ralph

<Use_When>

• A plan (.dt-handoff/<slug>/plan.md) is approved and execution must complete every story before stopping.
• User says "ralph", "/ralph", "keep going until done", "끝까지 가줘", "finish this".
• The task suits PRD-driven progress (multiple independently verifiable stories).
• autopilot Phase 3 (Execution) invokes this skill as a sub-step.
• A prior run left an incomplete .dt-handoff/<slug>/prd.json to resume. </Use_When>

<Do_Not_Use_When>

• The user wants requirements capture (WHAT, not HOW), not execution → establish a spec first; ralph runs an existing plan/PRD.
• The user wants consensus planning → ralplan.
• The user wants parallel multi-agent decomposition → team.
• The user wants a one-shot / trivial change → delegate to executor directly, no PRD overhead.
• The user wants automatic commit/PR → refuse; ralph stops at "ready for commit".
• No spec / plan exists and the task is vague → route to ralplan first to scope the work. </Do_Not_Use_When>

<Why_This_Exists> Complex tasks fail silently when an agent declares "done" on a partial implementation, skips tests, or forgets an edge case. The PRD loop structures work into stories with testable acceptance criteria and refuses to mark a story complete without fresh per-criterion evidence from the verifier agent — not inline self-verification. The reviewer pass adds an independent second opinion before declaring the batch ready.

The TDD Iron Law (failing test before production code) is enforced inside every story so tests encode intent rather than mirror implementation. The code-simplifier cleanup (Step 7.5) removes slop introduced during the loop, scoped to changed files; the post-cleanup regression re-verify exists because even small cleanup edits can regress. The 3-failure escalation branches by failure type — root cause unclear → debugger, design fundamentally wrong → architect — to avoid grinding on the wrong specialist.

Auto commit/PR is prohibited because this marketplace's git-commit / github-pr skills are user-triggered by design; chaining them would skip review opportunities the user expects.

The loop stays thin: handoff mechanics (@handoff-in/@handoff-out, descriptors, events.jsonl) are defined once in the handoff protocol; this skill references them rather than re-spelling them per step. </Why_This_Exists>

<Execution_Policy> Output language: status reports, story descriptions, and progress.txt use $LANGUAGE. prd.json field VALUES (descriptions, AC text) use $LANGUAGE; field NAMES stay English.

TDD Iron Law (non-negotiable): for each story, the test-engineer pass MUST produce a Red test that runs and fails for the expected reason BEFORE executor writes production code. Skipping Red is forbidden. Exception: pure-refactor stories tagged refactor may proceed without a new Red test if existing tests cover the behavior — note the exception in the story's progress entry.

Verdict routing (single source — steps reference this, never re-describe): verifier and the approver (reviewer or critic) return a machine-readable @handoff-out block with a verdict field. Route on the enum, NEVER on prose keyword matching:

• APPROVE / ACCEPT_WITH_RESERVATIONS → proceed (note reservations in the report).
• REVISE / ITERATE → mark affected stories passes: false, dispatch executor to fix, re-invoke the same agent; loop to Step 2. Do NOT self-verify.
• REJECT (3rd consecutive from the approver) → stop REVIEW_BLOCKED.