merge-flow

name	merge-flow
description	Merge a batch of approved workbuddy PRs into main: order them, resolve conflicts against design intent, validate post-merge health, and report. Use when the user asks to 'merge the batch', 'merge approved PRs', 'integrate finished issues', '批量合并', '处理冲突'; or when the merge-agent is invoked.

Batch Merge Flow

This skill codifies how to merge multiple approved workbuddy PRs into main safely, with design-intent judgment on conflicts.

When to use

• A batch of workbuddy issues has completed: their PRs are approved by review-agent and CI is green
• User asks to merge a list of PRs
• Typical invocation is manual — a human decides a batch is ready and hands the PR list to the merge-agent. The operator may detect a batch-ready condition and surface it, but the decision to merge belongs to a human because this skill involves design-intent judgment that operator autonomy shouldn't override.

Input contract

The caller provides:

prs: [#101, #102, #104]      # PR numbers to merge (in no particular order)
repo: Lincyaw/workbuddy       # target repo
base: main                    # base branch (default: main)
mode: auto | review-required  # auto = proceed on safe resolutions; review-required = escalate all conflicts

Procedure

Phase 0 — Batch coherence analysis

Before touching any branch, analyze the batch as a whole. This is where design-intent reasoning happens — do it once upfront, not hunk-by-hunk during rebase.

0.1 Issue-level duplicate detection

For each PR, fetch its linked issue:

gh pr view <n> --json body,closingIssuesReferences
gh issue view <issue-num> --json title,body,labels

Compare issues pairwise. Duplicate signals:

• Title near-match: same key terms, similar phrasing
• Symptom overlap: both describe the same user-facing problem or error signature
• AC overlap: acceptance criteria describe the same behavior
• File overlap: both PRs modify the same files in the same region
• Root-cause overlap: both diagnose the same underlying cause

For suspected duplicates, consult decisions.md and project-index.yaml to see if one is actually a re-filing of an older issue or a subset of a larger epic.

Output: a list of duplicate groups. Example:

group A: #87 ("heartbeat lease bug") + #88 ("worker submit failure")
  — #88's root cause IS #87. Duplicates.
group B: #91 (dynamic config reload) — standalone

0.2 PR-level feature overlap detection

For each pair of PRs (within AND across duplicate groups), analyze diffs:

gh pr diff <n>

Overlap signals:

• Same files changed with overlapping line ranges
• Same symbol modified (function, type, const) — git diff --name-only, then grep for symbol names
• Same behavior implemented (even if in different files) — e.g., both add a retry loop around the same API call
• Contradictory changes — one PR adds X, another removes X

Classify each overlap:

Class	Example	Default resolution
Identical intent, one better	Both fix the same bug; PR A is 3 lines, PR B is 50 lines with refactor	Adopt PR A entirely, close PR B
Identical intent, equivalent quality	Both fix same bug with equally clean code	Adopt oldest-approved PR, close the other
Complementary	Both add distinct features to same file; changes don't semantically conflict	Merge both in order
Partially overlapping	PR A adds features X+Y; PR B adds Y+Z; both need Y	Cherry-pick the non-overlapping parts into a merge order: take A's X, take one side's Y, take B's Z
Contradictory design	PR A uses pattern P1; PR B uses pattern P2; incompatible architecturally	Escalate — human must pick the approach

0.3 Decision authority

For each non-trivial overlap, the skill prescribes:

• "Adopt one" decisions (classes 1, 2): the merge-agent decides based on: code quality heuristics (lines changed, test coverage delta, complexity) + design-intent anchors (CLAUDE.md conventions, north-star alignment). Record reasoning.
• "Cherry-pick" decisions (class 4): the merge-agent proposes the split, but if the split requires non-trivial code rearrangement (not just a git cherry-pick -x of individual commits), escalate to human — don't rewrite authors' commits.
• Contradictory design (class 5): always escalate.

Record all decisions in the report (Phase 6) with rationale. For "close PR B" decisions, post a comment on the closing PR explaining why, linking to the kept PR, and thanking the author.

0.4 Output of Phase 0

A revised merge plan:

plan:
  merge:
    - pr: 101       # after cherry-pick: take only files A.go, B.go
      cherry_pick_from: [abc123, def456]
      reason: "partial overlap with #102; keeping non-overlapping commits"
    - pr: 102       # full merge
      reason: "standalone"
  close:
    - pr: 105
      reason: "duplicate of #101 with inferior implementation; fewer tests, larger diff"
      comment_to_post: "..."
  escalate:
    - prs: [103, 104]
      reason: "contradictory design choices for internal/store package layout"
      question_for_human: "..."

If any item lands in escalate: with a class 5 (contradictory design), stop and surface the question to the invoker before touching any branch. Resolving design-level conflicts is a human decision.

If only merge: and close: items remain, proceed to Phase 1 with the revised plan.

Phase 1 — Preconditions (per PR)

For each PR in prs, verify:

1. Approved: gh pr view <n> --json reviewDecision returns APPROVED
2. CI green: gh pr checks <n> shows all required checks passing
3. Base is current: PR's base branch is main (no stale PRs targeting old branches)
4. No status:blocked label on linked issue
5. Author branch exists and is not force-pushed recently (check commit SHA hasn't changed since approval timestamp)

Skip PRs that fail any precondition. Record reason in the report.

Phase 2 — Ordering

Determine merge order using a topological sort:

1. Explicit dependencies: parse PR bodies and linked issue bodies for "Blocked by #X" / "Depends on #X". PRs with unresolved deps go after their deps.
2. File-overlap heuristic: PRs that touch the same files are ordered by created_at (oldest first) so the second PR rebases onto the first.
3. Disjoint PRs: no specific order, merge in created_at order for reproducibility.

Output a linear order. Cycles (A blocks B and B blocks A) are an error → escalate.

Phase 3 — Serial merge

For each PR in order:

# 1. Fetch latest main
git fetch origin main
git checkout main
git pull --ff-only origin main

# 2. Try to rebase PR branch onto main (without pushing)
git fetch origin pull/<n>/head:pr-<n>-rebase
git checkout pr-<n>-rebase
git rebase main

If rebase succeeds cleanly → proceed to Phase 4 validation, then gh pr merge <n> --squash --rebase --delete-branch.

If rebase has conflicts → enter Phase 4 conflict resolution on the rebase, then:

• If resolved automatically: push the rebased branch (git push -f origin pr-<n>-rebase:<original-branch>), wait for CI, then merge.
• If escalated: abort rebase (git rebase --abort), comment on the PR with the conflict report, skip to next PR.

Phase 4 — Conflict resolution (when rebase conflicts)

Classify each conflict hunk:

Class	Signature	Strategy
Trivial	Whitespace, import order, comment-only diffs	Prefer the PR's side; re-run goimports or equivalent formatter
Additive	Both sides add distinct code to the same region (e.g., both add a new case to a switch)	Merge both additions, preserve order by PR created_at
Same-intent	Both PRs do the same thing slightly differently (e.g., both fix the same bug)	Prefer the approach that aligns with CLAUDE.md conventions / decisions.md entries. Flag if unclear.
Competing-intent	PRs implement overlapping features with incompatible designs	Escalate. Do not silently pick one.
Structural	Conflicts span multiple files and involve architectural choices (new package layout, interface changes)	Escalate.

Design-intent consultation

Before resolving same-intent or competing-intent hunks, consult in this order:

1. CLAUDE.md — project conventions (e.g., "squash + rebase only", "2-agent first principle")
2. decisions.md — recent architectural decisions with rationale
3. project-index.yaml — requirement text for both PRs' linked issues; the resolution must satisfy both AC sets
4. North-star targets in CLAUDE.md — if two resolutions are plausible, pick the one that advances primary targets (requirements coverage, test coverage, build health)

Record your reasoning inline in the resolution commit message:

resolve: PR #101 vs PR #102 conflict in cmd/coordinator.go

Context: both PRs modify handleTaskResult. #101 adds tokenized auth check,
#102 adds metrics emission. Changes are orthogonal — merged both.
Design-intent basis: north-star reliability favors both; no competition.

Phase 5 — Post-merge validation

After each successful merge (before proceeding to next PR):

git checkout main
git pull --ff-only origin main
go build ./...
go vet ./...
go test ./... -count=1

If validation fails:

• Immediate action: git revert <merge-sha> and push, restoring main to green state
• Comment on the just-merged PR explaining the revert with the failing output
• Mark the PR's linked issue with status:developing to route back through dev-agent
• Continue to next PR in batch (do not abort the entire batch)

Phase 6 — Report

Output a structured report:

## Merge Batch Report — <timestamp>

### Merged (N)
- #101 — clean rebase
- #102 — additive conflict auto-resolved (switch case)
- #104 — trivial conflict (imports)

### Reverted (M)
- #103 — merged then reverted: go test failed in internal/store
  (revert SHA: abc123; issue re-opened for developing)

### Escalated / Skipped (K)
- #105 — competing-intent conflict with #101 in internal/launcher/permissions.go:
  #101 uses env-var token only; #105 adds keyring fallback.
  Design intent unclear — human decision needed. PR left open.

### Validation
- Final main SHA: def456
- go build: ok
- go test: ok (42 packages, 0 failures)
- go vet: ok

Constraints

• Never force-push to main. Always use gh pr merge --squash --rebase.
• Never skip Phase 5 validation. Each merge is validated before the next begins.
• Never silently resolve competing-intent conflicts. Escalate and leave the PR open.
• Never delete a PR's source branch unless --delete-branch was specified by GitHub merge (which is fine).
• Never amend merged commits. If something's wrong, revert + new PR.
• Never bypass review. If a PR somehow lost its approval (e.g., new commits pushed), re-trigger review-agent instead of merging.

Edge cases

All PRs fail preconditions

Report and exit. Do not attempt to salvage.

Main drifts during batch (new commit lands from outside)

Re-rebase remaining PRs against new main. This is safe because main is always green (Phase 5 invariant).

Revert chain

If merging PR A causes failure, then reverting A also breaks (highly unusual):

• Reset batch: do not continue
• Alert: "main is broken in a way the batch flow cannot recover from" → escalate to human immediately

PR branch deleted mid-merge

The original author may have deleted their branch. Fetch via refs/pull/<n>/head (always available on GitHub).

Success criteria

A batch merge succeeded if:

• Every PR in prs is either: merged, reverted-and-re-queued, or escalated-with-report
• main is in a green state (build + test + vet pass) at the end
• The report accurately reflects all outcomes
• No manual intervention was required for same-intent or trivial conflicts