Claude Code¶

Run Claude Code inside an isolated Git worktree, with every session captured as a reviewable attempt.

Why wrap Claude Code with ait¶

Claude Code is fast and capable, but a single prompt can edit many files across your repository. ait keeps that work contained:

The agent edits an isolated worktree instead of your root checkout.
Every Claude Code run becomes one attempt — prompt, files, status, output, and commits stay linked.
You can inspect, recover, rebase, or apply an attempt with normal Git concepts.
A repo-local memory layer surfaces what previous Claude runs already tried.

Setup¶

ait init                       # detects `claude` on PATH, auto-installs the wrapper + hook
ait adapter doctor claude-code # optional sanity check

ait init detects every supported agent CLI on $PATH and wires it up in one shot — for Claude Code that means installing a repo-local claude wrapper plus merging the Claude Code hook config into .claude/settings.json. To re-run setup explicitly (e.g. after upgrading Claude Code), use ait adapter setup claude-code.

Run Claude Code under ait¶

Use Claude Code exactly as you normally would:

claude -p --permission-mode bypassPermissions \
  "Shorten the README and improve the quickstart"

Or wrap the call explicitly:

ait run --adapter claude-code --intent "Refactor query parser" -- claude

Set explicit intent and commit text via environment variables:

AIT_INTENT="Update README" \
AIT_COMMIT_MESSAGE="update README with Claude" \
claude -p --permission-mode bypassPermissions \
  "Shorten the README and improve the quickstart"

Review and apply¶

ait status
ait attempt list
ait attempt show <attempt-id>

Apply when the diff is good:

ait apply <attempt-id> --mode current

Discard when it is not:

ait attempt discard <attempt-id>

Use Claude Code as an adversarial reviewer¶

light review mode is a local deterministic risk scan. It does not call Claude Code. Use adversarial mode when you want Claude Code to review a finished attempt:

ait review attempt latest-reviewable --mode adversarial --review-adapter claude-code

The built-in claude-code review adapter resolves to the local claude -p CLI. AIT sends the structured reviewer brief on stdin, runs the reviewer outside the target attempt worktree, captures stdout/stderr, and expects a structured JSON findings object.

AIT removes ANTHROPIC_API_KEY from this child process environment. This keeps the reviewer path aligned with local Claude Code authentication and prevents a silent fallback to provider API credits. If the local claude binary is not available or not logged in, the review fails closed instead of switching to an API-key path.

Check local auth status:

ait adapter doctor claude-code --json

Expected local CLI mode reports will_use_api_key: false and will_fallback_to_credits: false.