cache: scope response entries by operation version

TASKS.md

@@ -62,6 +62,7 @@ Progress:
   - `ExecutionConfig.response_cache` is now part of the public operation model, and `admin-api` validates the safe Community baseline: only explicit `REST GET` operations without `auth_profile_ref`
   - cache boundary and namespace rules are now documented for `platform / coordination cache` versus `response cache`
   - runtime executor now performs actual response cache reads/writes for eligible `REST GET` operations with keys isolated by `workspace + agent + operation + request fingerprint`
+  - response cache keys now also include `operation version`, so a new published version does not reuse stale cached payloads from the previous one
   - mcp-server published tool catalog now uses shared coordination snapshots across instances, while preserving the same `workspace + agent` isolation model
 - pending:
   - preserve the same isolation model while extending response cache beyond the first `REST GET` hot path

crates/crank-runtime/src/executor.rs

@@ -789,10 +789,11 @@ fn response_cache_key(
     let fingerprint_hash = URL_SAFE_NO_PAD.encode(Sha256::digest(fingerprint_bytes));
 
     Some(format!(
-        "crank:response:workspace:{}:agent:{}:operation:{}:request:{}",
+        "crank:response:workspace:{}:agent:{}:operation:{}:version:{}:request:{}",
         scope.workspace_key,
         scope.agent_key,
         operation.operation_id.as_str(),
+        operation.operation_version,
         fingerprint_hash
     ))
 }
@@ -1062,6 +1063,32 @@ mod tests {
         assert_eq!(request_count.load(Ordering::SeqCst), 2);
     }
 
+    #[tokio::test]
+    async fn invalidates_rest_get_cache_on_operation_version_change() {
+        let request_count = Arc::new(AtomicUsize::new(0));
+        let base_url = spawn_cached_rest_server(Arc::clone(&request_count)).await;
+        let executor = RuntimeExecutor::new()
+            .with_response_cache_store(Arc::new(InMemoryResponseCacheStore::default()));
+        let operation_v1 = test_cached_rest_get_operation(&base_url);
+        let mut operation_v2 = operation_v1.clone();
+        let context = RuntimeRequestContext::from_request_id("req_cache_version")
+            .with_response_cache_scope("ws_cache", "agent_sales");
+
+        let first_output = executor
+            .execute_with_context(&operation_v1, &json!({ "city": "msk" }), Some(&context))
+            .await
+            .unwrap();
+        operation_v2.operation_version = 2;
+        let second_output = executor
+            .execute_with_context(&operation_v2, &json!({ "city": "msk" }), Some(&context))
+            .await
+            .unwrap();
+
+        assert_eq!(first_output, json!({ "city": "msk", "request_count": 1 }));
+        assert_eq!(second_output, json!({ "city": "msk", "request_count": 2 }));
+        assert_eq!(request_count.load(Ordering::SeqCst), 2);
+    }
+
     #[tokio::test]
     async fn emits_runtime_tracing_with_request_context() {
         let base_url = spawn_runtime_server().await;

crates/crank-runtime/src/model.rs

@@ -9,6 +9,7 @@ use serde_json::Value;
 #[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
 pub struct RuntimeOperation {
     pub operation_id: OperationId,
+    pub operation_version: u32,
     pub tool_name: String,
     pub protocol: Protocol,
     pub target: Target,
@@ -24,6 +25,7 @@ impl From<Operation<Schema, MappingSet>> for RuntimeOperation {
     fn from(value: Operation<Schema, MappingSet>) -> Self {
         Self {
             operation_id: value.id,
+            operation_version: value.version,
             tool_name: value.name,
             protocol: value.protocol,
             target: value.target,

docs/implementation-plan.md

@@ -195,7 +195,7 @@
 - Enterprise может подключать свой `Valkey/Redis` на уровне single-node или cluster deployment.
 - cache namespaces документированы так, чтобы:
   - platform / coordination state не смешивался с response cache;
-  - response cache по умолчанию изолировался по `workspace + agent + operation + request fingerprint`;
+  - response cache по умолчанию изолировался по `workspace + agent + operation + operation version + request fingerprint`;
   - внешний cache backend можно было безопасно шарить между несколькими рабочими областями и агентами.
 - shared coordination cache уже применяется не только для rate limiting, но и для multi-instance snapshots published MCP catalogs.
 

docs/mcp-interface.md

@@ -241,7 +241,7 @@ Crank должен корректно работать, если MCP client де
   - `REST`
   - `GET`
   - без `auth_profile_ref`
-  - cache keys изолируются минимум по `workspace + agent + operation + request fingerprint`
+  - cache keys изолируются минимум по `workspace + agent + operation + operation version + request fingerprint`
 - streaming operations публикуются как bounded tools или tool families;
 - reload published tools без пересборки сервиса;
 - никакой draft-логики или admin CRUD в MCP слое.

docs/runtime-config.md

@@ -195,7 +195,7 @@ Demo/deployment:
 
 Стартовая модель namespace для response cache:
 
-- `workspace + agent + operation + request fingerprint`
+- `workspace + agent + operation + operation version + request fingerprint`
 
 Стартовая модель namespace для platform / coordination cache: