Mk:cook

Cook — Full Implementation Pipeline

Implementation entry adapter. .claude/workflow.yaml is the only lifecycle phase-sequence authority; this skill selects a profile, invokes the relevant owners, and presents the required gates. TDD is opt-in via --tdd.

Usage

/cook <natural language task OR plan path>
/cook "Add user auth" --fast
/cook "Build payment processor" --tdd       # Strict TDD enforced
/cook tasks/plans/260329-feature/plan.md --auto

Profiles: quick/--fast (inspect → edit → focused verify → summary), standard (plan if needed → implement → verify → review), release (standard + explicit mk:ship request), and high-assurance (--tdd --strict, with human gates). Flags: --interactive (default) | --fast | --parallel | --auto | --no-test | --tdd | --verify | --strict | --no-strict.

Quick runs do not create a plan, wiki candidate, memory entry, or ship action by default. They still stop if a risk flag or changed contract requires the standard profile.

Modifier flags (layer on any mode): --verify [LIGHT] (light browser/artifact check) | --strict [HEAVY] (full evaluator pass) | --no-strict (suppress auto-strict)

Concrete cost depends on the inner harness, model tier, and target surface; treat [LIGHT] vs [HEAVY] as relative ordering only.

--verify is advisory (does not block ship). --strict is a hard gate (FAIL blocks ship and routes back to Phase 3).

--advice (composable, off by default) — opt-in strategic supervision for one run. At named checkpoints (GUIDE after Gate 1, RESCUE on a stall or contradiction, REVIEW after Verify and before the reviewer, RECHECK after a correction; hard cap 5) the athena agent assesses the situation, recommends an operational path inside the locked scope, and may return the work for correction. It never approves: Gate 1 and Gate 2 remain human, the reviewer keeps its verdict, and supervision counts as no verification. Without the flag there are zero calls and no state is written. Checkpoints: references/workflow-steps.md → Advice Checkpoints. Contract: .claude/rules-conditional/advice-supervision-rules.md (load only when the flag is present).

TDD mode (--tdd flag)

When --tdd is detected in the invocation, the cook skill MUST write on to .claude/session-state/tdd-mode via a Bash tool call BEFORE any other workflow step:

mkdir -p .claude/session-state && echo on > .claude/session-state/tdd-mode

This sentinel file is read by pre-implement.sh, tdd-detect.sh, and downstream agents to detect TDD mode. Without --tdd, the sentinel is absent and the workflow runs in default mode (Phase 2 optional, no RED-phase gate).

MEOWKIT_TDD=1 env var (set in CI / shell rc) is the highest-precedence opt-in and overrides the sentinel.

When loading an existing plan, scan plan.md and phase files for tdd: true, ## Tests Before, or ## Regression Gate. If any are present but this invocation lacks --tdd and MEOWKIT_TDD is not enabled, warn:

Plan contains TDD sections but cook is not in TDD mode. Re-run with --tdd to enforce RED-first execution, or continue in default mode with TDD guidance only.

HARD GATE

Do NOT write implementation code until a plan exists and Gate 1 is approved. In TDD mode (--tdd / MEOWKIT_TDD=1): do NOT skip Test RED phase — write failing tests BEFORE implementation. In default mode: Phase 2 is optional; the developer may implement directly per the approved plan. Exception: --fast mode skips research but still requires plan + (in TDD mode) TDD-flavored tests. User override: Planning may be skipped only when the user explicitly says "just code it" or "skip planning" and Phase 0 found zero matched risk flags. Record the human override and rationale per .claude/rules/intervention-recording-rules.md; otherwise explain why Gate 1 still applies.

Anti-Rationalization

Generic implementation-phase rationalizations live in .claude/rules/anti-rationalization.md — read before any non-trivial implementation step. Cook-specific addition (TDD mode):

Before starting work, read references/failure-catalog.md for common implementation failure modes (cross-linked to ## Gotchas below).

Smart Intent Detection

See references/intent-detection.md for full detection logic.

Green-field product build (new kanban app, full SaaS from scratch, multi-sprint autonomous build)? Use mk:autobuild instead. Cook handles single features, fixes, and refactors; harness owns the generator↔evaluator loop and adaptive scaffolding density.

Lifecycle Adapter

flowchart TD
    A[Phase 0: Orient] --> B{Has plan?}
    B -->|Yes| D[Load Plan]
    B -->|No| SP[Present scout summary to user]
    SP --> C[Phase 1: Research + Plan]
    C --> RQ{5 dims captured?}
    RQ -->|No| C
    RQ -->|Yes| G1[GATE 1: Human Approval]
    G1 -->|approved| D
    G1 -->|rejected| C
    D --> E[Phase 2: Test - RED if --tdd, optional otherwise]
    E --> F[Phase 3: Build GREEN]
    F --> S[mk:simplify — MANDATORY]
    S --> V[mk:verify — MANDATORY]
    V --> G2[Phase 4: Review — GATE 2]
    G2 -->|approved| BV{--verify or --strict?}
    BV -->|"--verify"| P45A["Phase 4.5 --verify advisory"]
    BV -->|"--strict"| P45B["Phase 4.5 --strict blocking"]
    BV -->|neither| SR[Report completion]
    P45A --> SR
    P45B --> BG{evaluator FAIL?}
    BG -->|no| SR
    BG -->|yes| F
    G2 -->|rejected| F
    SR --> STOP[Stop and await user direction]
    STOP -->|explicit ship request| H[Phase 5: Ship]
    STOP -->|explicit close request| I[Phase 6: Reflect]
    H --> I

The YAML specification is authoritative. This diagram is a convenience rendering only; phase details and transitions live in .claude/workflow.yaml and references/workflow-steps.md.

Scout-First Contract (Phase 0)

Before any clarifying question or plan generation, present a 3–6 bullet codebase-context summary to the user:

1. Project type / language / framework
2. Existing modules/files relevant to the task
3. Current patterns/conventions for similar features
4. In-flight plans (tasks/plans/) or docs (docs/) covering this area
5. Public APIs / schemas / contracts the task could affect

Skip when input is a plan.md / phase-*.md path — the plan already encodes scout output.

In --fast mode the summary is abbreviated (1–3 bullets) but still presented; "skip research" reduces external research depth, not codebase-context handoff to the user.

Exact-Requirements Contract (Phase 1)

mk:plan-creator MUST be able to answer all 5 dimensions in concrete sentences before returning a plan:

1. Expected output — concrete artifact(s) the user will see (paths, behavior, endpoint, CLI flags)
2. Acceptance criteria — specific inputs → outputs / edge cases
3. Scope boundary — what is explicitly OUT of scope this round
4. Non-negotiable constraints — stack, file locations, naming, compat, deadlines, performance
5. Touchpoints — which existing files (from scout) will be modified / which contracts must stay stable

Every clarifying-question option MUST cite scout findings (e.g., file paths). Vague abstract options are a failure mode — see references/failure-catalog.md.

Skip when input is a plan.md / phase-*.md path (plan encodes the 5 dimensions).

Workflow Modes

Gate 1 routing, parallelism, and full per-phase progression live in references/intent-detection.md (canonical Mode Behaviors table).

Gate 2: human approval mandatory in all modes — see .claude/rules/gate-rules.md for the full contract. If the reviewer surfaces a regression / side effect / broken workflow, follow the Regression Recovery Options pattern in references/review-cycle.md rather than silently patching.

Required Subagents

Concrete subagent name and HTTP tool depend on installed skill set; cook treats both as interfaces, not specific implementations. | 5 Ship | shipper via mk:ship | Only after an explicit ship request — runs full pre-ship pipeline (git-manager invoked inside shipper for commit + PR) | | 6 Reflect | analyst | MUST spawn — cost + pattern analysis | | 6 Reflect | documenter | MUST spawn — sync-back + docs | | 6 Reflect | mk:memory session-capture | MUST spawn — 3-category learning extraction |

During iterative build-test-fix cycles, follow references/loop-safety-protocol.md for checkpoint tracking, stall detection, and escalation triggers.

See references/subagent-patterns.md for Task() invocation patterns. See references/workflow-steps.md for detailed per-phase instructions. See references/review-cycle.md for review gate logic.

Simplify Step (Mandatory)

After Phase 3 (Build GREEN) completes, run mk:simplify to reduce complexity before Phase 4 (Review). This is mandatory — do not skip.

Rationale: Review quality improves when code is already simplified. Reviewers catch logic errors better when complexity is low. Running simplify after tests are green (but before review) ensures the reviewed code is the final, clean version.

Verify Step (Mandatory)

After mk:simplify completes, run mk:verify for a unified build→lint→type-check→test→coverage check. This confirms simplification didn't break anything and gives the reviewer a clean signal. By default mk:verify scopes lint and tests to the changed files (build and type-check stay whole-program); the complete whole-project run happens at ship time. Pass --full only when a full sweep is needed here.

If mk:verify FAILS after simplify: send back to developer to fix, then re-run verify before proceeding to Phase 4. Do not skip verify or proceed to review with a failing verify result.

Status Report (Post-Gate 2)

After a Gate 2 verdict PASS, delegate to project-manager per .claude/rules/post-phase-delegation.md Rule 1 (background — include "Run in the background" in the prompt). Status report is co-located at {plan-dir}/status-reports/{YYMMDD}-status.md. Stop after reporting completion; invoke mk:ship only when the user explicitly requests shipping. Skipped automatically when MEOWKIT_PM_AUTO=off.

Durable Task State (when an active task record exists)

When this run drives an active durable task (a tasks/active/<id>.json exists), emit status/step at phase transitions and record each acted-on mewkit capabilities resolve outcome (selected|skipped|unavailable|unsupported) via mewkit task-state update — per .claude/rules/task-state-emission.md. Advisory + best-effort: a failed or unavailable command is surfaced but never blocks the pipeline, and one-off work with no record emits nothing.

Related Rules

• .claude/rules/gate-rules.md — Gate 1 (Plan) and Gate 2 (Review) hard-stop conditions this skill enforces across all modes
• .claude/rules/post-phase-delegation.md — PM delegation fire points and skip conditions
• .claude/rules/task-state-emission.md — when/what/how to emit durable task-state events (advisory; active durable tasks only)
• .claude/rules-conditional/workflow-evidence-rules.md — workflow evidence index (traceability over existing Phase 0-6 outputs; mirrors the gate scripts, never approves). See references/workflow-steps.md → Workflow Evidence Index

Gotchas

Operational gotchas live here; pre-build failure modes and rationalizations live in references/failure-catalog.md.

• Skipping mk:simplify before review: Tests pass but code is still complex → run /mk:simplify between Phase 3 and Phase 4 every time, no exceptions
• Skipping Gate 1 on "simple" features: Features that seem simple grow during implementation. Always create a plan file; cancel it if truly trivial

• Auto-approve sneaking bugs past Gate 2: "auto" means automatic execution between gates, NOT automatic approval. Auto mode auto-fixes but NEVER auto-approves Gate 2 — gate-rules.md says NO exceptions. The workflow evidence index must be complete before Gate 2 is presented
• Context loss between phases: Long multi-phase workflows exceed the context window — the external-memory write after each completed plan phase is the load-bearing action that lets a fresh session resume the correct phase. This is a mandatory step, not a habit: see the Per-Phase Checkpoint in references/workflow-steps.md (Phase 6) — flip the phase-file checkboxes and update plan.md Agent State before any next-phase transition
• Parallel mode deadlocks: Phase dependencies cause deadlock when phase-03 waits for phase-02 results. Map dependency graph before spawning parallel agents
• Code mode on stale plans: Running old plan against changed codebase. Warn if plan.md is >14 days old
• Fast mode shallow test coverage: Skipping research means tests capture plan-level intent, not edge cases. Document: "fast mode = TDD-flavored, coverage may be lower"
• Missing model tier declaration: Expensive models on trivial tasks, cheap models on security-critical work. Always declare tier in Phase 0
• Forgetting memory read/write: Prior session learnings lost. Phase 0 reads canonical .meowkit/memory/*.json stores only for fix memory; Phase 6 writes canonical JSON stores and regenerates generated views when needed
• Subagent patterns using Agent() not Task(): Task() enables tracking, blocking, and progress. Always use Task() for phases 2-6
• --strict cost surprise: --strict spawns the full evaluator (mk:evaluate) — a [HEAVY] pass. Auto-triggered by scale-routing level=high. Use --no-strict to suppress, or --verify for the [LIGHT] check
• --strict vs --verify confusion: --verify = light browser check (advisory). --strict = full evaluator with rubrics (FAIL blocks ship). --strict supersedes --verify