bootstrap plugin

Read a codebase and generate a project-specific suite of domain-expert agents plus a critic gates config.

The bootstrap plugin analyzes a project and generates a custom suite of agents that know your codebase. It does not use static templates — agents write agents.

What it does

Given a project, bootstrap:

Ingests docs (CLAUDE.md, README, ARCHITECTURE.md, manifests) into discovery/project-context.md.
Detects languages by file count (discovery/languages.json).
Identifies domains (api-backend, data-persistence, caching, auth, analytics, testing, and so on) (discovery/domains.json).
Maps domains and languages to candidate agent roles (synthesis/expert-agents.json), always including test-strategist and security-privacy.
Decomposes complex agents into specialists and emits critic-gates.json (synthesis/decomposed-agents.json).
Validates agent specs pre-generation (synthesis/generation-validation.json).
Generates each agent with an LLM (colors assigned by domain) → .claude/agents/<name>.md, .closedloop-ai/bootstrap-metadata.json.
Runs a 6-step validation per generated agent (synthesis/agent-validation.json).
Writes the final validation-report.json and bootstrap-report.md.

Command

claude /bootstrap:start

Or the underlying slash command:

claude /bootstrap:agent-bootstrap

Flags

--target-command <name> — target command name
--depth quick|medium|deep — how deeply to explore
--focus <area> — focus on a specific domain
--output-dir <path> — override output directory
--dry-run — plan without writing files
--interactive — step through each phase
--update — update agents when project-context hash changes
--minimal / --enhance / --add-domain <domain>
--strategy backup|skip|overwrite|interactive — conflict strategy (default backup)

web-frontend, mobile-native, api-backend, data-persistence, caching, real-time-communication, authentication, authorization, analytics, testing, build-deployment, monitoring-observability.

Universal agents

Bootstrap assumes these already exist and does not regenerate them:

prd-analyst
feature-locator
plan-writer
plan-stager
plan-verifier
agent-trainer

Colors by domain

database=blue, API=green, frontend/state=purple, mobile=cyan, security=red, testing=yellow, performance=orange, analytics/monitoring=pink.

Architecture agents

Agents whose names end with -architect, -expert, or -specialist get a two-phase relevance check: a 30-second relevance gate first, then focused analysis — so they do not burn tokens on out-of-scope work.

Output locations

File	Purpose
`.claude/agents/<name>.md`	Generated agent file.
`.closedloop-ai/bootstrap-metadata.json`	SHA-256 hashes for `--update` detection.
`.closedloop-ai/settings/critic-gates.json`	Critic wiring used by the `code` plugin.
`.closedloop-ai/bootstrap/<timestamp>/discovery/`	Raw discovery artifacts.
`.closedloop-ai/bootstrap/<timestamp>/synthesis/`	Synthesis artifacts.
`.closedloop-ai/bootstrap/<timestamp>/validation-report.json`	Final validation report.
`.closedloop-ai/bootstrap/<timestamp>/bootstrap-report.md`	Human-readable summary.

Resilience

Retryable with 1-second base, 8-second max exponential backoff plus jitter, 2 attempts per agent.
Continues if at least 80% of agents generate successfully (min_success_ratio).

When to run

Run bootstrap once per new repository, then again whenever project-context.md changes materially (detected via its SHA-256 hash in bootstrap-metadata.json). The --update flag limits regeneration to agents whose inputs have changed.