Greenlight

Open a pull request for an issue from your fork branch. Verifies the branch exists, is ahead of upstream, and (on DCO repos) that every commit carries a Signed-off-by line. Composes the PR title and body (Closes #N plus verdict and gates at open time). DRY RUN BY DEFAULT: returns the composed PR without opening it unless execute is true; opening requires GITHUB_TOKEN.

Contributor passport for a GitHub login: the claim-ledger account profile, merged PR count, honor ratio, and open claims, returned as markdown plus structured JSON. Agent-callable surface of `greenlight passport`.

Verified review brief for ANY pull request (not just your own): what changed grouped by area/risk, CI triage with the base-vs-head classifier, review findings classified real vs false-positive with evidence, test evidence present or explicitly absent, merge-readiness blockers (conflicts, unresolved change requests, missing changesets), ending with a one-line look-at-this-first. Deterministic classifiers only; unverified items are marked as such.

Maintainer-side duplicate/competing PR radar for a repo: clusters open PRs working the same issue (closing-strength links) or touching the same files (Jaccard file overlap with title-prefiltered budget). Each cluster reports PR numbers, authors, evidence, and a heuristic suggested-keep pick; unclustered PRs are counted so the picture stays honest. API budget is measured and reported.

The one-call answer to 'can I work on this issue right now': the Claim Protocol issue status (claim holder, expiry, queue, competing PRs, gates) plus the repo friction summary (how hard this repo is to land a PR in, 0-100). Call this before committing to any issue; it combines claim_status and repo_friction.

Prepare to work an issue: check/create your fork and a branch named from the issue title (learning the repo's branch convention from recent PRs when detectable), and write a local PREP.md checklist (verdict, gates, DCO reminders, next commands). Never clones. DRY RUN BY DEFAULT: no fork or branch is created unless execute is true; creating them requires GITHUB_TOKEN.

Keyword race check for one issue: find open PRs that fix the same problem WITHOUT ever linking the issue, which link-based checks (claims, timelines, X-Ray) cannot see. Extracts the issue's significant problem keywords (filenames like $types.d.ts, identifiers, distinctive terms, quoted phrases), searches and scans the repo's open PRs for overlap, and returns RACE / SUSPECT / CLEAR with per-PR evidence (number, title, author, created date, matched keywords). Field case: sveltejs/kit issue #17465 was raced by PR #17453, which never mentioned the issue. Call this before writing code for an issue …

Repo-level maintainer report: what needs attention across the whole repo? Reports PR races (two or more open PRs fixing the same issue, with first-claim arbitration), claimed-by-other conflicts (an open PR whose linked issue was claimed by someone else first), stale claims (issues claimed 14+ days ago with no open PR), and assignment-gate conflicts (repos that require assignment before a PR). Pass a --limit to bound the API budget on large repos; sections that cannot be computed say so instead of showing empty. Call this when triaging a repo, not just one PR.

Friction score 0-100 for a repo (100 = easiest to land an outside PR). Based on detected contribution gates and computed outcome signals.

Triage the review feedback on your PR: which comments actually need you? Gathers bot (CodeRabbit, cubic, Greptile, Copilot) and human review comments, classifies each finding as likely-real, likely-noise, or needs-decision with the reason stated, and drafts a reply for each. Human reviewer comments are never called noise. Drafts only; nothing is posted. Call this after pushing and getting review feedback.

Rank a repo's open issues by verdict, GO first. Starts with label-tagged good-first-issue / help-wanted issues; when the repo does not use those labels, falls back to the most recent open unassigned issues (flagged as fallback). Use this to pick a winnable issue instead of a taken one.

Campaign mode: scan repos for winnable issues, rank survivors by friction, merge history, freshness, and verdict strength, and append them to a campaign file as SCOUTED. The campaign file dedupes against everything already claimed or PR'd. Budget-aware: stops with a resume cursor when the API budget is hit.

Release-readiness check for a shipped release: confirms the GitHub release exists, the npm package version is live, the MCP Registry record is active, and the GHCR image tag is published, returning a GO / NO-GO verdict per surface. Agent-callable surface of `greenlight shipcheck`.

Reconcile local tracking with the API census: PRs the registry does not know (another agent opened them), state mismatches vs API truth, and stale claims with no PR. Reads only. For multi-agent campaigns where several agents open PRs in parallel.

Honest base-vs-head verification for a checked-out repo: runs the detected test suite at the base ref and at HEAD and reports the red-first delta (newly failing, newly passing), never raw counts. Refuses a dirty working tree unless asked. Agent-callable surface of `greenlight verify`; run it before and after a fix to prove the fix did the work.

Draft the verified-finding comment for a maintainer-only repo (modelcontextprotocol/inspector, modelcontextprotocol/servers): outside PRs are closed on sight there, so the contribution that lands is a detailed, verified issue comment. Assembles the comment from your notes: claim, environment, reproduction steps, observed vs expected, root cause, fix sketch, and a verification footer. Sections you leave empty render as explicit gaps, never plausible filler. DRAFT ONLY: nothing is posted, ever. Copy the markdown into the issue thread yourself.

Build queue in one call: filters the Agent Policy Index for winnable open repos, timeline x-rays starter issues with the existing xray engine (CLEAR / RISKY / TAKEN with competing PRs named), excludes repos with behavioral gates (maintainers who reject AI-built PRs in practice), and skips repos where you already have an open PR. Stops with partial results under the API budget floor. Use this to pick what to build today instead of hand-vetting repos.

Timeline X-Ray: deep-verify candidate issues against their GitHub timelines before writing code. Fetches each issue's timeline, classifies every cross-referenced PR with link-strength scoring, and flags reporter-owned fix PRs and comment claims. Returns GO / RISKY / TAKEN per issue with PR evidence (number, state, link strength, URL). List views hide in-flight PRs; run this on scan results first.