git-repo

Git Repo

Git repository management and SourceGit GUI client integration.

Topics

Topic Dependencies

worktree (entry point — inventory + decision)
  └─→ rename-worktree (reuse registered worktree)
  └─→ move-worktree (register unregistered or relocate)

worktree-register (shared metadata register/relink mechanism)
  ├─← fix-worktree (recover a broken bare-worktree link)
  └─← to-bare (link the working tree after a regular→bare conversion)

to-bare  ←inverse→  to-ghq (formerly `migrate`)

ssh-key (SSH multi-account)  ──counterpart──  credential-helper (HTTPS multi-account)

Worktree decision tree (HARD STOP — every time a worktree is needed)

Whenever a worktree is needed (plan/build/test without switching branches, PR verification, isolated work environment, new commit while main branch has another in-flight task), check for inactive worktree reuse before creating a new one.

Flow

1. git -C worktree list to enumerate existing worktrees
2. Operation-state gate (HARD STOP — before any inactive classification): for each candidate worktree ``, check for an in-progress git operation:

   gitdir=$(git -C  rev-parse --absolute-git-dir)
   ls "$gitdir"/CHERRY_PICK_HEAD "$gitdir"/MERGE_HEAD "$gitdir"/REBASE_HEAD "$gitdir"/BISECT_LOG "$gitdir"/rebase-merge "$gitdir"/rebase-apply 2>/dev/null
   git -C  status --porcelain | grep -E '^(DD|AU|UD|UA|DU|AA|UU)'   # unmerged index entries
   

   Any hit = a cherry-pick/merge/rebase/bisect is mid-flight (likely the user's active operation) → the worktree is NOT an inactive candidate, its dirty files are the operation's payload (never offer discard / git add resolution / stash), and reuse is forbidden — report to the user instead. Merge-status heuristics (branch merged + ahead=0) do NOT override this gate. (see failed-attempts.md "cherry-pick in progress misclassified as abandoned")
3. Identify inactive candidates (only among worktrees that passed the gate):
   • Merged-PR worktrees (commit hash equals the base branch's merge commit)
   • Stale fix/refactor branch worktrees (confirm with the user)
4. If an inactive worktree exists → reuse via the rename-worktree topic (rename the directory + metadata, switch branch)
5. If no inactive worktree exists or the user opts for new → git worktree add

New-commit-start trigger (HARD STOP — paired with git.md)

When committing new work on a main/master/develop branch, if git status shows uncommitted changes from another task, worktree splitting is mandatory.

Entry conditions (any one)

• Current branch = main/master/develop
• git status shows 1+ changed files outside the new work (e.g., another task's unstaged files)
• Unpushed commits include another task's work

Decision tree on entry

git status (check for other changes)
  ├─ 0 other changes, branch = PR/feature → commit in place
  ├─ 0 other changes, branch = main/master/develop
  │    → AskUserQuestion: (a) create PR branch from current (b) create PR branch in a new worktree
  ├─ 1+ other changes, branch = main/master/develop
  │    → worktree split required. Leave other changes in place, isolate new work in a new worktree
  └─ Another task running on a PR branch with stray changes on main
       → Report to the user (confirm intent of stray changes) → decide worktree split + stray cleanup

Worktree split procedure

1. Leave the new work's diff in place — do NOT stash (preserve the working directory)
2. git -C worktree list to inspect inactive worktrees (per the decision tree above)
3. Inactive candidate present → reuse via rename-worktree; otherwise git -C worktree add .claude/worktrees/ (or the worktree path defined by the active environment context — see worktree.md "Default Path Rules")
4. cd into the new worktree, then git checkout -b (or use the existing inactive branch)
5. Re-apply the same diff in the new worktree via Edit (the original main changes stay on main)
6. add → commit → push → PR inside the new worktree

Don't / Do

Don't / Do

Recommended placement rule (HARD STOP — prevent 2nd recurrence)

When building AskUserQuestion options for worktree selection:

The Recommended marker must attach to the reuse option. "Safer to default to new because purpose is unclear" is an autonomous judgment — instead, document the "verify current commit/branch before reuse" procedure in the rename option description so the user can decide.

Self-check (before any worktree work)

1. Did you call git -C worktree list?
2. Did you run the operation-state gate (Flow step 2) on every candidate? A worktree with CHERRY_PICK_HEAD/MERGE_HEAD/REBASE_HEAD or unmerged status codes (DU/UU/AA…) is mid-operation → excluded from candidates, no discard/stash options for its files
3. Does any gated worktree match (a) the same commit as HEAD or (b) a merge commit hash? → inactive candidate
4. If inactive candidates exist, include both the new-add option and the reuse option in AskUserQuestion
5. Is the Recommended marker attached to the reuse option? (apply the "Recommended placement rule" table)
6. Use new-add only when 0 inactive candidates exist OR the user explicitly chose new

Typical failure mode

When a worktree is needed for plan verification, defaulting to "create new + remove later" wastes resources. git worktree list often shows merged-PR worktrees stuck on the merge commit hash — these are reuse candidates. Always inventory first before creating a new worktree.

Quick Reference

ghq Clone (automatic SourceGit registration)

When ghq get is executed, the following happens automatically:

1. Clone the repository
2. Register in SourceGit (under the appropriate group)
3. Auto-create the group if it doesn't exist

Proceeds automatically without user confirmation

Detailed guide

SourceGit Management

Directly edit the SourceGit GUI client's configuration file to add repositories, create workspaces, rename folders, etc.

Key features:

• Add/remove repositories
• Create workspaces
• Sync ghq repositories
• Update paths on folder rename

Detailed guide

ghq Migration (to-ghq) / Bare conversion (to-bare)

• to-ghq (formerly migrate): bare+worktree → regular .git at the ghq path (~/ghq/host/group/repo/). Auto bare+worktree conversion, optional symlink, nested groups. Detailed guide
• to-bare (inverse): regular repo → bare + worktree at a custom location (e.g. .claude/worktrees/), preserving uncommitted changes. Includes a Windows lock pre-check (SourceGit/VS Code handle on .git blocks the rename). Detailed guide

Repo Patrol (batch inspection)

Batch inspect and clean up the status of repositories under ghq.

Key features:

• Parallel collection of status, stash, unpushed for all repositories
• Status-based processing (commit-splitter integration, stash pop, push)
• Optional fetch all at the end

Detailed guide

Common Workflow

1. Repository migration: Migrate to ghq structure with to-ghq topic (or to-bare for the inverse)
2. SourceGit update: Register new paths with sourcegit topic
3. Batch inspection: Clean up uncommitted/unpushed changes with patrol topic

Scripts

• ./scripts/repo-to-ghq.sh - Move a repository to the ghq path (bare+worktree → regular)
• ./scripts/repo-to-bare-worktree.sh - Convert a regular repo → bare + worktree (inverse)
• skills/git-repo/scripts/local-to-staging-pr.sh - Cherry-pick local commit to a staging branch; prints manual push/PR commands by default, or pass --push [--body-file ] to also push and open the draft PR.