Bootstrap

The bootstrap plugin is what makes ClosedLoop aware of your project. It reads your codebase and docs, identifies the work domains that exist in your system, and writes a tailored set of expert agents plus a critic-gates.json that tells the code plugin which critics to run for which modules.

When to run it

• On a new project, after installing the plugins but before your first serious loop.
• After a significant architectural change (new service, new domain, new language).
• When you onboard a new team and want everyone's Claude sessions to inherit the same project context.

The command

/agent-bootstrap [options]

Useful options:

• --target-command <name> – which consumer plugin to write for (default code).
• --depth quick|medium|deep – how hard to explore.
• --focus frontend|backend|infra|mobile|web – bias the identification.
• --output-dir <path> – where the generated agent files go (default .claude/agents/).
• --dry-run – show what would be written without writing.
• --interactive – prompt for each decision.
• --update – re-run only for domains whose hashes have changed.
• --minimal – only write test-strategist and security-privacy.
• --enhance – augment an existing agent rather than overwrite.
• --add-domain <domain> – add a single named domain.
• --strategy backup|skip|overwrite|interactive – conflict resolution.

The eight-phase flow

Bootstrap runs its own orchestrator via commands/agent-bootstrap.json:

1. Project doc ingestion (project-doc-ingestor) – read CLAUDE.md, README.md, ARCHITECTURE.md, manifests.
2. Language detection (language-detector) – file-counting, not heuristics.
3. Domain identification (domain-identifier) – 12 recognized domains. Domain-first, not framework-first.
4. Expertise mapping (expertise-mapper) – always adds test-strategist and security-privacy as universal agents.
5. Agent decomposition (agent-decomposer) – emits critic-gates.json alongside the agent roster.
6. Generation validation (generation-validator) – pre-flight before writing.
7. Agent prompt generation (agent-prompt-generator) – fan-out up to 5 in parallel; LLM-driven (no static templates).
8. Validation and reporting (agent-prompt-validator, bootstrap-validator) – YAML header, structure, artifact contracts, content budget, anti-patterns, file quality.

Outputs

Durable:

• .claude/agents/<name>.md – one expert agent per identified domain.
• .closedloop-ai/bootstrap-metadata.json – tracks generated agents for --update.
• .closedloop-ai/settings/critic-gates.json – per-module critic selection.

Audit trail (under .closedloop-ai/bootstrap/<timestamp>/$RUN/):

• discovery/project-context.md
• discovery/languages.json
• discovery/domains.json
• synthesis/expert-agents.json
• synthesis/decomposed-agents.json
• synthesis/generation-validation.json
• validation-report.json
• bootstrap-report.md
• open-questions.md (when the bootstrapper needs input)

What bootstrap does not write

Agents that are assumed to already exist in the code plugin are not regenerated:

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

Why critic-gates.json matters

Generic code review is expensive and noisy. critic-gates.json is a small map from modules to the set of domain critics the code plugin should invoke at phase 2.5. For example, the mobile code paths might route only to a mobile-testing critic and an accessibility critic, while the backend routes to a security critic and a database-migration critic. This keeps critic runs focused and cheap.

Related reading