
Escalation when developer is stuck. Invoked by orchestrate via Task when operator asks. Diagnose root cause, implement fix. No preflight. Hand back to orchestrator when blocker fixed.
Use for UI/design work. Creates elegant, accessible, production-ready user interfaces. Execute only stages with Owner: frontend-dev.
Security review with confidence gating; FastAPI, Supabase, Next.js App Router, JWT
Break a plan, spec, or PRD into independently-grabbable GitHub issues using tracer-bullet vertical slices. Use when the user wants to convert a plan into issues, create implementation tickets, or break work into GitHub issues.
Close a spec-driven feature after all implementation repos finished — cross-repo issue rollup, PR links on spec parent issue, close PRD parent.
Issue state machine — transition GitHub issue labels per docs/agents/triage-labels.md. Batch helpers via lib/triage.sh.
Planning specialist that produces behavior-preserving refactor plan content
Optional pre-planning cache: write .research/<slug>.md (question, sources, findings, implications, stale-by). grill-me and to-prd load if present.
Verify docs and examples against current source
Minimal emergency fix from main, hotfix branch, [HOTFIX] PR
Optional teaching overlay that enriches any active workflow
Prototype code generator. Generates responsive, accessible HTML-only prototypes into .prototype/<slug>/ from design briefs. Strict Tailwind-only styling.
Performance review: DB, cache, React, Next.js; confidence + impact gating
Use when the model needs to see the UI. Image/layout reviewer for screenshots. Invoked by orchestrate only when developer, frontend-dev, or verifier explicitly request image review—not on every test run.
Failure paths: helper triggers, env blockers, loops, senior-dev escalation, review remediation, manual Task handoff.
Last-resort problem solver. Reviews failures and proposes solutions for orchestrator to convert into plan updates or developer tasks.
Planning specialist that produces design brief content for website prototypes
Compact the current conversation into a transfer document so another agent or session can continue. Use when the user asks for a handoff, session summary for a fresh agent, or context compaction before starting over.
Planning specialist that produces high-signal review plan content
Scoped feature planning specialist that produces a sub-problem report for one slice of a larger feature
Planning specialist that produces diagnosis-first debug plan content
Ultra-compressed communication mode. Drops filler and hedging while keeping full technical accuracy. Use when the user says "caveman mode", "less tokens", "be brief", or invokes caveman.
Step back and map how unfamiliar code fits the whole system before changing it. Use when you need a higher-level view, module/caller map, or how a path relates to the rest of the repo.
Post-implementation sign-off: Mode F (GitHub feature:<slug> — verify, close issues, docs on PR) or Mode B (.plan — review, docs, archive_plan).
Install OpenCode templates into one implementation repo path; optional legacy .plan/docs archive moves. Invoked only from setup-project via architect.
Spec-repo-only stack bootstrap — discover siblings, interview, legacy .plan/docs audit, configure all implementation repos via stack-bootstrap and scribe. Replaces per-repo setup-skills for vertical stacks.
Linked git worktree: symlink .env and .env.local from main checkout (bash/ln only)
Cross-repo companion to to-prd: after PRD frontmatter is filled, run bin/fanout <slug> from this spec repo to create child GitHub issues (one per ticket or legacy slice).
AI-powered code review using CodeRabbit CLI. Use only on explicit user request or when parent passes execution_mode orchestrate_coderabbit_gate. Do not run during orchestrated stage/issue work.
Planning mode: feature decomposition, specialists (debug/refactor/review/document/design), Mode A scribe handoff, artifact schema.
Disciplined debug loop: build feedback loop → reproduce → ranked hypotheses → instrument → fix + regression test → cleanup. Use for bugs, failures, and performance regressions.
Strict Red → Green → Refactor; commit after each green cycle
Steady execution: bootstrap, plan selection, stage loop, grading, difficulty completion gates, completion handoff to architect.
Environment readiness checks for runtime, toolchain, and test commands
Use for logic/backend implementation. Unified executor for plan artifacts using bounded TDD slices. Execute only stages with Owner: developer.
Reviews work and verifies completeness
Implementation technical planning on GitHub issues — codebase-backed plans, readable markdown, YAML stages for orchestrate. Spec fanout only captures requirements.
Orchestrate GitHub feature:<slug> backlog — next-runnable discovery, state transitions, flat or stage execution loop. Delegate gh and lib scripts to developer.
Synthesise a PRD from grill-me / research context, write docs/prd/<slug>.md, publish a GitHub issue with prd + state:ready-for-agent + feature:<slug>. Halt after publish — do not invoke fanout.
Commit → push → PR with explicit confirmation at each step
Generates documentation content from completed plans or GitHub feature sign-off. Read-only; returns changelog (required on sign-off), optional guides/architecture for scribe.
Grilling session (grill-with-docs) that challenges your plan against the existing domain model, sharpens terminology, and updates CONTEXT.md and ADRs as decisions crystallise — persisted via scribe. Architect Mode A loads this after plan type + first substantive requirements and before architect-plan. Also when the user wants stress-test against project language, documented decisions, or says "grill me".
Scaffold per-repo agent context — issue tracker, triage labels, domain doc layout — under docs/agents/ and an AGENTS.md or README section. Run once per repo before relying on to-issues, grill-me CONTEXT persistence, or cross-session portability.
Writes and updates markdown artifacts, docs, domain CONTEXT/ADR, README.md, AGENTS.md, and .env.example in allowed paths