mrvladd-d/memobank/cold-start

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-analysis (optional idea discovery), mb-from-prd (greenfield), mb-map-codebase (brownfield), mb-review (review), mb-execute (implementation), mb-verify (UAT), mb-red-verify (semantic adversarial verification).

• 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:

• Idea-only: repo has a raw idea, but no stable PRD yet; optionally route through /analysis, /brainstorm, and /brief.
• Clear concept: repo has enough direction for a product brief; normally run /brief, then /constitution only if project principles are not already ratified|partial, before /write-prd.
• 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)

2) 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)
• states/ (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/
• schemas/ (JSON schemas, including task records)
• tasks/ (empty JSON task index until `/prd-to-tasks`, indexed task records, and 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/constitution.md
• .memory-bank/mbb/index.md
• .memory-bank/spec-index.md
• .memory-bank/glossary.md
• .memory-bank/invariants.md
• .memory-bank/product.md
• .memory-bank/requirements.md
• .memory-bank/schemas/task.schema.json
• .memory-bank/tasks/index.json
• .memory-bank/testing/index.md

Also create the command specs under .memory-bank/commands/ (use references/commands/*). This includes optional Analysis commands (analysis.md, brainstorm.md, brief.md) when those command specs are present in the package. It also includes constitution.md for the /constitution command, which creates or updates .memory-bank/constitution.md.

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, /constitution, /prd, /execute, etc.) in all three tools.

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

Agents read .memory-bank/constitution.md early during priming. It is the short governing-principles layer for project decisions, not a replacement for .memory-bank/invariants.md, .memory-bank/contracts/*, .memory-bank/spec-index.md, or tier/workflow policy.

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); recommend /constitution before /write-prd if project principles are not ratified/partial.
• If repo is mostly empty and you only have an idea or loose concept → optionally run Analysis first: /analysis, /brainstorm when the idea is raw, /brief, then /constitution before /write-prd only if project principles are not already ratified|partial.
• 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 Write clarified PRD

• Run /constitution before /write-prd when project principles are still framework-default/skipped/missing. If they are already ratified/partial, continue directly to /write-prd. /constitution reads .memory-bank/analysis/product-brief.md when present and asks up to 5 contextual governance questions per pass.
• If the user explicitly skips /constitution, continue with framework-default/skipped principles and note that it can be ratified later.
• Run /write-prd to turn Product Brief / PRD text + Constitution into .memory-bank/prd.md.
• /write-prd handles PRD-level ambiguity with up to 5 targeted questions per pass.
• 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/task 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)
• continue with /autonomous after /write-prd is complete

3A.2 Route PRD to L1–L3

Run /spec-init to update .memory-bank/spec-index.md as a lightweight SDD route map from PRD/brief/existing-spec evidence, without architecture interview or invented authoritative specs. Then run /prd to decompose .memory-bank/prd.md into product, requirements, epics, features, testing, and index updates. After /prd, always run /spec-design. For small independent T0/T1 features it may record a minimal backbone with irrelevant areas marked not_applicable; for shared/T2/T3 concerns it creates or updates the needed backbone SDD specs and spec-index. This is not another mandatory heavy phase: it creates no tasks and no feature-local implementation design. /spec-improve FT-<NNN> remains the feature-level gate after backbone.

/prd owns:

• .memory-bank/product.md
• REQ-IDs
• RTM table mapping REQ → Epic → Feature → Test
• .memory-bank/epics/EP-*.md
• .memory-bank/features/FT-*.md

Each generated 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.
• /prd / mb-from-prd do not create tasks; canonical planning path is /write-prd → /spec-init → /prd → /spec-design → /spec-improve FT-<NNN> → /prd-to-tasks FT-<NNN>.
• Add feature clarification_status: pending|blocked only for explicit feature-level blockers.

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

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

Instead: 1) Ensure .memory-bank/schemas/task.schema.json and .memory-bank/tasks/index.json exist. 2) For each selected feature, use /clarify-feature FT-<NNN> only if the feature is explicitly pending/blocked. 3) Run /spec-design after /prd; use minimal backbone for simple T0/T1 scope and fuller backbone specs for shared/T2/T3 concerns. 4) Run /spec-improve FT-<NNN>, then /prd-to-tasks FT-<NNN> to produce:

• .memory-bank/tasks/plans/IMPL-FT-<NNN>.md
• atomic .memory-bank/tasks/TASK-*.task.json records grouped by wave, each with mandatory tier: T0|T1|T2|T3

Task routing is authoritative only through task.tier; the old risk / risk.level model is invalid.