workflow-plugin

Mechanical 1-2-file edit worker. Bumps a vendor pin, adds/removes a label, regenerates a CI matrix, fixes a single typo, wires a new doc into a TOC, or renames a symbol with bounded callers. No design, no abstractions, no test authoring beyond a smoke test. Frequently nested as a child by other agents to offload mechanical work.

Implementation role for one type:plan leaf (or a Maintenance / Integration / Bug leaf). Branches from main in an isolated worktree, implements the declared scope, adds the named Rust tests, opens a PR. Drives review-response cycles on its own PR via /review-respond.

Design role for one design spike. Re-derives aggregates, public interfaces, persistence schemas, hot-reload contracts, class structure, test cases. Authors spec sections under specs/<ctx>/SPEC.md and ADRs under specs/decisions/*.md. Owns its PR end-to-end via /review-respond.

Planning role for one breakdown / planning / testing spike. Decomposes filled specs into testable user-stories, leaf plans, and E2E .glibre-trace files. Authors TDD test lists, combines capability primitives into plan bodies, sets the Definition of Done. Owns its PR end-to-end via /review-respond.

Product role for one body of work. Elicits requirements, defines features and scope, prioritizes the backlog. Authors initiatives, epics, sub-epics, user-stories with personas, acceptance criteria, and phase/points labels. Writes spec stubs (§1 Purpose, §2 Ubiquitous Language, §11 Acceptance) when no design exists yet. Does not write implementation code or ADR design body.

Generalized review-respond loop for any artifact — code PR, design PR, plan issue, doc PR. Loops `review` → original-author re-dispatch on the same artifact until the reviewer verdict is APPROVE, or until the loop escalates to an `[SPIKE] iterate-*` issue. Replaces /review-loop. Author bucket adapts to artifact type. CI gate fires only for code-pr.

Effort/model modifier — boosts the next agent dispatch (or current reasoning turn) to opus + xhigh + extended-thinking. Use when reasoning depth is the blocker — a hard root-cause question, a recurring blocker, an ambiguous spec point, a refutation pass before pushing back on review, a "should I split this scope" judgement call. Pins the ranked-hypothesis output contract. Callable from chat or from inside any role agent.

This skill should be used when the user asks to "advance the plan", "pick up unblocked items", "what should I work on", "do the next thing", "execute the next leaf", or any variation of "make progress on the backlog". Picks ONE unblocked leaf issue, dispatches a single foreground role agent (product / design / plan / code / test / review / chore), blocks until done. Caller-decides parallelism — invoke /work multiple times in one chat turn for parallel leaves; each /work itself stays single-leaf.