Agentic Codebase

What it builds for you

• Turns your WTFs into guardrails.
• Every "no, not like that" becomes a rule.
• Turns "why did you do that?!" into a hook that stops it.
• Every .env your agent peeked at becomes a guardrail.
• Every prompt you typed three times becomes a skill.
• Every PR you reviewed by hand becomes a reviewer agent.
• Every "yes, run it" you clicked becomes a permission.
• Everything you re-explain each session lands in CLAUDE.md.
• Every service you keep mentioning becomes an MCP draft.
• Your f-bombs, refactored into rules.

It read every time you yelled at your agent. Then fixed why.

[!IMPORTANT]

🔒 Runs 100% on your machine

• This tool sends nothing, anywhere. No servers. No account. No telemetry.
• Zero network calls. Python standard library only. Nothing to pip install.
• Transcripts need your yes, every run. Your prompts only. Secrets scrubbed before anything is written.
• Never reads secrets. No .env values. No credential files.
• Never touches your branch. Builds on agentic-setup/<date>. Never pushes.

Your agent still talks to its own model, same as always. This tool adds nothing on top.

Install

One command. Works for both.

npx skills add Agentic-Studio-Pvt-Ltd/agentic-codebase

Then, in your agent:

run agentic-codebase on this repo

Needs: Python 3.9+. git optional.

/plugin marketplace add Agentic-Studio-Pvt-Ltd/agentic-codebase
/plugin install agentic-codebase@agentic-codebase

[!WARNING] Codex: trust the repo first. Untrusted repo → AGENTS.md, hooks and rules silently don't load. Set trust_level = "trusted" for the repo in ${CODEX_HOME:-~/.codex}/config.toml.

Before / After

BEFORE                         AFTER
my-app/                        my-app/
├── CLAUDE.md                  ├── CLAUDE.md             + indexed section
└── src/                       ├── .claude/
                               │   ├── rules/            one per code zone
                               │   ├── skills/           setup-manager, review-pr, qa …
                               │   ├── agents/           pr-reviewer, db-inspector …
                               │   └── settings.json     hooks + permissions
                               ├── .mcp.json             drafts, you authenticate
                               ├── docs/agentic-setup/   plan.md · report.md
                               └── src/

_{Example on Claude Code. Codex gets the same shape. See the table below.}

What it generates

Every file: an ID · a date · the evidence behind it · "Safe to delete or edit."

How much it builds

• No caps. Everything your code and history support.
• You cut at the plan. Nothing written before you approve.
• Every file traces to a count. No evidence → not built.
• Skipped only if: no evidence · already covered · low confidence.
• setup-manager: keeps the setup current
• pr-reviewer + review-pr: PR review
• qa: any web app
• db-inspector + query-db: any database, read-only
• security-auditor: security review
• designer + new-component: UI repos
• product-analyst + analytics skills: repos with analytics
• 3 guardrail hooks: env leak · destructive commands · wrong package manager
• Permissions: your commands allowed, .env* denied
• Rules: one per code zone
• Workflow skills: anything you ask for 3+ times

How it works

Safety

Won't do

• Set up credentialed MCP servers. Drafts only, you authenticate
• Cursor, Windsurf, Gemini CLI
• Hosted or web version
• Drift detection (on the roadmap)
• Rewrite, move or rename your existing setup
• Push or open a PR

Uninstall

[!NOTE] Undo from report.md, not from here. It's written for your exact run, including gitignored and untracked files a branch delete won't touch.

Contributing

• Read CLAUDE.md first.
• scripts/: analyzers. Stdlib only. No network. JSON out.
• SKILL.md + references/: the workflow. Agent-agnostic.
• adapters/: the only place agent-specific paths and formats live.

for s in discover mine_git mine_transcripts verify_artifacts; do
  python3 skills/agentic-codebase/scripts/$s.py --selftest
done
python3 -m unittest discover -s tests -v

License

MIT. Attribution welcome, never required. Strip it and everything still works.