rw-loop

name: rw-loop description: "Lite+Contract loop: DAG-aware dispatch, TDD evidence checks, dual quality gates (state/runtime), env preflight, and phase review contracts. Optional flags: --auto, --no-hitl, --hitl, --parallel, --max-parallel=<1..4>"

RW Loop Skill

DAG-aware task dispatch → TDD coder subagent → mandatory verification gates → phase/review contracts.

When to Use

• Implementing tasks from an existing plan (NEXT_COMMAND=rw-loop)
• Resuming implementation after a gate failure or pause
• Re-running loop after fixing a blocked task

Prerequisites

The following must exist before running this skill (created by rw-planner):

• .ai/CONTEXT.md — language policy
• .ai/PROGRESS.md — task status table
• .ai/tasks/ — atomic task files (TASK-XX-*.md)
• .ai/runtime/rw-active-plan-id.txt + .ai/plans/<PLAN_ID>/task-graph.yaml — DAG + node status source
• .github/skills/rw-loop/assets/ — 5 subagent prompt files
• .github/skills/rw-loop/scripts/check_state_sync.py + env_preflight.py — pre-dispatch guards

If any prerequisite is missing, the skill prints a failure token and stops.

Procedure Overview

Step 0: Guard     → verify prerequisites, resolve HITL/parallel mode
Loop:   Dispatch  → lock task → coder subagent → validate → gates → next
End:    Review    → final review gate → success output

Step 0: Guard

1. Read .ai/CONTEXT.md. If missing: print LANG_POLICY_MISSING, stop.
2. Verify .ai/PROGRESS.md and .ai/tasks/ exist. If missing: print TARGET_ROOT_INVALID, stop.
3. Verify runSubagent is available. If not: print RW_ENV_UNSUPPORTED, stop.
4. This mode never writes product code directly.
5. Verify required subagent prompt files exist in skill assets:
   • rw-loop-coder.subagent.md
   • rw-loop-task-inspector.subagent.md
   • rw-loop-security-review.subagent.md
   • rw-loop-phase-inspector.subagent.md
   • rw-loop-review.subagent.md
   • If any file is missing: print RW_SUBAGENT_PROMPT_MISSING, stop.
6. If .ai/memory/shared-memory.md exists, read it before loop start.
7. Read .ai/runtime/rw-active-plan-id.txt, then read matching .ai/plans/<PLAN_ID>/task-graph.yaml as primary dependency graph. If missing/unreadable: print TARGET_ROOT_INVALID, stop.
8. Verify required scripts exist:
   • .github/skills/rw-loop/scripts/check_state_sync.py
   • .github/skills/rw-loop/scripts/env_preflight.py
   • If any file is missing: print TARGET_ROOT_INVALID, stop.

Mode Resolution

When PARALLEL_MODE=ON and --max-parallel is omitted, MAX_PARALLEL=4.

Main Loop

Load full loop contract: loop-contract.md

Quick Reference

1. Check .ai/PAUSE.md. If exists: print PAUSE_DETECTED, stop. 1a. Run state sync checker before dispatch:
   • python .github/skills/rw-loop/scripts/check_state_sync.py
   • If checker returns fail tokens, stop and fix state artifacts first. 1b. Run environment preflight before dispatch:
   • python .github/skills/rw-loop/scripts/env_preflight.py
   • Require ENV_PREFLIGHT=PASS. On fail, stop and fix env/dependencies first.
2. Select dispatchable task(s) from DAG (in-progress first, then pending, never blocked).
   • Single mode: 1 task. Parallel mode: up to MAX_PARALLEL independent tasks (default 4).
3. Dispatch to coder subagent via runSubagent.
4. Validate (Gate A): completion delta (exactly N tasks), correct task IDs, evidence count increased, and state sync across PROGRESS + task frontmatter + task-graph.
   • Task-level verification should stay scoped/fast. Full regression belongs to Phase/Review gates via TASK-00 policy.
5. Task Inspector Gate (Gate B 포함): TASK_INSPECTION=PASS|FAIL, USER_PATH_GATE=PASS|FAIL, RUNTIME_GATE=PASS|FAIL.
6. Security Gate: SECURITY_GATE=PASS|FAIL.
7. Phase Inspector (when phase complete): run TASK-00 phase-gate full verification commands, then require PHASE_REVIEW_STATUS=APPROVED|NEEDS_REVISION|FAILED.
   • [HITL MANDATORY] If HITL_MODE=ON: output structured phase summary (completed tasks + evidence + inspector findings + 직접 확인 방법) THEN call askQuestions — unconditionally, never skip.
8. 3-strike rule: same task strike.active reaches 3 → blocked + escalate to planner. strike.total/security.total stay cumulative for history IDs, and active counters reset on successful completion.
9. Review Gate (all tasks complete): run TASK-00 final-gate full verification commands, then require REVIEW_STATUS=OK|FAIL|ESCALATE.
10. User Acceptance Gate: if PLANNING_PROFILE=UX_STRICT, user-acceptance-checklist.md becomes blocking and must pass USER_ACCEPTANCE_GATE=PASS.

State Transitions

pending → in-progress → completed
               |
               v
            blocked (3-strike or security critical)

• Single mode: 1 dispatch = exactly 1 task completed.
• Parallel mode: N dispatched = exactly N completed.
• completed requires VERIFICATION_EVIDENCE count increase.
• Runtime-visible behavior tasks require runtime evidence artifact(s) and RUNTIME_GATE=PASS.
• completed → pending forbidden unless explicit review rollback.

Loop Output Contract (success)

HITL_MODE=<ON|OFF>
PARALLEL_MODE=<ON|OFF>
PARALLEL_BATCH_SIZE=<1-4>
PLANNING_PROFILE=<STANDARD|FAST_TEST|UX_STRICT>
RUNSUBAGENT_DISPATCH_COUNT=<n>
RUN_PHASE_NOTE_FILE=<path|none>
PHASE_REVIEW_STATUS=<APPROVED|NEEDS_REVISION|FAILED|NA>
REVIEW_STATUS=<OK|FAIL|ESCALATE>
USER_ACCEPTANCE_GATE=<PASS|FAIL|NA>
ARCHIVE_RESULT=<SKIPPED|DONE|LOCKED>
NEXT_COMMAND=<done|rw-planner|rw-loop>

Failure Handling

Language Policy

• 모든 .ai/** 아티팩트 작성 전 .ai/CONTEXT.md를 먼저 읽을 것.