rse-project

rse-project is a small Node.js + TypeScript CLI that analyzes a local repository, writes reviewable .skill-extraction artifacts, creates a new branch from main, and runs local codex exec to generate repo-specific Codex skills and AGENTS.md.

This MVP is intentionally narrow:

• no OpenAI API or external LLM API integration;
• no fixtures or test suite yet;
• no multi-command CLI;
• no skill installation or synchronization;
• no hosted service.

Usage

npm install

# Run from source:
npm run dev -- /path/to/project

# Or build and run the compiled CLI:
npm run build
node dist/index.js /path/to/project

# After publishing, the lead-facing command is:
npx rse-project /path/to/project

During the run, the tool:

1. verifies the target is a clean local Git repository;
2. creates a new branch from main named like rse/skills-main-20260506-104500;
3. analyzes project structure, package scripts, frameworks, tests, docs, CI, config files, and repeated workflow signals;
4. writes .skill-extraction/analysis.json, facts.json, candidates.json, rejected-candidates.json, evidence.md, report.md, and codex-prompt.md;
5. invokes:

codex exec --cd <repo> --sandbox danger-full-access --dangerously-bypass-approvals-and-sandbox -

The full-access Codex mode is deliberate for this local workflow. Use it only on trusted local repositories.

If Codex is installed but not visible to Node, set the executable path explicitly:

CODEX_BINARY=/full/path/to/codex npm run dev -- /path/to/project

If a run fails after .skill-extraction/codex-prompt.md is created, rerun the same command while the target repository is still on the generated rse/skills-* branch. The CLI will resume by reusing the existing prompt instead of creating another branch.

Output In The Target Repository

The static analysis phase creates:

.skill-extraction/
  analysis.json
  facts.json
  candidates.json
  rejected-candidates.json
  evidence.md
  report.md
  codex-prompt.md
  codex-final-response.md

Codex is then instructed to create or update:

AGENTS.md
.agents/
  skills/
    <skill-name>/
      SKILL.md
      references/
        evidence.md
        review-example.md

Skill Quality Bar

The generated prompt tells Codex to keep only skills that are specific to the target repository. Good skills should encode concrete workflows, conventions, architecture, validation commands, and file references. Generic skills such as "React best practices", "TypeScript basics", or "write tests" are rejected.

Security Model

The analyzer respects .gitignore and always ignores common build outputs, dependency directories, environment files, private keys, and certificates. Secret-like values are masked in generated artifacts. Repository content is not sent to any external API by this tool; the only agent execution is the local codex exec command configured on your machine.

Limitations

v1 works best for JavaScript and TypeScript repositories, especially Node.js, React, Next.js, Vite, Express, Fastify, NestJS, and common monorepo layouts. Other stacks get generic fallback analysis with lower-confidence candidates.