How do I install the Merge Train skill?

Run npx skillmds@latest add ronniepinnell/merge-train in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Merge Train safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Merge Train?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Merge Train free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Merge Train?

ronniepinnell (@ronniepinnell) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Merge Train by ronniepinnell — Agent Skill

Name: Merge Train
Author: ronniepinnell

Merge Train

Merge every open PR that is mechanically safe — CI green, approved, mergeable, autonomy label present, no unacknowledged override markers — in base-first order; everything else is reported with the exact fact it is missing. Use at end of week, before a release, or whenever open PRs have piled up. Dry-run by default.

ronniepinnell Updated Sep 17, 2026

File contents

/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)

CI green — every status check concluded SUCCESS/NEUTRAL/SKIPPED.
Approved — review decision is APPROVED, or an evidence-backed adversarial-review:passed label (step 2b).
Mergeable — no conflicts with the target branch.
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.
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

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.
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 ", 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.
On --execute: merge the SAFE set in base-first order (PRs targeting the default branch before stacked PRs). A merge failure stops the train.
Verify after: target-branch CI still green.
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.

---name: merge-traindescription: Merge every open PR that is mechanically safe — CI green, approved, mergeable, autonomy label present, no unacknowledged override markers — in base-first order; everything else is reported with the exact fact it is missing. Use at end of week, before a release, or whenever open PRs have piled up. Dry-run by default.---# /merge-train — Merge What's Proven, Report the Rest"Safe to merge" is a claim, and claims need receipts. This skill reduces theclaim to five mechanical facts and merges only where all five hold. A held PRis 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; thejudgment 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```## Procedure1. 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.

Frequently asked questions