Babysit PR(s) to merge

Drive PR(s) open → merged. Handle transient failures only: rerun flaky CI, enqueue when green + approved, re-enqueue after flaky merge-queue removal. Real failure → hand back to human.

Starts after review. Reviews are a human gate — never wait on them. Normal PR must be approved before entering the loop; unapproved → report not-ready, stop. Backports exempt (bot approves them, step 2).

Owns its own loop. Invoke once; tends PR(s) across turns until all merged. No /loop wrapper.

Acts (reruns, enqueues) — unlike ci-fix-failure, which only diagnoses. Never edits code, addresses comments, masks a failure, or lowers the bar to force a merge.

Inputs

PR numbers in camunda/camunda or URLs. Shapes:

• Single: babysit #12345.
• Merge + backports: original + backport stable/X.Y PRs from the bot. Pass the set, or pass

the original and discover backports via gh pr list --repo camunda/camunda --search "<orig> in:body is:open" or the bot's linked-PR comment.

No PR → ask; don't infer from context. Non-camunda/camunda → ask human for confirmation, to assume monorepo conventions (GitHub Merge Queue, backport-action).

Prerequisites

• gh auth status works against camunda/camunda; CWD = repo root.

Hard rules

1. Rerun/re-enqueue = unblock, not fix. Transients toward main/stable aren't acceptable

long-term; real fix is hardening (retries-in-timeout, caching, runner sizing). Note every rerun; a check that flakes repeatedly → stop retrying, surface as a flake to fix.

1. Never edit code/config/workflows to pass CI. Code fix = author's call → report, stop that PR.
2. Never bump a timeout, disable a test, or bypass the queue. Timeouts are contracts; bypass is

incidents-only.

1. A passing rerun does NOT clear a Flaky Test Gate alert — sticky

(docs/monorepo-docs/flaky-test-gate.md). Needs a code change or ci:flaky-test-bypass label — author's call.

1. Cancelled = real until proven otherwise. Fetch annotations first; usually GHA timeouts.
2. Cap the loop. Rerun any check ≤ 3×; re-enqueue after flaky removal ≤ 3×. main/

stable paths: escalate sooner.

Flaky (rerun) vs real (stop)

Classify before every rerun. Unsure → real, stop.

Rerun (transient): known flake (dashboard); infra noise (network, registry, runner provisioning); failure unrelated to diff AND green on target; merge-queue failure on the temp branch, not PR head.

Stop (real): compile/assertion/lint/spotless/license failure in touched code; same check fails again after rerun; BEHIND/CONFLICTING (author rebase); merge-queue removal from a concurrently merged dependency (ci.md §6); Flaky Test Gate alert.

Non-obvious failure → delegate diagnosis to ci-fix-failure, act on its verdict. Don't reimplement log-reading.

Loop

One row per PR. Repeat passes until every PR terminal, then report and stop.

Terminal: MERGED ✅ · not-approved normal PR (needs review) 🕓 · needs-author ⛔ (real failure, conflict, BEHIND, Flaky Test Gate) · cap hit 🔁✋.

Each pass (per non-terminal PR)

1. State

bash
gh pr view <pr> --repo camunda/camunda \
  --json number,title,state,isDraft,mergeable,mergeStateStatus,reviewDecision,autoMergeRequest,headRefName,baseRefName,statusCheckRollup

• MERGED → ✅.
• isDraft → bot's conflict-backport (committed conflict markers) or not-ready PR. Don't resolve

unless asked → ⛔.

• CONFLICTING / BEHIND → author rebase (don't rebase their branch unless asked) → ⛔.

2. Approval gate (reviewDecision) — check before any CI work

• Normal PR, not APPROVED → not ready to babysit. Report 🕓 needs-review and mark terminal.

Do NOT rerun CI or loop waiting for review — reviews are a human gate.

• Backport → bot auto-approves + auto-merges once green; no human gate. Proceed.
• Approved → proceed.

3. Checks (statusCheckRollup / gh pr checks <pr>)

• Running → skip this pass.
• Failed → classify. Transient & under cap → rerun failed jobs only:

bash gh run rerun <run-id> --failed --repo camunda/camunda Log "rerun N/3 of <check>". Real or cap hit → ⛔ + link.

4. Enqueue (normal PR, green + approved)

bash
# = "Merge when ready": adds to the branch's merge queue, which applies its own
# configured merge method. Author squashes commits before enqueueing (CONTRIBUTING.md).
gh pr merge <pr> --repo camunda/camunda --auto

autoMergeRequest non-null = queued. Backports self-merge via the bot — don't enqueue one unless automation clearly stalled and the engineer asks.

5. Merge queue — re-enqueue on flaky removal

Queue builds temp branch (target + PR) and reruns CI. On removal, inspect:

bash
gh run list --repo camunda/camunda --event merge_group --branch <baseRefName> --limit 10

Flaky on temp branch & under cap → re-enqueue (repeat step 4) — sanctioned retry (CONTRIBUTING.md). Real → ⛔.

Between passes

Don't busy-wait. When every non-terminal PR is just waiting on CI, block until CI moves, then next pass. Background poll → re-invoked on change:

bash
while :; do
  for pr in <pr-list>; do
    gh pr checks "$pr" --repo camunda/camunda 2>/dev/null | grep -q pending || exit 0
  done
  sleep 300
done

Floor of a few minutes. Stop the loop once all PRs terminal.

Reporting

Live status line per PR:

#12345 (main)        📥 enqueued
#12346 (stable/8.7)  🔁 reran unit-tests (flake 1/3)
#12347 (stable/8.6)  ✅ merged

Final summary: what merged, what needs the engineer, any check that flaked repeatedly (hardening candidate, rule 1). Link runs by URL; code refs use the repo's stable-permalink convention.