A spec-driven pipeline that makes building with AI fast, accountable, and cumulative.
Agent
Stress-tests a REQ's architecture and task plan before the human reviews them. Assumes the design is wrong, broken, or incomplete, tries to prove it, and reports only the findings that survive its own refutation attempts. Hunts what was omitted entirely, not just flaws in what was written. Read-only — reports…
Agent
Reviews code changes for layering compliance, separation of concerns, API contracts, and integration with existing patterns. Read-only — reports findings without modifying code. Dispatched by /review during Phase 4.
Agent
Explores the codebase for a given REQ: finds similar existing code, the files the change will touch (blast radius), where the new code hooks in, and what tests exist. Read-only. Dispatched by /architect and /bugfix.
Agent
Reviews code changes for logic errors, race conditions, error handling gaps, and security vulnerabilities. Read-only — reports findings without modifying code. Dispatched by /review during Phase 4.
Agent
Decides a single pipeline gate during an autonomous run — /autopilot, or /sprint adjudicating its gate queue (autonomy.gates set to assisted or auto). Reads the gate's evidence packet and returns one verdict — APPROVE, REWORK, or HALT — with a confidence score and cited evidence. Read-only on source; writes only its…
Agent
Audits the UI source for design-system problems without running the app — hardcoded values that should be tokens, too many font sizes or spacing values, duplicated components, inconsistent names, screens that disagree with each other, and code that breaks the documented design system (with the cause classified). When…
Agent
Standalone codebase health audit — tech debt, code smells, dead code, complexity hotspots, missing tests. Operates on the whole codebase, not a single REQ. Read-only. Dispatched by /analyze.
Agent
Standalone performance and cost audit — API cost hotspots, database performance issues (N+1, missing indexes, full scans), latency drivers (sync I/O, unbounded loops, blocking calls). Read-only. Dispatched by /optimize.
Agent
Runs the complete /proceed pipeline for a single REQ inside an isolated worktree. All phases sequential within this agent's own context — CANNOT dispatch sub-agents. Pauses at every gate; surfaces gate-claims to the /sprint orchestrator. Dispatched only by /sprint.
Agent
Reviews code changes for convention compliance, naming, duplication, and test coverage. Read-only — reports findings without modifying code. Dispatched by /review during Phase 4.
Agent
Self-review against the captured knowledge vault. Checks whether the new code repeats any known mistake (lessons), respects any codebase quirk (gotchas), and conflicts with any accepted decision (ADRs), and sweeps user-facing repo docs for staleness against the change. Read-only — reports findings. Dispatched by…
Agent
Implements a single task end-to-end — writes code, updates tests, verifies it runs. Reads the task spec, the surrounding code, and the vault. Drafts a commit message; does NOT commit. Dispatched by /implement.
Agent
Runtime UI/UX review of a change. Starts the app's dev server and drives a browser to confirm the changed UI renders, the flows work, the interaction states are correct (disabled/loading/error/empty, not just the happy view), and the result matches the design and UI acceptance criteria — the things static review…
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: