Roster

Validate and fact-check another agent's output against evidence and the task's success criteria. Use when: verifying a report, fact-checking claims, cross-checking research findings, confirming a result meets its criteria. Not for browser/E2E testing (see e2e-runner).

Procedure the Researcher agent follows to answer a question with web search + page fetch: plan queries, search, open the promising pages, cross-check, and return a source-cited synthesis. Use when a task needs current or external facts.

Example skill scaffold. Replace this with a real, project-specific procedural playbook. The description must clearly state WHEN to use the skill (trigger phrases and scope) so the model can match it to user intent. USE FOR: <list the situations this skill applies to>. DO NOT USE FOR: <list out-of-scope situations and point to the right skill>.

The Planner's operating procedure. Use when: turning a principal goal into a Plan; revising a plan after rejection; dispatching ratified tasks; supervising sub-agent results; deciding whether to convene the Council; finalizing a run.

Classify a proposed action's risk tier (T0–T4), apply the gate policy, and (if required) summarize the proposal into a human-reviewable approval request. Use whenever an agent is about to perform an action that touches anything outside its own scratch state.

Append a structured event to the run's provenance log. Every agent calls this for every meaningful state transition (plan drafted, task dispatched, action proposed/executed, approval decided, etc.) so the run is fully replayable and auditable.

Implement one task on a feature branch. Use when: the Planner has dispatched a `task_assignment` to the Coder role; reading existing code; producing a diff; running a local build.

Run E2E test suites using playwright-cli. Use when: running test suites, executing smoke tests, running E2E/bug-bash tests, generating test reports, testing UI with plain-English test definitions. Dispatched after the Coder lands a change.

Review a diff against a task's success criteria. Use when: a `task_assignment` for role `reviewer` lands; you need to lint, summarize a diff, or produce structured review comments.

Run a cross-model deliberation: pose the same question to multiple model bindings (e.g. Claude, GPT, Gemini), have them propose independently, then critique each other, then synthesize — with explicit anti-conformity safeguards. Use ONLY for high-uncertainty / high-stakes decisions; not for every step.