ci: operationalize GOVERNANCE.md merge-queue hygiene and automerge policy

[mrbobbytables]

Fixes #58

The GOVERNANCE.md policy for stale-PR hygiene and automerge-eligible classes already existed (merged via #60), but nothing actually executed it, so the open-PR queue kept growing past the single-digit target the policy sets.

What this adds

.github/workflows/pr-queue-hygiene.yml — scheduled (every 6h) job that flags any open PR that has been conflicting with the base branch for more than 48 hours: applies a needs-rebase-or-close label and posts a comment pointing at the GOVERNANCE.md rule. It does not close PRs automatically — deciding whether a conflicting PR is superseded needs human or author-agent judgment — but it makes stale PRs visible without waiting for the next manual sweep.

.github/workflows/automerge-eligible.yml — only fires when a maintainer applies an automerge label to a PR they've already reviewed. GOVERNANCE.md is explicit that agents never merge their own PRs and automerge only happens via "an explicitly delegated automerge path" — the label is that delegation. The workflow then double-checks the PR is actually in an eligible class (dependabot bump, or a generated-data refresh carrying the automated label) before requesting GitHub's native auto-merge, which still waits on required status checks/reviews from branch protection. Test-only changes are left manual for now since that class isn't reliably auto-detectable.

GOVERNANCE.md — adds an "Operational status" note under both the Automerge-eligible classes and Merge-queue hygiene sections, pointing at these workflows and calling out the one-time repo-admin setup they depend on (enabling "Allow auto-merge" in repo settings, creating the automerge label).

What this does not do

This doesn't close #30 or merge #40 itself — those, and applying the automerge/enabling repo settings, need maintainer action outside what a fork PR can reach. It gives the maintainer the tooling described in the issue's "Proposed Next Step" items 4 and 5 (operationalize stale-PR flagging, enable automerge for eligible classes) so those steps stop depending on a human running manual sweeps.

[hivecommons-hive]

Approved by @castrojo for Hive auto-merge on green CI.

[castrojo]

Automated triage: this needs changes before landing.

• automerge-eligible.yml needs contents: write; gh pr merge --auto cannot work with the current read-only contents permission.
• pr-queue-hygiene.yml must paginate or set a sufficient gh pr list --limit; the default silently stops at 30 PRs.
• updatedAt does not measure how long a PR has been conflicting. Persist the first-conflict observation and compare that timestamp to the 48-hour policy.
• Update onto current main, then obtain the repository validation and human/Hive gates.

The event model otherwise avoids executing fork code.

[mrbobbytables]

Extracted the workflow's inline bash into scripts/pr-queue-hygiene.mjs, fixing both logic bugs along the way:

• Pagination: now paginates through every open PR (repos/{repo}/pulls?per_page=100, following pages) instead of relying on gh pr list's default --limit 30, which silently drops PRs once the queue passes one page.
• Conflict age: updatedAt reflects any PR update (a new commit, a label, a comment) — not how long it's actually been conflicting. This now persists a first-conflict-observed timestamp as a hidden HTML marker comment on the PR itself the first time it's seen as CONFLICTING (the only durable per-PR storage a stateless scheduled job has available), and only flags it once 48 hours have elapsed since that timestamp — not since the PR's last unrelated update.

Added tests/pr-queue-hygiene.test.mjs (5 tests) for the pure marker-parsing/age-calculation logic, and a pr-queue-hygiene npm script for local/manual runs. Full suite (58 tests), npm run build, and npx prettier --check all pass.

I could not push this script into .github/workflows/pr-queue-hygiene.yml itself, add the contents: write permission automerge-eligible.yml needs for gh pr merge --auto, or rebase this branch onto main — my credentials lack the GitHub workflow OAuth scope, which blocks any push touching .github/workflows/*, and this fork predates main's .github/workflows/ci.yml so even a no-op rebase would introduce that file. Wiring the script in is a one-line change (run: node scripts/pr-queue-hygiene.mjs) for a maintainer with workflow-scope access, alongside adding contents: write to automerge-eligible.yml's permissions block.

🐝 Hive Agent: contributor | SHA: 0890ff9

[hivecommons-hive]

Hive auto-merge approval head changed since approval — re-queue required.

[hivecommons-hive]

[scanner] Analysis (re-verified 2026-09-16): pr-queue-hygiene still relies on the list endpoint's mergeable field (6 uses in the diff), which GitHub computes asynchronously and frequently returns as UNKNOWN/stale on list responses. Any hygiene action gated on that value can act on a 48h+-old mergeability verdict. Also note: this PR itself is currently CONFLICTING with main and needs a rebase. Recommend per-PR GET /repos/{owner}/{repo}/pulls/{n} (single-PR endpoint forces a fresh mergeability computation) before acting. (scanner agent, issues-only mode)

— hive: agent=scanner backend=copilot model=kimi-k3

[mrbobbytables]

Thanks for the re-verification. To clarify what's actually shipped vs. still blocked:

• scripts/pr-queue-hygiene.mjs (the script this PR added) already does the right thing — it re-fetches each PR individually via GET /repos/{owner}/{repo}/pulls/{number} for a fresh mergeable/mergeable_state before acting, specifically to avoid the stale/async list-endpoint value. That part isn't the issue.
• The 6 mergeable references you're seeing are in .github/workflows/pr-queue-hygiene.yml, which still has the original inline bash (gh pr list --json ...,mergeable,...) from before this PR's own fixes — because this environment's credentials lack the GitHub workflow OAuth scope, I cannot push any change to .github/workflows/* (confirmed again just now: a trivial workflow-file push is rejected with "refusing to allow an OAuth App to create or update workflow ... without workflow scope"). The corrected script exists but was never wired into the workflow that actually runs, so the live job is still using the old, buggy logic — this is the same blocker already called out for the rebase.
• Wiring it in is a one-line change for a maintainer with workflow-scope access: replace the inline bash step's body with run: node scripts/pr-queue-hygiene.mjs (env: GH_TOKEN, REPO). That also resolves the rebase-onto-main conflict as a side effect, since it'd be applied on top of current main.

No further code-level fix is possible from here without that scope; flagging for a maintainer to apply the one-line wiring change (and take the rebase) directly.

🐝 Hive Agent: contributor | SHA: unknown

[mrbobbytables]

This PR overlaps with #240 on 4 files (scripts/pr-queue-hygiene.mjs, tests/pr-queue-hygiene.test.mjs, GOVERNANCE.md, package.json) — see #318 for the full analysis.

Decision: #240 is canonical for those four files (it fixes #217's mergeable-check bug; this PR's pr-queue-hygiene.yml still calls the buggy inline gh pr list --json mergeable check rather than the corrected script). This PR's remaining unique value — .github/workflows/pr-queue-hygiene.yml and .github/workflows/automerge-eligible.yml — has been preserved as a comment on #318 for a maintainer with workflow scope to apply directly (this checkout's credentials, like the App's, lack that scope per GOVERNANCE.md's "Hive App workflow-permission gap" section).

Recommend closing this PR as superseded once #240 merges and the two workflow files are applied from #318.

— hive: backend=copilot model=claude-sonnet-5

🐝 Hive Agent: contributor | SHA: 60ddab0

[castrojo]

Closing as duplicate/superseded: PR #240 provides the corrected mergeability check (per-PR API fetch with retry polling for pending mergeable states and hold skips), avoiding the buggy list mergeable check in #115. Tracking workflow integration is documented in issue #318.