Cheatsheet Dense quick reference for all MeowKit commands, agents, skills, and workflow phases.
Command Phase What it does /mk:meow [task]0 Entry point — classifies and routes to right agent /mk:breakdown <source>pre-1 Spec/PRD → user stories + optional Jira tickets + tech breakdown — stops before plan /mk:plan [feature]1 Premise challenge + two-lens plan + Gate 1 /mk:cook [feature]1→5 Full pipeline: plan → test → build → review → ship /mk:fix [bug]varies Auto-detect complexity, route to fix strategy /mk:review4 Multi-pass code review + adversarial analysis /mk:ship5 Pre-ship → commit → PR → CI verify → rollback doc /mk:test2 TDD enforcement — write or run tests /mk:audit4 Full security scan across all platforms /mk:validateany Run deterministic Python validation scripts /mk:arch [action]1 Generate, list, or analyze ADRs /mk:design [system]1 System design consultation (docs only) /mk:docs-init6 Initial codebase scan → doc skeleton /mk:docs-sync6 Diff-aware doc updates after feature work /mk:canary5 Staged deployment with monitoring /mk:retro6 Sprint retrospective with trend tracking /mk:budgetany Token cost tracking report /mk:party [topic]1 Multi-agent deliberation — 2-4 agents debate, forced synthesis /mk:spawn [agent]any Launch parallel agent session /mk:statusany Delivery status report via project-manager /mk:upgradeany Self-update MeowKit /mk:helpany Scan project state, recommend next pipeline step /mk:autobuild "build a X"0→5 Autonomous green-field build with generator/evaluator loop /mk:evaluate4 Behavioral rubric grading with active verification /mk:sprint-contract3 Draft contract before harness sprint /mk:rubric4 Load/compose rubric preset /mk:trace-analyzeany Scatter-gather analysis of trace log /mk:benchmarkany Canary suite (quick tier default; --full for 6-task tier)
Agent Auto-activates when Phase Key output orchestrator Every task 0 Routing decision + model tier planner Standard/complex tasks 1 Plan file at tasks/plans/ brainstormer Explicit or complex planning 1 Trade-off analysis researcher Research tasks 0,1,4 Structured findings architect Schema/API/infra changes 1 ADR + system design tester When invoked (always in --tdd mode, on-request otherwise) 2 Tests (failing tests in --tdd mode) security Auth/payments/security changes 2,4 BLOCK/PASS verdict developer After planner (TDD off) or tester (TDD on) 3 Implementation reviewer After implementation 4 5-dimension verdict shipper After Gate 2 5 PR + rollback docs documenter After ship 6 Updated docs + changelog analyst Session start/end 0,6 Cost report + patterns journal-writer On failure/escalation 6 Root cause documentation
Phase Name Gate? What happens 0 Orient — Read memory, load skills, route model tier 1 Plan ✋ Gate 1 Challenge premises, create plan, human approval 2 Test — Write failing tests first (TDD mode --tdd/MEOWKIT_TDD=1); optional otherwise 3 Build — Implement per plan; in TDD mode, until tests pass 4 Review ✋ Gate 2 5-dimension audit, human approval 5 Ship — Commit, PR, CI verify, rollback docs 6 Reflect — Update docs, memory, retrospective
Mode When Agents Gate behavior Sequential (default) All tasks 1 per phase Full gates Parallel COMPLEX + independent subtasks Up to 3, worktree isolation Gates always sequential Party Architecture decisions 2-4 deliberation agents No code, synthesis only
Tier Examples Model TRIVIAL Rename, typo, format Cheapest (Haiku) STANDARD Feature, bug fix, tests Default (Sonnet) COMPLEX Architecture, security, auth Best (Opus)
Domain override: mk:scale-routing at Phase 0. Fintech, healthcare, auth → force COMPLEX regardless of manual classification.
Mode Gates Security Use for default Both Full Most work strict Both (no WARN) Every file Production, auth, payments fast Gate 1 only BLOCK only Prototypes architect N/A N/A Design-only sessions audit N/A Comprehensive Security reviews
Hook Type What it does Blocks? post-write.sh PostToolUse Security scan on Edit/Write Yes post-session.sh Stop Capture session to memory No pre-task-check.sh Skill-invoked Injection pattern detection Yes pre-implement.sh Skill-invoked TDD gate — opt-in via --tdd / MEOWKIT_TDD=1; no-op otherwise Only when TDD enabled pre-ship.sh Skill-invoked Test + lint + typecheck Yes privacy-block.sh PreToolUse Block .env / *.key / credential reads Yes gate-enforcement.sh PreToolUse Block source writes before Gate 1 approval Yes project-context-loader.sh SessionStart Auto-load project-context.md into context No
Skill Trigger Output mk:cook "build feature" Full pipeline execution mk:fix "fix bug", "debug" Root cause + fix + regression test mk:ship "ship", "deploy", "create PR" PR URL + rollback docs mk:review "review", "check diff" Verdict: APPROVE/BLOCK mk:scout "find files", "search codebase" File map + architecture fingerprint mk:investigate "debug this", "why broken" Root cause + evidence mk:docs-finder "docs for [lib]" Structured documentation mk:multimodal Image/video/audio file Analysis via Gemini mk:qa-manual "test this flow" QA report or Playwright .spec.ts mk:cso "security audit" OWASP + STRIDE findings mk:party "decide architecture" Multi-agent decision brief mk:scale-routing Phase 0 (automatic) Domain complexity level mk:worktree Parallel execution setup Isolated git worktree mk:autobuild "build a X from scratch" Autonomous generator/evaluator build loop mk:evaluate Harness Phase 4 Rubric-graded behavioral verdict with evidence mk:sprint-contract Harness FULL density Acceptance criteria contract before sprint mk:rubric Evaluator setup Load/compose rubric preset mk:trace-analyze Trace log available Root cause analysis across trace spans mk:benchmark Canary testing Quick or full canary suite with scored output
Command Purpose Ends with mk:plan-creatorCreate plan from scratch Print & Stop mk:plan-ceo-reviewProduct lens review Print & Stop mk:plan-ceo-reviewEngineering lens review Print & Stop /mk:cook [path]Begin implementation Runs pipeline
All three review skills stop after printing — you control when to run /mk:cook [plan path].
Command What it does /mk:plan "feature"Create plan via mk:plan-creator npx mewkit task new --type feature "desc"Create task from template npx mewkit task listList active tasks
Template Use when Create with feature Adding new functionality npx mewkit task new --type featurebug-fix Fixing broken behavior npx mewkit task new --type bug-fixrefactor Restructuring code npx mewkit task new --type refactorsecurity Security review npx mewkit task new --type securityguideline Team standards Copy from tasks/templates/guideline.md
draft → in-progress → blocked → review → done
Create a task file for any change affecting > 2 files OR > 30 minutes.
Source of truth: .claude/.env.example. Most defaults live in .claude/settings.json env block (team-shared, git-committed); override here per-project / per-user. Uncommented values in .claude/.env win over settings.json; shell exports win over both.
Variable Values Effect MEOWKIT_AUTOBUILD_MODEMINIMAL | FULL | LEANOverride auto-detected scaffolding density MEOWKIT_MODEL_HINTe.g. opus-4-6 Tell harness which model is running (enables LEAN auto-detect) MEOWKIT_BUDGET_WARNe.g. 30 USD warn threshold (prints warning, continues) MEOWKIT_BUDGET_BLOCKe.g. 100 USD hard-block threshold (halts the run, sets final_status=TIMED_OUT) MEOWKIT_BUDGET_CAPe.g. 50 User-set cap; overrides BUDGET_BLOCK (can be lower OR higher)
Variable Values Effect MEOWKIT_TDD0 | 11 = strict red-green-refactor (pre-implement.sh enforces failing-test-first); 0 = optional (default)MEOWKIT_HOOK_PROFILEstandard | fast | strictHook profile selector. Legacy alias MEOW_HOOK_PROFILE still accepted; new name wins if both set
Variable Values Effect MEOWKIT_MEMORY_BUDGETchars (default 4000, ~1K tokens) Max on-demand memory context selected per prompt MEOWKIT_MEMORY_STALENESS_MONTHSmonths (default 6) Skip memory entries older than N months
Set to off to disable an individual hook; default on.
Variable Effect MEOWKIT_BUILD_VERIFYPhase 4 build / lint / typecheck verification hook MEOWKIT_LOOP_DETECTLoop / regression detection hook MEOWKIT_PRECOMPLETIONPre-completion guard hook MEOWKIT_TDD_FLAG_DETECTORDetects --tdd flag and toggles strict mode MEOWKIT_PM_AUTOoff disables silent project-manager post-phase fires; /mk:status still works
shell export > .claude/.env > .claude/settings.json "env" block Unquoted values strip inline comments after #. Quote values that contain a literal # (e.g. API keys): VAR="abc#123".