devflow-engine 1.2.1__tar.gz → 1.2.3__tar.gz

{devflow_engine-1.2.1 → devflow_engine-1.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: devflow-engine
-Version: 1.2.1
+Version: 1.2.3
 Summary: DevFlow v2 CLI-first engine (projects, stories, execution store, review packets)
 Author: DevFlow
 License: Proprietary

{devflow_engine-1.2.1 → devflow_engine-1.2.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "devflow-engine"
-version = "1.2.1"
+version = "1.2.3"
 description = "DevFlow v2 CLI-first engine (projects, stories, execution store, review packets)"
 readme = "README.md"
 requires-python = ">=3.11"

{devflow_engine-1.2.1 → devflow_engine-1.2.3}/src/devflow_engine/errors/error_solver_dag.py

@@ -151,6 +151,34 @@ def _proof_candidate(verification: dict[str, Any] | None) -> dict[str, Any]:
     return {}
 
 
+def _observation_fix_proof(context: dict[str, Any]) -> dict[str, Any]:
+    build_observability = context.get("BuildObservability")
+    if not isinstance(build_observability, dict):
+        return {}
+    output = build_observability.get("output")
+    if not isinstance(output, dict):
+        return {}
+    observation = output.get("observation")
+    if not isinstance(observation, dict):
+        return {}
+    proof = observation.get("fix_proof")
+    return proof if isinstance(proof, dict) else {}
+
+
+def _verification_with_observed_fix_proof(
+    verification: dict[str, Any] | None,
+    context: dict[str, Any],
+) -> dict[str, Any] | None:
+    if _proof_candidate(verification):
+        return verification
+    observed_proof = _observation_fix_proof(context)
+    if not observed_proof:
+        return verification
+    merged = dict(verification or {})
+    merged["fix_proof"] = observed_proof
+    return merged
+
+
 def _resolve_artifact_path(repo_root: Path, artifact_path: Any) -> Path | None:
     raw = str(artifact_path or "").strip()
     if not raw:
@@ -216,7 +244,23 @@ def _validate_ui_fix_proof(
     return True, normalized, None
 
 
-def _validate_ui_observability_repro(observation: dict[str, Any]) -> tuple[bool, str | None]:
+def _artifact_under_observability_sandbox(artifact: str, *, error_task_id: str) -> bool:
+    normalized = artifact.replace("\\", "/").strip()
+    allowed_prefixes = (
+        f".devflow/observability/{error_task_id}/",
+        ".devflow/observability/<id>/",
+    )
+    if any(normalized.startswith(prefix) for prefix in allowed_prefixes):
+        return True
+    marker = f"/.devflow/observability/{error_task_id}/"
+    return marker in normalized
+
+
+def _validate_ui_observability_repro(
+    observation: dict[str, Any],
+    *,
+    error_task_id: str,
+) -> tuple[bool, str | None]:
     artifact = str(observation.get("repro_artifact_path") or "").strip()
     command = str(observation.get("repro_run_command") or "").strip()
     output = str(observation.get("repro_output") or "").strip()
@@ -228,6 +272,8 @@ def _validate_ui_observability_repro(observation: dict[str, Any]) -> tuple[bool,
         return False, "ui_observability_missing_playwright_repro_artifact"
     if "playwright" not in command.lower():
         return False, "ui_observability_requires_playwright_repro_command"
+    if not _artifact_under_observability_sandbox(artifact, error_task_id=error_task_id):
+        return False, "ui_observability_repro_artifact_outside_sandbox"
     if not output:
         return False, "ui_observability_missing_repro_output"
     if not isinstance(oracles, list) or not any(str(o or "").strip() for o in oracles):
@@ -848,7 +894,10 @@ class BuildObservabilityNode(Node):
                     else {}
                 )
                 if observed_flag and _ui_proof_required({"runtimeContext": user_report_context_for_ui}):
-                    ui_repro_ok, ui_repro_failure_reason = _validate_ui_observability_repro(observation)
+                    ui_repro_ok, ui_repro_failure_reason = _validate_ui_observability_repro(
+                        observation,
+                        error_task_id=error_task_id,
+                    )
 
                 # Split deferred questions via LLM classifier (v1.1.22).
                 # An injectable hook is used when present so tests can
@@ -1138,6 +1187,10 @@ class ResolveNode(Node):
         error_task_id = task_context.event.error_task_id
 
         verification = task_context.metadata.get("verification")
+        effective_verification = _verification_with_observed_fix_proof(
+            verification if isinstance(verification, dict) else None,
+            context,
+        )
         # Red-before-fix invariant: ResolveNode refuses to mark resolved
         # unless BuildObservability persisted observable=True at intake.
         # Closes the "already-green-at-intake" leak — a green-gate True with
@@ -1147,14 +1200,14 @@ class ResolveNode(Node):
         if isinstance(bo, dict):
             bo_out = bo.get("output") if isinstance(bo.get("output"), dict) else {}
             observation_was_red = bool(bo_out.get("observable"))
-        green_gate = bool((verification or {}).get("green_gate"))
+        green_gate = bool((effective_verification or {}).get("green_gate"))
         proof_verified = True
         fix_proof: dict[str, Any] | None = None
         proof_failure_reason: str | None = None
         if _ui_proof_required(context):
             proof_verified, fix_proof, proof_failure_reason = _validate_ui_fix_proof(
                 repo_root=persistor.repo_root,
-                verification=verification if isinstance(verification, dict) else None,
+                verification=effective_verification if isinstance(effective_verification, dict) else None,
             )
         verified = green_gate and observation_was_red and proof_verified
         runtime_context = context.get("runtimeContext") if isinstance(context.get("runtimeContext"), dict) else None
@@ -1165,13 +1218,13 @@ class ResolveNode(Node):
             fingerprint = runtime_context.get("fingerprint") or current_record.get("fingerprint")
         title = context.get("metadata", {}).get("title") if isinstance(context.get("metadata"), dict) else None
 
-        resolve_output: dict[str, Any] = {"verification": verification, "verified": verified}
+        resolve_output: dict[str, Any] = {"verification": effective_verification, "verified": verified}
         if proof_failure_reason:
             resolve_output["proof_failure_reason"] = proof_failure_reason
         if fix_proof:
             resolve_output["fix_proof"] = fix_proof
         persistor.persist(context=context, node_name="Resolve", output=resolve_output)
-        self.save_output(ResolveNode.OutputType(verified=verified, verification=verification))
+        self.save_output(ResolveNode.OutputType(verified=verified, verification=effective_verification))
 
         if verified:
             if fix_proof:
@@ -1185,7 +1238,7 @@ class ResolveNode(Node):
                 persistor.repo_root,
                 error_task_id=error_task_id,
                 title=str(title) if title is not None else None,
-                verification=verification,
+                verification=effective_verification if isinstance(effective_verification, dict) else None,
             )
             task_context.metadata["commit_result"] = commit_result
             # v1.1.19: persist resolution_commit_sha + portal_token so the
@@ -1228,7 +1281,7 @@ class ResolveNode(Node):
             outcome=str(task_context.metadata.get("outcome") or "in_progress"),
             title=str(title) if title is not None else None,
             fingerprint=str(fingerprint) if fingerprint is not None else None,
-            verification=verification,
+            verification=effective_verification if isinstance(effective_verification, dict) else None,
             reason=str(task_context.metadata.get("reason")) if task_context.metadata.get("reason") else None,
             context_path=persistor.context_path,
             runtime_context=runtime_context,

{devflow_engine-1.2.1 → devflow_engine-1.2.3}/tests/area8_errors/test_us85_ui_playwright_proof_gate.py

@@ -43,8 +43,10 @@ def _observed_ui_repro() -> dict:
         "deferred_questions": [],
         "confidence": "high",
         "observed": True,
-        "repro_artifact_path": "admin-frontend/tests/industry-dropdown.spec.ts",
-        "repro_run_command": "npx playwright test admin-frontend/tests/industry-dropdown.spec.ts --project=chromium",
+        "repro_artifact_path": ".devflow/observability/<id>/industry-dropdown.spec.ts",
+        "repro_run_command": (
+            "npx playwright test .devflow/observability/<id>/industry-dropdown.spec.ts --project=chromium"
+        ),
         "repro_output": "expect(locator).toBeVisible failed before fix",
         "repro_oracles": [
             "Clicking the Industry / sector combobox should reveal the options list.",
@@ -93,6 +95,46 @@ def test_ui_user_report_does_not_reach_fix_without_playwright_observability_repr
     assert "Create or extend a Playwright test" in stored["required_next_action"]
 
 
+def test_ui_user_report_does_not_reach_fix_with_repro_artifact_outside_sandbox(tmp_path: Path) -> None:
+    repo_root = tmp_path / "repo"
+    repo_root.mkdir()
+    _init_repo(repo_root)
+    store = ExecutionStore(repo_root / ".devflow" / "execution.sqlite")
+    etid = _seed_ui_report(store, external_id="ui-product-path-repro")
+    first_pass_called = {"n": 0}
+
+    observation = _observed_ui_repro()
+    observation["repro_artifact_path"] = "admin-frontend/tests/industry-dropdown.spec.ts"
+    observation["repro_run_command"] = (
+        "npx playwright test admin-frontend/tests/industry-dropdown.spec.ts --project=chromium"
+    )
+
+    def fix_fn(_ctx):
+        first_pass_called["n"] += 1
+        return {"fixed": True, "verification": {"green_gate": True}}
+
+    result = store.run_error_solver_dag(
+        error_task_id=etid,
+        repo_root=repo_root,
+        build_observability_fn=lambda _ctx: {"observable": True, "observation": observation},
+        first_pass_fix_fn=fix_fn,
+        iterative_fix_fn=lambda _ctx: {"fixed": False},
+        needs_user_input_fn=lambda _ctx: False,
+    )
+
+    assert first_pass_called["n"] == 0
+    assert result["outcome"] == "observation_inconclusive"
+    with store._connect() as conn:
+        row = conn.execute(
+            "SELECT status, observability_json FROM error_tasks WHERE error_task_id=?",
+            (etid,),
+        ).fetchone()
+    assert row["status"] == "observation_inconclusive"
+    stored = json.loads(row["observability_json"])
+    assert stored["observability_error"] == "ui_observability_repro_artifact_outside_sandbox"
+    assert stored["observed"] is False
+
+
 def test_ui_user_report_does_not_resolve_without_playwright_video_proof(tmp_path: Path) -> None:
     repo_root = tmp_path / "repo"
     repo_root.mkdir()
@@ -135,6 +177,66 @@ def test_ui_user_report_does_not_resolve_without_playwright_video_proof(tmp_path
     assert "fix_proof" not in json.loads(row["observability_json"])
 
 
+def test_ui_user_report_resolves_with_observed_playwright_video_proof_on_fix_replay(tmp_path: Path) -> None:
+    repo_root = tmp_path / "repo"
+    repo_root.mkdir()
+    _init_repo(repo_root)
+    store = ExecutionStore(repo_root / ".devflow" / "execution.sqlite")
+    etid = _seed_ui_report(store, external_id="ui-observed-proof-replay")
+    video_path = Path(".devflow") / "observability" / etid / "proof" / "video.webm"
+    artifact = repo_root / video_path
+    artifact.parent.mkdir(parents=True, exist_ok=True)
+    artifact.write_bytes(b"webm proof")
+    observation = _observed_ui_repro()
+    observation["fix_proof"] = {
+        "kind": "playwright_video",
+        "tool": "playwright",
+        "artifact_path": str(video_path),
+        "run_command": "npx playwright test admin-frontend/tests/industry-dropdown.spec.ts --project=chromium",
+        "oracles": [
+            "The Industry / sector combobox opens the options list.",
+            "Selecting Healthcare leaves Healthcare visible in the form.",
+        ],
+        "result": "passed",
+    }
+    with store._connect() as conn:
+        conn.execute(
+            "UPDATE error_tasks SET status='observed', observability_json=? WHERE error_task_id=?",
+            (json.dumps(observation), etid),
+        )
+
+    def fix_fn(_ctx):
+        (repo_root / "fix.ts").write_text("// fix\n", encoding="utf-8")
+        return {
+            "fixed": True,
+            "verification": {
+                "green_gate": True,
+                "repro_run_command": "npx playwright test admin-frontend/tests/industry-dropdown.spec.ts --project=chromium",
+                "verification_output": "1 passed",
+            },
+        }
+
+    result = store.run_error_solver_dag(
+        error_task_id=etid,
+        repo_root=repo_root,
+        build_observability_fn=lambda _ctx: {"observable": False},
+        first_pass_fix_fn=fix_fn,
+        iterative_fix_fn=lambda _ctx: {"fixed": False},
+        needs_user_input_fn=lambda _ctx: False,
+        worker_role="fix",
+    )
+
+    assert result["outcome"] == "resolved"
+    row = store.get_resolved_with_portal(etid)
+    assert row is not None
+    assert row["status"] == "resolved"
+    assert row["resolution_commit_sha"]
+    proof = row["observability"]["fix_proof"]
+    assert proof["kind"] == "playwright_video"
+    assert proof["artifact_path"] == str(video_path)
+    assert proof["resolved_artifact_path"] == str(artifact.resolve())
+
+
 def test_ui_user_report_resolves_and_persists_playwright_video_proof(tmp_path: Path) -> None:
     repo_root = tmp_path / "repo"
     repo_root.mkdir()

{devflow_engine-1.2.1 → devflow_engine-1.2.3}/uv.lock

@@ -172,7 +172,7 @@ wheels = [
 
 [[package]]
 name = "devflow-engine"
-version = "1.2.1"
+version = "1.2.3"
 source = { editable = "." }
 dependencies = [
     { name = "cryptography" },