Codex ACP startup fails when acp_model triggers session/set_model

[morganmcg1]

Summary

Agent Canvas can save Codex ACP settings with acp_model set, for example gpt-5.5/medium. When a user submits a prompt, the conversation starts and then fails with a UI error because the backend calls session/set_model on @zed-industries/codex-acp@0.16.0, which returns JSON-RPC Method not found.

What I observed

Versions:

• @openhands/agent-canvas: 1.0.0
• OpenHands SDK/server: 1.28.1
• @zed-industries/codex-acp: 0.16.0

Saved settings included:

{
  "agent_kind": "acp",
  "acp_server": "codex",
  "acp_model": "gpt-5.5/medium"
}

Server log on submit:

ACP server initialized: agent_name='codex-acp', agent_version='0.16.0'
Authenticating with ACP method: chatgpt
ERROR Failed to start ACP server: Method not found
ERROR Error during conversation run

A direct ACP smoke test showed the failing call is session/set_model: both gpt-5.5/medium and gpt-5.5 return Method not found. new_session and set_config_option(configId="reasoning_effort") work.

Relevant code path

• openhands/sdk/agent/acp_agent.py::_start_acp_server: after conn.new_session(...), it calls _maybe_set_session_model(...) when self.acp_model is set.
• openhands/sdk/agent/acp_agent.py::_maybe_set_session_model: for known providers with supports_set_session_model, it calls conn.set_session_model(model_id=acp_model, session_id=session_id).
• acp/client/connection.py::set_session_model: sends the ACP method mapped as session/set_model.
• codex-acp@0.16.0 exposes model and reasoning as session config options (model, reasoning_effort) but does not implement session/set_model.

Workaround / fix

Clearing acp_model and passing Codex config directly avoids the failing ACP method:

{
  "acp_args": [
    "-c",
    "model=\"gpt-5.5\"",
    "-c",
    "model_reasoning_effort=\"medium\""
  ],
  "acp_model": null
}

With that config, codex-acp starts successfully and reports:

model gpt-5.5
reasoning_effort medium

Suggested product fix: for Codex ACP, do not persist combined values like gpt-5.5/medium into acp_model or call session/set_model. Split model and reasoning, then apply them through Codex config args or ACP set_config_option for model and reasoning_effort.

[jamiechicago312]

Great issue for new contributors! 🎉

Why this is a good first issue:
The report already narrows the bug down to a specific ACP/Codex flow, includes a concrete failing value (gpt-5.5/medium), and points at the relevant behavior difference between Codex and the other ACP providers. That makes it a bounded bugfix with a clear before/after outcome instead of an open-ended design task.

Instructions for tackling this issue:
Start by tracing how Codex ACP settings are saved and then turned into a conversation payload. The most relevant places are src/routes/agent-settings.tsx, src/constants/acp-providers.ts, and src/api/agent-server-adapter.ts. There is already useful regression coverage nearby in __tests__/routes/agent-settings.test.tsx, __tests__/api/agent-server-adapter.test.ts, and the onboarding ACP tests. Keep the fix narrow and Codex-specific: avoid regressing Claude/Gemini/custom ACP behavior, and update the tests so the saved/payload model handling matches the new Codex-safe behavior.

A strong PR here will prove two things: the UI no longer persists a combined Codex value in a way that triggers session/set_model, and starting a Codex conversation still preserves the intended model/reasoning choice.

Getting started:
You can work on this issue without asking for it to be assigned to you. Feel free to open a pull request directly!

Need help?
Join our Slack community: https://go.openhands.dev/slack

This comment was created by an AI agent (OpenHands) on behalf of the user.

[neubig]

I think this is fixed on main. Pending release