cstage 0.2.0

McKinsey-style AI business consulting tool — Socratic interview to strategy report

cstage

McKinsey-style AI business consulting tool. Socratic interview to strategy report, fully automated.

API key 불필요 -- Claude Code Max Plan 구독으로 동작합니다.

---

Install

git clone https://github.com/ZEP-Biz/cstage ~/Projects/cstage
bash ~/Projects/cstage/install.sh
# Claude Code 재시작

설치 스크립트가 자동으로:

• Python 의존성 설치 (uv)
• MCP 서버 등록 (~/.claude/.mcp.json)
• 필요 MCP 체크 (Tavily, Slack 등)

Usage

cstage "매출이 정체되고 있다"
cstage "동남아 시장 진출 전략"
cstage "일본 부등교 사업 전략을 세우려고 해"

Workflow (6 Steps)

[1] 인터뷰      참고자료 수집 → 가설 기반 소크라틱 인터뷰 (ambiguity <= 0.2까지)
[2] RFI 확인     SCR + 리서치 범위 확인
[3] 보고서 설정   형태(markdown/slide_deck) + 이해관계자 설정
[4] 중간 보고     초안 보고서 + 추가 질문
[5] 자동 개선     evaluate → evolve → stakeholder review (최대 3회)
[6] 최종 확인     승인/수정/거부

사용자가 하는 것: 인터뷰 답변 + 중간 승인. 나머지 자동.

Architecture (host-driven dispatch — ouroboros v0.44 contract)

MCP Server = state management + payload building + validation gates
  ↕ tool responses carry meta.host_action=spawn_subagents + payload arrays
  ↕ Claude Code (the host) spawns subagents, does ALL LLM work,
    folds structured results back via work_results
  ↕ No API key, no server-side subprocess (CSTAGE_DISPATCH=internal for legacy)

Quality is code-enforced, not model-promised:

• 인터뷰 완료 = 점수 게이트 + 컴포넌트 floor + 연속 2라운드 + ledger 구조 완결
• deterministic floor가 LLM의 낙관 점수를 코드로 교정
• 보고서 승인 = verify-by-default 게이트 (기준 미달이면 모델이 승인 불가)
• 이해관계자 리뷰에 adversarial probe 9종, 트리거 시 ADVOCATE/DEVIL/JUDGE 심의

Core Modules

Module	Purpose
bigbang/	Socratic interview, ambiguity scoring, consulting ledger
dispatch/	Host-driven work payloads (SubagentPayload, fold-back contracts)
verification/	Typed verdicts, adversarial probes, report criteria gate
core/	SCR builder, RFI generator, LLM adapter, model pins
frameworks/	Porter, SWOT, BCG, McKinsey 7S
data/	Deep research planner, Tavily client, MCP adapters
report/	Storyline report models, prompts, renderer
evaluate/	3-stage evaluation pipeline
evolve/	Wonder/reflect/convergence loop
ralph/	Stakeholder review, Slack interviews
agents/	Persona prompts (strategy-consultant, deep-researcher, etc.)

Report Format

McKinsey consulting deck structure:

Executive Summary (SCR)
  Situation → Complication → Resolution
Chapter 1 — Head message (one governing sentence)
  Sub-point ← backup data [확인됨/추정/가설]
  Sub-point ← backup data
Chapter 2 — Head message
  ...
Sensitivity Analysis — base/upside/downside scenarios
Action Items — human-required vs AI-completed
Implementation Roadmap — Week 1-2 through Month 7-12

MCP Tools

Tool	Description
start_consulting	Start session (with reference_materials, stakeholders)
submit_interview_answer	Answer Socratic question
complete_interview	Generate SCR + RFI (ambiguity gate: <= 0.2)
align_seed	Approve analysis plan
plan_collection	Get deep research plan (5-10 queries/RFI)
inject_collection	Inject research results
run_analysis	Execute frameworks + generate report
get_report	Retrieve report (markdown/slide_deck)
run_stakeholder_review	5-stakeholder evaluation
confirm_report	Approve/revise/reject
suggest_slack_interviews	Suggest people to interview via Slack
send_slack_interview	Send interview DM
check_slack_responses	Check Slack interview responses

Configuration

# Required
# (none — uses Claude Code Max Plan)

# Optional
export TAVILY_API_KEY="tvly-..."    # External research (recommended)
export CSTAGE_REPORT_FORMAT="markdown"  # markdown | slide_deck
export CSTAGE_LOG_LEVEL="INFO"

Development

uv run pytest tests/ -q                            # Run tests (5,300+)
uv run cstage-server                               # Start MCP server locally
uv run python scripts/e2e_host_driven_driver.py    # 39-check full-pipeline E2E
uv run python scripts/benchmark_gates.py .         # Gate defense benchmark

Update

cd ~/Projects/cstage && git pull && uv sync
# Claude Code 재시작

---

Forked from Ouroboros — ouroboros C pattern