mk:workflow-orchestrator
Auto-invoked 7-phase workflow coordinator for complex features — phase routing, token budget management, gate enforcement, fast-track mode.
| Source | .claude/skills/workflow-orchestrator/SKILL.md |
| Owner | lifecycle |
| Runtime | claude-code |
| Risk | high |
What This Skill Does
The central workflow coordinator. Auto-invoked when a session starts with complex-feature intent. Routes through the seven phases, managing gate enforcement, token budgets, agent team composition, and state persistence. Defers to mk:cook for single-task invocations.
When to Use
- Auto-invoked on session start with complex-feature intent (multi-file, multi-step tasks)
- NOT invoked when
mk:cookis explicitly called (cook owns the full pipeline for single tasks) - Never run both in the same session — if cook is active, orchestrator skips
Core Capabilities
- Phase routing: each phase dispatched to its owning skill — agent-detector and scale-routing at Orient, plan-creator at Plan, testing at Test, developer plus the mandatory simplify and verify stages at Build, review, ship, and memory after. The phase sequence itself is on Workflow Phases.
- Gate enforcement: Gate 1 after Phase 1 (plan approval), Gate 2 after Phase 4 (review verdict)
- Token budget: warns at 75%, handoff at 85%, forces context reset at 90%. Target ≤30K per workflow.
- State persistence:
workflow:handoff/workflow:resume/workflow:statuscommands for cross-session continuity - Agent Teams: parallel agent execution per phase with file ownership boundaries and task claim pattern
- Fast-track mode: skips phases for pre-approved specs with automated testing. Auto-stops on: tests pass in RED, 3 failed builds, critical security findings, coverage <80%
Arguments
| Flag | Effect |
|---|---|
| (no flag) | Auto-invoked on complex feature intent |
--fasttrack | Fast-track mode for pre-approved specs |
--teams | Agent Teams parallel execution mode |
Workflow
- Pre-execution checklist — run agent-detector, load memory, show agent banner, verify complexity, challenge requirements
- Phase 0: Orient — detect task agent, complexity tier, model selection
- Phase 1: Plan — route to plan-creator, enforce Gate 1
- Phase 2: Test — if TDD active, write failing tests
- Phase 3: Build — developer implements, plan-first gate enforced
- Phase 3.5: Simplify — mandatory simplification pass before review
- Phase 4: Review — reviewer verdict, enforce Gate 2
- Phase 5: Ship — shipper creates PR, pushes branch
- Phase 6: Reflect — memory captures learnings
Usage
# Auto-invoked — no explicit command needed
# Fast-track mode:
/mk:cook --fasttrack "add rate limiting middleware to the API"Example Prompt
Build a user dashboard with real-time metrics, role-based access control, and export to CSV. Use the existing auth system and PostgreSQL.This triggers auto-invocation of the orchestrator with complex-feature intent.
Common Use Cases
- Greenfield feature builds with multiple files
- Cross-cutting concerns (auth, logging, monitoring)
- Multi-phase features requiring plan → implement → review → ship
- Parallel agent execution for independent subtasks
Pro Tips
- Explicit
/mk:cookoverrides the orchestrator. Use cook when you know exactly what you want to build and don't need the full phase orchestration. - Phase 3.5 (Simplify) is mandatory. Code must be simplified before review — the orchestrator enforces this.
- Invalid transitions are blocked: you cannot skip Phase 1, ship with failing tests (Phase 5), or build without tests when TDD is active (Phase 3).
Canonical source:
.claude/skills/workflow-orchestrator/SKILL.md
mk:loop
Bounded, git-tracked optimization loop. Improve one scalar metric (coverage, bundle size, lint count, latency) by iterating: modify one scoped change, verify, keep or revert. Boundary-gated autonomy.
mk:orchestrate
Coordinates staged or parallel jobs across live-verified coding-agent runtimes and in-session subagents — capability/risk routing, worktree isolation, redacted capture, resume, arbiter review.