Cut auto-rebase fan-out thrash: restrict to review-ready PRs + instrument the agentic-conflict-resolution rate

[github-actions]

Problem

The Auto-rebase non-Dependabot PRs workflow fans out on every push to main, calling update-branch on every behind non-Dependabot PR. With the repo's measured throughput (~12.3 merges/day, ~14-18 concurrently-open PRs), this produces an estimated ~170-220 auto-triggered branch-update CI runs per day — the overwhelming majority of which are immediately re-staled by the next merge before a human reviews them (discussion #735, owner findings comment). Time-to-merge shows the matching tail: median 5h but mean 31h, p90 71h.

The owner's recommendation has two tiers: an immediate, free, plan-independent mitigation (stop fanning out to PRs nobody will look at — restrict to review-ready PRs), and a plan-gated follow-on (GitHub Merge Queue, which requires Team/Enterprise for private repos and would trade away the unique agentic conflict-resolution path). Before that trade can be evaluated, the agentic-conflict-resolution frequency — which the owner could not count because the trigger is an un-indexed HTML-comment sentinel — must be instrumented.

Scope (this epic)

1. Instrument auto-rebase: measure the agentic-conflict-resolution (sentinel) rate and the fan-out CI-run volume, on a recurring report.
2. Restrict the auto-rebase fan-out to a review-ready eligibility predicate instead of all behind PRs (the free mitigation).
3. Record the measured before/after outcome and a documented Merge Queue go/no-go gate as an initiative decision record.

Out of scope (plan-gated follow-on)

GitHub Merge Queue adoption itself — it requires a GitHub Team or Enterprise plan for private repos and the loss of agentic conflict resolution + proactive background branch updates. That is a billing + product decision for a human, surfaced as an open question, not decided here.

Success metrics

• Fan-out reduction: auto-rebase-triggered branch-update CI runs drop by >= 50% from the ~170-220/day baseline (discussion Auto-Rebase Workflow vs. GitHub Merge Queue #735), measured by the new observability report before vs. after the restriction lands.
• Conflict-rate established: the agentic-conflict-resolution rate (auto-rebase conflict sentinels fired / week, and dev-lead rebase intents emitted) is published as a concrete weekly number where it was previously unmeasurable.

Cost cap

Zero new paid spend. Everything in this epic stays on the free GitHub plan and within existing Actions minutes; adopting Merge Queue (which would require a paid plan tier) is explicitly excluded. The new observability report adds at most one scheduled run per day (mirroring daily-pr-review-health.yml / token-report.yml) and introduces no new agent/LLM token cost beyond an optional report-summary call bounded to a single run.

---

Planned from idea discussion #735 by the BMAD Scrum Master initiative-planner. Inert until a maintainer adds initiative:auto.

[don-petry]

Open questions resolved (owner decisions)

Recording maintainer decisions on the three epic open questions so the initiative can proceed.

1. Merge Queue plan eligibility — ✅ RESOLVED

GitHub Merge Queue is available on this repo's plan (the repo is public, and public repos have the feature). The plan-tier blocker noted in the epic does not apply. The Merge Queue go/no-go itself remains a human decision recorded later in Story #739 — this only removes the eligibility uncertainty.

2. Review-ready predicate (Story #738) — ✅ RESOLVED (recommended)

Predicate: update a behind PR only if it is non-draft AND (has an APPROVED review OR carries an auto-rebase:ready label), exposed as a tunable workflow input so a "front-of-queue N" variant can be swapped in without a code change.

Rationale:

• Targets exactly the auto-merge-on-approve PRs that branch protection will block when behind (scripts/dev-lead-intent.sh:253) — the update is necessary work, not churn.
• isDraft==false reuses the org's existing "consider this PR" gate (scripts/list-prs.sh --draft=false).
• auto-rebase:ready (namespaced per the initiative:auto / dev-lead:hands-off convention) is the human escape hatch for keeping a not-yet-approved PR warm.
• Preserves the merge-method update (approval-safe) and the <!-- auto-rebase-conflict: sentinel path.

Implementer note: read the actual APPROVED review state (as scripts/repair-pr-approvals.sh does), not only reviewDecision, which can be null when required-reviews aren't configured but bot advisory approvals exist.

3. Cross-repo access for Story #738 — ⏳ IN PROGRESS

The fan-out logic lives in petry-projects/.github (auto-rebase-reusable.yml), outside this repo's automation reach. A dedicated implementation issue is being filed in that repo carrying the predicate spec above. Story #738 stays dev-lead:hands-off / initiative:hold here and is tracked by that central-repo issue.

---

Generated by Claude Code