Business-requirements agent — authors business-rules.md (a BR-NN decision table) and golden-scenarios.md (SCEN-NN), the domain logic and worked examples that pin down "correct" before anyone builds. Runs standalone, as the driver of /sdlc-rules (Phase 1), and composed by /sdlc-evals and the /sdlc-review Bizreq lens.
Design-discipline agent for conversational surfaces (the voice and chat channels) — authors the user journey, surface layout (turn-script or message-flow), and the voice/chat interaction contract. Runs standalone, as the /sdlc-experience designer for voice/chat, or as the /sdlc-review Design lens (Phase 2).
Data-discipline agent — authors the PII-classified data-contract, the data-readiness assessment (advisory), and the lineage-audit for a feature or spec. PII is a risk-tier driver; readiness gaps become decision-log items. Runs standalone, as the driver of /sdlc-data, or as the /sdlc-review Data lens (Phase 2).
Cross-document analysis for discovery — finds contradictions across an intake corpus and generates the pre-workshop question list. Runs standalone or as Phase 0 Step 0d.
Product-discipline agent for the featuring hop — decomposes one epic into channel-aware features and specs, drafts feature-brief.md, proposes the decomposition and risk tiers, and opens decision-log items. Runs standalone or as the driver of /sdlc-feature (Phase 1).
Design-discipline agent for web/visual surfaces (the ag-ui channel) — authors the user journey, surface layout (screens), and the channel-interaction-spec that is the AG-UI event contract. Runs standalone, as the /sdlc-experience designer for ag-ui, or as the /sdlc-review Design lens (Phase 2).
System-design and architectural decisions: new services, cross-cutting changes, scaling and data-model calls, public API contracts. Use for HIGH-risk design work BEFORE a spec is built against it. Produces a design with trade-offs and a recommendation — it does not write code.
Fixes build, compile, type, and lint errors with minimal diffs. Use PROACTIVELY (auto-spawn) the moment a build or lint step fails. Gets the build green fast — no architectural changes, no scope expansion. Pairs with the Stop hook, which won't let a turn end on a broken build.
Root-causes a failing test, exception, or wrong behavior via hypothesis elimination, in an isolated context. Use when the cause isn't obvious. Returns a distilled diagnosis + a minimal reproduction — it does not apply the fix (that happens under the bug-fix procedure).
Grades a change against its spec, check by check, and posts an advisory verdict. Use before a PR is approved, or invoke locally before pushing. The grader NEVER grades its own work — it runs in a fresh context and did not write the code.
Produces an implementation plan for a spec before code is written. Use for any MEDIUM or HIGH spec, or any feature touching 5+ files / a new service. Returns a sequenced plan with files to touch, order of changes, risks, and unknowns — it does NOT write code.
Reviews changes on gated paths or HIGH-risk specs for security defects. Use PROACTIVELY when a change touches auth, identity, payments, PII, migrations, infra, the pipeline, or prompt/model/tool definitions. BLOCKS on a HIGH-severity finding; advises otherwise.
Reviews user-facing Angular changes for UX quality: every state exists (loading, empty, error, success), accessibility basics hold, components follow the codebase's existing patterns, and the flow works end-to-end. Angular-aware — also catches the framework's own failure modes. Use PROACTIVELY when a change adds or…