ywc-finish-branch

name: ywc-finish-branch version: 1.0.0 description: >- (ywc) Use when delivering a single completed feature branch to the base branch — covers mark-PR-ready, CI wait + bot review polling, merge (PR or local), post-merge verification, Mark Task Complete bookkeeping, and local branch cleanup. Triggers: "finish branch", "deliver branch", "finish-branch", "ywc-finish-branch", "merge feature branch", "branch 마무리", "branch 마감", "ブランチ完了", "deliver task", "task 마감". Do not use for branch creation (handled by upstream executor), worktree management (caller of ywc-parallel-executor handles), draft PR creation alone (use ywc-create-pr), or PR review comment handling (use ywc-handle-pr-reviews). category: delivery phase: execution requires: [ywc-create-pr] advisor_budget: 1

ywc-finish-branch

Announce at start: "I'm using the ywc-finish-branch skill to deliver this feature branch."

Encapsulates the post-implementation delivery flow that ywc-sequential-executor and ywc-parallel-executor previously inlined separately. One feature branch, one task — given a verified branch with passing local tests, this skill takes it the rest of the way to "done": optionally creates the PR, marks it ready, waits for CI, handles automated bot reviews, merges (PR or local), runs the post-merge hard gate, moves the task directory to completed/, and deletes the local feature branch. Worktree-specific cleanup is intentionally not in scope — that stays with ywc-parallel-executor because the lifecycle differs.

Rationalization Defense

When tempted to skip a step, check this table first:

Violating the letter of these rules is violating the spirit. Delivery is the step where the user trusts the executor most — earning that trust requires honesty about partial completion.

Arguments

Mode-flag mutual exclusivity: --mode is a single value; conflicts are impossible by construction. The caller is responsible for resolving ambiguity (e.g., the upstream executor's --draft and --local-merge flags must be collapsed into one --mode value before invoking this skill).

Worktree-path mode

When --worktree-path <path> is set, every git command this skill issues — the Step 1 git branch --list / git rev-parse validations, the Step 5 merge/checkout/pull/push/branch-delete sequence, the Step 6 post-merge git log, the Step 7 git check-ignore / git mv / git commit / post-move check, and the Step 8 checkout/pull/branch-delete — runs as git -C <path> …. The delivery semantics are unchanged: CI wait, bot-review polling, the merge-readiness gate, and Mark Task Complete behave identically; only the working directory the git commands act on differs (the run worktree instead of the main checkout). gh commands (gh pr ready / gh pr checks / gh pr merge / gh pr view) operate on the remote PR, not a working tree, so they are unaffected by --worktree-path.

HEAD invariant (per-task feature branch deletion is safe). At the moment Step 5 runs, the run worktree's HEAD is on the integration branch (the worktree's permanent checkout). finish-branch creates/merges feature/<task-name> and then checks out the integration branch before deleting feature/<task-name>, so the per-task branch is never the worktree's current checkout when git branch -d runs. The integration branch (and the trunk base branch) is never a deletion target here — finish-branch only deletes the per-task feature/<task-name>. This is why per-task branch deletion succeeds inside a worktree without needing --keep-branch: only a branch that is checked out in a worktree is undeletable, and the per-task branch is not.

Modes

per-task-pr is the special mode used inside ywc-parallel-executor's wave loop — this skill creates and pushes the PR but stops short of merging or marking complete, because wave-level orchestration handles those steps. For all other modes, this skill executes the full lifecycle.

Definition of Done

A --mode normal-pr or --mode local-merge invocation is done only when all of the following have happened, in this order:

1. Verification gate passed (caller's responsibility — this skill does not re-run tests; the upstream executor must have completed its verification before invoking).
2. Merge succeeded: either the PR was merged via gh pr merge --delete-branch (normal-pr) or git merge --no-ff succeeded locally and was pushed (local-merge).
3. Post-merge verification passed: git log -1 --format="%s" on the base branch begins with Merge branch 'feature/.
4. Task directory moved: git mv <tasks-dir>/<task-name> <tasks-dir>/completed/<task-name> and the chore: mark <task-name> as completed commit was created.
5. Local feature branch deleted: git branch -d feature/<task-name> succeeded (refuses to delete an unmerged branch — that is the safety check).

For --mode draft and --mode skip-ci-wait, conditions 2–5 are intentionally not performed; the skill's responsibility ends at PR creation (and ready-marking for skip-ci-wait). For --mode per-task-pr, only the PR creation portion runs; the caller handles the rest within the wave loop.

Execution Steps

Step 1: Validate Inputs

• Confirm --branch exists locally (git branch --list <branch> returns the branch).
• Confirm --base-branch is reachable (git rev-parse --verify <base-branch> succeeds).
• Confirm --task-name matches an entry under <tasks-dir>/ (not under <tasks-dir>/completed/).
• For PR-based modes: confirm gh auth status succeeds.

When --worktree-path <path> is set, run the git branch --list and git rev-parse validations as git -C <path> … (see Worktree-path mode).

If any check fails, return NEEDS_CONTEXT with the specific missing input. Do not proceed with partial information.

Step 1.5: Ubiquitous Language Update (optional)

Check whether docs/ubiquitous-language.md exists in the project root:

test -f docs/ubiquitous-language.md

• Exists: Invoke ywc-ubiquitous-language --mode update. This runs before PR creation so that any new domain terms introduced by this branch are captured in the glossary before the PR description is written.
• Not exists: Skip silently — the project has not yet established a ubiquitous language document.

ywc-create-pr is called in Step 2 with --skip-ubiquitous-update to prevent a second invocation of the same update.

Step 2: PR Creation (PR-based modes only)

For --mode ∈ {normal-pr, draft, skip-ci-wait, per-task-pr}: construct the PR title per the format below, then invoke ywc-create-pr passing:

• --title "<constructed-title>" — the [task-number] description string built from --task-name and --pr-lang
• --lang <pr-lang> — so the description is written in the correct language
• --base-branch <base-branch> — the resolved target branch
• --skip-post-ci-check — prevents ywc-create-pr from running its own post-PR CI + bot check; this skill handles CI verification in Step 4
• --skip-ubiquitous-update — prevents ywc-create-pr from running the UL update a second time; Step 1.5 already ran it

ywc-create-pr handles security check, CI pre-push validation, push, and PR creation as a draft. Because --title is provided explicitly, ywc-create-pr will use it verbatim and skip its own title-generation step.

PR title format: [<task-number>] <human-readable description in --pr-lang>

Run the bundled script to extract the task number and English slug without regex parsing:

python claude-code/skills/ywc-finish-branch/scripts/build-pr-title.py <task-name>
# TASK_NUMBER=000001-010
# SLUG_EN=Db Create Users Table

If the script exits 1 (TASK_NUMBER is empty): stop immediately and return NEEDS_CONTEXT — report the task name that failed to parse. Do not construct a title with an empty task number.

For English PRs (--pr-lang en), use --format title to get the complete title directly:

python claude-code/skills/ywc-finish-branch/scripts/build-pr-title.py <task-name> --format title
# [000001-010] Db Create Users Table

For other languages, translate only SLUG_EN to --pr-lang, then compose: [<TASK_NUMBER>] <translated-slug>.

The script supports both task-name formats (new 000001-010-slug and legacy 001010-slug). Examples of final titles: [000001-010] Create users table (en) · [000001-010] ユーザーテーブル作成 (ja) · [001010] DB 사용자 테이블 생성 (ko).

Before calling ywc-create-pr: verify the constructed title string starts with [. If it does not, stop — the [TASK_NUMBER] prefix was lost somewhere in the construction. Do not call ywc-create-pr without a valid title.

For --mode local-merge: skip this step entirely. Local-merge does not produce a PR.

Step 3: Mark PR Ready (normal-pr only)

Invoke gh pr ready <pr-number> to convert the draft into an open PR. Skip for --mode draft, --mode skip-ci-wait, --mode per-task-pr (those modes leave the PR as draft or hand control back to the caller).

Step 4: CI Verification + Bot Review Polling (normal-pr only)

Action required: Read ../references/pr-bot-polling.md now before proceeding. The 60-second initial wait, the 300-second polling window, the BOT_COUNT jq query, and the merge condition are all defined there. Do not implement your own polling logic without reading it first.

Run the CI wait + fix loop and the bot review polling per ../references/pr-bot-polling.md. The polling reference defines the canonical loop; this skill applies its --bot-action parameter to choose the post-bot behavior:

• --bot-action sequential: after ywc-handle-pr-reviews runs, re-run the CI verification step before re-polling. Bot fixes may trigger a new CI run that must pass before merging.
• --bot-action parallel: after ywc-handle-pr-reviews runs, re-poll without re-gating on CI. The wave loop in ywc-parallel-executor handles CI at the wave level.

CI failure handling (up to 2 fix attempts per failure cycle):

1. Identify failing checks:

   gh pr checks <pr-number>
   

2. Get failure logs:

   # Get the most recent failed run ID for the feature branch
   gh run list --branch <feature-branch> \
     --json databaseId,name,conclusion \
     --jq '.[] | select(.conclusion == "failure") | .databaseId' | head -1
   gh run view <run-id> --log-failed
   

3. Categorize and fix by failure type:

   Failure type	Fix action
   Lint / format	Run the project's auto-fix command (eslint --fix, prettier --write, ruff --fix, biome check --apply), commit, push
   Type errors	Read compiler output, fix type mismatches in affected files, commit, push
   Test failures	Analyze failing test output, fix the implementation (never disable or weaken tests), commit, push
   Build errors	Read compiler/bundler output, fix compilation or import errors, commit, push

4. After each push, re-run the CI wait and bot review polling loop (return to the top of Step 4).

5. If CI still fails after 2 fix attempts, return BLOCKED with: the failing check name(s), a log excerpt (last 30 lines), the PR URL, and the failure category so the user has enough context to resolve it manually.

Step 4 final: Merge-Readiness Gate (normal-pr only)

CI green is necessary but not sufficient — the base may have advanced into a conflict while CI ran. Before Step 5's gh pr merge, confirm the PR can actually merge:

Action required: Read ../references/pr-conflict-resolution.md — it defines the mergeable / mergeStateStatus table and the merge-not-rebase update procedure.

gh pr view <pr-number> --json mergeable,mergeStateStatus --jq '{mergeable, mergeStateStatus}'

• MERGEABLE / CLEAN → proceed to Step 5.
• BEHIND (branch protection requires up-to-date; no textual conflict) → merge the base into the feature branch (never rebase), push, and re-verify CI (counts as one CI cycle). Then re-check the gate.
• CONFLICTING / DIRTY that auto-resolves → merge the base into the feature branch (never rebase), push, and re-verify CI (counts as one CI cycle). Then re-check the gate.
• CONFLICTING / DIRTY with real textual conflicts → surface the conflicting files + PR URL and return BLOCKED. Do not auto-resolve or force-push.
• BLOCKED (a required check or review is missing — not a conflict) → do not run the base-merge procedure; surface the outstanding required check/review and return BLOCKED so the user can resolve the gate.
• UNKNOWN → poll briefly per the reference, then re-read.

Skip this step entirely for --mode local-merge, --mode draft, --mode skip-ci-wait, --mode per-task-pr.

Step 5: Merge

For --mode normal-pr:

gh pr merge <pr-number> --delete-branch

For --mode local-merge:

# Caller has already verified its Step 4 (Task Verification); this skill does not re-verify.
git checkout <base-branch>
git pull origin <base-branch>
git merge --no-ff feature/<task-name> -m "Merge branch 'feature/<task-name>'"
git push origin <base-branch>          # omit when --defer-push is set
git branch -d feature/<task-name>      # omit when --keep-branch is set (parallel wave loop case)

When --worktree-path <path> is set, prefix every git command above with -C <path> (e.g. git -C <path> checkout <base-branch>, git -C <path> merge --no-ff feature/<task-name> …, git -C <path> branch -d feature/<task-name>); here <base-branch> is the run's integration branch (the worktree's checkout). Per the Worktree-path mode HEAD invariant, the integration branch is checked out before the per-task feature/<task-name> is deleted, so the delete succeeds.

For all other modes: skip — merge happens later (manual for draft, caller for skip-ci-wait and per-task-pr).

Why git merge --no-ff specifically (not cherry-pick, not direct commit on main): preserves original commit SHAs, marks the task boundary in git log --graph, and lets git branch -d act as a built-in safety check that refuses to delete an unmerged branch. Full rationale lives in the upstream executor's references/branch-lifecycle.md.

Error handling: git merge conflict → stop and ask the user to resolve manually (do not auto-abort; the user may want to inspect). gh pr merge failure due to conflict/out-of-date → apply the Step 4 final Merge-Readiness Gate (references/pr-conflict-resolution.md): merge the base into the feature branch and re-verify, or return BLOCKED on a real textual conflict. Other gh pr merge failures → stop and report. Push rejected (remote moved) → stop; never force-push.

Step 6: Post-Merge Verification (hard gate)

For --mode normal-pr and --mode local-merge:

git log -1 --format="%s"

The output must begin with Merge branch 'feature/. If not, the merge did not execute correctly — investigate and retry before proceeding to Step 7. Do not write a completion marker for a merge that did not happen; downstream tasks would resolve dependencies against a broken contract. When --worktree-path <path> is set, run this check as git -C <path> log -1 --format="%s".

Step 7: Mark Task Complete

For --mode normal-pr and --mode local-merge:

Use the shared marker script — it handles the .gitignore branch, the mandatory marker commit, and the post-move verification in one deterministic step (exits non-zero if anything failed):

bash claude-code/skills/scripts/mark-complete.sh <tasks-dir> <task-name> [--push | --defer-push]

Why a script, not inline git: git mv cannot stage a path inside a gitignored <tasks-dir> (the move produces no diff), so that case needs a plain mv plus an --allow-empty marker commit, while a tracked <tasks-dir> uses git mv. The script detects which applies. The chore: mark <task-name> as completed commit is mandatory in both cases — it is the git log audit boundary that humans, audit tooling, downstream skills, the Completion Report, and resume / replay all rely on to verify task completion at a specific commit. The script then verifies the move (destination exists, source gone, marker commit at HEAD) and exits non-zero if any check fails; do not declare DONE on a non-zero exit — investigate and retry.

When --worktree-path <path> is set, run mark-complete.sh so its git check-ignore / git mv / git commit and post-move git log operate on the worktree (invoke it with the worktree as the working directory, e.g. cd <path> && bash <repo>/claude-code/skills/scripts/mark-complete.sh …); the <tasks-dir> move acts on <path>/<tasks-dir>/… directly, and the verification checks resolve against the worktree's HEAD (see Worktree-path mode).

Push strategy: pass --push to mark-complete.sh for --mode local-merge and for --mode normal-pr when --defer-push is not set. Otherwise pass --defer-push (the script's default) so the caller can batch completion-marker commits and push them once at the end of the range. The caller is responsible for the eventual push when deferral is requested.

For --mode draft, --mode skip-ci-wait, --mode per-task-pr: skip Step 7 entirely. The task is not yet merged; marking it complete would be incorrect.

Step 8: Local Branch Cleanup

For --mode normal-pr:

git checkout <base-branch>
git pull origin <base-branch>
git branch -d feature/<task-name>      # omit when --keep-branch is set

The local branch must be deleted explicitly because gh pr merge --delete-branch only removes the remote. Skip the deletion when --keep-branch is set — the caller (typically ywc-parallel-executor) will release the worktree and delete the branch in its own cleanup step. When --worktree-path <path> is set, run all three commands as git -C <path> … (here <base-branch> is the run's integration branch); the per-task feature/<task-name> deletion is safe because the worktree HEAD is on the integration branch per the Worktree-path mode invariant.

For --mode local-merge: already handled inside Step 5's command sequence (which honors --keep-branch likewise).

For all other modes: skip. The branch stays alive for the user or caller to handle.

Worktree prune (when caller used a worktree) — if the caller created a per-task worktree (ywc-parallel-executor does this), the post-cleanup prune is delegated to the ywc-worktrees skill:

ywc-worktrees --mode prune --task-name <task-name>

ywc-worktrees --mode prune runs the bundled cleanup-worktree.sh against the resolved worktree path — git worktree remove + local git branch -d + git worktree prune + post-removal verification, refusing to operate on dirty worktrees without --force. This skill does not inline the prune logic; the worktree priority chain (.worktrees/ > CLAUDE.md worktree_root > --root fallback) and the dirty-tree safety check live in one place. Skip this delegation when no worktree was created (the typical --mode local-merge / --mode normal-pr flow that ran inside the main checkout).

Output Format

## Finish Branch Result: <task-name>

### Mode
<mode>

### Outcome
- PR: <pr-url-or-"none-local-merge">
- Merge: <commit-sha-or-"deferred">
- Mark Complete commit: <commit-sha-or-"skipped">
- Local branch: <"deleted"|"preserved">

### Verification
- Post-merge gate: <passed|skipped|failed>
- Post-move gate: <passed|skipped|failed>

### Completion Status
<DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT>

Completion Status rules:

Banned Output Patterns

This skill produces operational side effects (commits, pushes, merges), not source code. The "Banned Output Patterns" idea applied here means: never declare DONE while leaving any of the following:

• A merge commit that the post-merge verification step did not confirm
• A chore: mark <task-name> as completed commit that has not been verified by the post-move test
• A local feature branch that should have been deleted (modes normal-pr and local-merge, when --keep-branch is not set)
• An unpushed commit when push was required (modes local-merge always; normal-pr when --defer-push is not set)

Each of these is a hidden defect that breaks the next task's dependency resolution. Surface them honestly via DONE_WITH_CONCERNS or BLOCKED rather than declaring DONE.

Integration

• upstream: ywc-sequential-executor (replaces its Steps 5–8), ywc-parallel-executor (replaces the merge + mark-complete portion of Step 4e–4f; worktree cleanup at Step 4g stays in parallel).
• internal delegation: ywc-create-pr (Step 2), ywc-handle-pr-reviews (Step 4 inside the bot polling loop).
• shared references: ../references/pr-bot-polling.md (CI + bot polling), ../references/local-merge-permissions.md (pre-authorized git/gh patterns the caller must set), ../references/subagent-status-actions.md (how the caller responds to this skill's status return), ../references/pr-conflict-resolution.md (merge-readiness gate and conflict resolution procedure).

Notes

This skill is a single-task delivery unit. It does not iterate over a range, manage a wave, or maintain checkpoint state — those concerns belong to the calling executor. Worktree creation and removal are intentionally out of scope, because the lifecycle differs between callers; placing it here would force one of them to special-case its cleanup. What this skill does support is operating inside a caller-supplied worktree via --worktree-path (see Worktree-path mode): ywc-sequential-executor --worktree runs its whole range inside one run worktree and passes that path so finish-branch's git commands target the worktree HEAD, while the worktree's own create/prune lifecycle stays with that caller. The historical "sequential (no worktree)" assumption now reads: sequential delivers in the main checkout unless --worktree is active, in which case --worktree-path redirects the git commands; parallel still owns its per-task worktrees.