steward-os

Capture issue-shaped reports that have not reached the tracker yet. Divert suspected vulnerabilities, dedupe across every surface, and stage genuinely new actionable items in the pre-tracker queue. Never files or replies on the public tracker. Watcher role, Band A.

Independently fact-check the autonomous actions the system took (issue closes, labels, capture marks, etc.) against live ground truth, and alarm on any mismatch. Run on a schedule. This is what makes Band-A public actions safe.

Triage tracker issues — initial reply under a confidence gate, label, milestone, sprint-candidate, and the dropped-ball "reporter waiting on us" surfacer. Use for ongoing issue-tracker hygiene.

Deep-review a routed PR and run the authoritative quality gate before merge. Reads the whole diff, reproduces, runs the layered gates fresh, bounces with a fix-spec or fixes mechanical blockers. Stages 3-4 of the PR lifecycle.

Screen and route an open pull request — fit/scope screen, marginal-benefit screen, and lane routing. Use at the front of the PR pipeline before any deep review. Stages 0-2 of the PR lifecycle.

Aggregate the contributor trust ledger into a shrunk-and-decayed reliability score that weights review priority and gate intensity — never a reason to skip a gate. Band A, no LLM: classification happens upstream at close, this is arithmetic over an existing event log.

Merge an all-clear PR and ship it — rebase, merge preserving attribution, tag, deploy, verify it's actually live, close PR/issues with credit, update the trust ledger. Stages 5-6 of the PR lifecycle. Runs only after pr-deep-review's gate is all-clear.

Compute and render the living, ranked triage board of open work — mechanical + judgment dimensions into a composite rank, with evidence-readiness tags. Run on a schedule so the board is always current.

Close an issue only when its fix is provably shipped and fully resolved — a merged closing PR whose commit is in a released tag, single-surface, no still-broken pushback. Band A with watchdog, deterministic, no LLM. Under-close rather than mis-close; anything short stays open for a human.

Deterministically reconcile size/area/type labels on issues and PRs against a declared taxonomy — add-only, never overriding a human's label. Band A, no LLM. The mechanical labeling floor beneath the judgment-gated pr_labeling.

Post a templated release announcement to a fixed channel via a secret-isolating helper, gated on a last-announced marker so it fires once per release and never re-posts. Band A, downstream of release-pipeline.

The craft-level discipline for writing and editing code well — load before any non-trivial code change. Surgical edits, define-done-up-front, read-before-write, fail-loud, tests-assert-intent. The Builder's companion to pr-deep-review.

Run the setup interview to wire Steward to a specific project. Use when first installing the system, or when the project's repos/platforms/CI change. Produces config.yaml.