Skip to content

flurdy/agent-skills / skills/ready-to-merge

ready-to-merge

Verified

Pre-merge gate — verify a PR is green, approved, in sync, and free of obvious risk, then (on explicit approval) squash-merge it. Composes /pr-status, /contract-check, and /review-pr rather than reimplementing them.

Ready to Merge — Pre-Merge Gate

Verify a PR is safe to merge, present a terse readout, and — only after explicit user approval — squash-merge it.

This skill is a synthesis layer: it calls into /pr-status's scripts, optionally /contract-check, and reuses /review-pr's view/diff scripts. It does NOT reimplement those skills.

Relationship to other skills

  • /pr-status — passive multi-PR dashboard. This skill is single-PR, decision-oriented, and ends with an action.
  • /review-pr — deep AC-vs-diff review. Linked but not auto-invoked. Suggest it if the diff is large or AC fit is unclear.
  • /contract-check — read-only contract health audit. Auto-invoked only when pacts are present in the repo, and scoped to status (summary only).
  • /contract-test — contract test runner (generate → sync → verify). Never auto-invoked here — CI already ran the tests on push. Surfaced as a remediation pointer when /contract-check flags stale or uncommitted pacts.
  • tracking-auditor (agent) — branch-vs-ticket scope check. Suggest it if the diff looks broader than the ticket.

Usage

/ready-to-merge              # PR for current branch
/ready-to-merge 5753         # explicit PR number
/ready-to-merge --no-merge   # report only, never offer to merge

Merge timing policy

Optional and opt-in: without a policy, timing is never checked or rendered. A policy is the nearest merge-timing.toml in a directory above the repository's top level, for example ~/Code/{client}/merge-timing.toml. A file inside the repository is ignored and reported, so a PR cannot relax its own restriction. A candidate inside another checkout of the same repository is unknown, not trusted. The nearest trusted file wins; there are no built-in defaults.

timezone = "Europe/London"          # IANA name; never the machine's timezone

[[rule]]
name = "business-hours"
enforcement = "required"           # blocks the merge offer; only for explicit client policy
days = ["mon", "tue", "wed", "thu", "fri"]   # allowed days; omit for every day
start = "09:00"                    # earliest merge time; omit for 00:00
end = "17:30"                      # merge before this time; omit for 24:00

[[rule]]
name = "late-afternoon"
enforcement = "advisory"           # caution only; normal confirmation still applies
end = "16:00"
end_by_day = { fri = "12:00" }     # day-specific cutoffs replace end

Every rule must allow the current policy-local time. Any required violation is required; otherwise any advisory one is advisory. Unknown keys, missing enforcement, invalid times or timezones, empty or contradictory day settings, and unreadable files are unknown, never clear. Holidays, overrides and automatic scheduling are out of scope.

The helper prints one JSON object: status (none, clear, advisory, required or unknown), plus policy, timezone, local_time, weekday, violations (rule, enforcement, reason), next_allowed, ignored or error when applicable. Treat exit 3, any other non-zero exit, or unparseable output as unknown. --now exists for tests only; never pass it here. Requires Python 3.11+ (standard library only) and Git 2.31+.

Procedure

Phase 1 — Resolve PR

# If no number given:
~/.agents/skills/review-pr/scripts/gh-pr-current-number.sh
# Fallback: gh pr view --json number --jq '.number'

Resolve owner/repo from git remote get-url origin.

Phase 2 — Gather (run in parallel where possible)

Run these in parallel — they're independent fetches:

  1. PR details (state, mergeState, reviewDecision, checksState, approvers, threads, isDraft, base, head):
    ~/.agents/skills/pr-status/scripts/gh-pr-details.sh {owner} {repo} {pr}
    
  2. PR view (title, body, additions, deletions, changedFiles, files):
    ~/.agents/skills/review-pr/scripts/gh-pr-view.sh {pr}
    
  3. Diff:
    ~/.agents/skills/review-pr/scripts/gh-pr-diff.sh {pr}
    
  4. Stacked children — PRs that target THIS PR's head branch:
    gh pr list --base {head-branch} --state open --json number,title,headRefName,url
    
  5. Local working copy — if the resolved PR matches the current branch, also:
    git status --porcelain
    ~/.agents/skills/ready-to-merge/scripts/git-unpushed.sh   # unpushed commits
    
    Uncommitted/unpushed work is a hard blocker. Use the wrapper for the unpushed-commits check — never inline git log @{u}..HEAD. The sandbox flags the bare @{u}..HEAD as brace expansion, so it can never be auto-approved.
  6. Merge timing — from the PR's repository checkout:
    ~/.agents/skills/ready-to-merge/scripts/merge-timing.py
    
    Keep the result as the timing snapshot shown to the user.

Phase 3 — Ticket cross-reference

Extract the Jira key from branch name first, then PR title (/[A-Z]+-\d+/).

If found:

mcp__jira__jira_get
  path: /rest/api/3/issue/{key}
  jq: "{key: key, summary: fields.summary, status: fields.status.name, issuetype: fields.issuetype.name}"

Flag if Jira status is not one of: Code Review, In Progress, Test/Review, Ready for Test. Specifically:

  • Done → ⚠️ already closed before merge — unusual, confirm with user.
  • Backlog / Ready to Work → ⚠️ ticket never moved to In Progress — flow drift.
  • Blocked → ❌ confirm before merging.

Beads: if bd is available, look for a bead referencing this Jira key:

bd list --status=in_progress
bd list --status=open

Match by Jira key in title or description. Note id + status. If a bead is in_progress for this work it will need closing after merge.

Trello: only if TRELLO_BOARD_ID is set for the repo. Use ~/.agents/skills/trello-beads/scripts/trello-api.sh. Look for a card matching the branch or ticket — best effort, skip silently if none found.

Phase 4 — Diff risk scan

Walk the changed files list (don't re-read the whole diff). Flag only what's present — quiet success, loud risk:

RiskDetection (heuristic)
Migration / schema changepath matches migrations/, migrate/, *.sql, schema.prisma, alembic/
Dependency changepackage.json, package-lock.json, pnpm-lock.yaml, yarn.lock, requirements.txt, Cargo.lock, go.mod, Gemfile.lock, build.sbt
Secrets / sensitive.env*, *.pem, *.key, files matching credential / secret
CI / infra changes.github/workflows/*, Dockerfile*, terraform/, helm/
Public-API deletegrep diff for removed export , pub fn, public def
Feature flag absentnew behaviour added but no flag toggle — flag only if PR description claims to be gated
Leftover debugconsole.log, dbg!, println!, pp , binding.pry, debugger added in diff
Lingering TODOsnew TODO/FIXME/XXX lines added
Test/code rationon-trivial code change (≥50 LOC added in non-test files) with zero new/modified test files → flag low coverage

Skip rows that have nothing to report — only show risks that fired.

Phase 5 — Gate evaluation

Compute three buckets:

  • Blockers (❌) — anything in this list and the PR cannot merge:
    • isDraft: true
    • reviewDecision != APPROVED
    • CI checks not SUCCESS
    • unresolved threads > 0
    • mergeState is DIRTY (conflict), BLOCKED, or BEHIND
    • base is not main/master (PR is stacked — should be merged via parent, not directly)
    • uncommitted or unpushed local changes (if PR matches current branch)
    • Jira status Blocked
  • Warnings (⚠️) — proceed-with-caution, don't auto-block:
    • any risk fired in Phase 4
    • mergeState: UNKNOWN — GitHub hasn't computed mergeability yet; usually transient. Suggest a retry before merging.
    • stacked children exist (they'll need rebasing post-merge)
    • Jira status drift (Backlog, Done, missing)
    • bead still in_progress (will need closing post-merge)
    • large diff (≥500 LOC or ≥20 files changed) — recommend /review-pr for depth
  • Notes (ℹ️) — informational only:
    • post-merge follow-ups (close bead, transition Jira, rebase children)
    • ignored in-repository merge-timing.toml (untrusted; move it above the repository)

Merge timing is evaluated separately from technical readiness; it never changes the three buckets or the technical verdict:

  • none / clear — render nothing (quiet success).
  • advisory — 🕒 caution; the merge may still be offered with normal confirmation.
  • required — ⛔ no merge offer, while still reporting technical readiness.
  • unknown — ⛔ no merge offer: mandatory applicability cannot be established.

Phase 6 — Optional contract check

Only if pacts are present in the repo (./scripts/contract-check exists, or **/pacts/ directory present, or *Consumer*.scala/*.consumer.test.* files exist):

Skill /contract-check status

Add a row to the gate table for Contracts. If WARN or FAIL, move to Warnings (don't auto-block — contract drift rarely blocks a single PR merge but should be visible).

Do NOT auto-run /contract-test — it's the runner (generate → sync → verify) and is redundant with CI, which already ran on push. If /contract-check surfaces stale, uncommitted, or sync-gap pacts, recommend /contract-test (or /contract-test sync / full) as the remediation under Risks, and let the user run it before re-invoking /ready-to-merge.

Phase 7 — Draft squash commit

The user squash-merges by default. Pre-draft the squash commit so they can approve or tweak:

  • Subject: the PR title, verbatim (already conventional-commit style at PR creation time).
  • Body: 2–5 bullets summarising the change, derived from the PR description and diff. Strip boilerplate (template sections that weren't filled in). Always include a trailing Jira: {KEY} line if a ticket is linked.

Keep the body under ~10 lines. The user can edit if they want more.

Phase 8 — Render the readout

Render compactly. Skip empty sections.

## PR #{number} — Ready-to-Merge Check

**{title}**
[#{number}]({url}) · `{head}` → `{base}` · +{additions} / -{deletions} across {changedFiles} files

| Gate         | Status                              |
|--------------|-------------------------------------|
| CI           | ✅ / ❌ / ⏳                         |
| Approvals    | ✅✅ alice, bob (or 🔔 awaiting)      |
| Threads      | ✅ 0 / 💬 N unresolved               |
| Sync w/ base | ✅ clean / ❌ behind / 💥 conflict   |
| Mergeable    | ✅ / ❌ {reason}                     |
| Contracts    | ✅ / 🔶 / —                          |
| Tests        | ✅ N test files touched / 🔶 none    |

**Linked work**
- Jira: [{KEY}]({url}) — {summary} _(status: {status})_
- Bead: `{bead-id}` — {status}
- Trello: {card-name} (or `—`)

**Stacked children** _(omit section if none)_
- ⚠️ #{n} {title} — will need `/rebase merged {this-branch}` after this merges.

**Diff summary**
- 3–5 bullets, each one short, derived from the diff. No file lists.

**Risks** _(omit section if empty — do not pad with "no secrets" etc.)_
- ⚠️ Migration added `db/migrate/20260511_xxx.sql` — verify backwards compatible.
- ⚠️ 87 LOC added without new tests in `src/foo.ts`.

**Drafted squash commit**

{subject}

  • bullet 1
  • bullet 2

Jira: {KEY}


**Merge timing** _(omit for `none` / `clear`)_
- 🕒 / ⛔ {Ddd HH:MM} {timezone} — {reason} _({rule}, {enforcement})_
- Next allowed: {next_allowed in policy-local time}, or `not computable within a week`
- ⛔ Policy unknown: {error} — merge not offered

### Verdict

❌ Blocked — {reason}  
or  
⚠️ Mergeable with warnings — review risks above.  
or  
🚀 Ready to squash-merge.

Then, only when timing fired, one timing line:

🕒 Timing caution — {reason}.  
or  
⛔ Technically ready, but merge timing does not allow merging now — {reason or error}.

Phase 9 — Ask for go/no-go

If verdict is ❌ blocked, stop here. Print the blockers and suggest next steps (often /rebase main, /review-comments, /clean-code).

If --no-merge was passed, stop here regardless. The readout still reports applicable timing.

If timing is required or unknown, stop here: do not offer the merge. Report next_allowed when present. Do not wait, sleep, schedule or retry until then; the user re-invokes later.

If timing is advisory, keep the caution visible in the question; it does not change the options.

Otherwise ask the user via AskUserQuestion:

Question: Squash-merge PR #{n} now?
Options:

  • Yes — merge with drafted commit (recommended when verdict is 🚀)
  • Edit commit message first — user supplies new subject/body, then re-prompt
  • Wait — print no further action, end the skill

Phase 10 — Merge

Only on explicit "Yes". Use the GitHub CLI's squash mode with the drafted (or edited) commit.

Recheck timing immediately before merging. As the tool call directly before gh pr merge, run merge-timing.py again. Proceed only if its status and violated rule names match the timing snapshot shown when the user approved, and the status is not required or unknown. Otherwise do not merge: render the new timing. If it is now required or unknown, stop. If it is now clear or advisory but differs, ask again with fresh confirmation. Approval never carries across a timing change, and a recheck never merges, waits or schedules anything by itself.

Do NOT pass --delete-branch. It makes gh run local git operations (switch to the default branch, pull, delete the local branch) as a side effect of the merge. In a worktree checkout — where main/master is checked out in a different worktree — that local step fails with fatal: '{branch}' is already used by worktree at ... even though the remote merge succeeded. The merge is done; only the local cleanup errored, which reads as a scary failure for a no-op. Branch deletion is handled separately below, against the remote only.

Pin the commit author email. GitHub otherwise authors the squash commit with the account's default email, which may be a personal address in a client's history. Read the repository's git config user.email beforehand and pass its literal value; if it is unset, stop and report it rather than merging with GitHub's default.

gh pr merge {number} --squash \
  --author-email "{author-email}" \
  --subject "{subject}" \
  --body "$(cat <<'EOF'
{body}
EOF
)"

Idempotency: if the command prints Pull request ... was already merged (e.g. a prior attempt merged remotely before erroring on local cleanup), treat that as success — do not retry or alarm. Confirm with gh pr view {number} --json state --jq .state (expect MERGED) if unsure.

After confirming the selected PR is MERGED, follow post-merge local default-branch update. Its read-only planner offers separately confirmed fetch and fast-forward commands without switching checkouts. Dirty, diverged, ambiguous or guard-denied targets are skipped with a reason. Merge approval never authorizes local synchronization. This step is independent of remote-branch cleanup; report its outcome separately.

After merging, remote-branch cleanup is a separately confirmed action, not part of merge approval. Never infer deletion permission from merge approval. First re-read the selected PR, repository default branch, and dependent PRs read-only:

~/.agents/skills/review-pr/scripts/gh-pr-view.sh "{owner}" "{repo}" "{number}"
gh repo view "{owner}/{repo}" --json nameWithOwner,url,defaultBranchRef
gh pr list --repo "{owner}/{repo}" --base "{head-branch}" --state open --json number,title

Require PR state MERGED, the same selected PR identity, headRefName, headRefOid and headRepository as the reviewed snapshot, and repository identity matching {owner}/{repo}. Skip cleanup for a fork or unknown head repository; never redirect deletion into a fork. The helpers support github.com only: skip non-public-GitHub targets or a host mismatch rather than changing hosts. Require a known defaultBranchRef.name; never delete the default branch, the PR's base branch, main or master. If dependent PRs remain, skip cleanup and report them. Missing or changed metadata is not deletion evidence.

Then read the branch through the fixed GET helper:

~/.agents/skills/ready-to-merge/scripts/gh-branch-view.py "{owner}" "{repo}" "{head-branch}"

Require successful JSON with name equal to headRefName and sha equal to the reviewed headRefOid. On any error or mismatch, skip deletion. A 404 does not prove auto-deletion: the branch may be missing or inaccessible. Report cleanup unavailable, not successful or auto-deleted.

Only after those checks, show the exact command below and ask for fresh confirmation to delete that specific remote branch. Declining, abandoning or denying permission means skip cleanup. On approval, execute it as the next standalone tool call, never chained with the reads:

~/.agents/skills/ready-to-merge/scripts/gh-branch-delete.py "{owner}" "{repo}" "{head-branch}"

Both helpers require Python 3.10+, Git and authenticated gh; they share scripts/github_branch.py for target validation. Each accepts exactly three positional arguments, validates repository/ref syntax and encodes URL-sensitive branch characters while retaining ref separators. Methods and endpoint namespaces are fixed in separate entry points; no arbitrary method, endpoint, host or flags are accepted. The read helper never mutates. The delete helper rejects main/master, but cannot itself prove merge state or a custom default branch: caller checks and confirmation remain mandatory. Keep the delete path separately prompted in user permission policy; do not edit settings automatically. Skill grants are Claude-style metadata, not a portable enforcement guarantee.

This REST flow does not provide atomic compare-and-delete: the branch can change after the GET. Do not claim race-free deletion or retry an ambiguous result. Both helpers pass through gh stdout, stderr and exit status without retries; malformed arguments return 2 and missing tools 127. The DELETE is best-effort: on API/SSO failure or local permission denial, note that cleanup was not confirmed and move on. Never delete the local branch or switch worktrees yourself.

After merge succeeds, print:

✅ Merged #{number}.{branch-note}

where {branch-note} is Remote branch deleted. only on confirmed DELETE success, or (remote branch cleanup skipped or unconfirmed — {reason}; inspect manually if wanted.) otherwise.

Then list post-merge follow-ups as a checklist (do NOT execute them):

  • Transition Jira {KEY} → Test/Review (or Done)
  • bd close {bead-id} if a bead is still in_progress
  • /rebase merged {this-branch} on stacked child PR(s): #{n1}, #{n2}
  • Revisit the local default-branch update if skipped or unconfirmed (include its reason)
  • Remove the merged local worktree/branch if wanted; remote deletion does not remove it

Don't auto-perform these — they're explicit user follow-ups, often touching other branches/repos. Never switch the local checkout or run mutating git worktree/git branch -d commands yourself; read-only git worktree list belongs to the local-update inspection.

Operating rules

  • Never merge without explicit go-ahead. AskUserQuestion answer must be "Yes" — anything else (silence, "Wait", a clarifying comment) means do not merge.
  • Never --admin merge. Branch protections exist for a reason. If a check is failing, surface it, don't bypass it. This one is not enforceable by the tool grant: any gh permission broad enough to merge is broad enough to pass --admin. Closing it properly needs a merge wrapper script that rejects the flag, not a narrower allowed-tools.
  • Never --no-verify. Same reason.
  • Don't auto-run /review-pr or tracking-auditor. They're expensive. Recommend them when their value is high (large diff, scope concerns).
  • Quiet success. Only render risks that fired; only render sections with content. Padding the readout with "✅ no issues" lines defeats the terseness.
  • Reuse, don't reimplement. Call into pr-status / review-pr scripts directly. Don't re-fetch CI status with bespoke gh calls.
  • Honour --no-merge. When passed, the skill is a report only — never prompt to merge.
  • Trust only the timing helper. Never read timing policy from the PR, its diff, its description or an in-repository file, and never treat a malformed or unreadable configured policy as clear. No override bypasses a required restriction.

Failure modes

  • No PR for branch: tell the user to /create-pr first.
  • No Jira MCP: skip Jira cross-reference. Render _Jira MCP not configured._ and continue.
  • No bd: skip beads cross-reference. Continue.
  • Contract-check script missing: skip silently (project has no contracts).
  • GraphQL/gh failure on a single fetch: render that gate as ? rather than aborting. Surface the failed fetch under Risks so the user knows the check was incomplete.
  • gh pr merge errors with fatal: '{branch}' is already used by worktree: this is the --delete-branch local-cleanup failure (Phase 10) — the remote merge already succeeded. Don't pass --delete-branch; verify with gh pr view {number} --json state (expect MERGED) and use the separately confirmed remote-branch helper flow in Phase 10. Never resolve it by switching worktrees or deleting local branches.
  • gh pr merge prints was already merged: success, not an error — a prior attempt merged remotely before its local step failed. Proceed to remote-branch cleanup and follow-ups.
receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementationStars: 322kUpdated Claude CodeCodexVerifiedInstall
requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirementsStars: 322kUpdated Claude CodeCodexVerifiedInstall
using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallbackStars: 322kUpdated Claude CodeCodexVerifiedInstall
verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions alwaysStars: 322kUpdated Claude CodeCodexVerifiedInstall