Ctrl+T "steer now" → 400 "Instructions are required" on OpenAI Responses models

Summary

A now-mode steer aborts the run with a 400 from the model API when the active model uses the OpenAI Responses API. The same steer with mode="queue" works fine. It looks like the mid-turn steer history-processor appends a ModelRequest without instructions, so the Responses request goes out with empty instructions.

Environment

• code-puppy 0.0.550
• pydantic-ai 1.56.0
• Python 3.14.4 (macOS, arm64)
• Model served via the OpenAI Responses API (model name gpt-5.5; endpoint returns {'detail': 'Instructions are required'}, i.e. it requires a non-empty instructions).

Steps to reproduce

1. Configure a model that uses the OpenAI Responses API.
2. Give the agent a multi-step task so it's mid-run.
3. While it's running, Ctrl+T and submit a steering message in "steer now" mode (equivalently: get_message_bus().provide_response(SteerAgentCommand(text="…", mode="now"))).
4. The next model call fails.

Expected

The steer is injected as a user turn and the run continues (as it does for mode="queue").

Actual

🎯 Injecting steer mid-turn — model will see: 'Actually make it a unit converter …'
Unexpected error: status_code: 400, model_name: gpt-5.5, body: {'detail': 'Instructions are required'}
...
  File ".../pydantic_ai/models/openai.py", line 1386, in request_stream
    response = await self._responses_create(...)
  File ".../pydantic_ai/models/openai.py", line 1681, in _responses_create
    raise ModelHTTPError(status_code=status_code, model_name=self.model_name, body=e.body) from e
pydantic_ai.exceptions.ModelHTTPError: status_code: 400, model_name: gpt-5.5, body: {'detail': 'Instructions are required'}

The run aborts; the session is auto-saved and returns to the prompt.

Likely root cause

code_puppy/agents/_steer_processor.py → make_steer_history_processor injects each now-mode steer by appending a bare ModelRequest to the end of the message list:

injected.append(ModelRequest(parts=[UserPromptPart(content=steer_text)]))
new_messages = list(messages) + injected
return new_messages

That ModelRequest carries no instructions. On the OpenAI Responses path, pydantic-ai derives the request's instructions from the (final) ModelRequest, so the call goes out with empty instructions and a Responses endpoint that requires them rejects it with 400.

mode="queue" doesn't hit this because it's drained between turns and replayed via a normal follow-up run, which carries the agent's instructions.

Suggested fix (one of)

• Carry the agent's instructions onto the injected request, e.g. ModelRequest(parts=[UserPromptPart(content=steer_text)], instructions=<agent instructions>), or
• Append the steer as a UserPromptPart to the existing final ModelRequest instead of adding a new bare one (so the original request's instructions is preserved).

Workaround

Use mode="queue" for steering on Responses-API models.

Context

Found while prototyping a remote-control plugin (idea discussed in #395). Happy to open a PR for the fix if that's welcome.

Start in code_puppy/agents/_steer_processor.py at make_steer_history_processor and trace the now-mode path reached by SteerAgentCommand. Compare the injected ModelRequest with the existing message history and the pydantic-ai OpenAI Responses request shown in the traceback. Done means a now-mode steer continues successfully on Responses API models while queue mode remains working.