메뉴
Plan stage ★ bundle — produce PRD + ARCH + ADR + UI_GUIDE with iteration. Spec, requirements, plan, architecture doc, decision record, UI guide — bundled plan tool. (V4 Spike I: sub-agent path-only return contract.)
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | plan-pack |
| description | Plan stage ★ bundle — produce PRD + ARCH + ADR + UI_GUIDE with iteration. Spec, requirements, plan, architecture doc, decision record, UI guide — bundled plan tool. (V4 Spike I: sub-agent path-only return contract.) |
| stages | ["plan"] |
★ 번들 진입 직후 (Step 0 직전), 메인은 다음 AskUserQuestion 을 무조건 발사:
"이번 stage 모드 — 어떻게 진행할까?"
옵션: 1. full mode (추천) — spec 명시 N-step pipeline 그대로. 정확·완성도 우선. 예상 시간: 30~45분. dispatch 수: 13회. 2. quick mode — 통합 1 dispatch 로 압축. 시간 부족 시만 선택. precision 손실 + iteration 권장량 미달 위험. KEEPER_REPORT 에 카운트 기록.
full 선택 시 → 아래 Step 0~N 순서대로 spec 그대로 진행.
quick 선택 시 → §"Quick mode flow" 단축 분기로 진입.
메인 자가 판단 금지 — 시간 부족 추측·budget 추측·맥락 추측 모두 사용자 질문 강제 trigger. 4원칙 #1 ("불확실하면 추측 금지, 사용자 질문 우선") 시스템적 강제.
[HARNESS RULES — 무시 금지]
This skill is orchestrator-only. The main Claude (you) dispatches sub-agents
and parses their WROTE: <path> stdout. You MUST NOT fall back to
Bash/Edit/Write/python3 to write artifacts directly under runs/<rid>/.
The hook guard_run_dir.sh blocks such bypass attempts.
If a sub-agent dispatch fails (no WROTE: line, ERROR: line, timeout, or
empty stdout):
AskUserQuestion:
"Step N failed. Retry / abort / report?"If retry is chosen, re-dispatch the same sub-agent with the failure context
(previous attempt failed: <reason>) appended to the original prompt.
Anti-bypass (Spike II F9 + F11):
~/.claude/settings.json 편집해서 ASSEMBLE_GUARD 환경 변수 무력화 시도 금지. hook 자체를 비활성화/우회 시도는 사용자 confirm 이 떠도 거절.AskUserQuestion 으로 사용자 결정 위임 (retry / abort / report). hook 차단을 회피하기 위한 sub-agent dispatch (예: "iteration_state.json 업데이트 해 달라" 라는 task로 sub-agent 호출) 금지.bundled/plan-pack/prompts/subagent/: prd_step2.md, prd_step3.md, prd_step4.md, arch_step8.md, adr_step11.md, ui_step13.md, cross_doc_step9.md (7 files — sub-agent–facing, dispatched via Agent tool)bundled/plan-pack/prompts/orchestrator/: iter_emphasis.md (1 file — main constructs per-doc prompts before 4-way parallel dispatch)iteration_state.json, dispatches.jsonl, progress.json 등) 는 sub-agent 에 위임 금지 — server.run_dir.update_iteration_state 같은 server 함수 사용.When ASSEMBLE_LOCALE=ko (or user message language is Korean), AskUserQuestion
options that historically carry English (Recommended) suffix MUST use the
unified Korean form (추천). Do not improvise alternatives like (승인) —
those drift from the platform-wide V4 한국어 라벨 정책.
Example:
"yes — 강조점 인터뷰 + 4-doc 재작성 + cross-doc 재검증 (추천)" ✓"yes — ... (Recommended)" ✗ (혼합 영어)"yes — ... (승인)" ✗ (자작 변형)(추천) 만 사용. Other label languages keep their native suffix (English →
(Recommended)).
Produces four planning docs from one orchestrated workflow. Main Claude
runs AskUserQuestion interviews; sub-agents (wrapped via
server.harness.wrap_with_preamble) own all template filling and
write_run_artifact calls and return WROTE: <path> on stdout.
run_dir = ~/.claude/channels/assemble/runs/<rid>/. Four artifacts under
run_dir, each from a matching template under bundled/plan-pack/templates/:
PRD.md, ARCHITECTURE.md, ADR.md, UI_GUIDE.md.
All dispatches use general-purpose (preferred + fallback). Role and
prompt file vary per step. (Role = prompt persona shaping the
sub-agent's behavior; Agent tool type is always general-purpose.)
| Step | Role | Prompt file |
|---|---|---|
| 2 | plan-implementation | prompts/subagent/prd_step2.md |
| 3 | plan-implementation | prompts/subagent/prd_step3.md |
| 4 | second-opinion | prompts/subagent/prd_step4.md |
| 8 | plan-implementation | prompts/subagent/arch_step8.md |
| 9 | second-opinion | prompts/subagent/cross_doc_step9.md |
| 11 | plan-implementation | prompts/subagent/adr_step11.md |
| 13 | plan-implementation | prompts/subagent/ui_step13.md |
Steps 0/1/7/10/12 are user-facing AskUserQuestion interviews. Steps 2+3
fire as a single parallel message (B-1 surface). Steps 8/11/13 are
single-dispatch first-pass; Step 6 yes-path promotes 8/11/13 + 2/3 to
4-way parallel (B-5).
0 → 1 → (2 + 3 in parallel) → 4 → 7 → 8 → 10 → 11 → 12 → 13 → 9 → 6
Step numbers reflect historical addition order, not execution order. Step 4
(PRD second-opinion) sub-agent writes the updated PRD.md itself, so no
separate main-write step is needed (the prior Step 5 has been deleted in
Spike I). Step 9 runs after all four docs are written. Step 6 (iteration) is
final. Do NOT run steps in numeric order.
Read <rid> from the active assemble run. The artifact path is
~/.claude/channels/assemble/runs/<rid>/PRD.md. If the file already
exists, treat the workflow as iteration mode (Step 6 yes-path entry).
사용자 명시 동의 없이 단축 금지 — N-step pipeline 의 각 step 은 별도 sub-agent dispatch 로 진행. 메인이 단축 결정 시 4원칙 #1 위반. Mode-gate 가 quick 으로 답한 경우만 §"Quick mode flow" 분기 허용.
For each dispatch step:
prompt_text = server.dispatch_prompt("<file>.md") — loads the prompt
under bundled/plan-pack/prompts/{subagent,orchestrator}/ (resolver
checks both subdirs + flat fallback) and prepends the harness preamble
via server.harness.wrap_with_preamble (byte-identity contract
surface). Unknown <file>.md raises ValueError (anti-bypass guard,
Spike III §1.2). The function does not substitute placeholders —
{{KEY}} tokens come back intact.prompt_text yourself, e.g.
prompt_text = prompt_text.replace("{{TASK}}", task). The orchestrator
owns which {{KEY}} tokens are caller-substituted (Inputs section)
vs. preserved for the sub-agent's own .replace instructions inside
the canonical save block. A naive global .replace over both classes
would corrupt the latter — Spike III §1.2 rationale.server.harness.record_dispatch for the on-disk evidence trail
(runs/<rid>/dispatches.jsonl) — invoke as
server.record_dispatch(run_id, step, prompt_text, subagent_type="general-purpose", prompt_file="<file>.md", description="...", wrote_path=...)
to append the hash-only audit row. The prompt_file= kwarg matches
the allowlist (Spike III §1.2).general-purpose via the Agent tool with prompt_text.WROTE: <path> on stdout — parse with regex
^WROTE: (.+)$. On ERROR: or missing WROTE:, follow §CRITICAL.Sub-agents MUST replace every <TBD: ...> sentinel inside their
canonical save-block triple-quoted strings with concrete content before
printing WROTE:. A literal <TBD: ...> (or bare ...) reaching
write_run_artifact is a contract violation — surface as ERROR: and
follow §CRITICAL. Spike III §2.1 + tests/unit/test_prompts_no_bare_ellipsis.py.
record_dispatch 시그니처 (verbatim — from server import record_dispatch):
record_dispatch(
run_id: str,
step: str,
prompt_text: str,
*,
subagent_type: str = "",
description: str = "",
wrote_path: Optional[str] = None,
prompt_file: Optional[str] = None,
) -> Path
role kwarg 없음. 위 표 (Step → Role) 의 Role 은 prompt 본문에 자연어로 박히는 페르소나 라벨이고 subagent_type 은 항상 "general-purpose" 다. record_dispatch(..., role=...) 로 호출하면 TypeError. prompt_file 은 ALLOWED_PROMPT_FILES 외 값이면 stderr 경고 (default) 또는 ValueError (ASSEMBLE_DISPATCH_STRICT=1).
8 questions across two AskUserQuestion calls of 4 questions each
(platform maxItems: 4). Treat as single batch. Call 1 (Q1–Q4): (1) What
building? (2) Who uses it? (3) Three core features? (4) Three exclusions
from MVP (harness #2). Call 2 (Q5–Q8): (5) One-line success criterion?
(6) One AC bash command for external verification? (7) One-line design
direction? (UI_GUIDE seed) (8) One risk / open question?
Prompt: prompts/subagent/prd_step2.md. Placeholders: {{TASK}},
{{INTERVIEW_ANSWERS}}, {{RUN_ID}}. Fired in the same parallel message
as Step 3 (true 2-way parallel — see
docs/research/2026-04-29-platform-limit.md). Sub-agent partially
populates PRD.md; Step 3 fills the AC bash placeholder.
Prompt: prompts/subagent/prd_step3.md. Placeholders: {{RUN_ID}},
{{SUCCESS_CRITERION}}, {{AC_REQUEST}}.
Prompt: prompts/subagent/prd_step4.md. Placeholders: {{RUN_ID}}.
Single Agent call as second-opinion (preferred general-purpose per v4
swap commit 85366f1). Sub-agent reads PRD.md, triages each critique
bullet via the Step 4b verify-before-appending protocol (1-shot Bash for
runtime claims; drop speculation), prepends an audit header, appends
## Review notes, rewrites via write_run_artifact. Step 4 is skipped on
the iteration yes-path — Step 9 provides second-opinion coverage there.
After Step 4 returns, 6 questions across two AskUserQuestion calls of
3 questions each. Call 3 (A1–A3): (1) Primary tech stack? (2) Top-level
directory structure? (3) Architectural patterns (MVC, microservices,
event-driven, monolith, CQRS — name + rationale)? Call 4 (A4–A6):
(4) Main data flow ≤3 steps? (5) External services / third-party APIs?
("none" valid) (6) Main modules + boundaries?
Prompt: prompts/subagent/arch_step8.md. Placeholders: {{TASK}},
{{INTERVIEW_ANSWERS}}, {{RUN_ID}}. Sub-agent reads PRD, builds the 6
ARCH sections (Stack, Directory tree, Architectural patterns, Data flow,
External dependencies, Module boundaries), fills
templates/ARCHITECTURE.md.template, writes. Proceed to Step 10.
After Step 8 returns, two phases of AskUserQuestion:
Call 5 — single AskUserQuestion call with 3 sub-questions (D1, D2, D3 — title only). Each sub-question shape: PRD Q1–Q4 batch shape (multi-select: choose top decisions). multi-select schema MUST include minSelected: 3, maxSelected: 5 (gate B3.2 minimum). After response, main MUST verify the user selected ≥ 3 decisions; if fewer, re-prompt the same Call 5 with the message "최소 3개 결정이 필요합니다" — do NOT proceed with 2 decisions even if AskUserQuestion returns successfully.
Call 6 — 3 separate AskUserQuestion calls, one per decision chosen in Call 5, each with 3 sub-questions: (a) strongest rejected alternative, (b) main tradeoff, (c) decision-specific risks/unknowns. The three sub-questions are numbered 1/2/3 within each call. Three calls in sequence, NOT parallel — each subsequent call's wording references the prior decision title for user context.
Three decisions = minimum (gate B3.2); volunteer more → N ≥ 3, accept up to 5.
(Spike II F5/F6: spec 모호함 차단 — "1-question multi-select" 또는 single 6-question call 자유 재해석 금지.)
Prompt: prompts/subagent/adr_step11.md. Placeholders: {{TASK}},
{{INTERVIEW_ANSWERS}}, {{RUN_ID}}. Sub-agent reads ARCH + PRD,
synthesizes ### Context and ### Reasoning per decision (these two
sub-headings are the sub-agent's job, never the user's), emits
## Decision N: blocks with all five sub-headings, fills
templates/ADR.md.template, writes. Proceed to Step 12.
After Step 11 returns, 6 questions across two AskUserQuestion calls of
3 questions each.
Call 7 (U1–U3): (1) visual identity / aesthetic, (2) 3 priority flows, (3) 3 required component patterns. Sub-questions U2 and U3 are multi-select with minSelected: 3, maxSelected: 3 — exactly 3 items each. Main MUST verify the count after response; if user supplied 4+ items via free text or the schema returned fewer than 3, re-prompt that single sub-question only (don't restart Call 7). Spike II F7: B-6 dogfood received 4 answers for U2/U3 — schema 강제 + main 검증으로 차단.
Call 8 (U4–U6): (4) ≤5 brand color tokens (hex/named with role; "no preference" valid), (5) primary + 1 supporting font family ("no preference" valid), (6) project-specific antipattern emphasis beyond template baseline ("none" valid). Q6 surfaces project-specific antipattern signal for gate B4.3.
Prompt: prompts/subagent/ui_step13.md. Placeholders: {{TASK}},
{{INTERVIEW_ANSWERS}}, {{RUN_ID}}. Sub-agent reads PRD ## Design direction + ## Core features, builds the 5-section UI body (Visual
identity, Color tokens, Typography, Component patterns ≥3, Priority
screens ≥3), respects the template's antipattern keyword exclusion list
(gate B4.3), fills templates/UI_GUIDE.md.template, writes. Proceed to
Step 9.
Prompt: prompts/subagent/cross_doc_step9.md. Placeholders: {{RUN_ID}},
{{ITERATION_COUNT}} (read from runs/<rid>/iteration_state.json; 0 for
first-pass). Single Agent call as second-opinion role.
Sub-agent reads all four artifacts, produces verified findings under 7
categories (PRD↔ARCH gap, ARCH↔ADR decision integrity, PRD↔ADR motivation
traceability, PRD↔UI_GUIDE design audit, ARCH↔UI_GUIDE component coverage,
ADR↔UI_GUIDE UX integrity, numerical/unit consistency), applies the
Step 4b triage, picks heading (## Cross-doc review for first-pass,
## Cross-doc review (iteration N) for iterations), runs the precondition
assert heading not in adr_text, and rewrites ADR.md.
Sub-agent prints two stdout lines — WROTE: <path> and
COUNTS: resolved=<N> unresolved=<N> new=<N>. Main parses the second
stdout line with regex ^COUNTS: resolved=\d+ unresolved=\d+ new=\d+$
(Spike II F10). 매칭 실패 시 (다른 키, 키 누락, 비정수, 추가 토큰)
Step 9 dispatch failure 처리 — §CRITICAL 분기 (retry/abort/report).
Parse both. To update
runs/<rid>/iteration_state.json (drives the multi-iteration stop
condition), call server.update_iteration_state directly — do not
dispatch a sub-agent for this:
from server import update_iteration_state
update_iteration_state(rid, {
"resolved_pct": resolved / max(resolved + unresolved + new, 1),
"new_count": new,
"stopped": False,
"reason": "",
})
orchestrator metadata 는 §CRITICAL 룰 (Spike II F11) 에 의해 sub-agent 위임 금지. hook 화이트리스트 (Spike II F8) 가 main 직접 write 를 허용한다. If the COUNTS: line is missing or unparseable (non-integer values, missing key), treat as Step 9 dispatch failure per §CRITICAL — surface to user, do NOT advance to Step 6.
If the precondition assert fires (heading already in ADR.md from a failed
prior iteration overwrite), the sub-agent prints
ERROR: contract violation: heading ... already in ADR.md. Surface to the
user via AskUserQuestion ("Step 9 contract violation — retry / abort?")
— do NOT auto-recover or edit ADR.md directly. On any other ERROR: or
missing WROTE:, follow §CRITICAL. After Step 9 returns, proceed to Step 6.
The orchestrator picks one of two AskUserQuestion prompts based on
iteration_count (read from runs/<rid>/iteration_state.json; 0 if file
absent). The two prompts NEVER both fire on the same iteration boundary.
| iteration_count | prompt | options (Korean) |
|---|---|---|
| 0 | "네 문서 작성 완료 — PRD.md, ARCHITECTURE.md, ADR.md, UI_GUIDE.md. 한 차례 반복 진행할까?" | yes — 강조점 인터뷰 + 네 문서 재작성 + 문서 간 재검증 / no — 종료 |
| ≥ 1 | "반복을 계속할까?" | yes — 강조점 인터뷰 + 네 문서 재작성 + 문서 간 재검증 한 라운드 더 / no — 여기서 종료 |
no from either prompt terminates the loop (V4 identity rule — user is
never forced into another pass). yes from iteration_count == 0
enters the §"Step 6 yes-path detail" flow; yes from iteration_count ≥ 1
re-enters the same flow (the §"User exit override" block below documents
the exit-side wording).
Entry (iteration 0): per the selector table above.
각 prompt 호출 직전, orchestrator 가 iteration_state.json 의
counts (resolved, new, unresolved) 와 iteration_count 를 읽고
다음 알고리즘으로 default-recommended option 선택:
def recommend_iter_continue(iteration_count, resolved, new, unresolved):
if iteration_count >= 7:
return "no" # max iteration reached, ★ guideline
if iteration_count <= 3:
# 초기 iteration — 거의 항상 yes 권장
if new > 0 or unresolved > 0:
return "yes"
if resolved < 5: # 이슈 자체가 적으면 더 고민할 거리 없음
return "no"
return "yes"
# iteration_count 4~6
if new > 0:
return "yes"
if unresolved > 0:
return "yes"
if resolved >= 8: # 충분히 정제됨
return "no"
return "yes" # 의심스럽다면 yes
AskUserQuestion options 빌드 시:
(추천 — 사유: <iteration_count=N, resolved=X, new=Y>) 추가iteration_state.json 이 missing 또는 malformed 일 때 default = yes
(보수적 — multi-iteration 약속 보존).
After user picks "yes — 강조점 인터뷰 + 네 문서 재작성 + 문서 간 재검증 (추천)":
For each of PRD.md, ARCHITECTURE.md, ADR.md, UI_GUIDE.md,
AskUserQuestion: "이번 라운드에서 <DOC>의 강조할 측면 한 줄
요약 — 변경 없으면 (no change)".
Build the iter1 dispatch list. For each doc:
If user emphasis is empty or literal (no change):
server.dispatch_and_record(
run_id,
prompt_file="iter_emphasis.md",
step=f"step6.iter{N}.{DOC}",
status="skipped",
note=f"user emphasis: (no change)",
)
# No Agent dispatch. Audit row written.
Else, prepare the wrapped prompt (caller substitutes per-doc placeholders before sending to Agent):
prompt = server.dispatch_and_record(
run_id,
prompt_file="iter_emphasis.md",
step=f"step6.iter{N}.{DOC}",
description=f"iter{N} {DOC} emphasis redraft",
)
# Substitute {{RUN_ID}} / {{DOC_NAME}} / {{EMPHASIS}} / etc. in
# `prompt` before invoking Agent. Audit row already written by
# dispatch_and_record.
Send a single message with one Agent call per non-skipped doc
(parallel dispatch, V4 결정 #11). Each Agent call uses
subagent_type="general-purpose" and the substituted prompt.
docs/research/2026-04-29-platform-limit.md confirms 4-way + 5-way
works. Sequential fallback only on documented input dependency or
detected retry-after.
Each substituted prompt embeds the iteration scope discipline rule verbatim — do not paraphrase:
Scope discipline: PRD
## Core featuresis the authoritative scope. Do not introduce new features, modules, components, screens, or token sets that have no counterpart in the existing PRD## Core features. Items the ADR has explicitly deferred (> **Future ADRs**) MUST NOT be pre-emptively decided. Existing sections that are not the explicit target of the iteration emphasis MUST be returned verbatim — do not reword Reasoning/Tradeoffs/ Rejected-alternatives blocks. Pre-existing identifiers (variable, token, module, component names) MUST NOT be renamed unless the rename IS the requested change.
Parse WROTE: <path> lines from each sub-agent. Re-dispatch
cross_doc_step9.md for the iter1 cross-doc review.
Audit invariant (Spike IV §1.1): every iteration produces exactly
4 rows in dispatches.jsonl for step6.iter{N}.<DOC> — one per doc,
each with status set to "dispatched" or "skipped". There must be
no missing rows.
UI_GUIDE.md is always re-run alongside PRD, ARCH, ADR — produced as a quadruple, must stay consistent.
Phase B-5 replaced the 1-iteration cap with an explicit stop-condition loop (B-2/B-3/B-4 each showed iteration resolving findings but introducing NEW ones at the cap).
Stop condition (verbatim — do not paraphrase in implementations):
The orchestrator continues iterating while either condition holds: (a) Step 9 review reports
NEW ≥ 1for the just-completed iteration, OR (b) the most recent two iterations have not both satisfiedRESOLVED ≥ 80% AND NEW ≤ 0. Iteration stops when two consecutive iterations both satisfyRESOLVED ≥ 80% AND NEW ≤ 0, or when the iteration counter reaches 7, whichever comes first.
Iteration state tracking (verbatim):
The orchestrator maintains a per-run state file at
runs/<rid>/iteration_state.jsonwith shape{"iterations": [{"index": N, "resolved_pct": F, "new_count": N, "stopped": bool, "reason": "..."}, ...]}. The file is updated after each Step 9 cross-doc review. Termination reason is one ofstop-condition-met,cap-reached,user-requested-stop.
After each Step 9 review the main Claude parses the
COUNTS: resolved=N unresolved=N new=N second-stdout line, computes
resolved_pct = resolved / (resolved + unresolved + new), calls
update_iteration_state(rid, {...}) (server function, see Step 9), and
decides whether to continue. The function appends to iterations[] with
auto-assigned index.
Exit (iteration ≥ 1): per the selector table above. On no, record
reason: "user-requested-stop". The user is never forced through
additional iterations (V4 identity rule).
If the iteration counter reaches 7 without the stop condition firing and
without a user no-answer, emit a one-line warning to the user citing the
cap and the unresolved count (e.g. "iteration cap (7) reached with 2
unresolved findings; exiting"), record reason: "cap-reached", and stop.
docs/specs/2026-04-30-v4-spike-i-design.md (commit eeb6c96) — §3.4 anti-fallback verbatim, §4.3 Step body templatesdocs/dogfood/phase-b-5.md (run 20260429-135600-3b6d)bundled/_shared/harness-preamble.md (rules 5/6, commit 6598788)bundled/plan-pack/prompts/subagent/{prd_step2,prd_step3,prd_step4,arch_step8,adr_step11,ui_step13,cross_doc_step9}.md, bundled/plan-pack/prompts/orchestrator/iter_emphasis.mdMode-gate 가 quick 으로 답한 경우만 진입 (full 이면 이 section 미사용).
server.dispatch_prompt('plan_pack_quick.md') 로 단일 sub-agent dispatch:
prompt = server.dispatch_and_record(
run_id,
prompt_file="plan_pack_quick.md",
step="plan-pack.quick",
description="plan-pack quick mode — single-dispatch fallback",
)
# Substitute placeholders in prompt, then dispatch via Agent (general-purpose).
# Sub-agent must produce the FULL artifact schema (all sections that full mode
# would write across N steps), in a single pass.
dispatches.jsonl audit row 의 description 필드에 mode=quick 메타 명시
(KEEPER_REPORT 가 카운트 집계 시 사용).
Quick mode 라도 산출물 (예: PRD.md / IMPL_REPORT.md / DEBUGGER_LOG.md 등) 의 sections schema 는 full mode 와 동일. precision 만 떨어질 뿐 schema 미준수 X.
KEEPER_REPORT § "Mode usage" 가 quick 카운트 표시. 다음 run 에서 시간 확보 권장.
Tool concierge — scan installed Claude Code skills/plugins/agents, recommend a per-stage workflow, and let the user pick which tool to use at each stage. Use when the user types `/assemble <task description>` or `/assemble resume` or `/assemble list`.
디자인 시스템 + AI 슬롭 안티패턴 + UI 가이드를 1페이지에 정리하는 표준 번들 (V4 design stage)
모호한 아이디어를 사용자/문제/wedge 1페이지로 정리하는 표준 번들 (V4 discover stage)
Meta-stage ★ bundle. Audits finished run artifacts (parsed_scope, dispatches.jsonl, REPORT.md) for 5-rule violations, summarizes via bounded LLM, appends to global learnings ledger with deterministic prune. Observational, not gating.
Execute stage ★ bundle — TDD-first red→green pipeline (test_first.sh → impl → verify.sh). SCOPE.md contract + IMPL_REPORT.md audit trail. (V4 Spike V: parallel to debugger ★, sub-agent path-only return contract.)
Debug stage ★ bundle — systematic hypothesis → reproducer → bisect → root cause → fix workflow with audit trail. (V4 Spike IV: parallel to plan-pack ★, sub-agent path-only return contract.)