feat(worktree): resume into a preserved worktree on re-invocation

[pro-vi]

Problem

When gnhf preserves a worktree on exit (any run that lands a commit), subsequent invocations with the same prompt fail:

gnhf: Command failed: git worktree add -b 'gnhf/...' '.../gnhf-worktrees/...'
Preparing worktree (new branch 'gnhf/...')
fatal: a branch named 'gnhf/...' already exists

The branch is still registered against the preserved worktree, so git worktree add -b cannot re-create it. The user's only options today are to merge the branch and remove the worktree, or delete the branch manually. There's no way to resume and keep iterating on the same preserved worktree.

Change

Detect the "worktree already preserved" case and resume the existing run instead of re-creating.

• New worktreeExists(baseCwd, worktreePath) in src/core/git.ts reads git worktree list --porcelain and returns true when the path is already registered.
• initializeWorktreeRun in src/cli.ts checks for the preserved worktree plus the matching .gnhf/runs/<runId>/ directory. When both are present it calls resumeRun rather than createWorktree + setupRun, and returns a resumed: true flag.
• The call site reads startIteration from the preserved run via getLastIterationNumber and skips installing the worktreeCleanup handler on the resume path, so Ctrl-C on a resumed invocation cannot take down a worktree that already carries earlier iterations' commits.

The non-worktree resume path (existing behavior when running gnhf from an already-checked-out gnhf/* branch) is untouched.

Example

$ cat prompt.md | gnhf --worktree --agent claude
... run 1 lands 3 commits ...
  gnhf: worktree preserved at /path/gnhf-worktrees/xyz

$ cat prompt.md | gnhf --worktree --agent claude
  gnhf: resuming preserved worktree at /path/gnhf-worktrees/xyz
  gnhf: continuing run xyz from iteration 3
... run 2 appends commits on the same branch ...

Test plan

• npm run typecheck
• npm run lint
• npm run format:check
• npm test (354/354 passing, including new e2e)
• npm run test:e2e (7/7 passing)

New e2e test runs --worktree twice with the same prompt and asserts: the second invocation does not mention "already exists", logs "resuming preserved worktree", lands an additional commit on the same branch, and continues iteration numbering (gnhf #2:).

Compatibility

• First invocation behavior unchanged (branch does not exist yet, falls through to current createWorktree + setupRun path).
• Second invocation with a different prompt still spawns a fresh worktree under a different slug, same as before.
• If the worktree directory exists on disk but is not registered as a git worktree (e.g. removed from git but left on disk), worktreeExists returns false and the current path runs, matching existing behavior as closely as possible.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9a130f76d2

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[pro-vi]

Thanks for the review. Both items addressed in e3625a7:

P1 (path normalization): worktreeExists now calls path.resolve on both the git worktree list --porcelain output and the target path before comparing, so Windows slash/backslash mismatches no longer miss a real match. Added unit tests for the match, non-match, normalize (trailing slash + .. segment), and command-failure paths.

P2 (branch verification): initializeWorktreeRun now reads the preserved worktree's current branch before resuming and throws a clear error if it is not gnhf/<runId>. The message points at the exact git -C / git worktree remove commands the user can run to resolve. Added an e2e test that switches the preserved worktree to a sideways branch between invocations and asserts the second run fails with the branch name in the error.

All 359 tests + lint + typecheck + prettier remain clean.

[kunchenguid]

thanks for the solid work here, especially the P1/P2 follow-ups. reviewed the diff and ran everything locally - all green (363 tests, lint, typecheck, format:check).

ready to merge once the conflict with main is sorted. it's a trivial one: #77 added closeSync / createReadStream / createWriteStream to the same fs import block in src/cli.ts where this PR adds existsSync. a merge or rebase against main should resolve it cleanly.