Merge branch 'staging' into codex/fix-daily-user-spend-alert-dedupe

Author: kyle-compute

backend/api/services/agent_capabilities.py

@@ -468,6 +468,7 @@ def is_stale(
     current_hash: str,
     schema_version: int,
     task_version_id: str,
+    experiment_id: str | None = None,
 ) -> bool:
     """Whether a stored block must be regenerated. See ``stale_reason``."""
     return (
@@ -476,6 +477,7 @@ def is_stale(
             current_hash=current_hash,
             schema_version=schema_version,
             task_version_id=task_version_id,
+            experiment_id=experiment_id,
         )
         is not None
     )

backend/tests/test_agent_capabilities_endpoint.py

@@ -46,6 +46,38 @@ def test_not_stale_when_all_match():
     )
 
 
+def test_is_stale_when_experiment_changed():
+    block_meta = {
+        "cohort_hash": "aaa",
+        "schema_version": 1,
+        "task_version_id": "v1",
+        "experiment_id": "exp-1",
+    }
+    assert is_stale(
+        block_meta,
+        current_hash="aaa",
+        schema_version=1,
+        task_version_id="v1",
+        experiment_id="exp-2",
+    )
+
+
+def test_not_stale_when_experiment_matches():
+    block_meta = {
+        "cohort_hash": "aaa",
+        "schema_version": 1,
+        "task_version_id": "v1",
+        "experiment_id": "exp-1",
+    }
+    assert not is_stale(
+        block_meta,
+        current_hash="aaa",
+        schema_version=1,
+        task_version_id="v1",
+        experiment_id="exp-1",
+    )
+
+
 def test_missing_metadata_is_stale():
     assert (
         is_stale(None, current_hash="aaa", schema_version=1, task_version_id="v1")

docs/swe-marathon-eval-runbook.md

@@ -79,6 +79,7 @@ oddish run -p <task-dir> --agent <agent> --model <provider/model> \
   API host and disables server-side web tools for restricted agent phases.
   Optional overrides still exist (`--allow-agent-host`, `--disable-web-tools`)
   if you need an extra host or want to force the web-tool disable.
+- DO NOT RUN QA. DO NOT USE THE QA FLAG.
 
 ## 3. Load-bearing gotchas (each cost real debugging time)