/merge-train — Merge What's Proven, Report the Rest

"Safe to merge" is a claim, and claims need receipts. This skill reduces the claim to five mechanical facts and merges only where all five hold. A held PR is not a failure — its reason line tells a human exactly what to supply (a review, a label, a green run). No judgment happens at merge time; the judgment was banked here once.

The five facts (all must hold)

1. CI green — every status check concluded SUCCESS/NEUTRAL/SKIPPED.
2. Approved — review decision is APPROVED, or an evidence-backed adversarial-review:passed label (step 2b).
3. Mergeable — no conflicts with the target branch.
4. Autonomy label — the PR carries autonomy:green, applied by a human. The label IS the standing approval; the model executing this skill never applies it.
5. Overrides acknowledged — any JUDGMENT-OVERRIDE: marker in the PR body is matched by an override-acked label.

Invocation

/merge-train              # dry-run: SAFE/HELD report, no merges
/merge-train --execute    # merge the SAFE set, base-first

Procedure

1. Enumerate open PRs (GitHub: python3 scripts/merge_train.py ships the whole check; other forges: gather the same five facts via their CLI/API). No forge access → every PR is UNVERIFIED, reported as such — never assumed safe.
2. Classify each PR SAFE or HELD; a HELD PR lists every missing fact. 2b. Adversarial review — the review a bot can't do. Read the diff trying to find a reason it must NOT merge, framing it as an external contributor's. Findings carry file:line evidence; post them, or "adversarial pass: no blocking findings — checked <dimensions>", as a PR comment. Run it on any substantive PR (behaviour-changing; docs, lockfile bumps, generated files and pure formatting are exempt) — measured at ~15% overhead for two fabrications CI structurally could not see.
   • It does not replace fact 2. A PR whose review bot silently never ran may substitute the label adversarial-review:passed, and that label counts ONLY when a comment backs it — merge_train.py holds a labelled PR that has no evidence comment. The one route to SAFE without a real approval is the one route that must not be self-asserting.
   • Blocking findings → fix first. The label is never applied over open findings.
3. On --execute: merge the SAFE set in base-first order (PRs targeting the default branch before stacked PRs). A merge failure stops the train.
4. Verify after: target-branch CI still green.
5. End with ONE verdict line and log it to the ledger:

MERGE-TRAIN: <n> merged (#ids) | <m> HELD (<top missing fact>) | <k> UNVERIFIED | target CI: green|red

Rules

• Never widen the safe set inline. "That red check is flaky" is a judgment call — the PR waits for a human, or the check gets fixed. Loosening the five facts is a one-way-door decision (/door), not an edit.
• HELD and UNVERIFIED are never silently upgraded.

Composes with

• /refute — how a PR earns its evidence before approval.
• /gate — the five facts are a standing gate; the ledger line is its record.
• /verdict — MERGE-TRAIN: lines land in the append-only ledger, so the next train diffs against the last.