`1403a6a`

Persist TodoWrite resumes

Authored by mfwolffe <wolffemf@dukes.jmu.edu> 2 weeks ago

SHA: 1403a6a1b81c3f2952cc48c1b44ba639e894d896
Parents: fa5d5a8
Tree: d76c856

2 changed files

Status	File	+	-
M	`src/loader/runtime/tool_batches.py`	3	3
M	`tests/test_tool_batches.py`	38	25

src/loader/runtime/tool_batches.pymodified

                          "Perform the mutation now instead of spending another turn on "
                          "planning, rereads, or verification."
+                     )
 -                    self.context.queue_ephemeral_steering_message(concrete_message)
 +                    self.context.queue_steering_message(concrete_message)
                      return
 -                self.context.queue_ephemeral_steering_message(
 +                self.context.queue_steering_message(
                      "Todo tracking is updated. Continue with the next pending item: "
                      f"`{next_pending}`. Use the current output files as the source of "
                      "truth, and do not reopen reference materials unless one specific "
              if next_pending
              else ""
+         )
 -        self.context.queue_ephemeral_steering_message(
 +        self.context.queue_steering_message(
              "Todo tracking is updated. A declared output artifact is still missing."
              + next_pending_suffix
              + _missing_artifact_resume_suffix(

tests/test_tool_batches.pymodified

          verify_action=verify_action,
          auto_recover=False,
+     )
 -    queued_messages: list[str] = []
 -    context.queue_steering_message_callback = queued_messages.append
 +    persistent_messages: list[str] = []
 +    ephemeral_messages: list[str] = []
 +    context.queue_steering_message_callback = persistent_messages.append
 +    context.queue_ephemeral_steering_message_callback = ephemeral_messages.append
      runner = ToolBatchRunner(context, DefinitionOfDoneStore(temp_dir))
      dod = create_definition_of_done("Create a multi-file nginx guide.")
      dod.implementation_plan = str(implementation_plan)
+     )
      runner._queue_duplicate_observation_nudge(tool_call, dod=dod)  # type: ignore[attr-defined]
 -    assert queued_messages
 -    message = queued_messages[-1]
 +    assert persistent_messages
 +    message = persistent_messages[-1]
      assert "06-ssl-configuration.html" in message
      assert "Do not switch into review or consistency-check mode" in message
      assert (
          verify_action=verify_action,
          auto_recover=False,
+     )
 -    queued_messages: list[str] = []
 -    context.queue_steering_message_callback = queued_messages.append
 +    persistent_messages: list[str] = []
 +    ephemeral_messages: list[str] = []
 +    context.queue_steering_message_callback = persistent_messages.append
 +    context.queue_ephemeral_steering_message_callback = ephemeral_messages.append
      runner = ToolBatchRunner(context, DefinitionOfDoneStore(temp_dir))
      dod = create_definition_of_done("Create a multi-file nginx guide.")
      dod.implementation_plan = str(implementation_plan)
      assert any(
          "All explicitly planned artifacts now exist." in message
 -        for message in queued_messages
 +        for message in persistent_messages
+     )
      assert any(
          "Ensure all files are properly linked and formatted consistently" in message
 -        for message in queued_messages
 +        for message in persistent_messages
+     )
      assert any(
          "Move to verification once no specific mismatch remains." in message
 -        for message in queued_messages
 +        for message in persistent_messages
+     )
          verify_action=verify_action,
          auto_recover=False,
+     )
 -    queued_messages: list[str] = []
 -    context.queue_steering_message_callback = queued_messages.append
 +    persistent_messages: list[str] = []
 +    ephemeral_messages: list[str] = []
 +    context.queue_steering_message_callback = persistent_messages.append
 +    context.queue_ephemeral_steering_message_callback = ephemeral_messages.append
      runner = ToolBatchRunner(context, DefinitionOfDoneStore(temp_dir))
      dod = create_definition_of_done("Create a multi-file nginx guide.")
      dod.implementation_plan = str(implementation_plan)
          consecutive_errors=0,
+     )
 -    assert queued_messages
 -    message = queued_messages[-1]
 +    assert persistent_messages
 +    message = persistent_messages[-1]
      assert "Next step: create `01-getting-started.html`." in message
      assert (
          f"Prefer one `write(file_path=..., content=...)` call for `{chapter_one.resolve(strict=False)}` now."
          verify_action=verify_action,
          auto_recover=False,
+     )
 -    queued_messages: list[str] = []
 -    context.queue_steering_message_callback = queued_messages.append
 +    persistent_messages: list[str] = []
 +    ephemeral_messages: list[str] = []
 +    context.queue_steering_message_callback = persistent_messages.append
 +    context.queue_ephemeral_steering_message_callback = ephemeral_messages.append
      runner = ToolBatchRunner(context, DefinitionOfDoneStore(temp_dir))
      dod = create_definition_of_done("Create a thorough nginx guide.")
      dod.implementation_plan = str(implementation_plan)
      assert any(
          "Resume by creating `06-performance-tuning.html` now." in message
 -        for message in queued_messages
 +        for message in ephemeral_messages
+     )
      assert not any(
          "All explicitly planned artifacts now exist." in message
 -        for message in queued_messages
 +        for message in ephemeral_messages
+     )
          verify_action=verify_action,
          auto_recover=False,
+     )
 -    queued_messages: list[str] = []
 -    context.queue_steering_message_callback = queued_messages.append
 +    persistent_messages: list[str] = []
 +    ephemeral_messages: list[str] = []
 +    context.queue_steering_message_callback = persistent_messages.append
 +    context.queue_ephemeral_steering_message_callback = ephemeral_messages.append
      runner = ToolBatchRunner(context, DefinitionOfDoneStore(temp_dir))
      dod = create_definition_of_done("Create a thorough nginx guide.")
      dod.implementation_plan = str(implementation_plan)
          consecutive_errors=0,
+     )
 -    assert queued_messages
 -    message = queued_messages[-1]
 +    assert ephemeral_messages
 +    message = ephemeral_messages[-1]
      assert "Resume by creating `05-advanced-features.html` now." in message
      assert "No TodoWrite, no verification, no rereads until that artifact exists." in message
      assert "refresh `TodoWrite`" not in message
          verify_action=verify_action,
          auto_recover=False,
+     )
 -    queued_messages: list[str] = []
 -    context.queue_steering_message_callback = queued_messages.append
 +    persistent_messages: list[str] = []
 +    ephemeral_messages: list[str] = []
 +    context.queue_steering_message_callback = persistent_messages.append
 +    context.queue_ephemeral_steering_message_callback = ephemeral_messages.append
      runner = ToolBatchRunner(context, DefinitionOfDoneStore(temp_dir))
      dod = create_definition_of_done("Create a multi-file nginx guide.")
      dod.implementation_plan = str(implementation_plan)
          consecutive_errors=0,
+     )
 -    assert queued_messages
 -    message = queued_messages[-1]
 +    assert persistent_messages
 +    message = persistent_messages[-1]
      assert "Todo tracking is updated. A declared output artifact is still missing." in message
      assert "Resume by creating `02-installation.html` now." in message
      assert "refresh `TodoWrite`" in message
      assert "Do not spend the next turn on TodoWrite alone" in message
 +    assert ephemeral_messages == []
  @pytest.mark.asyncio