`4e907a5`

Persist first child handoffs after recovery

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

SHA: 4e907a503608e52c8e1123a6cc1a7315ac1d843f
Parents: 6bf8210
Tree: c193a13

3 changed files

Status	File	+	-
M	`src/loader/runtime/tool_batches.py`	31	1
M	`tests/test_tool_batches.py`	4	4
M	`tests/test_turn_completion.py`	90	0

src/loader/runtime/tool_batches.pymodified

              project_root=self.context.project_root,
+         )
          session_messages = list(getattr(self.context.session, "messages", []) or [])
 -        if use_persistent_handoff and _recent_recovery_prompt(session_messages):
 +        if (
 +            use_persistent_handoff
 +            and _recent_recovery_prompt(session_messages)
 +            and not _should_preserve_first_child_handoff_after_recovery(
 +                tool_call=tool_call,
 +                resume_target=resume_target,
 +                dod=dod,
 +                project_root=self.context.project_root,
 +            )
 +        ):
              use_persistent_handoff = False
          queue_message = (
              self.context.queue_steering_message
      ) == 0
 +def _should_preserve_first_child_handoff_after_recovery(
 +    *,
 +    tool_call: ToolCall,
 +    resume_target: Path | None,
 +    dod: DefinitionOfDone,
 +    project_root: Path,
 +) -> bool:
 +    if resume_target is None or not resume_target.suffix or _is_summary_artifact_path(resume_target):
 +        return False
 +    raw_target = str(tool_call.arguments.get("file_path", "")).strip()
 +    if not raw_target:
 +        return False
 +    written_target = Path(raw_target).expanduser().resolve(strict=False)
 +    if not _is_summary_artifact_path(written_target):
 +        return False
 +    return _confirmed_substantive_file_artifact_count(
 +        dod,
 +        project_root=project_root,
 +    ) == 0
++
++
  def _is_summary_artifact_path(path: str | Path) -> bool:
      return Path(path).name.lower() in _SUMMARY_ARTIFACT_NAMES

tests/test_tool_batches.pymodified

  @pytest.mark.asyncio
 -async def test_tool_batch_runner_softens_first_file_handoff_after_recovery_prompt(
 +async def test_tool_batch_runner_preserves_first_file_handoff_after_recovery_prompt(
      temp_dir: Path,
  ) -> None:
      async def assess_confidence(
          consecutive_errors=0,
+     )
 -    assert persistent_messages == []
 -    assert ephemeral_messages
 -    message = ephemeral_messages[-1]
 +    assert persistent_messages
 +    assert ephemeral_messages == []
 +    message = persistent_messages[-1]
      assert "Next step: create `01-introduction.html`." in message
      assert "Write a compact but real initial version of that file now" not in message

tests/test_turn_completion.pymodified

      assert "write a compact but real initial version of that file now" in agent.session.messages[-1].content.lower()
 +@pytest.mark.asyncio
 +async def test_turn_completion_interrupts_first_chapter_narration_from_declared_index_graph(
 +    temp_dir: Path,
 +) -> None:
 +    backend = ScriptedBackend()
 +    config = non_streaming_config()
 +    config.reasoning.completion_check = False
 +    agent = Agent(
 +        backend=backend,
 +        config=config,
 +        project_root=temp_dir,
 +    )
 +    runtime = ConversationRuntime(agent)
 +    events = []
++
 +    async def capture(event) -> None:
 +        events.append(event)
++
 +    prepared = await runtime.turn_preparation.prepare(
 +        task=(
 +            "Create a multi-file nginx guide under ~/Loader/guides/nginx "
 +            "with an index and chapter files."
 +        ),
 +        emit=capture,
 +        requested_mode="execute",
 +        original_task=None,
 +        on_user_question=None,
 +    )
 +    await runtime.phase_tracker.enter(
 +        TurnPhase.ASSISTANT,
 +        capture,
 +        detail="Requesting assistant response",
 +        reason_code="request_assistant_response",
 +    )
++
 +    guide_root = temp_dir / "Loader" / "guides" / "nginx"
 +    chapters_dir = guide_root / "chapters"
 +    chapters_dir.mkdir(parents=True)
 +    index_path = guide_root / "index.html"
 +    index_path.write_text(
 +        "\n".join(
 +            [
 +                "<!DOCTYPE html>",
 +                '<a href="chapters/01-introduction.html">Chapter 1: Introduction to Nginx</a>',
 +                '<a href="chapters/02-installation.html">Chapter 2: Installation and Setup</a>',
 +                "",
 +            ]
 +        )
 +    )
++
 +    implementation_plan = temp_dir / "implementation.md"
 +    implementation_plan.write_text(
 +        "# Implementation Plan\n\n"
 +        "## File Changes\n\n"
 +        f"- `{index_path}`\n"
 +        f"- `{chapters_dir}/`\n"
 +    )
++
 +    prepared.definition_of_done.implementation_plan = str(implementation_plan)
 +    prepared.definition_of_done.touched_files.append(str(index_path))
 +    prepared.definition_of_done.mutating_actions.append("write")
 +    prepared.definition_of_done.pending_items.append(
 +        "Develop the nginx guide content following the same structure and cadence as the fortran guide"
 +    )
++
 +    content = "Now I'll create the first chapter of the nginx guide."
 +    decision = await runtime.turn_completion.handle_text_response(
 +        content=content,
 +        response_content=content,
 +        task=prepared.task,
 +        effective_task=prepared.effective_task,
 +        iterations=1,
 +        max_iterations=agent.config.max_iterations,
 +        actions_taken=[],
 +        continuation_count=0,
 +        dod=prepared.definition_of_done,
 +        emit=capture,
 +        summary=prepared.summary,
 +        executor=prepared.executor,
 +        rollback_plan=prepared.rollback_plan,
 +    )
++
 +    assert decision.action == TurnCompletionAction.CONTINUE
 +    assert decision.continuation_count == 1
 +    assert prepared.summary.completion_decision_code == "in_progress_transition_continue"
 +    assert agent.session.messages[-1].role.value == "user"
 +    assert agent.session.messages[-1].content.startswith("[CONTINUE CURRENT STEP]")
 +    assert "01-introduction.html" in agent.session.messages[-1].content
++
++
  @pytest.mark.asyncio
  async def test_turn_completion_handles_fake_tool_narration_without_reroute(
      temp_dir: Path,