clavity

Use when a spec, plan, or other high-leverage artifact is drafted or materially edited and needs an adversarial multi-seat review before it is acted on. Convenes a team panel of distinct expert seats, each hunting a different defect-class, over one or more rounds; escalates high-leverage artifacts to the live agy peer; folds verified findings; emits a PANEL VERDICT.

Use ONLY before declaring a plan or implementation COMPLETE - never on routine intermediate commits. Runs a convergent, rounds-until-green adversarial review of the already-COMMITTED code and tests, not the plan artifact: the peer cites file:line, the driver measures every finding before folding. A hard round cap plus human-adjudicated GREEN gate the completion claim. Ends with one ASCII [VERDICT] token.

Use when facing a design, scope, approach, or sequencing fork in subproject work (typically at the brainstorming approaches step). Runs a divergent, review-only consult of the live agy peer under forcing functions, verifies every factual claim by measurement before folding, negotiates on material disagreement, and ends with one ASCII [VERDICT] token. Best-effort prompt-discipline, manually invokable; auto-fire is added separately.

Use ONLY after the AGY-CAPSTONE gate has passed - a GREEN or an owner round-cap waiver - and before declaring a development branch done; never mid-implementation. Audits the committed TEST SUITES for coverage exhaustiveness: untested behaviours, vacuous assertions, missing edge cases, each gap verified by measurement. Asks what the capstone does not: would the tests catch the next regression? Ends with one ASCII [VERDICT] token.

Use to drive a paired agy peer via the clavity-ls MCP tools (agy_look / agy_status / agy_ask) - when to look vs ask, write/quota semantics, handling waiting/modal results, and convening multi-lens review panels.

Orientation for agy when it is paired with a Claude peer over the clavity-ls Language Server - keep one active conversation, avoid blocking modals, and emit precise, parseable output a programmatic reader can consume.

Use when an agent notices a defect, tool misbehavior, or operational blocker that is NOT the task at hand, and when triaging the anomalies already captured. Capture is one appended line; triage promotes each entry to a tracked item or deletes it with a reason.