team-shinchan-start

name	team-shinchan:start
description	Use when you want to start a new task with the integrated workflow.
user-invocable	true

MANDATORY EXECUTION - DO NOT SKIP

When this skill is invoked, execute immediately. Do not explain.

Injected Context (bigproject phase mode)

This skill is normally invoked standalone. It can also be driven by the bigproject phase loop (skills/bigproject/SKILL.md Step 5), which executes these same steps on the main thread with the following inputs already resolved:

• DOC_ID — pre-set to {PROJECT_ID}-phase-{N}. Use it verbatim; do NOT generate one.
• parent_doc_id — the owning {PROJECT_ID}.
• phase_number — the 1-based phase index.
• PHASE_CONTEXT — project title + this phase's title/domain/acceptance criteria + summaries of completed phases.

When invoked with injected context:

• Skip Step 0 entirely (expiry/archive is the project's concern, run once by bigproject).
• In Step 1, use the injected DOC_ID and write parent_doc_id / phase_number into the state (see Step 1).
• In Step 2A, pass PHASE_CONTEXT to Misae so the interview is scoped to this phase.

When invoked standalone (no injected context), behave exactly as before.

Step 0: Expire and Archive Stale Workflows

Read workflow_expiry_days from:

1. .shinchan-config.yaml in the current project root (key: workflow_expiry_days) — takes priority
2. Otherwise use the plugin default: 7 (from plugin.json settings)
3. If workflow_expiry_days is 0 or cannot be read → skip expiry entirely

For each .shinchan-docs/*/WORKFLOW_STATE.yaml where status: active:

Expiry check (skip if workflow_expiry_days == 0):

1. Read the updated timestamp from WORKFLOW_STATE.yaml
2. Parse the timestamp (ISO 8601). If parsing fails → skip this entry
3. Calculate elapsed days: (now - updated) / 86400000
4. If elapsed >= workflow_expiry_days: a. Set status: expired in WORKFLOW_STATE.yaml b. Add event to history:

   - timestamp: "{ISO now}"
     event: auto_expired
     agent: shinnosuke
     archived_at: "{ISO now}"
     archived_reason: auto_expiry
   

   c. Calculate archive path: .shinchan-docs/archived/{YYYY-MM}/ where YYYY-MM comes from the current date d. Attempt: mkdir -p .shinchan-docs/archived/{YYYY-MM}/ && mv .shinchan-docs/{DOC_ID}/ .shinchan-docs/archived/{YYYY-MM}/{DOC_ID}/ e. If mv fails: silently continue (status stays expired, folder stays in place) f. Do NOT output any paused/expired notification to the user g. Continue to next workflow

Non-expired active workflows are left as-is. Multiple workflows can be active simultaneously. The workflow guard (workflow-guard.sh) protects the most recently updated active workflow. Use /team-shinchan:resume to switch the guard target to a different workflow.

Step 1: Setup (Folder + State)

1. DOC_ID: If a DOC_ID was injected (bigproject phase mode), use it verbatim and skip generation. Otherwise: if args contains ISSUE-xxx use it; else {branch}-{next_index} from git branch + ls. When computing {next_index}, ignore directories matching *-phase-* — those belong to a bigproject and are not standalone workflows. Truncate + warn if args > 2000 chars.
2. mkdir -p .shinchan-docs/{DOC_ID}
3. Create WORKFLOW_STATE.yaml:

version: 1
doc_id: "{DOC_ID}"
created: "{timestamp}"
updated: "{timestamp}"
current:
  stage: requirements
  phase: null
  owner: misae
  status: active
  answer_mode: human              # /start default — the human answers every AskUserQuestion. `/autopilot` sets this to `proxy` (see "Answer Mode" below + agents/_shared/proxy-user-panel.md). Absent ⇒ human.
  execution_mode: micro-execute   # /start default — pins Stage 3 to the per-task (spec→quality→skeptic) review chain. Set to `dag` ONLY when explicitly opting into parallel dag-executor dispatch.
  # parent_doc_id / phase_number: ONLY when injected by bigproject phase mode; omit standalone.
  # parent_doc_id: "{parent_doc_id}"
  # phase_number: {phase_number}
  interview: { step: 0, collected_count: 0, last_question: null }
  ak_gate:
    requirements:
      status: pending          # pending | in_review | approved | rejected | escalated
      retry_count: 0           # 0, 1, or 2 (max)
      last_rejection_reasons: []  # list of strings — most recent rejection points
    design:
      status: pending
       
       []
    
       
       
       []

    
     
     

History entry format for AK review (appended after each AK review):

- timestamp: "{ISO timestamp}"
  event: ak_review
  agent: action_kamen
  stage: requirements          # or planning
  verdict: REJECTED            # or APPROVED
  retry_count: 0               # which attempt this was (0 = first, 1 = first retry, 2 = second retry)
  rejection_reasons:
    - "Problem Statement lacks quantified success metrics"
    - "FR coverage missing error-handling scenarios"

ak_gate schema notes:

• status values: pending (not yet reviewed) | in_review (AK review in progress) | approved (AK approved) | rejected (AK rejected, retries remaining) | escalated (max retries reached, waiting for user)
• retry_count persists across session restarts (NFR-2: session-restart safe)
• Existing workflows without ak_gate field continue to function (backwards-compatible)

Stage rules and transition gates are defined in CLAUDE.md and hooks/workflow-guard.md.

4. Linear Sync — START (In Progress). If this is a standalone invocation (NOT bigproject phase mode — i.e. no injected parent_doc_id), run the START → In Progress transition per agents/_shared/linear-sync.md: detect a Linear issue in args/DOC_ID/branch, confirm it with get_issue, and if real move it to In Progress and persist it to current.linear_issue in WORKFLOW_STATE.yaml. When invoked with injected bigproject phase context, skip this — the parent project owns the issue lifecycle. No-op if no Linear issue / Linear MCP unavailable.

Answer Mode — human (default) vs proxy (autopilot)

Read current.answer_mode from WORKFLOW_STATE.yaml (absent ⇒ human). It controls who answers every AskUserQuestion this skill raises — the requirements interview questions (2A.1), the REQUESTS approval (2A.3), the design decisions (2B.1), and the DESIGN approval (2B.3):

• human (default) — behave exactly as written below: call AskUserQuestion and wait for the user.
• proxy (set by /autopilot) — do NOT call AskUserQuestion. Instead route the identical {question, options, header, context_hint} through the proxy-user panel (agents/_shared/proxy-user-panel.md): a K=3 diverse panel votes, a cautious judge breaks ties, and the winning option is used as if the user had picked it. Append a proxy_answer history entry per that spec. Everything else in the flow — Misae's clarity gate, Hiroshi's design-completeness gate, the materiality/closure re-entry loops, and every Action Kamen review — runs unchanged. Proxy mode replaces the human at the seam, nothing before it.

ROUTING RULE: wherever a step below calls AskUserQuestion(...), that call is answer_mode-aware. Under proxy, substitute a proxy-user-panel selection over the same options and continue with the returned choice. This applies to 2A.1, 2A.3, 2B.1, and 2B.3. (Options-overflow pagination is a human-UI concern and is skipped under proxy — the panel sees the full options[] at once.)

ESCALATION UNDER PROXY: the 3-way ESCALATE prompts (2A.1b / 2B.1b) must never block autopilot on a human. Under proxy, do NOT call AskUserQuestion there — auto-select B (Open Questions로 기록 후 진행): record the unresolved items and move on. This is the hands-off, non-destructive default (A risks looping to hard_cap, C discards prior answers). Log it as a proxy_answer history entry with seam: escalation.

Step 2: Greeting + Agent Invocation

Output greeting (adapt to user's language):

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
👦 [Shinnosuke] Hey! Let's build something great~ 💪
📁 Project: {DOC_ID} | 🎯 Stage: Requirements
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Step 2A-pre: Visual Input Detection

If args contain image/PDF paths (.png, .jpg, .jpeg, .gif, .svg, .pdf, .webp) or reference visual content:

Task(subagent_type="team-shinchan:ume", model="sonnet",
  prompt="Analyze visual content for requirements.\nDOC_ID: {DOC_ID}\nExtract: UI components, layout, design patterns, user flows, ambiguities.\nUser request: {args}")

Store result as {vision_context}. Skip if no visual input.

Step 2A: Requirements — Parent-Orchestrated Interview Loop

CRITICAL: Sub-agents cannot call AskUserQuestion for the user. The main thread (this skill) drives the interview; Misae designs each question. See agents/misae.md § "Parent-Orchestrated Interview Protocol".

2A.0 — Read interview config (FR-6):

# Read .shinchan-config.yaml if present (project root), else use defaults.
config = read_yaml(".shinchan-config.yaml") if file_exists else {}
skip_threshold    = config.get("interview.skip_threshold",    0.85)
done_threshold    = config.get("interview.done_threshold",    0.75)
hard_cap          = config.get("interview.hard_cap",          10)
gate_loop_enabled = config.get("interview.gate_loop_enabled", True)        # NEW
gate_threshold    = config.get("interview.gate_threshold",    0.8)         # NEW
stagnation_delta  = config.get("interview.stagnation_delta",  0.05)        # NEW
stagnation_window = config.get("interview.stagnation_window", 2)           # NEW
soft_cap          = config.get("interview.soft_cap",          6)           # NEW
ak_double_check   = config.get("interview.ak_double_check",   False)       # NEW
project_type      = config.get("interview.project_type",      "greenfield")# NEW
problem_framing_gate = config.get("interview.problem_framing_gate", True)   # NEW (main-079, FR-7 default true)

# Sanity check (FR-6, HR-2, HR-3): invalid values → fall back to defaults + warn
warnings = []
if not (0 < done_threshold < skip_threshold <= 1.0) or not (1 <= hard_cap <= 50):
  warnings.append("skip/done/hard_cap out of range"); skip_threshold = 0.85; done_threshold = 0.75; hard_cap = 10
if gate_threshold <= done_threshold:
  warnings.append("gate_threshold must be > done_threshold"); gate_threshold = 0.8
if soft_cap >= hard_cap:
  warnings.append("soft_cap must be < hard_cap"); soft_cap = 6
if stagnation_window < 2:
  warnings.append("stagnation_window must be >= 2 (HR-3 DoS protection)"); stagnation_window = 2
if stagnation_delta <= 0:
  warnings.append("stagnation_delta must be > 0"); stagnation_delta = 0.05
if project_type not in ("brownfield", "greenfield"):
  warnings.append("project_type must be brownfield|greenfield"); project_type = "greenfield"
if warnings:
  print("⚠️ [start] .shinchan-config.yaml validation errors — using defaults: " + "; ".join(warnings))

# Write resolved gate_loop_enabled + gate_threshold into WORKFLOW_STATE.current so
# mechanical-check Check D can enforce the gate without reading .shinchan-config.yaml.
write_to_state({"current.gate_loop_enabled": gate_loop_enabled, "current.gate_threshold": gate_threshold})

# main-079 FR-4/FR-6/FR-7: resolve problem-framing gate live-ness ONCE before the 2A.1 loop.
brainstorm_exists = file_exists(".shinchan-docs/{DOC_ID}/brainstorm-output.md")  # FR-4 one-shot cap
gate_live = (problem_framing_gate
             and current.answer_mode != "proxy"    # FR-6 autopilot exclusion
             and not brainstorm_exists)             # FR-4 one-shot

2A.1 — Interview loop (clarity-gated, hard_cap is the only ceiling):

answers       = []
exit_reason   = null   # populated by Misae's status:done payload
for turn in 1..hard_cap:
  result = Task(subagent_type="team-shinchan:misae", model="sonnet", prompt=
    "mode: DESIGN_NEXT_QUESTION
    DOC_ID: {DOC_ID} | WORKFLOW_STATE: .shinchan-docs/{DOC_ID}/WORKFLOW_STATE.yaml
    turn: {turn}
    prior_answers: {answers}
    user_request: {args}
    vision_context: {vision_context or 'None'}
    phase_context: {PHASE_CONTEXT or 'None'}   # bigproject phase mode: scope questions to this phase
    skip_threshold: {skip_threshold}
    done_threshold: {done_threshold}
    hard_cap: {hard_cap}
    gate_loop_enabled: {gate_loop_enabled}
    gate_threshold: {gate_threshold}
    stagnation_delta: {stagnation_delta}
    stagnation_window: {stagnation_window}
    soft_cap: {soft_cap}
    ak_double_check: {ak_double_check}
    project_type: {project_type}
    gate_live: {gate_live}   # main-079: Layer-1 gate live-ness; false ⇒ Misae skips solution-smell (NFR-4)
    Return the interview-question JSON block per agents/misae.md contract.")

  Parse the last ```interview-question ... ``` fenced block in result.

  # FR-1 / AC1: zero-turn fast path on first invocation
  if turn == 1 and parsed.status == "done" and parsed.reason in {"pre_interview_clear", "user_skip_override"}:
    exit_reason = parsed.reason
    break  # skip the loop entirely; jump to FINALIZE_DRAFT with answers == []

  # main-079 FR-3 / AC-2: problem-framing gate — needs_reframe interceptor.
  # Placed BEFORE the ask/done GUARD because `needs_reframe` is neither "ask" nor "done";
  # the GUARD (b) set is intentionally left unchanged (additive, minimal seam).
  if turn == 1 and parsed.status == "needs_reframe":
    # HR-8: turn counter NOT incremented and clarity_score.history NOT written — pre-scoring detour.
    # (1) Reuse skills/brainstorm VERBATIM (NFR-3): Hiroshi writes .shinchan-docs/{DOC_ID}/brainstorm-output.md.
    invoke Skill team-shinchan:brainstorm with args = {args}   # HR-9: PHASE_CONTEXT flows through as it already does
    # Ignore brainstorm's own Step-5 "proceed with /requirements?" suggestion — the 3-way below supersedes it.
    # (2) Extract the reframed problem: the prose UNDER the "## Problem Reframe" header of
    #     .shinchan-docs/{DOC_ID}/brainstorm-output.md (the 2-3 sentence restatement — see
    #     skills/brainstorm/SKILL.md Output Format). This is the ONLY section used for ①.
    reframed_problem = read_section(".shinchan-docs/{DOC_ID}/brainstorm-output.md", from="## Problem Reframe", to="## Alternative Approaches")
    # (3) 3-way choice — answer_mode-aware ROUTING RULE (under proxy this branch is unreachable: gate_live=false via FR-6)
    choice = AskUserQuestion(questions=[{
      question: "brainstorm 결과를 검토했습니다. 어떻게 진행할까요?",
      header:   "문제 프레이밍",
      options: [
        {label: "① 재정의된 문제로 진행", description: "brainstorm의 Problem Reframe로 요구사항 시작"},
        {label: "② 원래 방향 유지",        description: "원래 요청 그대로 요구사항 시작"},
        {label: "③ 직접 편집",            description: "직접 수정한 방향으로 시작"}],
      multiSelect: false}])
    if choice == "③ 직접 편집":
      # AskUserQuestion is MULTIPLE-CHOICE ONLY — capture the edited request with a follow-up
      # open prompt (design-review nit): print "수정된 요청/방향을 한 문장 이상으로 입력해주세요:"
      # and read the user's next free-text message as effective_request.
      effective_request = <user's typed free-text reply to the follow-up prompt>
    elif choice == "① 재정의된 문제로 진행":
      effective_request = reframed_problem
    else:  # ② 원래 방향 유지
      effective_request = args
    # (4) audit trail — DEC-4 / NFR-5 / HR-3, same history write path as ak_review (HR-2, no new path)
    append_history({event: "reframe_choice", agent: "hiroshi",
                    choice: {"① 재정의된 문제로 진행": "reframed",
                             "② 원래 방향 유지":       "kept_original",
                             "③ 직접 편집":            "edited"}[choice],
                    timestamp: ISO_now()})
    # (5) Turn-safe re-entry (HR-8): re-invoke Misae at turn 1 (counter NOT incremented) with the
    #     chosen direction and gate_live=false (brainstorm-output.md now exists → FR-4 one-shot →
    #     needs_reframe CANNOT recur → guaranteed termination, HR-5). Then fall through to the
    #     zero-turn fast-path check + GUARD for this SAME iteration (turn still == 1).
    args      = effective_request
    gate_live = false
    result = Task(subagent_type="team-shinchan:misae", model="sonnet", prompt=<the 2A.1 prompt above, with turn=1, user_request=effective_request, gate_live=false>)
    Parse the last ```interview-question``` fenced block in result → parsed
    # Do NOT re-run the zero-turn fast-path block above for this fresh `parsed` — control simply
    # falls through to the GUARD below (this SAME iteration, turn still == 1). A `done` re-entry
    # (pre_interview_clear/user_skip_override) is already handled downstream: GUARD clause (d)
    # accepts those reasons and the general `if status == "done": break` exits with answers == [].

  GUARD (parsing / options integrity / FR-4 contract):
    Validate ALL of the following before calling AskUserQuestion:
    (a) An `interview-question` fenced JSON block exists and parses.
    (b) `status` is "ask" or "done".
    (c) If status == "ask":
        - `question` is a non-empty string (>= 5 chars).
        - `options` is an array with >= 2 entries (no upper bound).
        - Every option has a `label` that is a non-empty string
          (>= 2 chars, NOT whitespace-only, NOT just "A." / "B." prefix).
        - `header` is a non-empty string.
        - **(FR-4)** `targets_subscore` is one of {"goal_clarity", "constraint_clarity", "success_criteria"}.
        - **(FR-4)** `closes_unknown` is a non-empty string ≤ 80 chars.
    (d) If status == "done":
        - `reason` is one of {"pre_interview_clear", "clarity_threshold_met",
          "user_skip_override",
          "stagnation_escalate", "soft_cap_escalate",
          "no_more_actionable_gaps_escalate", "hard_cap_escalate",   # gate_loop_enabled: true
          "hard_cap_reached", "no_more_actionable_gaps"}.            # legacy (gate_loop_enabled: false)

    On ANY validation failure:
      Re-invoke Misae with mode=DESIGN_NEXT_QUESTION, appending to the prompt:
      "CRITICAL: Your previous response failed validation: {specific reason,
       e.g. 'targets_subscore missing or not in {goal_clarity|constraint_clarity|success_criteria}',
       'closes_unknown empty or > 80 chars', 'options[1].label was empty',
       'no interview-question block found', 'question string too short'}.
       Re-read agents/misae.md § Parent-Orchestrated Interview Protocol.
       Emit EXACTLY ONE fenced block tagged `interview-question` containing:
       non-empty question (>=5 chars), non-empty header, 2 이상 options each
       with substantive label, AND (FR-4) `targets_subscore` in
       {goal_clarity|constraint_clarity|success_criteria} AND `closes_unknown`
       string ≤ 80 chars."
      Retry up to 2 times.
      On 3rd failure: abort the interview, print the raw Misae output
      verbatim to the user, and STOP.

  If status == "done":
    exit_reason = parsed.reason  # propagated to FINALIZE_DRAFT
    break
  If status == "ask" (and guard passed):
    // OPTIONS OVERFLOW HANDLING:
    // AskUserQuestion tool allows max 4 options per call.
    // When Misae returns > 4 options, paginate:
    if options.length <= 4:
      user_answer = AskUserQuestion(questions=[{
        question, header, options, multiSelect
      }])
    else:
      // Split: first 3 options + "더 많은 선택지 보기"
      page1 = options.slice(0, 3)
      page1.push({label: "더 많은 선택지 보기", description: "추가 옵션을 확인합니다"})
      user_answer = AskUserQuestion(questions=[{
        question, header, options: page1, multiSelect: false
      }])
      if user_answer == "더 많은 선택지 보기":
        // Show remaining options (up to 4 per page, repeat if needed)
        remaining = options.slice(3)
        while remaining.length > 0:
          page = remaining.slice(0, 4)
          remaining = remaining.slice(4)
          if remaining.length > 0:
            page = page.slice(0, 3)
            page.push({label: "더 많은 선택지 보기", description: "추가 옵션을 확인합니다"})
            remaining = options.slice(3 + (pageNum * 3))  // adjust offset
          user_answer = AskUserQuestion(questions=[{
            question, header, options: page, multiSelect: false
          }])
          if user_answer != "더 많은 선택지 보기": break
    answers.push({turn, question, answer: user_answer})

2A.1b — ESCALATE 3-way prompt (FR-3, gate_loop_enabled: true):

When the loop exits with an ESCALATE reason, do NOT silently proceed and do NOT loop forever — hand the user a 3-way choice. (clarity_threshold_met / pre_interview_clear / user_skip_override skip this block and go straight to 2A.2.)

ESCALATE_REASONS = {"stagnation_escalate", "soft_cap_escalate",
                    "no_more_actionable_gaps_escalate", "hard_cap_escalate"}

while gate_loop_enabled and exit_reason in ESCALATE_REASONS:
  current_score = parsed.clarity_score.get("weighted_overall") or parsed.clarity_score.get("overall", "?")
  unresolved    = parsed.get("remaining_unknowns") or []
  choice = AskUserQuestion(questions=[{
    "question": "인터뷰가 [{exit_reason}]로 종료 기준({gate_threshold})을 충족하지 못했습니다.\n"
                "현재 weighted_overall: {current_score} (목표 {gate_threshold})\n미해결: {unresolved}",
    "header": "인터뷰 에스컬레이션",
    "options": [
      {"label": "A. 인터뷰 계속 진행", "description": "추가 질문으로 명확도를 더 끌어올립니다 (hard_cap까지)"},
      {"label": "B. Open Questions로 기록 후 진행", "description": "미해결 항목을 REQUESTS.md에 남기고 다음 스테이지로"},
      {"label": "C. 인터뷰 처음부터 다시", "description": "이전 답변을 초기화하고 재시작"}
    ],
    "multiSelect": false
  }])
  # HR-1 audit: persist choice + score
  write_to_state({"current.interview.escalation_choice": choice})
  append_history({"event": "escalation_prompt", "agent": "shinnosuke",
                  "exit_reason": exit_reason, "weighted_overall": current_score, "escalation_choice": choice})

  if choice == "A":      # continue — re-enter the interview loop up to hard_cap (NFR-3)
    exit_reason = null
    resume the 2A.1 loop from turn+1 (do NOT reset answers); re-evaluate on next done
  elif choice == "C":    # restart
    answers = []; exit_reason = null; turn = 0; restart the 2A.1 loop
  else:                  # B — fall through to 2A.2 with exit_reason intact (Open Questions written)
    break

2A.2 — Finalize draft (Misae writes REQUESTS.md + runs AK review):

Task(subagent_type="team-shinchan:misae", model="sonnet",
  prompt=`mode: FINALIZE_DRAFT
  DOC_ID: ${DOC_ID} | WORKFLOW_STATE: .shinchan-docs/${DOC_ID}/WORKFLOW_STATE.yaml
  answers: ${JSON.stringify(answers)}
  user_request: ${args}
  vision_context: ${vision_context or 'None'}
  exit_reason: ${exit_reason || 'clarity_threshold_met'}
  gate_loop_enabled: ${gate_loop_enabled} | ak_double_check: ${ak_double_check}
  Per agents/misae.md: run the materiality audit (step 0, only when exit_reason == clarity_threshold_met and gate_loop_enabled), write REQUESTS.md (include ## Open Questions if exit_reason in the ESCALATE/legacy-gap set), run mechanical check, run AK review loop (max 2 retries). Return finalize-result JSON block.`)

Parse the finalize-result JSON block.

• If next == "materiality_reject" (gate-loop materiality audit failed) → re-enter the 2A.1 interview loop from turn+1 with failed_item re-added to unresolved_unknowns (do NOT reset prior answers), then re-run 2A.2. This keeps the gate honest: a PASS score is necessary but not sufficient — material ambiguity sends it back.
• If next == "closure_reject" (WS-03 4a — Misae withheld analyst-acceptance with a new actionable gap, main-075 benchmark adoption) → treat exactly like materiality_reject: re-enter the 2A.1 loop from turn+1 with failed_item re-added to unresolved_unknowns (do NOT reset answers), then re-run 2A.2. The 2-loop cap in Misae's WS-03 guarantees this cannot loop forever — after 2 closure loops Misae proceeds and logs the residual gap to ## Open Questions instead of rejecting.
• If restated_goal is present on the result (WS-03 4b — OPTIONAL additive field) → carry it into the Phase E-2 approval prompt below so the user confirms the one-sentence goal.
• If ak_verdict == "APPROVED" → continue to 2A.3.
• If ak_verdict == "ESCALATED" → show rejection_reasons to user and stop (user decides next step per Misae Phase E-4).

2A.3 — Phase E-2 user approval (parent drives AskUserQuestion):

# WS-03 4b: if finalize-result carried restated_goal, prepend it so the user confirms the
# one-sentence goal alongside approval (main-075 benchmark adoption). Optional — omit if absent.
restate_line = ("목표 확인: " + restated_goal + "\n\n") if restated_goal else ""
user_decision = AskUserQuestion(questions=[{
  question: restate_line + "REQUESTS.md을 승인하시겠어요?",
  header: "최종 승인",
  options: [
    {label: "A. 승인 — Stage 1.5 (Design)로 진행", description: "요구사항 확정 → Hiroshi와 설계를 진행합니다"},
    {label: "B. 수정 필요 — 피드백 제공", description: "어떤 부분을 바꿔야 하는지 알려주세요"}
  ],
  multiSelect: false
}])

If user picked B:
  feedback = AskUserQuestion free-form or prompt user to describe changes
  Task(misae, mode: REVISE, user_feedback: feedback)  // loops back through AK
  Repeat 2A.3.

If user picked A (or equivalent):
  Task(subagent_type="team-shinchan:misae", model="sonnet",
    prompt=`mode: TRANSITION
    DOC_ID: ${DOC_ID}
    User approved REQUESTS.md. Apply your TRANSITION mode.`)

Single source of truth for the destination stage = the agent's Mode: TRANSITION block. Do NOT name a destination stage in this prompt and do NOT write current.stage from the skill — Misae's TRANSITION already targets the design stage. (Naming "planning" here was the v4.45.0 bug that skipped Stage 1.5.) The next step is the design stage (Step 2B); design is only skipped via the explicit Step 2B skip conditions, evaluated in Step 2B.

Step 2A-post: Requirements Complete

Misae has performed hidden requirements analysis as part of FINALIZE_DRAFT. Misae's TRANSITION moves the workflow to stage: design (owner: hiroshi). Once TRANSITION returns, proceed to Step 2B (Design Stage).

Step 2B: Design Stage — Parent-Orchestrated Design Interview

CRITICAL: Sub-agents cannot call AskUserQuestion. Exactly like the requirements interview (2A), the main thread drives the design interview; Hiroshi designs each decision. See agents/hiroshi.md § "Design Stage Interview Protocol".

Skip conditions — design is DEFAULT-ON; skip is the narrow exception. Skip ONLY when one of these is unambiguously true (when in doubt, do NOT skip — run the loop):

• user_request/args contains the literal token skip-design, OR
• the workflow entered via the Quick Fix Path (Shinnosuke classified it Lite: ≤3 files, no design decisions), OR
• REQUESTS.md has zero architecture/design choices: no competing approaches, no new component/interface/schema/data-flow/layout decision, and an explicit Design decisions: none — {reason} style waiver is warranted.

A multi-FR feature, anything with ## Open Questions that are design-shaped (layouts, ranges, contracts, new keys), or any "approach A vs B" is NOT skippable — run the design loop.

When (and only when) a skip condition holds: narrate "Design stage skipped ({which condition})" and run Task(subagent_type="team-shinchan:hiroshi", prompt="mode: TRANSITION\nDOC_ID: {DOC_ID}\nDesign skipped ({reason}). Apply your TRANSITION mode.") to advance to the next stage — do NOT write current.stage from the skill (the destination lives in hiroshi.md's TRANSITION). Then go to Step 2C. Otherwise run the loop below.

2B.0 — Read design interview config (.shinchan-config.yaml, else defaults): design.soft_cap (default 5), design.hard_cap (default 8).

2B.1 — Design decision loop (design-completeness-gated, hard_cap is the ceiling):

decisions    = []
exit_reason  = null
for turn in 1..design_hard_cap:
  result = Task(subagent_type="team-shinchan:hiroshi", model="opus", prompt=
    "mode: DESIGN_NEXT_DECISION
    DOC_ID: {DOC_ID} | WORKFLOW_STATE: .shinchan-docs/{DOC_ID}/WORKFLOW_STATE.yaml
    REQUESTS.md: .shinchan-docs/{DOC_ID}/REQUESTS.md
    turn: {turn}
    prior_decisions: {decisions}
    user_request: {args}
    vision_context: {vision_context or 'None'}
    soft_cap: {design_soft_cap} | hard_cap: {design_hard_cap}
    Return the design-question JSON block per agents/hiroshi.md contract.")

  Parse the last ```design-question ... ``` fenced block in result.

  GUARD (mirror 2A.1): a `design-question` block must parse; `status` ∈ {ask, done};
    if status == "ask": `question` ≥ 5 chars, `header` non-empty, `options` ≥ 2 each with a
    substantive `label`, `decision_id` present, `closes_decision` ≤ 80 chars.
    On failure: re-invoke Hiroshi with the specific reason appended, max 2 retries; on 3rd
    failure print raw output and STOP.

  if status == "done":
    exit_reason = parsed.reason; break
  if status == "ask" (guard passed):
    // Same options-overflow pagination as 2A.1 (>4 options → "더 많은 선택지 보기").
    user_answer = AskUserQuestion(questions=[{question, header, options, multiSelect}])
    decisions.push({turn, decision: parsed.closes_decision, choice: user_answer})

2B.1b — ESCALATE 3-way prompt (when exit_reason ∈ {soft_cap_escalate, hard_cap_escalate}): hand the user a choice exactly like 2A.1b — A. 설계 계속 (hard_cap까지), B. Open Questions로 기록 후 진행, C. 처음부터. (design_complete / user_skip_override skip this and go to 2B.2.)

2B.2 — Finalize design (Hiroshi writes DESIGN.md + runs AK review):

Task(subagent_type="team-shinchan:hiroshi", model="opus",
  prompt=`mode: FINALIZE_DESIGN
  DOC_ID: ${DOC_ID} | REQUESTS.md: .shinchan-docs/${DOC_ID}/REQUESTS.md
  decisions: ${JSON.stringify(decisions)}
  exit_reason: ${exit_reason || 'design_complete'}
  Per agents/hiroshi.md: write DESIGN.md (## Open Questions only if escalated), run the
  design-stage AK review loop (max 2 retries). Return finalize-design-result JSON block.`)

Parse the finalize-design-result block.

• If ak_verdict == "APPROVED" → continue to 2B.3.
• If ak_verdict == "ESCALATED" → show rejection_reasons to user and stop.

2B.3 — User approval of DESIGN.md (parent drives AskUserQuestion):

user_decision = AskUserQuestion(questions=[{
  question: "DESIGN.md(설계)를 승인하시겠어요?",
  header: "설계 승인",
  options: [
    {label: "A. 승인 — Stage 2 (Planning)로 진행", description: "이 설계대로 Nene가 계획 수립"},
    {label: "B. 수정 필요 — 피드백 제공", description: "어떤 설계를 바꿔야 하는지 알려주세요"}
  ],
  multiSelect: false
}])

If B: Task(hiroshi, mode: REVISE, user_feedback: feedback) → repeat 2B.3.
If A: Task(subagent_type="team-shinchan:hiroshi", model="opus",
        prompt="mode: TRANSITION\nDOC_ID: {DOC_ID}\nUser approved DESIGN.md. Apply your TRANSITION mode.")

Destination stage is defined by hiroshi.md's Mode: TRANSITION (→ planning) — do not restate it here or write current.stage from the skill.

Once TRANSITION returns, proceed to Step 2C.

Step 2C: Stage Transition Narration

Shinnosuke 호출 전에 사용자에게 직접 알린다:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
👦 [Shinnosuke] Stage 1.5 완료 ✅ 설계 확정됨 (DESIGN.md)
→ Stage 2: Planning 시작합니다. Nene가 설계대로 Phase를 구성합니다.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Then invoke Shinnosuke:

Task(subagent_type="team-shinchan:shinnosuke", model="opus",
  prompt="Continue from Stage 2 via /team-shinchan:start.
  DOC_ID: {DOC_ID} | REQUESTS.md: approved | DESIGN.md: approved (architecture settled in Stage 1.5).
  Stages 1 + 1.5 DONE. Start Stage 2 (Planning) via Nene, then Stage 3 (Execution), then Stage 4 (Completion).
  CRITICAL: After Stage 3, you MUST execute Stage 4 — write RETROSPECTIVE.md, IMPLEMENTATION.md, and run final Action Kamen review. See 'Stage 4: Completion' section in agents/shinnosuke.md.

  ## Learning-capture vs heavy completion (main-076 FR-7)
  Learning/eval capture is the EXECUTION stage's definition-of-done — hooks/session-wrap.sh writes
  a deterministic skeleton each execution Stop and nudges the retrospective ONCE per stage. The
  heavy IMPLEMENTATION.md + comprehensive Action Kamen final review is a SEPARATE, push-gated
  formal completion. Never lump the two: skipping the heavy completion must not skip learning
  capture, and a captured learning does not substitute for the formal review.

  ## Micro-Task Execution (RULE 2.7) — MANDATORY, this is the /start default
  WORKFLOW_STATE.current.execution_mode is `micro-execute`. Stage 3 MUST run via the
  micro-execute path — do NOT fall back to the standard Phase Loop and do NOT use the
  dag-executor (that path skips the per-task review chain and is opt-in only).

  When invoking Nene for Stage 2 planning, request MICRO-TASK FORMAT for PROGRESS.md
  (agents/nene.md 'Micro-Task Plan Format' — NOT the DAG Plan Schema). Nene plans AGAINST the
  approved DESIGN.md (architecture is already decided — do NOT re-decide it). Each phase should be
  broken into 2-3 minute tasks with exact file paths, complete code, and verification commands.

  In Stage 3, use the micro-execute pattern (RULE 2.7): for EACH micro-task, dispatch a fresh
  implementer subagent → spec-compliance review → code-quality review → independent skeptic
  refutation. Never skip any of these four steps. See skills/micro-execute/SKILL.md for the
  full execution protocol.

  Nene's summary: {nene_result_summary}")

Stage 3 (Executing): Dispatch Mode

The executing stage has two dispatch modes, selected by current.execution_mode in WORKFLOW_STATE.yaml (set in Step 1):

• micro-execute (DEFAULT for /start) — each task runs through the full per-task review chain: implementer → spec-compliance review → code-quality review → independent skeptic refutation (skills/micro-execute/SKILL.md). This is the accuracy backbone; do NOT bypass it. Stage 2B already requests MICRO-TASK FORMAT from Nene so PROGRESS.md feeds this mode directly.
• dag (OPT-IN ONLY) — the DAG-driven dispatch below. Faster via parallelism, but its only per-task gate is the verify shell command exit code — it does NOT run the spec/quality/skeptic review chain. Use ONLY when the user explicitly opts in (sets execution_mode: dag) and accepts the weaker per-task verification. Never the silent default.

⚠️ Routing rule: if execution_mode is micro-execute or absent → use micro-execute (RULE 2.7). Only when execution_mode: dag is explicitly set do you follow the DAG protocol below.

DAG-Driven Dispatch via dag-executor (execution_mode: dag only)

When opted in, the executing stage drives implementation as a topologically-ordered task DAG using src/dag-executor.js instead of an ad-hoc sequential loop (FR-9). The requirements → planning → executing → done stage machine is preserved; only the executing stage's dispatch logic changes.

Protocol for the executing stage:

1. Parse the plan's task DAG (src/dag-executor.js parsePlan() reads the six-field schema — id / depends_on / touches / verify / estimate / scope — from PLAN.md or a structured object).
2. topoSort() orders tasks by depends_on and throws on a circular dependency before any task is dispatched (no partial execution).
3. buildConflictGraph() + connectedComponents() compute the static conflict graph from touches[]: tasks in the same component serialize; disjoint components dispatch in parallel up to executor.concurrency_cap (.shinchan-config.yaml, default 4) — mapped onto the Workflow parallel() primitive.
4. Each task runs through the per-task verify gate: its verify command must exit 0 to be marked DONE; an absent/NL-only verify is FAILED (never a silent SKIP), and downstream dependents become BLOCKED.
5. On task error the recovery ladder branches by error type (transient → retry ≤ 3; deterministic → local-patch; scope-drift → replan, bounded) with counters persisted to WORKFLOW_STATE (no infinite loops, survives restart).
6. After all per-task gates pass, the post-merge integration test (plan_meta.integration_test) runs; if absent it warns but does not block.
7. The completion gate (evaluateCompletionGate()) enforces strict ALL-PASS — zero FAILED/SKIP, all DONE, integration passing — and blocks the executing → done stage transition otherwise.

Serial-fallback / escape hatch (high coupling): when the conflict graph collapses to a single connected component (every task shares a resource), execution is fully serial and the executor logs an NFR-5 warning (serial_fallback: true in .shinchan-config.yaml). This mirrors the gate_loop_enabled-style escape so operators can inspect plan granularity rather than chase phantom parallelism.

See src/dag-executor.js and docs/dag-executor.md for the module contract.

Prohibited

• Only explaining without executing
• Skipping folder/YAML creation
• Invoking Shinnosuke for Stage 1 (must use Misae directly)
• Gathering requirements without invoking Misae