mhpdigital/kerbe

Use when a bug is reported or a change touches a data model, shared guard, migration or data flow — runs the impact analysis that finds every downstream consumer before the fix is written, so one commit closes the whole class instead of three chasing it.

Use when you need to know whether a slice is finished — before implementing (does the plan task everything the design and spec promise?) or before sign-off (is everything promised actually built and wired?) — or when the user asks what is missing, whether anything was dropped between design, spec, plan and code, or for a to-build inventory.

Use when fetching or analysing a slice's Figma design — grading it against the handoff gates, extracting UI elements with measurements, comparing the design against UI_ELEMENTS.md, filling a slice's Design-sources block, or checking design provenance and freshness of built UI.

Use when a slice's plan draft lists open decisions, or its spec docs carry decisions nobody has made — runs the grilling rounds that put them to the person who owns them, then writes the rulings into DECISIONS.md and the spec docs, and — when the plan is already frozen — into PLAN.md as a dated amendment. kerbe's Specify step, between the plan draft and the freeze.

Use when a slice's frozen plan is ready to build — resolves the workspace, derives the live progress tracker from the plan, dispatches one isolated worker per task, and gates each task on real verification evidence. Also runs remediation from a coverage fix list.

Use when a slice's spec docs and design are in place and the next step is a task-by-task TDD implementation plan. Runs twice: first it drafts PLAN.md and lists the decisions the draft surfaced for /kerbe:grill; after grilling it freezes the plan with the answers folded in. Also writes a remediation plan for a fix list from a coverage run.

Use when a slice's diff is ready for review — categorises every changed file by risk tier, produces per-line open commands for the business-logic rows, runs an adversarial pass over its own review, and records it as a numbered QR in the slice's REVIEW.md.

Use when a recorded review needs working through with a human — walks a QR's rows one turn at a time, pre-reads and opens each item, records verdicts in place, and resumes from the first unresolved row.

Use when starting a new feature slice — create its folder, populate the spec docs from the lifecycle and stack-adapter templates, and register it in the slice index — or when asked for slice status, the slice list, or to advance a slice's lifecycle stage.