.bind(batch)
.fetch_one(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let expired_queued: i64 = expired
.try_get("expired_queued")
.map_err(Self::unavailable)?;
let purged = sqlx::query(
"delete from tasks
where task_id in (
select task_id from tasks
where state in ('completed', 'failed', 'cancelled') and retain_until <= now()
order by task_id
limit $1
)",
)
.bind(batch)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
transaction.commit().await.map_err(Self::unavailable)?;
Ok(TaskPurgeReport {
expired_queued: u64::try_from(expired_queued).unwrap_or_default(),
expired_terminal: purged.rows_affected(),
})
}
async fn claim_batch(
&self,
owner: &str,
batch_limit: u32,
lease_ms: i64,
) -> Result<Vec<TaskLease>, TaskError> {
let batch = (batch_limit.clamp(1, 64)) as usize;
let mut transaction = self.start().await?;
// A checkpointed attempt is never redispatched blindly: once its lease expires another worker
// must decide per Task whether the pinned safe-retry evidence, the budget and the deadline
// still allow exactly one bounded new attempt (Story 4.4). Everything else resolves to the
// non-terminal uncertain state and releases its admission slot.
let candidates = sqlx::query(
"select t.task_id, t.current_attempt_id, t.attempts, t.max_attempts,
t.execution_deadline > clock_timestamp() as deadline_left,
coalesce(a.retry_allowed, false) as retry_allowed,
a.resolved_at is null as attempt_open
from tasks t
left join task_attempts a
on a.task_id = t.task_id and a.attempt_id = t.current_attempt_id
where t.state = 'working' and t.lease_owner is not null
and t.current_attempt_id is not null
and t.lease_until < clock_timestamp()
order by t.task_id
limit 64
for update of t skip locked",
)
.fetch_all(&mut *transaction)
.await
.map_err(Self::unavailable)?;
for candidate in candidates {
let task_id: Uuid = candidate.try_get("task_id").map_err(Self::unavailable)?;
let attempt_id: String = candidate
.try_get("current_attempt_id")
.map_err(Self::unavailable)?;
let attempts: i32 = candidate.try_get("attempts").map_err(Self::unavailable)?;
let max_attempts: i32 = candidate
.try_get("max_attempts")
.map_err(Self::unavailable)?;
let deadline_left: bool = candidate
.try_get("deadline_left")
.map_err(Self::unavailable)?;
let retry_allowed: bool = candidate
.try_get("retry_allowed")
.map_err(Self::unavailable)?;
let attempt_open: bool = candidate
.try_get("attempt_open")
.map_err(Self::unavailable)?;
let safe_recovery =
retry_allowed && attempt_open && attempts < max_attempts && deadline_left;
let close = sqlx::query(
"update task_attempts
set resolved_at = now(), outcome = 'outcome_unknown', certainty = 'unknown',
phase = 'dispatched'
where task_id = $1 and attempt_id = $2 and resolved_at is null",
)
.bind(task_id)
.bind(&attempt_id)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let _ = close;
if safe_recovery {
// Interrupted, not failed: re-arm the same admission intent and let the bounded claim
// hand the Task to a new owner with a new fence and exactly one new attempt.
sqlx::query(
"update tasks
set state = 'queued',
lease_owner = null,
lease_until = null,
current_attempt_id = null,
next_attempt_at = null,
version = version + 1,
updated_at = now()
where task_id = $1 and state = 'working'
and lease_until < clock_timestamp()",
)
.bind(task_id)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
sqlx::query(
"update execution_admission_entries
set state = 'reserved', granted_at = null
where task_id = $1 and state = 'executing'",
)
.bind(task_id)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
} else {
sqlx::query(
"update tasks
set state = 'outcome_unknown',
lease_owner = null,
lease_until = null,
error_envelope = jsonb_build_object('code', 'outcome_unknown', 'phase', 'dispatched'),
version = version + 1,
updated_at = now()
where task_id = $1 and state = 'working'
and lease_until < clock_timestamp()",
)
.bind(task_id)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
sqlx::query("delete from execution_admission_entries where task_id = $1")
.bind(task_id)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
}
}
// A crash before the durable dispatch checkpoint left the Task queued for execution: its
// admission intent is re-armed as reserved and the Task becomes claimable again. Recovery is
// bounded by the <=30s lease expiry, inside the 60s story budget. The expiry condition is
// re-checked in the UPDATE so a concurrent renewal can never be clobbered by recovery.
sqlx::query(
"with candidates as (
select task_id from tasks
where state = 'working' and lease_owner is not null
and current_attempt_id is null
and lease_until < clock_timestamp()
order by task_id
limit 64
for update skip locked
), requeued as (
update tasks
set state = 'queued',
lease_owner = null,
lease_until = null,
next_attempt_at = null,
version = version + 1,
updated_at = now()
where task_id in (select task_id from candidates)
and state = 'working'
and current_attempt_id is null
and lease_until < clock_timestamp()
returning task_id
)
update execution_admission_entries
set state = 'reserved', granted_at = null
where task_id in (select task_id from requeued) and state = 'executing'",
)
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
transaction.commit().await.map_err(Self::unavailable)?;
let mut leases = Vec::new();
for _ in 0..batch {
let mut transaction = self.start().await?;
let workspace_cursor =
match super::admission::lock_admission_in_tx(&mut transaction, "*").await {
Ok(cursor) => cursor,
Err(error) => {
transaction.rollback().await.ok();
return Err(TaskError::unavailable(format!("{error:?}")));
}
};
let (task_id, _updated_cursor) = match super::admission::grant_next_task(
&mut transaction,
"*",
&workspace_cursor,
&self.limits,
self.scheduler,
)
.await
{
Ok(value) => value,
Err(error) => {
transaction.rollback().await.ok();
return Err(TaskError::unavailable(format!("{error:?}")));
}
};
// Persist the updated controller cursor even when no task is granted.
let Some(task_id) = task_id else {
transaction.commit().await.map_err(Self::unavailable)?;
break;
};
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"update tasks
set state = 'working',
lease_owner = $1,
lease_until = clock_timestamp() + ($2::bigint || ' milliseconds')::interval,
fence = fence + 1,
updated_at = now()
where task_id = $3
and (
(state = 'queued' and lease_owner is null)
or (state = 'working' and lease_owner is not null
and lease_until < clock_timestamp() and current_attempt_id is null)
)
and (next_attempt_at is null or next_attempt_at <= clock_timestamp())
returning {TASK_PROJECTION}"
)))
.bind(owner)
.bind(lease_ms)
.bind(task_id)
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
// Persist cursor update (grant_next_task already did it) regardless.
let Some(row) = row else {
transaction.rollback().await.ok();
break;
};
let record = record_from_row(&row).map_err(Self::unavailable)?;
leases.push(TaskLease {
owner: owner.to_owned(),
fence: record.fence,
lease_until: record.lease_until_unix_ms.unwrap_or_default(),
record,
});
transaction.commit().await.map_err(Self::unavailable)?;
}
Ok(leases)
}
async fn renew_lease(
&self,
owner: &str,
task_id: TaskId,
fence: u64,
lease_ms: i64,
) -> Result<TaskLease, TaskError> {
let mut transaction = self.start().await?;
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"update tasks
set lease_until = clock_timestamp() + ($1::bigint || ' milliseconds')::interval,
updated_at = now()
where task_id = $2 and lease_owner = $3 and fence = $4
and lease_until > clock_timestamp()
returning {TASK_PROJECTION}"
)))
.bind(lease_ms)
.bind(task_id.as_uuid())
.bind(owner)
.bind(i64::try_from(fence).unwrap_or_default())
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
};
let record = record_from_row(&row).map_err(Self::unavailable)?;
let lease = TaskLease {
owner: owner.to_owned(),
fence: record.fence,
lease_until: record.lease_until_unix_ms.unwrap_or_default(),
record,
};
transaction.commit().await.map_err(Self::unavailable)?;
Ok(lease)
}
async fn write_progress(
&self,
owner: &str,
task_id: TaskId,
fence: u64,
progress: TaskEnvelope,
) -> Result<TaskRecord, TaskError> {
let mut transaction = self.start().await?;
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"update tasks
set progress_envelope = $1, updated_at = now()
where task_id = $2 and lease_owner = $3 and fence = $4
and lease_until > clock_timestamp()
returning {TASK_PROJECTION}"
)))
.bind(progress.value().clone())
.bind(task_id.as_uuid())
.bind(owner)
.bind(i64::try_from(fence).unwrap_or_default())
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
};
let record = record_from_row(&row).map_err(Self::unavailable)?;
transaction.commit().await.map_err(Self::unavailable)?;
Ok(record)
}
async fn read_executable_input(
&self,
owner: &str,
task_id: TaskId,
fence: u64,
) -> Result<(String, String, i64), TaskError> {
let row = sqlx::query(
"select executable_input_ciphertext, executable_input_key_version, executable_input_master_key_epoch
from tasks
where task_id = $1 and state = 'working' and lease_owner = $2 and fence = $3
and lease_until > clock_timestamp()",
)
.bind(task_id.as_uuid())
.bind(owner)
.bind(i64::try_from(fence).unwrap_or_default())
.fetch_optional(&self.pool)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
return Err(TaskError::StateConflict);
};
let ciphertext: Option<String> = row
.try_get("executable_input_ciphertext")
.map_err(Self::unavailable)?;
let key_version: Option<String> = row
.try_get("executable_input_key_version")
.map_err(Self::unavailable)?;
let epoch: Option<i64> = row
.try_get("executable_input_master_key_epoch")
.map_err(Self::unavailable)?;
if let Some(epoch) = epoch {
let active = sqlx::query_scalar::<_, i64>(
"select epoch from master_key_identities where status = 'active' order by epoch desc limit 1",
)
.fetch_optional(&self.pool)
.await
.map_err(Self::unavailable)?;
if let Some(active) = active
&& active != epoch
{
return Err(TaskError::StateConflict);
}
}
match (ciphertext, key_version, epoch) {
(Some(ciphertext), Some(key_version), Some(epoch)) if !ciphertext.is_empty() => {
Ok((ciphertext, key_version, epoch))
}
_ => Err(TaskError::NotFound),
}
}
async fn checkpoint_dispatch(&self, request: BeginAttempt) -> Result<AttemptLease, TaskError> {
self.checkpoint_dispatch_attempt(request).await
}
async fn resolve_attempt(
&self,
lease: &AttemptLease,
outcome: AttemptOutcome,
) -> Result<AttemptResolution, TaskError> {
self.resolve_attempt_outcome(lease, outcome).await
}
async fn cancel(&self, request: TaskCancelRequest) -> Result<TaskCancelOutcome, TaskError> {
self.cancel_task(request).await
}
async fn attempts(
&self,
access: &TaskAccess,
task_id: TaskId,
) -> Result<Vec<TaskAttempt>, TaskError> {
self.attempt_history(access, task_id).await
}
async fn finish_leased(
&self,
request: TaskFinishRequest,
owner: &str,
fence: u64,
) -> Result<TaskRecord, TaskError> {
request.validate()?;
let mut transaction = self.start().await?;
let access = &request.access;
let row = sqlx::query(sqlx::AssertSqlSafe(format!(
"update tasks
set state = $1,
result_envelope = $2,
error_envelope = $3,
version = version + 1,
updated_at = now(),
lease_owner = null,
lease_until = null
where task_id = $4 and workspace_id = $5 and owner_kind = $6
and owner_id = $7 and authorization_scope = $8
and state = $9 and version = $10
and lease_owner = $11 and fence = $12
and lease_until > clock_timestamp()
and current_attempt_id is null
returning {TASK_PROJECTION}"
)))
.bind(request.target_state.as_str())
.bind(request.result.as_ref().map(|value| value.value().clone()))
.bind(request.error.as_ref().map(|value| value.value().clone()))
.bind(request.task_id.as_uuid())
.bind(&access.workspace_id)
.bind(access.owner_kind.as_str())
.bind(&access.owner_id)
.bind(&access.authorization_scope)
.bind(request.expected_state.as_str())
.bind(i64::from(request.expected_version))
.bind(owner)
.bind(i64::try_from(fence).unwrap_or_default())
.fetch_optional(&mut *transaction)
.await
.map_err(Self::unavailable)?;
let Some(row) = row else {
transaction.rollback().await.ok();
return Err(TaskError::StateConflict);
};
let record = record_from_row(&row).map_err(Self::unavailable)?;
sqlx::query("delete from execution_admission_entries where task_id = $1")
.bind(request.task_id.as_uuid())
.execute(&mut *transaction)
.await
.map_err(Self::unavailable)?;
transaction.commit().await.map_err(Self::unavailable)?;
Ok(record)
}
async fn refuse_attempt(&self, lease: &TaskLease, code: &str) -> Result<TaskRecord, TaskError> {
self.refuse_attempt_resolution(lease, code).await
}
}
fn policy_mode_str(mode: ToolExecutionMode) -> &'static str {
match mode {
ToolExecutionMode::Synchronous => "synchronous",
ToolExecutionMode::Background => "background",
ToolExecutionMode::Adaptive => "adaptive",
}
}
fn task_sql_error(error: sqlx::Error) -> AttemptError {
if error
.as_database_error()
.and_then(|database| database.code())
.as_deref()
== Some("23505")
{
return AttemptError::UniqueViolation;
}
AttemptError::from_sql(error)
}
fn record_from_row(row: &sqlx::postgres::PgRow) -> Result<TaskRecord, TaskError> {
let task_id: Uuid = row.try_get("task_id").map_err(unavailable_row)?;
let state: String = row.try_get("state").map_err(unavailable_row)?;
let owner_kind: String = row.try_get("owner_kind").map_err(unavailable_row)?;
let policy_mode: String = row.try_get("policy_mode").map_err(unavailable_row)?;
let trace_id: String = row.try_get("trace_id").map_err(unavailable_row)?;
let request_id: String = row.try_get("request_id").map_err(unavailable_row)?;
let envelope = |column: &str| -> Result<Option<TaskEnvelope>, TaskError> {
let value: Option<Value> = row.try_get(column).map_err(unavailable_row)?;
value.map(TaskEnvelope::from_persisted).transpose()
};
Ok(TaskRecord {
task_id: TaskId::parse(&task_id.to_string())?,
state: TaskState::parse(&state)?,
workspace_id: row.try_get("workspace_id").map_err(unavailable_row)?,
owner_kind: TaskOwnerKind::parse(&owner_kind)?,
owner_id: row.try_get("owner_id").map_err(unavailable_row)?,
authorization_scope: row
.try_get("authorization_scope")
.map_err(unavailable_row)?,
platform_api_key_id: row
.try_get::<Option<String>, _>("platform_api_key_id")
.map_err(unavailable_row)?
.map(PlatformApiKeyId::new),
operation_id: row.try_get("operation_id").map_err(unavailable_row)?,
input_fingerprint: row.try_get("input_fingerprint").map_err(unavailable_row)?,
pins: TaskPins {
operation_version: u32::try_from(
row.try_get::<i32, _>("operation_version")
.map_err(unavailable_row)?,
)
.unwrap_or_default(),
policy_mode: match policy_mode.as_str() {
"background" => ToolExecutionMode::Background,
"adaptive" => ToolExecutionMode::Adaptive,
_ => ToolExecutionMode::Synchronous,
},
execution_deadline_unix_ms: row
.try_get("execution_deadline_ms")
.map_err(unavailable_row)?,
retain_until_unix_ms: row.try_get("retain_until_ms").map_err(unavailable_row)?,
approval_reference: row.try_get("approval_reference").map_err(unavailable_row)?,
retry: TaskRetryPolicy::new(
u32::try_from(
row.try_get::<i32, _>("max_attempts")
.map_err(unavailable_row)?,
)
.unwrap_or(1),
row.try_get::<i64, _>("retry_backoff_step_ms")
.map_err(unavailable_row)?,
),
},
support: TaskSupportIds {
request_id: crank_core::RequestId::parse(&request_id)
.map_err(|_| TaskError::unavailable("stored request id is invalid"))?,
trace_id: crank_core::TraceId::parse(&trace_id)
.map_err(|_| TaskError::unavailable("stored trace id is invalid"))?,
traceparent: row.try_get("traceparent").map_err(unavailable_row)?,
trace_sampled: row.try_get("trace_sampled").map_err(unavailable_row)?,
},
input: envelope("input_envelope")?
.ok_or_else(|| TaskError::unavailable("stored task input envelope is missing"))?,
progress: envelope("progress_envelope")?,
result: envelope("result_envelope")?,
error: envelope("error_envelope")?,
attempts: u32::try_from(row.try_get::<i32, _>("attempts").map_err(unavailable_row)?)
.unwrap_or_default(),
fence: u64::try_from(row.try_get::<i64, _>("fence").map_err(unavailable_row)?)
.unwrap_or_default(),
lease_owner: row.try_get("lease_owner").map_err(unavailable_row)?,
lease_until_unix_ms: row.try_get("lease_until_ms").map_err(unavailable_row)?,
current_attempt_id: row.try_get("current_attempt_id").map_err(unavailable_row)?,
dispatch_started_at_unix_ms: row
.try_get("dispatch_started_at_ms")
.map_err(unavailable_row)?,
next_attempt_at_unix_ms: row.try_get("next_attempt_at_ms").map_err(unavailable_row)?,
cancel_requested_at_unix_ms: row
.try_get("cancel_requested_at_ms")
.map_err(unavailable_row)?,
cancel_reason: row.try_get("cancel_reason").map_err(unavailable_row)?,
version: u32::try_from(row.try_get::<i32, _>("version").map_err(unavailable_row)?)
.unwrap_or_default(),
created_at_unix_ms: row.try_get("created_ms").map_err(unavailable_row)?,
updated_at_unix_ms: row.try_get("updated_ms").map_err(unavailable_row)?,
})
}
fn unavailable_row(error: sqlx::Error) -> TaskError {
PostgresTaskStore::unavailable(error)
}
@@ -1,4 +1,5 @@
mod integration {
mod admission;
mod agent_resource_bindings;
mod agents_usage;
mod approval;
@@ -16,5 +17,6 @@ mod integration {
mod prompt_drafts;
mod prompt_publication;
mod resource_drafts;
mod tasks;
mod workspace_access;
}
@@ -0,0 +1,496 @@
//! ATDD Story 4.1 (FR-21/FR-24): authoritative PostgreSQL execution admission.
//!
//! These oracles drive the real store against a real PostgreSQL schema: scoped permits, bounded
//! queue, typed overload, durable enqueue decisions, deadline propagation, quota keys that ignore
//! the operation version, fairness order and fail-closed health.
use super::common;
use std::time::{Duration, Instant};
use crank_core::{
ADMIN_ORIGIN_KEY, AdmissionMode, AdmissionOutcome, AdmissionPermit, AdmissionRequest,
AdmissionSchedulerConfig, AdmissionScope, ApprovalEvidence, ExecutionAdmission,
NoTasksFallback, ToolExecutionMode, current_unix_ms,
};
use crank_registry::{AdmissionLimits, PostgresExecutionAdmission};
use tokio::sync::mpsc;
fn limits(global: usize, agent: usize, tool: usize, queue: usize) -> AdmissionLimits {
AdmissionLimits {
global_limit: global,
agent_limit: agent,
tool_limit: tool,
queue_limit: queue,
}
}
fn store(
pool: sqlx::PgPool,
limits: AdmissionLimits,
scheduler: AdmissionSchedulerConfig,
) -> PostgresExecutionAdmission {
PostgresExecutionAdmission::new(pool, limits, scheduler).expect("valid admission store")
}
fn scope(agent: &str, operation: &str, version: u32) -> AdmissionScope {
AdmissionScope {
workspace_id: "ws_default".to_owned(),
agent_key: agent.to_owned(),
operation_id: operation.to_owned(),
operation_version: version,
}
}
fn request(scope: AdmissionScope, deadline_in_ms: i64) -> AdmissionRequest {
AdmissionRequest {
scope,
mode: AdmissionMode::Synchronous,
policy_mode: ToolExecutionMode::Synchronous,
no_tasks_fallback: NoTasksFallback::Synchronous,
approval: ApprovalEvidence::NotRequired,
fingerprint: format!("fp_{}", current_unix_ms()),
expires_at_unix_ms: current_unix_ms() + deadline_in_ms,
}
}
fn background(scope: AdmissionScope, fallback: NoTasksFallback) -> AdmissionRequest {
AdmissionRequest {
mode: AdmissionMode::Background,
no_tasks_fallback: fallback,
..request(scope, 30_000)
}
}
fn permit(outcome: AdmissionOutcome) -> AdmissionPermit {
match outcome {
AdmissionOutcome::Permitted(permit) => permit,
other => panic!("expected a scoped permit, got {other:?}"),
}
}
async fn wait_for_metric_poll() {
tokio::time::sleep(Duration::from_millis(20)).await;
}
#[tokio::test]
async fn permits_are_scoped_released_once_and_respect_the_agent_budget() {
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let admission = store(
registry.pool().clone(),
limits(4, 1, 4, 8),
Default::default(),
);
let first = permit(
admission
.admit(request(scope("agent_a", "op_leads", 1), 2_000))
.await
.expect("admission store is reachable"),
);
assert_eq!(first.scope.agent_key, "agent_a");
// The same Agent is already at its scoped budget: the second request waits and expires instead of
// over-admitting, and it never dispatches.
let waiting = admission
.admit(request(scope("agent_a", "op_leads", 2), 120))
.await
.expect("admission store is reachable");
assert_eq!(waiting, AdmissionOutcome::DeadlineElapsed);
// Another Agent still has the global capacity it is entitled to.
let other = permit(
admission
.admit(request(scope("agent_b", "op_leads", 1), 2_000))
.await
.expect("admission store is reachable"),
);
assert_eq!(other.scope.agent_key, "agent_b");
assert!(admission.release(&other).await.unwrap());
// Repeated release is idempotent and never frees a foreign or already released permit.
assert!(!admission.release(&other).await.unwrap());
assert!(admission.release(&first).await.unwrap());
let snapshot = admission.snapshot().await.unwrap();
assert_eq!(snapshot.executing, 0);
assert_eq!(snapshot.waiting, 0);
database.cleanup().await;
}
#[tokio::test]
async fn full_queue_returns_a_typed_overload_with_retry_guidance() {
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let admission = store(
registry.pool().clone(),
limits(1, 1, 1, 2),
Default::default(),
);
let held = permit(
admission
.admit(request(scope("agent_a", "op_leads", 1), 5_000))
.await
.expect("admission store is reachable"),
);
// Two bounded waiting entries fill the queue.
let mut waiters = Vec::new();
for version in 1..=2 {
let admission = admission.clone();
waiters.push(tokio::spawn(async move {
admission
.admit(request(scope("agent_b", "op_leads", version), 5_000))
.await
}));
}
while admission.snapshot().await.unwrap().waiting < 2 {
wait_for_metric_poll().await;
}
let rejected = admission
.admit(request(scope("agent_c", "op_leads", 1), 5_000))
.await
.expect("admission store is reachable");
match rejected {
AdmissionOutcome::Overloaded { retry_after_ms } => {
assert!((1..=crank_core::MAX_RETRY_AFTER_MS).contains(&retry_after_ms));
}
other => panic!("expected typed overload, got {other:?}"),
}
// Cancelling the held permit lets the queued waiters make progress instead of leaking capacity.
assert!(admission.release(&held).await.unwrap());
for waiter in waiters.drain(..) {
let outcome = waiter.await.expect("waiter task");
if let Ok(AdmissionOutcome::Permitted(permit)) = outcome {
let _ = admission.release(&permit).await;
}
}
database.cleanup().await;
}
#[tokio::test]
async fn background_decisions_are_durable_pinned_and_released_once() {
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let admission = store(
registry.pool().clone(),
limits(4, 4, 4, 8),
Default::default(),
);
let decision = match admission
.admit(background(
scope("agent_a", "op_report", 3),
NoTasksFallback::Synchronous,
))
.await
.expect("admission store is reachable")
{
AdmissionOutcome::Enqueued(decision) => decision,
other => panic!("expected a durable enqueue decision, got {other:?}"),
};
assert_eq!(decision.scope.operation_version, 3);
assert_eq!(decision.policy_mode, ToolExecutionMode::Synchronous);
// A second process (fresh pool and store over the same schema) must read the committed decision:
// it occupies bounded queue capacity and is not a runnable execution.
let reconnected = store(
database.concurrent_registry(2).await.pool().clone(),
limits(4, 4, 4, 8),
Default::default(),
);
let snapshot = reconnected.snapshot().await.unwrap();
assert_eq!(snapshot.waiting, 1);
assert_eq!(snapshot.executing, 0);
assert!(
reconnected
.cancel(decision.ticket_id, "not_the_owner")
.await
.unwrap()
.eq(&false),
"a foreign cancel must never release the decision"
);
assert!(
reconnected.snapshot().await.unwrap().waiting.eq(&1),
"the foreign cancel must leave the decision intact"
);
database.cleanup().await;
}
#[tokio::test]
async fn background_without_a_task_transport_is_typed_unsupported() {
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let admission = store(
registry.pool().clone(),
limits(4, 4, 4, 8),
Default::default(),
);
let outcome = admission
.admit(background(
scope("agent_a", "op_report", 1),
NoTasksFallback::Unsupported,
))
.await
.expect("admission store is reachable");
assert_eq!(outcome, AdmissionOutcome::Unsupported);
assert_eq!(admission.snapshot().await.unwrap().waiting, 0);
database.cleanup().await;
}
#[tokio::test]
async fn unverified_approval_never_produces_a_decision() {
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let admission = store(
registry.pool().clone(),
limits(4, 4, 4, 8),
Default::default(),
);
let mut denied = background(
scope("agent_a", "op_report", 1),
NoTasksFallback::Synchronous,
);
denied.approval = ApprovalEvidence::NotGranted;
assert_eq!(
admission.admit(denied).await.unwrap(),
AdmissionOutcome::ApprovalRequired
);
let mut denied_sync = request(scope("agent_a", "op_report", 1), 1_000);
denied_sync.approval = ApprovalEvidence::NotGranted;
assert_eq!(
admission.admit(denied_sync).await.unwrap(),
AdmissionOutcome::ApprovalRequired
);
assert_eq!(admission.snapshot().await.unwrap().waiting, 0);
assert_eq!(admission.snapshot().await.unwrap().executing, 0);
database.cleanup().await;
}
#[tokio::test]
async fn quota_keys_ignore_the_operation_version() {
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let admission = store(
registry.pool().clone(),
limits(4, 4, 1, 8),
Default::default(),
);
let held = permit(
admission
.admit(request(scope("agent_a", "op_leads", 1), 2_000))
.await
.expect("admission store is reachable"),
);
// A different Published Version of the same Tool shares the same per-Tool budget.
let outcome = admission
.admit(request(scope("agent_a", "op_leads", 2), 120))
.await
.expect("admission store is reachable");
assert_eq!(outcome, AdmissionOutcome::DeadlineElapsed);
assert!(admission.release(&held).await.unwrap());
database.cleanup().await;
}
#[tokio::test]
async fn weighted_agents_are_served_their_quantum_in_order() {
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let admission = store(
registry.pool().clone(),
limits(1, 4, 4, 8),
AdmissionSchedulerConfig {
agent_weight: 2,
tool_weight: 1,
},
);
let held = permit(
admission
.admit(request(scope("agent_z", "op_leads", 1), 5_000))
.await
.expect("admission store is reachable"),
);
let (sender, mut receiver) = mpsc::channel(3);
let mut tasks = Vec::new();
for (agent, version) in [("agent_a", 1), ("agent_a", 2), ("agent_b", 1)] {
let admission = admission.clone();
let sender = sender.clone();
tasks.push(tokio::spawn(async move {
let outcome = admission
.admit(request(scope(agent, "op_leads", version), 5_000))
.await;
if let Ok(AdmissionOutcome::Permitted(permit)) = outcome {
let _ = sender.send(permit.scope.agent_key.clone()).await;
// Hold briefly so the grant order, not task scheduling, decides the sequence.
tokio::time::sleep(Duration::from_millis(30)).await;
let _ = admission.release(&permit).await;
}
}));
}
drop(sender);
assert!(admission.release(&held).await.unwrap());
let mut order = Vec::new();
while let Some(agent) = receiver.recv().await {
order.push(agent);
if order.len() == 3 {
break;
}
}
for task in tasks {
let _ = tokio::time::timeout(Duration::from_secs(5), task).await;
}
// agent_a carries weight 2 and therefore receives two consecutive grants before agent_b.
assert_eq!(order, vec!["agent_a", "agent_a", "agent_b"]);
database.cleanup().await;
}
#[tokio::test]
async fn mandatory_state_failure_fails_closed_and_reports_not_ready() {
let database = common::TestDatabase::new().await;
// No migration on this schema: the mandatory admission state does not exist yet.
let unmigrated = database.raw_pool().await;
let admission = store(unmigrated, limits(4, 4, 4, 8), Default::default());
assert!(admission.health().await.is_err());
assert!(
admission
.admit(request(scope("agent_a", "op_leads", 1), 1_000))
.await
.is_err()
);
database.cleanup().await;
}
#[tokio::test]
async fn held_permits_survive_a_new_process_view() {
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let admission = store(
registry.pool().clone(),
limits(1, 4, 4, 8),
Default::default(),
);
let held = permit(
admission
.admit(request(scope("agent_a", "op_leads", 1), 2_000))
.await
.expect("admission store is reachable"),
);
let reconnected = store(
database.concurrent_registry(2).await.pool().clone(),
limits(1, 4, 4, 8),
Default::default(),
);
assert_eq!(reconnected.snapshot().await.unwrap().executing, 1);
let outcome = reconnected
.admit(request(scope("agent_b", "op_leads", 1), 120))
.await
.expect("admission store is reachable");
assert_eq!(
outcome,
AdmissionOutcome::DeadlineElapsed,
"a held permit is not silently reused by another process"
);
assert!(reconnected.release(&held).await.unwrap());
database.cleanup().await;
}
#[tokio::test]
async fn published_execution_policy_is_pinned_across_a_new_draft_version() {
use crank_core::{NoTasksFallback, OperationStatus, ToolExecutionMode, ToolExecutionPolicy};
use crank_registry::{CreateVersionRequest, PublishRequest};
use time::{OffsetDateTime, format_description::well_known::Rfc3339};
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let workspace = common::test_workspace_id();
let policy = ToolExecutionPolicy {
mode: ToolExecutionMode::Background,
deadline_ms: 45_000,
no_tasks_fallback: NoTasksFallback::Unsupported,
};
let mut version_one = common::test_operation("op_policy", 1, OperationStatus::Draft);
version_one.execution_config.execution_policy = Some(policy);
registry
.create_operation(&workspace, &version_one, None)
.await
.unwrap();
registry
.publish_operation(PublishRequest {
workspace_id: &workspace,
operation_id: &version_one.id,
version: 1,
published_at: &OffsetDateTime::parse("2026-03-25T12:10:00Z", &Rfc3339).unwrap(),
published_by: Some("alice"),
})
.await
.unwrap();
let published = registry
.get_published_operation(&version_one.id)
.await
.unwrap()
.expect("published operation");
assert_eq!(published.execution_config.execution_policy, Some(policy));
// A later Draft with another policy never rewrites the already accepted publication.
let mut version_two = common::test_operation("op_policy", 2, OperationStatus::Draft);
version_two.execution_config.execution_policy = Some(ToolExecutionPolicy {
mode: ToolExecutionMode::Synchronous,
deadline_ms: 1_000,
no_tasks_fallback: NoTasksFallback::Synchronous,
});
registry
.create_version(CreateVersionRequest {
workspace_id: &workspace,
snapshot: &version_two,
change_note: Some("next draft"),
created_by: Some("alice"),
})
.await
.unwrap();
let unchanged = registry
.get_published_operation(&version_one.id)
.await
.unwrap()
.expect("published operation");
assert_eq!(unchanged.version, 1);
assert_eq!(unchanged.execution_config.execution_policy, Some(policy));
database.cleanup().await;
}
#[tokio::test]
async fn admin_origin_bucket_is_bounded_together_with_agents() {
let database = common::TestDatabase::new().await;
let registry = database.registry().await;
let admission = store(
registry.pool().clone(),
limits(1, 1, 1, 4),
Default::default(),
);
let held = permit(
admission
.admit(request(scope(ADMIN_ORIGIN_KEY, "op_leads", 1), 2_000))
.await
.expect("admission store is reachable"),
);
let started = Instant::now();
let outcome = admission
.admit(request(scope("agent_a", "op_leads", 1), 120))
.await
.expect("admission store is reachable");
assert_eq!(outcome, AdmissionOutcome::DeadlineElapsed);
assert!(started.elapsed() >= Duration::from_millis(100));
assert!(admission.release(&held).await.unwrap());
database.cleanup().await;
}
@@ -2,6 +2,7 @@ use crank_registry::{MigrationAuthority, MigrationPreflight, PostgresRegistry};
use sqlx::Row;
mod artifact_metadata;
mod durable_tasks_guard;
mod import_plan;
mod legacy_adoption;
mod preflight;
@@ -70,6 +71,12 @@ async fn controlled_authority_is_versioned_and_safe_under_concurrent_apply() {
(21, "prompts-v21"),
(22, "prompt-publications-v22"),
(23, "prompt-dependencies-v23"),
(24, "execution-admission-v24"),
(25, "durable-tasks-v25"),
(26, "durable-tasks-v26"),
(27, "durable-tasks-v27"),
(28, "durable-tasks-v28"),
(29, "durable-tasks-v29"),
];
assert_eq!(rows.len(), expected.len());
for (row, (version, name)) in rows.iter().zip(expected) {
@@ -180,7 +187,7 @@ async fn controlled_authority_is_versioned_and_safe_under_concurrent_apply() {
);
assert_eq!(
MigrationAuthority::preflight(first.pool()).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
}
#[tokio::test]
@@ -228,7 +235,7 @@ async fn v18_guard_rejects_missing_agent_resource_binding_trigger() {
.expect("the V18-local guard must be restorable");
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"restoring the exact V18 contract must return the current verdict"
);
}
@@ -354,7 +361,7 @@ async fn ledgerless_legacy_baseline_is_consolidated_without_data_loss() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 0,
target: 23,
target: 29,
}
);
MigrationAuthority::apply(&pool).await.unwrap();
@@ -467,13 +474,13 @@ async fn healthy_v2_is_reported_as_migration_required_and_upgrades_to_v3() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 2,
target: 23,
target: 29,
}
);
MigrationAuthority::apply(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
let trace_column: bool = sqlx::query_scalar(
"select exists (
@@ -516,7 +523,7 @@ async fn healthy_v3_upgrades_to_v4_with_honest_legacy_snapshot_provenance() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 3,
target: 23,
target: 29,
}
);
MigrationAuthority::apply(&pool).await.unwrap();
@@ -576,7 +583,7 @@ async fn healthy_v4_upgrades_to_v5_without_fabricating_legacy_outcomes() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 4,
target: 23,
target: 29,
}
);
MigrationAuthority::apply(&pool).await.unwrap();
@@ -830,6 +837,6 @@ async fn v3_upgrade_ignores_oversized_legacy_request_ids_in_partial_index() {
MigrationAuthority::apply(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
}
@@ -32,7 +32,7 @@ async fn healthy_v11_upgrades_to_v13_without_rewriting_prior_ledger() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 11,
target: 23,
target: 29,
}
);
@@ -177,7 +177,7 @@ async fn v12_guard_still_rejects_column_constraint_index_and_relation_drift() {
sqlx::raw_sql(restore).execute(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"restore: {restore}"
);
}
@@ -230,7 +230,7 @@ async fn v13_guard_rejects_wrong_cleanup_index_order_and_predicate() {
sqlx::raw_sql(restore).execute(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"restore: {restore}"
);
}
@@ -507,7 +507,7 @@ async fn assert_v12_drift(pool: &sqlx::PgPool, drift: &str, restore: &str) {
.unwrap();
assert_eq!(
MigrationAuthority::preflight(pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"restore: {restore}"
);
}
@@ -845,7 +845,7 @@ async fn v17_guard_rejects_publication_schema_drift() {
MigrationAuthority::apply(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
for (drift, restore) in [
@@ -879,7 +879,7 @@ async fn v17_guard_rejects_publication_schema_drift() {
sqlx::raw_sql(restore).execute(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"restore: {restore}"
);
}
@@ -289,7 +289,7 @@ async fn assert_primary_key_nullability(pool: &sqlx::PgPool, table: &str, column
);
assert_eq!(
MigrationAuthority::preflight(pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"a rejected primary-key nullability change must leave the schema current"
);
@@ -473,7 +473,7 @@ async fn assert_definition_only_drift(
assert_eq!(super::catalog_constraint_definitions(pool).await, before);
assert_eq!(
MigrationAuthority::preflight(pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
}
@@ -554,7 +554,7 @@ async fn story36_v16_guard_rejects_constraint_and_created_index_definition_drift
MigrationAuthority::apply(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"the baseline before the drift matrix must be the current V16 schema"
);
let mut observations: Vec<String> = Vec::new();
@@ -773,7 +773,7 @@ async fn story36_v16_guard_rejects_constraint_and_created_index_definition_drift
.await
.expect("restoring the exact definition must apply");
let restored = MigrationAuthority::preflight(&pool).await;
let restored_current = matches!(restored, Ok(MigrationPreflight::Current { version: 23 }));
let restored_current = matches!(restored, Ok(MigrationPreflight::Current { version: 29 }));
// Точное восстановление доказывается сравнением снятого снимка с baseline, а не
// мнением того же guard-а, который принимает неправильные определения (F4).
let restored_exact = catalog_definition_snapshot(&pool).await == before;
@@ -804,7 +804,7 @@ async fn story36_v16_guard_resolves_dependent_source_fk_in_the_current_schema()
MigrationAuthority::apply(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"baseline before inserting the shadow schema"
);
@@ -875,7 +875,7 @@ async fn story36_v16_guard_resolves_dependent_source_fk_in_the_current_schema()
// Тень не меняет вердикт guard-а по текущей схеме.
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"a same-named object in another schema must not change the verdict for the current schema"
);
@@ -947,7 +947,7 @@ async fn story36_v16_guard_resolves_dependent_source_fk_in_the_current_schema()
);
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 },
MigrationPreflight::Current { version: 29 },
"the exact restoration must return the current V16 schema"
);
@@ -958,7 +958,7 @@ async fn story36_v16_guard_resolves_dependent_source_fk_in_the_current_schema()
.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
assert!(
@@ -0,0 +1,83 @@
use super::*;
use crank_registry::MigrationApplyResult;
#[tokio::test]
async fn clean_legacy_ledger_upgrades_to_v28_after_v25_removal() {
let database_url = crank_test_support::postgres_schema_url("test_v43_clean_legacy").await;
let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
MigrationAuthority::apply(&pool).await.unwrap();
super::remove_v25_schema(&pool).await;
// remove_v25_schema drops `tasks` and ledger >= 25, but V25 also added the
// nullable dispatch-intent column/index/FK to execution_admission_entries;
// restore those to the V24 shape so the legacy ledger is self-consistent.
sqlx::raw_sql(
"alter table execution_admission_entries drop constraint if exists execution_admission_entries_task_fkey;
drop index if exists execution_admission_entries_task_uidx;
alter table execution_admission_entries drop column if exists task_id;",
)
.execute(&pool)
.await
.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 24,
target: 29,
}
);
assert_eq!(
MigrationAuthority::apply(&pool).await.unwrap(),
MigrationApplyResult::Applied { from: 24, to: 29 }
);
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 29 }
);
}
#[tokio::test]
async fn partial_v27_on_legacy_ledger_is_rejected_by_preflight() {
let database_url = crank_test_support::postgres_schema_url("test_v43_partial_v27").await;
let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
MigrationAuthority::apply(&pool).await.unwrap();
super::remove_v25_schema(&pool).await;
sqlx::raw_sql(
"alter table execution_admission_entries drop constraint if exists execution_admission_entries_task_fkey;
drop index if exists execution_admission_entries_task_uidx;
alter table execution_admission_entries drop column if exists task_id;",
)
.execute(&pool)
.await
.unwrap();
// Reintroduce a partial V27-shaped `tasks` table on top of the legacy (v24) ledger.
sqlx::raw_sql(sqlx::AssertSqlSafe(
"create table tasks (
task_id uuid primary key,
executable_input_ciphertext text,
executable_input_key_version text,
executable_input_master_key_epoch bigint
);",
))
.execute(&pool)
.await
.unwrap();
let error = MigrationAuthority::preflight(&pool).await.unwrap_err();
assert_eq!(error.code(), "partial_sequence");
// The rejection today is attributed to V25's absent-schema guard because the
// `tasks` relation already exists, so a behavioral RED pin for the missing
// V27 guard line cannot be reached; this test fixes the rejection contract
// instead of claiming new behavior.
assert_eq!(error.version(), Some(24));
assert_eq!(error.stage(), "preflight.schema");
let count: i64 =
sqlx::query_scalar("select count(*)::bigint from __crank_migrations where version >= 25")
.fetch_one(&pool)
.await
.unwrap();
assert_eq!(
count, 0,
"nothing at or above V25 should have been applied over the partial fixture"
);
}
@@ -71,7 +71,7 @@ async fn fresh_install_exposes_import_plan_objects_with_cascading_job_fk() {
MigrationAuthority::apply(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
let tables = sqlx::query(
"select table_name
@@ -224,7 +224,7 @@ async fn healthy_v13_upgrades_to_v14_without_changing_prior_ledger() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 13,
target: 23,
target: 29,
}
);
MigrationAuthority::apply(&pool).await.unwrap();
@@ -88,7 +88,7 @@ async fn published_ledgerless_baseline_upgrades_without_data_loss() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 0,
target: 23,
target: 29,
}
);
MigrationAuthority::apply(&pool).await.unwrap();
@@ -101,7 +101,7 @@ async fn published_ledgerless_baseline_upgrades_without_data_loss() {
assert_eq!(operation_count, 1);
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
}
@@ -194,7 +194,7 @@ async fn published_ledgerless_historical_column_layout_upgrades_without_data_los
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 0,
target: 23,
target: 29,
}
);
MigrationAuthority::apply(&pool).await.unwrap();
@@ -230,7 +230,7 @@ async fn published_ledgerless_historical_column_layout_upgrades_without_data_los
);
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
}
@@ -262,7 +262,7 @@ async fn published_initial_mcp_layout_is_accepted() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 0,
target: 23,
target: 29,
}
);
MigrationAuthority::apply(&pool).await.unwrap();
@@ -280,7 +280,7 @@ async fn published_initial_mcp_layout_is_accepted() {
assert!(supports_elicitation);
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
}
@@ -314,13 +314,13 @@ async fn published_in_place_mcp_upgrade_layout_is_accepted() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 0,
target: 23,
target: 29,
}
);
MigrationAuthority::apply(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
}
@@ -14,7 +14,7 @@ async fn fresh_install_creates_the_empty_revocation_authority() {
MigrationAuthority::apply(&pool).await.unwrap();
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
let rows: i64 = sqlx::query_scalar("select count(*) from resource_security_revocations")
.fetch_one(&pool)
@@ -42,12 +42,12 @@ async fn v18_upgrades_to_v19_without_rewriting_the_prior_ledger() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 18,
target: 23,
target: 29,
}
);
assert_eq!(
MigrationAuthority::apply(&pool).await.unwrap(),
MigrationApplyResult::Applied { from: 18, to: 23 }
MigrationApplyResult::Applied { from: 18, to: 29 }
);
let after: Vec<(i64, String, String)> =
sqlx::query_as("select version, name, checksum from __crank_migrations where version < 19 order by version")
@@ -57,7 +57,7 @@ async fn v18_upgrades_to_v19_without_rewriting_the_prior_ledger() {
assert_eq!(before, after, "V16–V18 ledger rows must not be rewritten");
assert_eq!(
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::Current { version: 23 }
MigrationPreflight::Current { version: 29 }
);
}
@@ -84,18 +84,24 @@ async fn a_database_newer_than_the_binary_fails_closed_at_startup() {
let database_url = crank_test_support::postgres_schema_url("test_v19_future_version").await;
let pool = sqlx::PgPool::connect(&database_url).await.unwrap();
MigrationAuthority::apply(&pool).await.unwrap();
// The same `version > CURRENT_VERSION` branch an old V23 binary takes when it
// meets a V24 database; a synthetic V24 row exercises it without a second binary.
// The same `version > CURRENT_VERSION` branch an older binary takes when it meets a newer
// database; a synthetic version above the current one exercises it without a second binary.
let current: i64 = sqlx::query_scalar("select max(version) from __crank_migrations")
.fetch_one(&pool)
.await
.unwrap();
let future = current + 1;
sqlx::query(
"insert into __crank_migrations (version, name, checksum, phase, compatibility)
values (24, 'future-v24', repeat('a', 64), 'expand', 'n-minus-one-readable')",
values ($1, 'future-synthetic', repeat('a', 64), 'expand', 'n-minus-one-readable')",
)
.bind(future)
.execute(&pool)
.await
.unwrap();
let error = MigrationAuthority::preflight(&pool).await.unwrap_err();
assert_eq!(error.code(), "future_version");
assert_eq!(error.version(), Some(24));
assert_eq!(error.version(), Some(future));
assert_eq!(error.recovery(), "install_matching_application");
assert!(
MigrationAuthority::require_current(&pool).await.is_err(),
@@ -155,7 +161,7 @@ async fn failed_v19_migration_rolls_back_schema_and_ledger() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 18,
target: 23,
target: 29,
}
);
}
@@ -63,7 +63,7 @@ async fn failed_artifact_metadata_migration_rolls_back_schema_and_ledger() {
MigrationAuthority::preflight(&pool).await.unwrap(),
MigrationPreflight::MigrationRequired {
current: 11,
target: 23,
target: 29,
}
);
}
@@ -38,6 +38,7 @@ pub(crate) async fn remove_v4_schema(pool: &sqlx::PgPool) {
.unwrap();
}
pub(crate) async fn remove_v5_schema(pool: &sqlx::PgPool) {
remove_v24_schema(pool).await;
remove_v6_schema(pool).await;
sqlx::raw_sql(
"drop index if exists invocation_logs_workspace_operation_version_idx;
@@ -154,6 +155,7 @@ pub(crate) async fn remove_v11_schema(pool: &sqlx::PgPool) {
.unwrap();
}
pub(crate) async fn remove_v14_schema(pool: &sqlx::PgPool) {
remove_v24_schema(pool).await;
remove_v15_schema(pool).await;
sqlx::raw_sql(
"drop table if exists import_plan_applications;
@@ -224,6 +226,7 @@ pub(crate) async fn remove_v18_schema(pool: &sqlx::PgPool) {
.unwrap();
}
pub(crate) async fn remove_v19_schema(pool: &sqlx::PgPool) {
remove_v24_schema(pool).await;
remove_v20_schema(pool).await;
sqlx::raw_sql(
"drop table if exists resource_security_revocations;
@@ -307,6 +310,7 @@ pub(crate) async fn remove_v20_schema(pool: &sqlx::PgPool) {
.unwrap();
}
pub(crate) async fn remove_v12_schema(pool: &sqlx::PgPool) {
remove_v24_schema(pool).await;
remove_v14_schema(pool).await;
sqlx::raw_sql(
"drop index if exists import_jobs_openapi_source_idx;
@@ -323,3 +327,28 @@ pub(crate) async fn remove_v12_schema(pool: &sqlx::PgPool) {
.await
.unwrap();
}
/// Story 4.2: снять V25 (durable Tasks и Task-linked dispatch intent) перед возвратом к < 25.
pub(crate) async fn remove_v25_schema(pool: &sqlx::PgPool) {
// V25 created `tasks`; V26–V28 only altered it, so dropping the table removes their schema too.
sqlx::raw_sql(
"drop table if exists tasks cascade;
delete from __crank_migrations where version >= 25;",
)
.execute(pool)
.await
.unwrap();
}
/// Story 4.1: снять V24 (authoritative execution admission) перед возвратом к < 24.
pub(crate) async fn remove_v24_schema(pool: &sqlx::PgPool) {
remove_v25_schema(pool).await;
sqlx::raw_sql(
"drop table if exists execution_admission_entries;
drop table if exists execution_admission_controllers;
delete from __crank_migrations where version = 24;",
)
.execute(pool)
.await
.unwrap();
}
Loaded 100 of 128 files, more files were not shown because too many files have changed in this diff. Show more