Skip to content

v1.58.1

Choose a tag to compare

@github-actions github-actions released this 10 Nov 08:22
· 300 commits to main since this release

Codex gets a reasoning effort per message

This release includes everything from v1.58.0, which was never published.

Model per message now covers Codex. In a Codex terminal opened from the sidebar, the router picks the reasoning effort for each message you type: low for a quick question, medium or high for building, high or xhigh for design and hard bugs. It never changes the model, because Codex's encrypted reasoning does not carry over from one model to another. The pane header shows each pick, for example "Codex · medium → low · 6 ms".

  • Termpolis starts Codex with -c openai_base_url=… pointing at its local proxy. The proxy changes only the request's effort and forwards it to the endpoint Codex would have used anyway: chatgpt.com when you sign in with ChatGPT, api.openai.com with an API key. Nothing is compressed.

  • The proxy turns down Codex's WebSocket, so Codex falls back to HTTP. Streaming still works.

  • Refusals: if OpenAI refuses an effort, the request is sent again exactly as Codex asked. That effort is then not tried again in that terminal (xhigh falls back to high). After three refused efforts, Termpolis stops routing that terminal.

  • It steps aside and launches Codex as it is when you route Codex yourself:

    • the command uses --oss or --profile, or names its own endpoint or provider;
    • the top level of config.toml names its own endpoint, provider or profile;
    • OPENAI_BASE_URL or a proxy variable is set;
    • the command sets model_reasoning_effort (unless the mode is "Every Claude and Codex terminal");
    • the compression proxy is off.

    Termpolis reads only config.toml, never Codex's sign-in. Codex's own subagent requests are left alone.

  • The proxy keeps port 47853 when it is free, so open terminals keep working after the proxy restarts.

  • Choose the mode in Settings → General → Default Models → Model per message: "Claude terminals started on Auto, and Codex" (the default), "Every Claude and Codex terminal", or "Off".

Fixes in 1.58.1

  • Switching a routed Codex terminal to another account works again. Termpolis now strips its own openai_base_url argument when it relaunches. A model change from the pane header keeps the terminal routed. An account switch does not, because that account may pick its own endpoint: open a new Codex terminal from the sidebar to route it again.
  • The gitWatcher real-filesystem test no longer flakes on late directory echoes (macOS FSEvents, Windows last-access updates).

Docs and privacy

PRIVACY.md, the documentation, the README and the Settings hint describe the Codex route. The proxy section no longer says that Codex never goes through it.

Full Changelog: v1.57.0...v1.58.1