Add bounded final-answer acceptance gate in regular runner

tests/test_bounded_execution.py

@@ -110,3 +110,58 @@ def test_villani_task_reports_completed_when_done(tmp_path: Path) -> None:
 
     assert result["execution"]["completed"] is True
     assert result["execution"]["terminated_reason"] == "completed"
+
+
+def test_completed_budget_does_not_bypass_finalization_gate(tmp_path: Path) -> None:
+    runner = _runner(
+        tmp_path,
+        [
+            {"role": "assistant", "content": [{"type": "text", "text": "I'll update x.py to fix this."}]},
+            {"role": "assistant", "content": [{"type": "text", "text": "done"}]},
+        ],
+    )
+    budget = ExecutionBudget(max_turns=20, max_tool_calls=20, max_seconds=100.0, max_no_edit_turns=20, max_reconsecutive_recon_turns=20)
+
+    result = runner.run("implement fix", execution_budget=budget)
+
+    nudges = [
+        block.get("text", "")
+        for msg in result["messages"]
+        if msg.get("role") == "user"
+        for block in msg.get("content", [])
+        if isinstance(block, dict) and block.get("type") == "text"
+    ]
+    assert any("no file was modified" in text for text in nudges)
+    assert result["execution"]["terminated_reason"] == "completed"
+
+
+def test_failed_bash_verification_uses_exit_code_for_followup_gate(tmp_path: Path, monkeypatch) -> None:
+    monkeypatch.setattr("villani_code.state_runtime.run_verification", lambda _runner, _trigger="edit": "<verification>status=pass</verification>")
+    runner = _runner(
+        tmp_path,
+        [
+            {"role": "assistant", "content": [{"type": "tool_use", "id": "1", "name": "Bash", "input": {"command": "python -c \"import sys; sys.exit(1)\""}}]},
+            {"role": "assistant", "content": [{"type": "text", "text": "done"}]},
+            {"role": "assistant", "content": [{"type": "text", "text": "done"}]},
+        ],
+    )
+    monkeypatch.setattr(
+        runner,
+        "_execute_tool_with_policy",
+        lambda tool_name, _tool_input, _tool_use_id, _msg_index: {
+            "content": "simulated bash run",
+            "is_error": False,
+            "exit_code": 1 if tool_name == "Bash" else 0,
+        },
+    )
+    budget = ExecutionBudget(max_turns=20, max_tool_calls=20, max_seconds=100.0, max_no_edit_turns=20, max_reconsecutive_recon_turns=20)
+
+    result = runner.run("implement fix", execution_budget=budget)
+    nudges = [
+        block.get("text", "")
+        for msg in result["messages"]
+        if msg.get("role") == "user"
+        for block in msg.get("content", [])
+        if isinstance(block, dict) and block.get("type") == "text"
+    ]
+    assert any("last verification failed" in text for text in nudges)

tests/test_final_answer_gate.py

@@ -0,0 +1,171 @@
+from villani_code.state import final_answer_block_reason
+
+
+def test_concrete_edit_claim_without_edits_blocks_once() -> None:
+    nudge = {}
+    reason = final_answer_block_reason(
+        instruction="implement fix",
+        final_text="I'll update x.py to handle this case.",
+        has_any_edit=False,
+        changed_files=[],
+        known_verification_command=True,
+        verification_ran_after_last_edit=True,
+        last_verification_failed=False,
+        had_corrective_action_after_last_failure=False,
+        nudge_state=nudge,
+    )
+    assert reason is not None
+    reason2 = final_answer_block_reason(
+        instruction="implement fix",
+        final_text="I'll update x.py to handle this case.",
+        has_any_edit=False,
+        changed_files=[],
+        known_verification_command=True,
+        verification_ran_after_last_edit=True,
+        last_verification_failed=False,
+        had_corrective_action_after_last_failure=False,
+        nudge_state=nudge,
+    )
+    assert reason2 is None
+
+
+def test_vague_suggestion_does_not_block() -> None:
+    reason = final_answer_block_reason(
+        instruction="implement fix",
+        final_text="One possible fix would be to update parser.py.",
+        has_any_edit=False,
+        changed_files=[],
+        known_verification_command=True,
+        verification_ran_after_last_edit=True,
+        last_verification_failed=False,
+        had_corrective_action_after_last_failure=False,
+        nudge_state={},
+    )
+    assert reason is None
+
+
+def test_new_concrete_edit_phrases_detected() -> None:
+    phrases = [
+        "The fix is straightforward: register it.",
+        "Found it. The markdown_handler is imported but never added to the REGISTRY dictionary. The fix is straightforward: register it.",
+        "The fix is simple: add it to the registry.",
+        "Need to register it.",
+        "Register it in the registry.",
+        "Add it to the registry.",
+        "Wire it into the registry.",
+        "Hook it up in the dispatch table.",
+        "Include it in the mapping.",
+        "Fix it by registering the handler.",
+    ]
+    for phrase in phrases:
+        reason = final_answer_block_reason(
+            instruction="implement fix",
+            final_text=phrase,
+            has_any_edit=False,
+            changed_files=[],
+            known_verification_command=True,
+            verification_ran_after_last_edit=True,
+            last_verification_failed=False,
+            had_corrective_action_after_last_failure=False,
+            nudge_state={},
+        )
+        assert reason is not None
+
+
+def test_vague_add_suggestions_still_do_not_trigger() -> None:
+    for phrase in [
+        "One possible fix would be to register it.",
+        "You could add it to the registry.",
+        "I recommend registering it.",
+        "It might need to be added.",
+        "Maybe wire it into the registry.",
+        "This looks like a registry issue.",
+        "No code changes are needed.",
+    ]:
+        reason = final_answer_block_reason(
+            instruction="implement fix",
+            final_text=phrase,
+            has_any_edit=False,
+            changed_files=[],
+            known_verification_command=True,
+            verification_ran_after_last_edit=True,
+            last_verification_failed=False,
+            had_corrective_action_after_last_failure=False,
+            nudge_state={},
+        )
+        assert reason is None
+
+
+def test_plan_only_task_does_not_block() -> None:
+    reason = final_answer_block_reason(
+        instruction="Plan only, do not implement.",
+        final_text="I'll update parser.py",
+        has_any_edit=False,
+        changed_files=[],
+        known_verification_command=True,
+        verification_ran_after_last_edit=False,
+        last_verification_failed=True,
+        had_corrective_action_after_last_failure=False,
+        nudge_state={},
+    )
+    assert reason is None
+
+
+def test_changed_files_without_verification_blocks_when_command_known() -> None:
+    reason = final_answer_block_reason(
+        instruction="implement fix",
+        final_text="Done",
+        has_any_edit=True,
+        changed_files=["src/a.py"],
+        known_verification_command=True,
+        verification_ran_after_last_edit=False,
+        last_verification_failed=False,
+        had_corrective_action_after_last_failure=False,
+        nudge_state={},
+    )
+    assert "have not verified" in str(reason)
+
+
+def test_changed_files_without_known_verification_does_not_block() -> None:
+    reason = final_answer_block_reason(
+        instruction="implement fix",
+        final_text="Done",
+        has_any_edit=True,
+        changed_files=["src/a.py"],
+        known_verification_command=False,
+        verification_ran_after_last_edit=False,
+        last_verification_failed=False,
+        had_corrective_action_after_last_failure=False,
+        nudge_state={},
+    )
+    assert reason is None
+
+
+def test_failed_verification_requires_followup_action() -> None:
+    reason = final_answer_block_reason(
+        instruction="implement fix",
+        final_text="Done",
+        has_any_edit=True,
+        changed_files=["src/a.py"],
+        known_verification_command=True,
+        verification_ran_after_last_edit=True,
+        last_verification_failed=True,
+        had_corrective_action_after_last_failure=False,
+        nudge_state={},
+    )
+    assert "last verification failed" in str(reason)
+
+
+def test_failed_verification_with_followup_does_not_block() -> None:
+    reason = final_answer_block_reason(
+        instruction="implement fix",
+        final_text="Done",
+        has_any_edit=True,
+        changed_files=["src/a.py"],
+        known_verification_command=True,
+        verification_ran_after_last_edit=True,
+        last_verification_failed=True,
+        had_corrective_action_after_last_failure=True,
+        nudge_state={},
+    )
+    assert reason is None

tests/test_state_runtime.py

@@ -199,7 +199,9 @@ def test_repeated_stale_verification_returns_compact_block(tmp_path: Path, monke
     _seed_repo(tmp_path)
     runner = Runner(client=DummyClient(), repo=tmp_path, model="m", stream=False, small_model=True)
     runner._verification_baseline_changed = set()
-    monkeypatch.setattr(state_runtime, "git_changed_files", lambda _repo: [])
+    (tmp_path / "tests").mkdir(exist_ok=True)
+    (tmp_path / "tests" / "test_x.py").write_text("def test_x():\n    assert False\n", encoding="utf-8")
+    monkeypatch.setattr(state_runtime, "git_changed_files", lambda _repo: ["tests/test_x.py"])
 
     first = runner._run_verification("edit")
     second = runner._run_verification("edit")
@@ -238,6 +240,37 @@ def test_verification_detail_event_keeps_raw_trace(tmp_path: Path, monkeypatch:
     assert "last_validation_summary:" in out
 
 
+def test_failed_verification_summary_includes_visible_failure_lines(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    from types import SimpleNamespace
+    from villani_code.autonomy import VerificationStatus
+
+    _seed_repo(tmp_path)
+    runner = Runner(client=DummyClient(), repo=tmp_path, model="m", stream=False, small_model=True)
+    runner._verification_baseline_changed = set()
+    (tmp_path / "tests").mkdir(exist_ok=True)
+    (tmp_path / "tests" / "test_x.py").write_text("def test_x():\n    assert False\n", encoding="utf-8")
+    monkeypatch.setattr(state_runtime, "git_changed_files", lambda _repo: ["tests/test_x.py"])
+
+    def fake_run(_cmd, **_kwargs):
+        return SimpleNamespace(
+            returncode=1,
+            stdout="E       AssertionError: expected 2, got 3\nFile \"src/a.py\", line 10\nassert x == y",
+            stderr="",
+        )
+
+    monkeypatch.setattr(state_runtime.subprocess, "run", fake_run)
+    runner._verification_engine.verify = lambda *args, **kwargs: SimpleNamespace(
+        status=VerificationStatus.FAIL,
+        confidence_score=0.1,
+        findings=[],
+        summary="failed",
+    )
+    out = runner._run_verification("edit")
+    assert "AssertionError" in out
+    assert "expected 2, got 3" in out
+    assert "Treat visible test assertions as authoritative evidence." in out
+
+
 def test_repeated_validation_updates_compact_state_without_repeated_prose(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ) -> None: