cold-start（冷启动）

Cold Start: Memory Bank + Agents bootstrap (greenfield & brownfield)

Summary

Note: cold-start is the package skill for all-in-one bootstrap. Do not confuse it with the generated project command /cold-start: that command is a lightweight router used after skeleton creation inside a target repo. For modular usage, prefer individual skills: mb-init (skeleton), mb-from-prd (greenfield), mb-map-codebase (brownfield), mb-review (review), mb-execute (implementation), mb-verify (UAT).

• What it does: creates the Memory Bank skeleton, writes agent entry points, and routes the repo into the right workflow.
• Use it when: you want one entry point for either a new project with a PRD or an existing codebase that needs mapping first.
• Input: repository root plus either prd.md / requirements text or an existing codebase.
• Output: .memory-bank/, .tasks/, .protocols/, agent entry files, and the next step for greenfield or brownfield work.

Supported scenarios:

• Greenfield: repo has prd.md or requirements text, but no code yet.
• Brownfield: repo already contains code and needs as-is documentation before change planning.

Non-negotiable principles

1) Orchestrator vs workers

• You (main agent) are the orchestrator. You plan and coordinate.
• Workers (subagents) do the heavy lifting (scanning many files, producing long reports).
• Max depth = 2: orchestrator → subagent. No sub-sub-agents.

2) .tasks/ is runtime memory (НЕ Memory Bank)

• Every agent process gets a unique task folder: .tasks/TASK-XXX/.
• Subagents write *detailed* outputs there, and return *only short summaries* to the orchestrator.

3) Avoid conflicts by design

• Split work by non-overlapping file sets.
• If overlap is unavoidable, use git branches/worktrees per agent and merge later.

4) Parallelism limits

• Run up to 5–7 subagents in parallel.
• If your runtime supports multi-tool calls: spawn parallel subagents in a single orchestrator turn.

5) Docs First

After completing any meaningful unit of work:

1. Update Memory Bank while context is fresh.
2. Then change code / commit.

Output you MUST create/maintain

Repo root

• AGENTS.md (canonical, short, ~100 lines)
• CLAUDE.md → symlink (or copy) to AGENTS.md
• *(optional)* GEMINI.md → symlink (or copy) to AGENTS.md

Repo folders

• .memory-bank/ — durable knowledge base
• .tasks/ — operational runtime memory
• .protocols/ — file-based protocols (plans / decision logs / resume)

Step 0 — Detect environment (don’t guess)

1. Check what tools you have:

- Codex available? (Codex CLI / MCP tool) - Claude Code available? (CLI claude, subagents) - Browser tools? (Playwright / agent-browser / CDP MCP)

1. Decide primary engine:

- If working inside Codex: use gpt-5.2-high for implementation and review (prefer fresh sessions for critical reviews). - If working inside Claude Code: use subagents for analysis + use Opus for synthesis; optionally call Codex via shell for second opinion or structured review. - If both Codex and Claude are available (dual-engine): use Claude for wide-context scanning and synthesis, Codex for structured review and implementation. Cross-validate critical outputs (MB compliance, architecture) by running the review step on the other engine. See Step 4 for the hybrid workflow.

If you don’t control Codex model selection via UI, create a project .codex/config.toml profile set (see Step 1.5).

Step 1 — Bootstrap the skeleton

1.1 Create directories

Create (if missing):

• .memory-bank/

- mbb/ - architecture/ - guides/ - adrs/ - tech-specs/ - domains/ *(optional but recommended)* - contracts/ *(optional but recommended)* - runbooks/ *(optional but recommended)* - workflows/ *(optional; can keep wfs/ if you already use it)* - quality/ *(optional but recommended)* - testing/ - skills/ - epics/ - features/ - tasks/ *(Memory Bank backlog & plans)* - commands/ *(slash-command specs used by humans/agents)* - agents/ *(subagent prompt library)* - archive/ - bugs/

• .tasks/
• .protocols/

1.2 Create core files (use the templates)

Use the templates in:

• ./references/shared-structure-template.md

At minimum you must create:

• AGENTS.md
• CLAUDE.md symlink/copy
• .memory-bank/index.md
• .memory-bank/mbb/index.md
• .memory-bank/product.md
• .memory-bank/requirements.md
• .memory-bank/testing/index.md
• .memory-bank/tasks/backlog.md

Also create the command specs under .memory-bank/commands/ (use references/commands/*).

1.2.1 Create native skills (proxy commands)

Create thin proxy skills so commands work natively in each runtime:

• .claude/skills/<name>/SKILL.md → Claude Code + OpenCode
• .agents/skills/<name>/SKILL.md → Codex CLI + OpenCode

Each proxy just says: Read and follow the instructions in.memory-bank/commands/<name>.md. This makes commands available natively (/mb, /prd, /execute, etc.) in all three tools.

The init-mb.js script creates both sets automatically.

1.3 Enforce frontmatter rule

Every markdown file inside .memory-bank/ must include YAML frontmatter with at least:

• description:...
• status: draft|active|deprecated|archived

1.4 Create .tasks/ protocol

Create a task folder for this run (pick a new id):

• .tasks/TASK-MB-INIT/

Inside it, create:

• TASK-MB-INIT-S-00-orchestrator-plan.md — what you will do + which subagents

1.5 Optional: Codex profile presets

If Codex is used, create .codex/config.toml with profiles:

• default: coding (gpt-5.2, high)
• profile deep-review: review (gpt-5.2, xhigh)

(If your repo is shared, consider keeping it local or documenting it in .memory-bank/runbooks/.)

Step 2 — Choose scenario

Decision rule

• If repo has substantial code (src/, package.json, go.mod, Cargo.toml, etc.) → Brownfield (Step 3B).
• If repo is mostly empty and you have prd.md → Greenfield (Step 3A).
• If both exist: treat as Brownfield + PRD delta (Step 3B).
• If repo is empty/new and no prd.md → Skeleton-only (Step 3C).

Record the scenario in:

• .tasks/TASK-MB-INIT/TASK-MB-INIT-S-00-orchestrator-plan.md

Step 3A — Greenfield workflow (PRD → Memory Bank)

3A.1 Read PRD and do Deep Questioning

• Read prd.md.
• If gaps exist, run deep questioning in rounds (3–5 questions each). Use ./references/shared-deep-questioning.md.
• If PRD mentions “use skills/tools/CLIs” — run /find-skills first (project-installed → marketplace).

If user is temporarily unavailable (“запуск и ушёл”):

• Record Open questions in .protocols/PRD-BOOTSTRAP/decision-log.md.
• Stop and wait (do not invent facts; do not proceed to EP/FT/backlog generation without answers).

If the user explicitly wants autonomous mode:

• record non-blocking gaps as Assumptions
• halt only on blocking gaps (security/compliance/external contract/data-loss risks)
• after L1–L3 + review gate, continue with /autonomous

3A.2 Write product brief (L1)

Update .memory-bank/product.md using user’s wording.

3A.3 Requirements + RTM

Update .memory-bank/requirements.md:

• REQ-IDs
• RTM table mapping REQ → Epic → Feature → Test

3A.4 Create epics and features

Create:

• .memory-bank/epics/EP-001-<slug>.md
• .memory-bank/features/FT-001-<slug>.md

Each feature MUST include:

• use cases
• acceptance criteria
• failure modes / edge cases
• test strategy pointers

Status policy:

• Default EP/FT frontmatter to status: draft until Open questions are resolved.
• Promote to status: active only when acceptance criteria + verification plan are stable.

3A.5 Tasks planning (per-feature, no “everything at once”)

Do not generate a full task backlog for all features in one pass.

Instead:

1. Create/refresh .memory-bank/tasks/backlog.md as a skeleton (waves + placeholders).
2. For each selected feature, run /prd-to-tasks FT-<NNN> to produce:

- .memory-bank/tasks/plans/IMPL-FT-<NNN>.md - atomic TASK-* items grouped by waves in backlog.md

3A.6 Identify key concepts and create duo docs

For every non-trivial concept, create a duo:

• .memory-bank/architecture/<concept>.md (WHAT/WHY)
• .memory-bank/guides/<concept>.md (HOW)

3A.7 Update index

Update .memory-bank/index.md with annotated links to all created docs.

Step 3B — Brownfield workflow (Repo → Memory Bank)

3B.1 Spawn repo-scanning subagents (parallel)

Create a new task folder:

• .tasks/TASK-MB-MAP/

Spawn up to 5 subagents in parallel with non-overlapping scopes:

1. S-01: build/tooling (package managers, scripts, CI)
2. S-02: backend/services
3. S-03: frontend/UI
4. S-04: data layer (DB, migrations, schema)
5. S-05: tests + quality gates

Each subagent MUST:

• verify its file glob targets exist ("smart calling")
• write a detailed report into .tasks/TASK-MB-MAP/ using naming: TASK-MB-MAP-S-0X-final-report-<code|docs>-YY.md
• return a 5–10 line summary + file list

Use ./agents/shared-repo-scanner.md as baseline prompt, but scope it.

3B.2 Synthesize Memory Bank from reports

Using the .tasks/TASK-MB-MAP/ reports, fill:

• .memory-bank/product.md — what the system is today
• .memory-bank/architecture/ — C4 L1–L3 overview + key invariants
• .memory-bank/runbooks/ — setup, dev, test, deploy
• .memory-bank/contracts/ — API/event contracts
• .memory-bank/testing/index.md — canonical gates + verification notes
• .memory-bank/index.md — annotated links

PRD-less rule (non-negotiable): if there is no prd.md, you MUST NOT create or populate: - .memory-bank/epics/* - .memory-bank/features/* - .memory-bank/tasks/backlog.md with waves/tasks Empty skeleton files/folders are allowed if they were created by bootstrap. Repo mapping is as-is documentation, not roadmap planning. If something cannot be derived from code/logs/tests, record it as a hypothesis or an open question.

3B.3 Ask user for PRD delta

After baseline MB exists:

• ask the user for prd.md describing what to change/add
• run /prd and /prd-to-tasks style decomposition against the existing baseline

Step 3C — Skeleton-only workflow (no PRD, no code)

When the repo is new/empty and no prd.md is available:

3C.1 Create skeleton only

Run Step 1 as usual — create all directories, core files from templates, AGENTS.md, CLAUDE.md symlink.

The skeleton provides a ready-to-fill structure: product.md, requirements.md, backlog.md, etc. remain as draft stubs.

3C.2 Ask for PRD

After skeleton is created, ask the user to provide a PRD:

"Memory Bank skeleton created. To fill it with product details, epics, features, and a backlog, please provide a prd.md file (or paste requirements text). You can do this now or later — run /prd when ready."

3C.3 Wait or proceed

• If user provides PRD now → continue to Step 3A (Greenfield workflow).
• If user defers → stop here. The skeleton is valid and usable. The user can invoke $mb-from-prd or /prd later to fill the Memory Bank.
• If user provides partial info → run deep questioning (./references/shared-deep-questioning.md) to extract enough for product.md, then stop and wait for full PRD.

Note: The skeleton-only state is a valid stopping point. AGENTS.md + .memory-bank/index.md + MBB rules are enough for agents to start navigating the repo.

Step 4 — Hybrid mode (Claude ↔ Codex)

If you’re in Claude Code but want Codex quality or long-horizon autonomy:

1. Use Claude subagents to produce the scan reports into .tasks/TASK-MB-MAP/.
2. Call Codex via shell to synthesize Memory Bank:

codex exec --ephemeral --full-auto -m gpt-5.2-high \
  'Read .tasks/TASK-MB-MAP and build/refresh .memory-bank per MBB. Keep AGENTS.md short. Produce a summary and run a self-review.'

1. Then run a Codex deep review profile (or a fresh Claude session) for 5-expert review.

Step 5 — Multi-expert review loop (fresh context)

Run fresh-context reviewers (do not reuse the writer context):

• Architect (C4 + dependencies)
• Scope analyst (REQ → Epic → Feature → Task coverage)
• MBB compliance reviewer (frontmatter, links, atomics, duo)
• Plan reviewer (backlog quality, waves, gates)
• Security reviewer (auth, sensitive data, OWASP risks)
• Code quality reviewer (conditional: if code exists — quality gates, conventions, hotspots)

Use prompts in ./agents/shared-review-*.md and ./agents/shared-mb-reviewer.md.

Rules:

• If any reviewer returns REJECT → fix MB and repeat review.
• Persist reviewer reports into .tasks/TASK-MB-REVIEW/.

Step 6 — Start executing tasks

After review gate passes (APPROVE):

1. Pick the highest-priority task from .memory-bank/tasks/backlog.md.
2. Run mb-execute for the task (plan → implement → quality gates → MB-SYNC).
3. Run mb-verify to check acceptance criteria and record evidence.
4. Repeat until the wave is complete or user stops.

If the intended mode is unattended end-to-end:

• do not stay in manual loop here
• switch to generated project command /autonomous

If mb-execute or mb-verify are not installed, follow their SKILL.md manually.

Definition of done

You are done when:

• AGENTS.md exists, short, points to .memory-bank/index.md.
• CLAUDE.md is a symlink/copy of AGENTS.md.
• .memory-bank/ contains at minimum: index + MBB + product + testing (requirements/backlog can remain as stubs until PRD exists).
• .tasks/ contains scan/review artifacts with naming + stage ids.
• Greenfield: epics/features/backlog created from PRD.
• Brownfield: repo mapped as-is into MB and user asked for PRD delta (no roadmap entities invented without PRD).
• Skeleton-only: skeleton created, user asked for PRD (valid stopping point).
• Multi-expert review passes (APPROVE) — for Greenfield/Brownfield; skip for Skeleton-only.
• Execution loop is available (mb-execute + mb-verify reachable or documented).
• Autonomous loop is available (/autonomous + /autopilot documented).

References in this skill

• ./references/shared-structure-template.md
• ./references/shared-deep-questioning.md
• ./agents/shared-repo-scanner.md
• ./agents/shared-mb-reviewer.md
• ./agents/shared-review-architect.md
• ./agents/shared-review-scope.md
• ./agents/shared-review-code.md
• ./agents/shared-review-plan.md
• ./agents/shared-review-security.md
• ./scripts/shared-init-mb.js (optional helper)