[rafaeelricco]

Motivation

Make commit generation faster. Every message and split plan waits on the main model at its configured effort, even though writing a commit message is simple summarization that a smaller model handles well.

This branch sits on top of #37, so the PR also shows #37's two commits until #37 merges. Only the last commit, "Add a fast model for commit messages and split plans", belongs to this PR.

What's New

Fast Model Config

• A new optional fast_model config field (a model id from the same provider and sign-in as ai). Older config files without it load as Nothing and behave exactly as before; commit model keeps it and commit setup resets it.
• When it is set, Commit.create swaps it in with withModel + withMinEffort, so the plain message, Regenerate, Adjust and split plans (including split-plan Regenerate in Split) all run on it at the provider's lowest effort. commit branch keeps the main model.

commit fast-model Command

• Picks a model from the provider's list, or "Use the main model" to turn it off; cancelling saves nothing. It saves the sign-in that resolveProvider just refreshed, so running it never writes stale OAuth tokens back.
• Listed in commit --help, the README and the dispatcher in index.ts; commit doctor shows it after the main model (claude-sonnet-5 (medium effort), fast: claude-haiku-4-5).

Anthropic Reasoning Retry

• Claude Haiku 4.5 rejects adaptive thinking and effort, which the Anthropic path always sent. A 400 whose message mentions either now gets one retry without both, on the API-key and setup-token paths alike, and the commit note reports "provider default" effort. Other 400s are not retried. This mirrors the retry-without-effort that the OpenAI and xAI paths already have.

Model Selection Flow

flowchart TB
    A["Commit.create()"] --> B["resolveProvider(config)"]
    B --> C{"fast_model set?"}
    C -->|no| D["main ai, configured effort"]
    C -->|yes| E["withModel + withMinEffort"]
    D --> F["messages, Adjust, split plans"]
    E --> F
    F --> G{"Anthropic rejects thinking/effort?"}
    G -->|no| H["reply"]
    G -->|yes| I["retry once without them"]
    I --> H

Loading

Changed Files

File	Change Type	Summary
src/domain/config/config.ts	Modified	Optional fast_model field.
src/cli/commit.ts	Modified	Commit.create swaps in the fast model at lowest effort.
src/cli/fast-model.ts	Added	commit fast-model pick / clear command.
src/cli/parser.ts, index.ts	Modified	fast-model command, help text, dispatch, header notes.
src/cli/doctor.ts	Modified	Shows the fast model next to the main one.
src/cli/setup.ts	Modified	New configs start with no fast model.
src/infra/llm/anthropic.ts	Modified	One retry without thinking/effort when the model rejects them.
src/domain/llm/auth-resolver.ts	Modified	Header lists FastModel among the refreshing commands.
README.md	Modified	commit fast-model usage and table row.
test/**	Modified/Added	Config default, fast model reaching commit and split calls, command cases, Anthropic retry; fixtures gain fast_model: Nothing().

Testing & Feedback

Check Commit.create in src/cli/commit.ts: every router call in commit and split should see the fast config, and commit branch should not. In src/infra/llm/anthropic.ts, check that the retry only fires on 400s that mention effort or thinking. The new commit, split and Anthropic-retry tests fail on the pre-change code.

Not verified: the exact wording Anthropic returns when Haiku 4.5 gets effort or adaptive thinking. To try it for real:

• Run commit fast-model and pick claude-haiku-4-5, then run commit on a staged change. The commit note should show claude-haiku-4-5 with provider default effort.
• Stage two or more files with split_commits on and confirm the split plan also comes from the fast model.
• Run commit fast-model again, choose "Use the main model", and confirm commit doctor no longer shows a fast model.

If you find any bugs or have recommendations for improvements, please open an issue and assign it to me.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-27T16:23:57.662373Z	1739ec9	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[rafaeelricco]

Closing: keeping only the prompt updates, which are in #37. The fast-model config isn't wanted.