saman-mb

document

Write or refresh documentation that actually works — the technical-writer drafts it from the real code, then a fresh reader agent follows the steps against the repo and must reach the stated result. Loops until the docs are drift-free and completable.

fix-bug

Fix a bug the honest way — reproduce it as a failing test first, root-cause it, apply the minimal fix, and prove it with the test flipping red→green while the suite stays green. Worktree-isolated, CI-gated, opens a PR.

harden

Security-harden a surface — threat-model it, rank the findings, and re-review until every finding has a fix or an explicit accepted-risk note (and secrets/dependency scans are clean). Read-only by default — it reports; remediation happens on a branch, opt-in.

migrate

Run a mechanical migration across a whole codebase — discover every call site, transform each in isolation, verify per-site, and gate on a clean sweep (no old-pattern remnants) with the suite green. For API/dependency/pattern/framework migrations.

onboard

Read an unfamiliar repo and write the agent-facing context file every other order depends on — conventions, commands, boundaries and the quality bar, proven by running them. Gated on a fresh agent answering the crew's real questions from the file alone.

plan-epics

Turn a brief (or several) into a tracked backlog — GitHub epics + linked, labelled user stories, with a product-manager subagent authoring each epic's stories in parallel.

refactor

Change the shape of the code without changing what it does — pin current behaviour in characterization tests first, transform, then prove equivalence by those tests passing unmodified and no existing test being deleted, skipped or loosened.

release

Cut a release safely — assemble the changelog from what actually merged, bump the version, gate on green CI at the tag, run an SRE pre-flight (rollback + migration safety), then tag and (opt-in) publish.

ship-issue

Take one or more GitHub issues/stories from open → reviewed PR (→ merged, opt-in) autonomously — worktree, subagent build, CI gate, specialist acceptance board, follow-up issues.

spike

De-risk an open technical decision with a time-boxed spike — frame the question, prototype several approaches in parallel as throwaways, judge them against the project's constraints, and return a recommendation as an ADR. Produces a decision, not production code.