feat: expose streaming test run results

This commit is contained in:
a.tolmachev
2026-04-07 18:30:08 +03:00
parent 35e9c8c754
commit 1a4f0ea6f3
9 changed files with 865 additions and 91 deletions
+420 -15
View File
@@ -20,14 +20,15 @@ use crank_core::{
use crank_mapping::{JsonPathRoot, MappingSet, infer_mapping_from_samples};
use crank_proto::{ProtoService, services_from_descriptor_set_bytes};
use crank_registry::{
AgentSummary, AgentVersionRecord, AsyncJobFilter, CreateAgentRequest, CreateInvitationRequest,
CreateInvocationLogRequest, CreatePlatformApiKeyRequest, CreateSecretRequest,
CreateVersionRequest, CreateWorkspaceRequest, InvitationRecord, InvocationLogRecord,
ListInvocationLogsQuery, MembershipRecord, OperationAgentRef, OperationSampleMetadata,
OperationSummary, OperationUsageSummary, OperationVersionRecord, Page, PlatformApiKeyRecord,
PostgresRegistry, PublishAgentRequest, PublishRequest, RegistryError, RegistryOperation,
RotateSecretRequest, SampleKind, SaveAgentBindingsRequest, SaveAuthProfileRequest,
SaveDescriptorMetadataRequest, SaveSampleMetadataRequest, StreamSessionFilter,
AgentSummary, AgentVersionRecord, AsyncJobFilter, CreateAgentRequest, CreateAsyncJobRequest,
CreateInvitationRequest, CreateInvocationLogRequest, CreatePlatformApiKeyRequest,
CreateSecretRequest, CreateStreamSessionRequest, CreateVersionRequest, CreateWorkspaceRequest,
InvitationRecord, InvocationLogRecord, ListInvocationLogsQuery, MembershipRecord,
OperationAgentRef, OperationSampleMetadata, OperationSummary, OperationUsageSummary,
OperationVersionRecord, Page, PlatformApiKeyRecord, PostgresRegistry, PublishAgentRequest,
PublishRequest, RegistryError, RegistryOperation, RotateSecretRequest, SampleKind,
SaveAgentBindingsRequest, SaveAuthProfileRequest, SaveDescriptorMetadataRequest,
SaveSampleMetadataRequest, StreamSessionFilter, UpdateAsyncJobStatusRequest,
UpdateWorkspaceRequest, UsageAgentBreakdown, UsageBucket, UsageOperationBreakdown, UsageQuery,
UsageSummary, UsageTimelinePoint, WorkspaceMembershipRecord, WorkspaceRecord,
};
@@ -128,9 +129,123 @@ pub struct TestRunPayload {
#[derive(Clone, Debug, Serialize)]
pub struct TestRunResult {
pub ok: bool,
pub mode: ExecutionMode,
pub request_preview: Value,
pub response_preview: Value,
pub errors: Vec<Value>,
#[serde(skip_serializing_if = "Option::is_none")]
pub window: Option<WindowTestRunView>,
#[serde(skip_serializing_if = "Option::is_none")]
pub stream_session: Option<StreamSessionStartView>,
#[serde(skip_serializing_if = "Option::is_none")]
pub async_job: Option<AsyncJobStartView>,
}
#[derive(Clone, Debug, Serialize)]
pub struct WindowTestRunView {
pub window_complete: bool,
pub truncated: bool,
pub has_more: bool,
pub cursor: Option<Value>,
}
#[derive(Clone, Debug, Serialize)]
pub struct StreamSessionStartView {
pub session_id: String,
pub status: StreamStatus,
pub expires_at: String,
pub poll_after_ms: u64,
pub preview: Value,
}
#[derive(Clone, Debug, Serialize)]
pub struct AsyncJobStartView {
pub job_id: String,
pub status: JobStatus,
pub progress: Value,
}
#[derive(Debug, Serialize, Deserialize)]
struct StoredSessionState {
input: Value,
summary: Value,
items: Vec<Value>,
next_index: usize,
batch_size: usize,
}
enum TestRunOutcome {
Unary {
output: Value,
},
Window {
output: crank_runtime::WindowExecutionResult,
},
Session {
output: StreamSessionStartView,
},
AsyncJob {
output: AsyncJobStartView,
},
}
impl TestRunOutcome {
fn into_result_views(
self,
) -> (
&'static str,
Value,
Option<WindowTestRunView>,
Option<StreamSessionStartView>,
Option<AsyncJobStartView>,
) {
match self {
Self::Unary { output } => ("admin test run completed", output, None, None, None),
Self::Window { output } => (
"admin window test run completed",
json!({
"summary": output.summary,
"items": output.items,
"cursor": output.cursor,
"window_complete": output.window_complete,
"truncated": output.truncated,
"has_more": output.has_more,
}),
Some(WindowTestRunView {
window_complete: output.window_complete,
truncated: output.truncated,
has_more: output.has_more,
cursor: output.cursor,
}),
None,
None,
),
Self::Session { output } => (
"admin session test run started",
json!({
"session_id": output.session_id,
"status": output.status,
"expires_at": output.expires_at,
"poll_after_ms": output.poll_after_ms,
"preview": output.preview,
}),
None,
Some(output),
None,
),
Self::AsyncJob { output } => (
"admin async job test run started",
json!({
"job_id": output.job_id,
"status": output.status,
"progress": output.progress,
}),
None,
None,
Some(output),
),
}
}
}
#[derive(Clone, Debug, Deserialize)]
@@ -1889,6 +2004,12 @@ impl AdminService {
.get_operation_version(workspace_id, operation_id, payload.version)
.await?;
let runtime = RuntimeOperation::from(record.snapshot.clone());
let mode = runtime
.execution_config
.streaming
.as_ref()
.map(|streaming| streaming.mode)
.unwrap_or(ExecutionMode::Unary);
let request_preview =
match build_request_preview(&record.snapshot.input_mapping, &payload.input) {
Ok(preview) => preview,
@@ -1910,11 +2031,15 @@ impl AdminService {
.await?;
return Ok(TestRunResult {
ok: false,
mode,
request_preview: Value::Null,
response_preview: Value::Null,
errors: vec![crate::error::runtime_test_failure(&RuntimeError::Mapping(
error,
))],
window: None,
stream_session: None,
async_job: None,
});
}
};
@@ -1924,16 +2049,39 @@ impl AdminService {
.await;
let started_at = std::time::Instant::now();
match match resolved_auth {
Ok(resolved_auth) => {
self.runtime
Ok(resolved_auth) => match mode {
ExecutionMode::Unary => self
.runtime
.execute_with_auth(&runtime, &payload.input, resolved_auth.as_ref())
.await
}
.map(|output| TestRunOutcome::Unary { output }),
ExecutionMode::Window => self
.runtime
.execute_window_with_auth(&runtime, &payload.input, resolved_auth.as_ref())
.await
.map(|output| TestRunOutcome::Window { output }),
ExecutionMode::Session => self
.start_stream_session_test(
workspace_id,
&record.snapshot,
&runtime,
&payload.input,
resolved_auth.as_ref(),
)
.await
.map(|output| TestRunOutcome::Session { output }),
ExecutionMode::AsyncJob => self
.start_async_job_test(workspace_id, &record.snapshot, &runtime, &payload.input)
.await
.map(|output| TestRunOutcome::AsyncJob { output }),
},
Err(error) => Err(error),
} {
Ok(output) => {
Ok(outcome) => {
let duration_ms =
u64::try_from(started_at.elapsed().as_millis()).unwrap_or(u64::MAX);
let (message, response_preview, window, stream_session, async_job) =
outcome.into_result_views();
self.record_invocation(InvocationRecordRequest {
workspace_id,
agent_id: None,
@@ -1941,19 +2089,23 @@ impl AdminService {
source: InvocationSource::AdminTestRun,
level: InvocationLevel::Info,
status: InvocationStatus::Ok,
message: "admin test run completed".to_owned(),
message: message.to_owned(),
status_code: None,
error_kind: None,
duration_ms,
request_preview: request_preview.clone(),
response_preview: output.clone(),
response_preview: response_preview.clone(),
})
.await?;
Ok(TestRunResult {
ok: true,
mode,
request_preview,
response_preview: output,
response_preview,
errors: Vec::new(),
window,
stream_session,
async_job,
})
}
Err(error) => {
@@ -1976,9 +2128,13 @@ impl AdminService {
.await?;
Ok(TestRunResult {
ok: false,
mode,
request_preview,
response_preview: Value::Null,
errors: vec![crate::error::runtime_test_failure(&error)],
window: None,
stream_session: None,
async_job: None,
})
}
}
@@ -2009,6 +2165,187 @@ impl AdminService {
.map(Some)
}
async fn start_stream_session_test(
&self,
workspace_id: &WorkspaceId,
operation: &RegistryOperation,
runtime: &RuntimeOperation,
input: &Value,
resolved_auth: Option<&ResolvedAuth>,
) -> Result<StreamSessionStartView, RuntimeError> {
let streaming = runtime.execution_config.streaming.as_ref().ok_or_else(|| {
RuntimeError::MissingStreamingConfig {
operation_id: runtime.operation_id.as_str().to_owned(),
}
})?;
let seed = self
.runtime
.execute_session_seed_with_auth(runtime, input, resolved_auth)
.await?;
let batch_size = streaming.max_items.unwrap_or(10).max(1) as usize;
let preview_count = seed.items.len().min(batch_size);
let preview_items = seed.items[..preview_count].to_vec();
let next_index = preview_count;
let created_at = now_string().map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?;
let expires_at = add_millis(
&created_at,
streaming.max_session_lifetime_ms.unwrap_or(60_000),
)
.map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?;
let session = StreamSession {
id: crank_core::StreamSessionId::new(new_prefixed_id("sess")),
workspace_id: workspace_id.clone(),
agent_id: None,
operation_id: operation.id.clone(),
protocol: operation.protocol,
mode: ExecutionMode::Session,
status: StreamStatus::Running,
cursor: (next_index < seed.items.len()).then(|| json!(next_index)),
state: json!(StoredSessionState {
input: input.clone(),
summary: seed.summary.clone(),
items: seed.items.clone(),
next_index,
batch_size,
}),
expires_at: expires_at.clone(),
last_poll_at: Some(created_at.clone()),
created_at,
closed_at: None,
};
self.registry
.create_stream_session(CreateStreamSessionRequest { session: &session })
.await
.map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?;
Ok(StreamSessionStartView {
session_id: session.id.as_str().to_owned(),
status: session.status,
expires_at,
poll_after_ms: streaming.poll_interval_ms.unwrap_or(1_000),
preview: json!({
"summary": seed.summary,
"items": preview_items,
}),
})
}
async fn start_async_job_test(
&self,
workspace_id: &WorkspaceId,
operation: &RegistryOperation,
runtime: &RuntimeOperation,
input: &Value,
) -> Result<AsyncJobStartView, RuntimeError> {
let created_at = now_string().map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?;
let expires_at =
add_millis(&created_at, 300_000).map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?;
let job = AsyncJobHandle {
id: crank_core::AsyncJobId::new(new_prefixed_id("job")),
workspace_id: workspace_id.clone(),
agent_id: None,
operation_id: operation.id.clone(),
status: JobStatus::Running,
progress: json!({ "pct": 0 }),
result: None,
error: None,
expires_at: Some(expires_at),
created_at: created_at.clone(),
updated_at: created_at.clone(),
finished_at: None,
};
self.registry
.create_async_job(CreateAsyncJobRequest { job: &job })
.await
.map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?;
let registry = self.registry.clone();
let secret_crypto = self.secret_crypto.clone();
let runtime_for_task = self.runtime.clone();
let workspace_for_task = workspace_id.clone();
let operation_for_task = runtime.clone();
let input_for_task = input.clone();
let job_id = job.id.clone();
tokio::spawn(async move {
let resolved_auth = resolve_runtime_auth_for_task(
&registry,
&secret_crypto,
&workspace_for_task,
&operation_for_task.execution_config,
)
.await;
let result = match resolved_auth {
Ok(resolved_auth) => {
runtime_for_task
.execute_with_auth(
&operation_for_task,
&input_for_task,
resolved_auth.as_ref(),
)
.await
}
Err(error) => Err(error),
};
let finished_at = match now_string() {
Ok(value) => value,
Err(_) => return,
};
let _ = match result {
Ok(output) => {
registry
.update_async_job_status(UpdateAsyncJobStatusRequest {
job_id: &job_id,
current_status: JobStatus::Running,
next_status: JobStatus::Completed,
progress: &json!({ "pct": 100 }),
result: Some(&output),
error: None,
expires_at: None,
updated_at: &finished_at,
finished_at: Some(&finished_at),
})
.await
}
Err(error) => {
registry
.update_async_job_status(UpdateAsyncJobStatusRequest {
job_id: &job_id,
current_status: JobStatus::Running,
next_status: JobStatus::Failed,
progress: &json!({ "pct": 100 }),
result: None,
error: Some(&json!({
"code": runtime_error_code(&error),
"message": error.to_string(),
})),
expires_at: None,
updated_at: &finished_at,
finished_at: Some(&finished_at),
})
.await
}
};
});
Ok(AsyncJobStartView {
job_id: job.id.as_str().to_owned(),
status: job.status,
progress: job.progress,
})
}
async fn resolve_auth_profile(
&self,
workspace_id: &WorkspaceId,
@@ -4267,6 +4604,74 @@ fn now_string() -> Result<String, ApiError> {
.map_err(|error| ApiError::internal(error.to_string()))
}
fn add_millis(timestamp: &str, millis: u64) -> Result<String, ApiError> {
let parsed = OffsetDateTime::parse(timestamp, &Rfc3339)
.map_err(|error| ApiError::internal(error.to_string()))?;
parsed
.checked_add(time::Duration::milliseconds(millis as i64))
.ok_or_else(|| ApiError::internal("failed to add millis to timestamp"))?
.format(&Rfc3339)
.map_err(|error| ApiError::internal(error.to_string()))
}
async fn resolve_runtime_auth_for_task(
registry: &PostgresRegistry,
secret_crypto: &SecretCrypto,
workspace_id: &WorkspaceId,
execution_config: &crank_core::ExecutionConfig,
) -> Result<Option<ResolvedAuth>, RuntimeError> {
let Some(auth_profile_id) = execution_config.auth_profile_ref.as_ref() else {
return Ok(None);
};
let auth_profile = registry
.get_auth_profile(workspace_id, auth_profile_id)
.await
.map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?
.ok_or_else(|| RuntimeError::MissingAuthProfile {
auth_profile_id: auth_profile_id.as_str().to_owned(),
})?;
let mut secrets = BTreeMap::new();
let used_at = now_string().map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?;
for secret_id in auth_profile.config.secret_ids() {
let secret = registry
.get_secret(workspace_id, secret_id)
.await
.map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?
.ok_or_else(|| RuntimeError::MissingSecret {
secret_id: secret_id.as_str().to_owned(),
})?;
let version = registry
.get_current_secret_version(workspace_id, secret_id)
.await
.map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?
.ok_or_else(|| RuntimeError::MissingSecretVersion {
secret_id: secret_id.as_str().to_owned(),
version: secret.secret.current_version,
})?;
let plaintext = secret_crypto.decrypt(&version.secret_version.ciphertext)?;
registry
.touch_secret(workspace_id, secret_id, &used_at)
.await
.map_err(|error| RuntimeError::SecretCrypto {
details: error.to_string(),
})?;
secrets.insert(secret_id.clone(), plaintext);
}
ResolvedAuth::from_profile(&auth_profile, &secrets).map(Some)
}
fn default_invitation_expiry() -> Result<String, ApiError> {
OffsetDateTime::now_utc()
.checked_add(time::Duration::days(7))