Merge Train

Merge Train

Process a whole list of open pull requests in one pass: order them, fix what is fixable, review them, then merge every PR that can be merged without breaking the order — and clearly report why each remaining PR is blocked.

This skill performs irreversible shared-repo actions (pushing fixes, merging PRs, posting comments). It follows a grouped-confirmation model: it does all read-only analysis first, presents one complete plan, and waits for a single explicit confirmation before performing any write. Until the user confirms, nothing is pushed, merged, or commented.

When to Activate

• User shares a GitHub pull-requests URL (e.g. https://github.com///pulls, a filtered query, or a milestone/label view) and asks to process, triage, order, or merge them.
• User says: "process all the PRs", "what's the merge order", "fix CI and conflicts then merge what can be merged", "review and merge the open PRs", "clear the PR queue".
• User invokes /merge-train (optionally with a GitHub PRs URL or --dry-run).

Inputs

• Primary: a GitHub pull-requests URL. Any of: /pulls, a search query (/pulls?q=...), a label/milestone filter. Honor the filter in the URL — only the PRs that the URL would list are in scope.
• If no URL is given, default to the current repo's open PRs (gh pr list), and ask the user to confirm the repo if it is ambiguous.
• Flags:
• --dry-run — produce the full plan and stop; never write. Overrides the confirmation step.
• --base — treat this as the integration branch when detecting stacked PRs (default: auto-detect the repo default branch).
• --ci-timeout — per-PR ceiling on how long to wait for CI to reach a definitive state (default: 30). Once the ceiling is hit on a PR, that PR is left for the next run and reported as ci-still-pending; the train moves on.
• --ci-poll — interval between gh pr checks polls while waiting for CI (default: 30).

Preconditions

Verify before any analysis. If any fail, stop and report the blocker.

1. gh auth status succeeds and has repo scope (needed for merge + comment).
2. git is available and the working tree is clean (git status --porcelain empty). A dirty tree risks contaminating CI/conflict fixes — stop and ask the user to commit or stash.
3. The target repo is reachable via gh (gh repo view /).
4. Resolve the default/integration branch: gh repo view --json defaultBranchRef -q .defaultBranchRef.name (overridden by --base).

Workflow

Phase 1 — Enumerate (read-only)

List every open PR in scope:

gh pr list --repo / --state open --limit 200 \
  --json number,title,headRefName,baseRefName,isDraft,author,url,labels,body,mergeable,mergeStateStatus

Apply the URL's filter (label/milestone/query) to the list. Drafts are reported but never merged.

Phase 2 — Per-PR data gathering (read-only)

For each PR collect:

• Base/head branches — for stacked-PR detection.
• Mergeability — mergeable (MERGEABLE / CONFLICTING / UNKNOWN) and mergeStateStatus (CLEAN, BLOCKED, DIRTY, BEHIND, UNSTABLE).
• CI status — gh pr checks --repo / → list failing/pending checks.
• Review state — gh pr view --json reviewDecision,reviews.
• Changed files — gh pr diff --name-only (for overlap detection and review scope).
• Size/risk — additions/deletions, number of files, whether it touches only docs/deps/config vs. core source.
• Dependency hints — scan title + body for: depends on #N, blocked by #N, stacked on #N, after #N, requires #N, base: .

Phase 3 — Compute merge order

Build a directed dependency graph (A → B means A must merge before B). Apply heuristics in priority order:

1. Stacked branches (strongest signal). If PR B's base branch is PR A's head branch (i.e. base ≠ integration branch and equals another open PR's head), then A → B. Chains can be multi-level.
2. Explicit mentions. Parse the dependency hints from Phase 2. B depends on/blocked by/after #A ⇒ A → B. Explicit mentions override weaker signals but never override a contradicting stacked-branch fact — if they conflict, stop and surface the contradiction to the user.
3. File overlap. For PRs with no stack/mention relationship that touch one or more of the same files, order the lower-risk / older PR first to minimize rebase churn, and flag the overlap so the user knows a conflict resolution will be needed on the later one.
4. Size / risk (tie-breaker only). Among otherwise-independent PRs, schedule small low-risk PRs (docs, dependency bumps, config) before large refactors.

Detect cycles. If the graph has a cycle (e.g. mutual depends on), do not guess — report the cycle and ask the user how to break it.

Produce a topologically-ordered list. Independent PRs keep their relative size/risk ordering.

Phase 4 — Per-PR triage + code review (read-only)

Walking the order, classify each PR:

• mergeable — CI green (or only fixable failures), no conflicts (or auto-resolvable), review acceptable, and every PR before it in its dependency chain is also mergeable.
• changes-required — code review finds correctness/security/blocking issues, OR CI fails for reasons that are not a mechanical fix (genuine test failures, design problems).
• blocked-by-order — the PR itself is fine, but an upstream PR in its dependency chain is not yet merged / not mergeable. Record the specific blocking PR number(s).

Code review per PR is a focused five-axis pass (correctness, readability, architecture, security, performance) scoped to the diff. Keep it proportionate to PR size; flag blockers explicitly, note nits separately.

A PR is only merge-eligible if it is mergeable and all PRs it depends on are already merged or earlier in the same confirmed batch.

Phase 5 — Present the consolidated plan (the single confirmation gate)

Output one report (see Output Format). It must show, for every PR: position in merge order, status, the blocking PR if any, the failing CI checks that would be fixed, whether a conflict resolution is needed, and the review verdict.

Then ask one grouped confirmation, e.g.:

> "Plan ready: N PRs mergeable (#a, #b, #c), M need changes, K blocked by order. I will, in order: fix the listed CI failures, resolve conflicts against ``, merge the N mergeable PRs, and post an order-blocking status comment on the rest. Proceed with the full batch? (yes / dry-run / pick a subset)"

• --dry-run (or a dry-run answer) ⇒ stop here, write nothing.
• The user may approve a subset; respect it and recompute eligibility (a skipped PR blocks everything downstream of it).
• Without an explicit "yes", do nothing.

Phase 6 — Execute (writes; only after confirmation)

Process strictly in merge order. For each merge-eligible PR:

1. Resolve conflicts first (if any): update the branch against `` (rebase or merge per the repo's convention — prefer the project's existing strategy; do not force-push shared branches without the user's go-ahead beyond the batch confirmation, and never force-push the integration branch). Resolve conflicts faithfully (never discard a side blindly), run the build/tests locally, push to the PR head branch.

1. Drive CI to green — loop until all required checks are SUCCESS. This is a hard gate: a PR is NEVER merged until its CI is green. The loop is bounded only by --ci-timeout (default 30 min); within that ceiling, keep fixing failures and pushing until CI is green.

Repeat:

a. Poll CI state: gh pr checks --repo / (or gh pr view --json statusCheckRollup). Categorize each required check as SUCCESS, FAILURE (FAILURE / CANCELLED / TIMED_OUT / ACTION_REQUIRED), or PENDING (QUEUED / IN_PROGRESS / WAITING).

b. If every required check is SUCCESS → CI is green. Exit the loop and go to step 3.

c. If any required check is PENDING → sleep --ci-poll seconds (default 30s) and re-poll. Track total wait time against --ci-timeout (default 30 min). When the ceiling is hit while checks are still pending, leave the PR untouched, mark it ci-still-pending, post the status comment (Phase 6 final step), and move on to the next PR in the queue — do not merge a PR whose CI never settled.

d. If any required check is FAILURE → fetch the failing check's logs (gh run view --log-failed --repo /), check out the PR head branch, reproduce locally where feasible, apply the minimal fix scoped to that failure, run the relevant lint/typecheck/test/build to confirm the fix locally, commit with a clear message (e.g. ci: fix ), push to the PR head branch. Then go back to (a) — the loop continues; do not move on to merge after a single push.

e. Keep looping as long as failures are mechanical (lint, formatting, type errors, lockfile drift, flaky-retry, env/setup issues). The only ways to break out of the loop without reaching green are:

• Real defect — a check fails because the diff actually contains a bug (a unit test catches a true regression, a security scanner flags a real issue, a design problem surfaces). Reclassify the PR as changes-required with the failing check + reason, and move on.
• Same failure recurs unchanged after two consecutive fix attempts targeting it — treat as a real defect (the obvious mechanical fix didn't take), reclassify changes-required with the evidence, and move on.
• --ci-timeout reached while checks are still pending — mark ci-still-pending and move on (see (c)).

Never weaken a check to make it pass (no --no-verify, no deleting/skipping tests, no loosening lint, no --admin merge). If you find yourself tempted, the failure is a real defect — reclassify and move on.

1. Merge (only reachable when CI is green from step 2): re-check mergeability one last time (gh pr view --json mergeable,mergeStateStatus); if still MERGEABLE / CLEAN, merge using the repo's merge method (gh pr merge --repo / -- — detect allowed method; default to squash if unspecified). After a successful merge, branches downstream of it may become mergeable — re-evaluate the remaining queue (rebase the next PR onto the new base if needed, which re-enters its own step 1 → step 2 loop).

1. If any step fails irrecoverably (push rejected by branch protection, merge API error, etc.), stop touching that PR, mark it changes-required (or ci-still-pending if (c) applied) with the reason, and continue with PRs not dependent on it.

For every PR not merged (changes-required, blocked-by-order, or ci-still-pending), post a single status comment via gh pr comment:

• blocked-by-order: "⏸ Blocked by merge order — must merge after #N (and #M). Will be mergeable once those land."
• changes-required: a concise summary of the blocking issues from the review / CI, with file:line references. If broken out of the CI loop because of a real defect, name the failing check and the symptom.
• ci-still-pending: "⏳ CI did not reach a definitive state within the --ci-timeout window. Last poll: checks still pending (). Re-run merge-train when CI has settled."

Phase 7 — Final report

Summarize: what merged (in order, with commit/merge URLs), what was fixed (CI + conflicts per PR), what remains and why, and the recommended next action for the blocked set.

Safety Rules

• One grouped confirmation before any write. No write in --dry-run.
• CI must be green before merging — no exceptions. Poll gh pr checks until every required check is SUCCESS. A PR with any FAILURE or PENDING required check is never merged. There is no "merge then watch" mode.
• Keep iterating on CI failures until CI is green (bounded by --ci-timeout). The fix-and-push loop only ends in one of three ways: (1) every required check is SUCCESS → merge, (2) the failure is a real defect (not mechanical) or the same failure recurs after a fix attempt → reclassify changes-required, (3) the timeout is hit while checks are still pending → mark ci-still-pending and move on. Never merge to "unblock" the queue.
• Never force-push the integration/default branch. Never --admin-merge or bypass branch protection unless the user explicitly asks.
• Never merge a draft PR.
• Conflict resolution must preserve both sides' intent — if a resolution is non-obvious or risks data loss, stop and ask.
• CI "fixes" are limited to mechanical issues (lint, formatting, type errors, flaky-retry, lockfile, env/setup). A failing test that reflects a real defect is changes-required, not something to silence (never delete/skip tests, never add --no-verify, never weaken a check to make CI pass).
• Contradictory dependency signals or a dependency cycle ⇒ stop and ask; do not guess an order.
• Respect the URL's filter — do not act on PRs outside the requested scope.
• Keep each fix surgical and scoped to making that PR mergeable; no opportunistic refactoring.

Output Format

A single Markdown report:

## Merge Train — / ( open PRs in scope)
CI gate: every PR must reach all-green before merge · ci-timeout: min · ci-poll: s

### Merge order
1. #12  feat: X            → mergeable        (CI: 1 failing — fixable · no conflicts)
2. #15  refactor: Y        → mergeable        (CI: green · conflicts vs base)
3. #18  feat: Z            → blocked-by-order  (after #15)
4. #20  fix: W             → changes-required  (review: 2 blockers)
...

### Per-PR detail
#12 …  status, reasoning, CI checks to fix, conflict?, review verdict (blockers/nits)
#15 …
...

### Plan
Mergeable & eligible: #12, #15  → resolve conflicts, loop on CI until green, merge in this order
Will comment (blocked): #18 (after #15)
Will comment (changes): #20 — 

> Proceed with the full batch? (yes / dry-run / subset)

After execution, append an Execution result section with merged PR URLs, fixes applied (per PR: every CI fix iteration — failing check → commit SHA → resulting CI state — until green or break-out reason), and the remaining queue (with break-out reason for each: changes-required / blocked-by-order / ci-still-pending).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: big-emotion
• Source: big-emotion/agent-atelier
• License: MIT

Install and usage instructions live in the source repository linked above.