Expose test failure evidence to retries

runtime_loop/autonomous_task_loop.py

@@ -92,6 +92,7 @@ class AutonomousTaskLoop:
 
             test_run = self.tests.run(task.project, task, Path(worktree.worktree_path), test_command)
             if test_run.status != "PASS":
+                self._attach_test_failure_evidence(attempt, test_run)
                 self.bus.publish(EventType.TEST_FAILED, project=task.project, task=task, payload={"test_run_id": str(test_run.id)})
 
             tools.git(["add", "-N", "."])
@@ -217,6 +218,23 @@ class AutonomousTaskLoop:
             "write_file_fallback_rate": 0 if write_file_operations == 0 else totals["write_file_fallbacks"] / write_file_operations,
         }
 
+    def _attach_test_failure_evidence(self, attempt: TaskAttempt, test_run) -> None:
+        content = test_run.output_artifact.content if test_run.output_artifact else {}
+        stdout = str(content.get("stdout", "")) if isinstance(content, dict) else ""
+        stderr = str(content.get("stderr", "")) if isinstance(content, dict) else ""
+        evidence = {
+            "test_run_id": str(test_run.id),
+            "status": test_run.status,
+            "stdout_excerpt": stdout[-12000:],
+            "stderr_excerpt": stderr[-4000:],
+        }
+        coder_result = dict(attempt.coder_result or {})
+        metadata = dict(coder_result.get("metadata", {})) if isinstance(coder_result.get("metadata", {}), dict) else {}
+        metadata["test_failure_evidence"] = evidence
+        coder_result["metadata"] = metadata
+        attempt.coder_result = coder_result
+        attempt.save(update_fields=["coder_result", "updated_at"])
+
     def _scrub_context(self, context: dict[str, object]) -> dict[str, object]:
         scrubbed = dict(context)
         scrubbed.pop("secrets", None)

tests/test_m2_autonomous_loop.py

@@ -6,7 +6,8 @@ from pathlib import Path
 from agents.providers import DeterministicCodingProvider
 from control_plane.agents.management.commands.seed_core_agents import Command as SeedAgentsCommand
 from control_plane.events.models import Event, EventType
-from control_plane.projects.models import CommitRecord, Milestone, Project, ProjectPlan, Task, TaskStatus
+from control_plane.agents.models import AgentRole, AgentVersion
+from control_plane.projects.models import Artifact, CommitRecord, Milestone, Project, ProjectPlan, Task, TaskStatus
 from control_plane.verification.models import Review, TestRun as ArtifexTestRun, Verification, VerificationResult
 from model_router.router import ModelRouter
 from runtime_loop.autonomous_task_loop import AutonomousTaskLoop
@@ -128,3 +129,27 @@ def test_m2_negative_path_rejected_without_commit(tmp_path: Path) -> None:
     assert Review.objects.filter(task=task).exclude(status="PASS").exists()
     assert task.attempts.filter(status="FAILED").exists()
     assert Event.objects.filter(task=task, event_type=EventType.TASK_FAILED).exists()
+
+
+def test_failed_test_output_is_attached_to_attempt_retry_context(tmp_path: Path) -> None:
+    repo = create_disposable_django_repo(tmp_path)
+    task = create_task(repo, "Add a feature that fails tests", ["tests pass"])
+    SeedAgentsCommand().handle()
+    coder = AgentVersion.objects.get(agent__role=AgentRole.CODER, promotion_status="CHAMPION")
+    attempt = task.attempts.create(task=task, attempt_number=1, coder=coder, status="RUNNING", coder_result={"metadata": {}})
+    artifact = Artifact.objects.create(
+        project=task.project,
+        task=task,
+        artifact_type="test_report",
+        name="failed-tests",
+        content={"stdout": "short traceback", "stderr": "error details", "returncode": 1},
+    )
+    test_run = ArtifexTestRun.objects.create(project=task.project, task=task, command="pytest", status="FAIL", output_artifact=artifact)
+
+    loop()._attach_test_failure_evidence(attempt, test_run)
+
+    attempt.refresh_from_db()
+    evidence = attempt.coder_result["metadata"]["test_failure_evidence"]
+    assert evidence["status"] == "FAIL"
+    assert evidence["stdout_excerpt"] == "short traceback"
+    assert evidence["stderr_excerpt"] == "error details"