`0544f0e`

Hand off repair support writes

Authored by mfwolffe <wolffemf@dukes.jmu.edu> 1 week ago

SHA: 0544f0e86a285ea4f924d37c1080c7700da4c2c5
Parents: 8b1407b
Tree: a713aaf

2 changed files

Status	File	+	-
M	`src/loader/runtime/tool_batches.py`	39	0
M	`tests/test_tool_batches.py`	81	0

src/loader/runtime/tool_batches.pymodified

                      dod=dod,
+                 )
              else:
 +                self._queue_active_repair_success_handoff_nudge(tool_call)
                  self._queue_post_mutation_self_audit_nudge(tool_call, dod=dod)
                  self._queue_completed_artifact_observation_handoff_nudge(
                      tool_call,
              f"and repair `{repair.artifact_path}` before widening the change set."
+         )
 +    def _queue_active_repair_success_handoff_nudge(self, tool_call: ToolCall) -> None:
 +        """After a repair mutation, hand back to finalization instead of ad hoc edits."""
++
 +        if tool_call.name not in {"write", "edit", "patch"}:
 +            return
 +        raw_path = str(tool_call.arguments.get("file_path", "")).strip()
 +        if not raw_path:
 +            return
++
 +        repair = extract_active_repair_context(self.context.session.messages)
 +        if repair is None or not repair.allowed_paths:
 +            return
++
 +        try:
 +            changed_path = str(Path(raw_path).expanduser().resolve(strict=False))
 +        except (OSError, RuntimeError, ValueError):
 +            changed_path = str(Path(raw_path).expanduser())
 +        allowed_paths = {
 +            str(Path(path).expanduser().resolve(strict=False))
 +            for path in repair.allowed_paths
 +        }
 +        if changed_path not in allowed_paths:
 +            return
++
 +        if changed_path == str(Path(repair.artifact_path).expanduser().resolve(strict=False)):
 +            self.context.queue_steering_message(
 +                "The active verification repair target was updated. "
 +                "Do not keep auditing or retarget nearby links by guesswork. "
 +                "Finish with a final response now so Loader can re-run verification."
 +            )
 +            return
++
 +        self.context.queue_steering_message(
 +            "The support file for the active verification repair now exists. "
 +            f"Do not retarget `{repair.artifact_path}` to a different missing path by guesswork. "
 +            "Finish with a final response now so Loader can re-run verification."
 +        )
++
      def _queue_blocked_late_reference_drift_nudge(
          self,
          event_content: str,

tests/test_tool_batches.pymodified

      assert "before widening the change set" in queued[0]
 +@pytest.mark.asyncio
 +async def test_tool_batch_runner_hands_off_after_active_repair_support_file_write(
 +    temp_dir: Path,
 +) -> None:
 +    async def assess_confidence(
 +        tool_name: str,
 +        tool_args: dict,
 +        context: str,
 +    ) -> ConfidenceAssessment:
 +        raise AssertionError("Confidence scoring should be disabled in this scenario")
++
 +    async def verify_action(
 +        tool_name: str,
 +        tool_args: dict,
 +        result: str,
 +        expected: str = "",
 +    ) -> ActionVerification:
 +        raise AssertionError("Verification should not run in this scenario")
++
 +    repair_target = temp_dir / "guide" / "index.html"
 +    stylesheet = temp_dir / "guide" / "style.css"
 +    repair_target.parent.mkdir(parents=True)
 +    repair_target.write_text('<link rel="stylesheet" href="style.css">\n')
 +    context = build_context(
 +        temp_dir=temp_dir,
 +        messages=[
 +            Message(
 +                role=Role.ASSISTANT,
 +                content=(
 +                    "Repair focus:\n"
 +                    f"- Fix the broken local reference `style.css` in `{repair_target}`.\n"
 +                    f"- Immediate next step: edit `{repair_target}`.\n"
 +                    f"- If the broken reference should remain, create `{stylesheet}`; otherwise remove or replace `style.css`.\n"
 +                ),
 +            )
 +        ],
 +        safeguards=FakeSafeguards(),
 +        assess_confidence=assess_confidence,
 +        verify_action=verify_action,
 +    )
 +    queued: list[str] = []
 +    context.queue_steering_message_callback = queued.append
 +    runner = ToolBatchRunner(context, DefinitionOfDoneStore(temp_dir))
 +    dod = create_definition_of_done("Repair a guide stylesheet link.")
 +    tool_call = ToolCall(
 +        id="write-style",
 +        name="write",
 +        arguments={
 +            "file_path": str(stylesheet),
 +            "content": "body { font-family: sans-serif; }\n",
 +        },
 +    )
++
 +    await runner.execute_batch(
 +        tool_calls=[tool_call],
 +        tool_source="assistant",
 +        pending_tool_calls_seen=set(),
 +        emit=_noop_emit,
 +        summary=TurnSummary(final_response=""),
 +        dod=dod,
 +        executor=FakeExecutor(
 +            [
 +                tool_outcome(
 +                    tool_call=tool_call,
 +                    output=f"Successfully wrote {stylesheet}",
 +                    is_error=False,
 +                )
 +            ]
 +        ),  # type: ignore[arg-type]
 +        on_confirmation=None,
 +        on_user_question=None,
 +        emit_confirmation=None,
 +        consecutive_errors=0,
 +    )
++
 +    assert queued
 +    assert any("support file for the active verification repair now exists" in message for message in queued)
 +    assert any("Do not retarget" in message for message in queued)
 +    assert any("Loader can re-run verification" in message for message in queued)
++
++
  def test_tool_batch_runner_blocked_late_reference_drift_nudge_points_to_missing_artifact(
      temp_dir: Path,
  ) -> None: