data-agent-lab

A versioned home for coding-agent extensions, skills, prompts, and benchmarks. Built around Pi (primary) with Claude Code as a fallback — both read the same shared artifacts, agent-specific bits stay separate.

Why this exists: coding-agent customizations tend to live as scattered, unversioned dotfiles, and benchmarks are ad-hoc. This repo collects both in one place, with one source of truth per artifact and agent-agnostic benchmarks.

How the two agents share artifacts (and what doesn't port) is in docs/architecture.md.

Quick start

# 1. clone
git clone https://github.com/sipgate/data-agent-lab.git ~/projects/data-agent-lab
cd ~/projects/data-agent-lab

# 2. enable the pre-commit hook (skill/symlink sanity checks)
git config core.hooksPath .githooks

# 3. install python deps for the web-distill extension
pip install --user beautifulsoup4

# 4. install this repo as a Pi package
pi install ~/projects/data-agent-lab

# 5. (if migrating from loose extensions) remove old copies in a fresh Pi session
rm ~/.pi/agent/extensions/web-distill.ts
rm -rf ~/.pi/agent/extensions/cortecs

# 6. start Pi
pi

Full setup (Browserless for Chromium fallback, env vars, verification) → docs/setup.md.

What's inside

data-agent-lab/
├── AGENTS.md                  # canonical project memory (CLAUDE.md symlinks here)
├── extensions/                # Pi TypeScript extensions
│   ├── web-distill/           #   overrides fetch_content + web_search
│   └── cortecs/               #   registers the cortecs provider
├── skills/                    # SKILL.md skills — portable across both agents
├── prompts/                   # Pi slash commands
├── .claude/
│   ├── skills → ../skills     # symlink (Claude reads the same SKILL.md)
│   └── commands/              # Claude Code slash commands
├── benchmarks/                # agent-agnostic task specs + results + runner
├── scripts/check-skills.sh    # pre-commit skill/symlink validation
└── docs/                      # deep context

Use with Pi

pi install ~/projects/data-agent-lab adds the repo to ~/.pi/agent/settings.json → packages. Pi reads the pi manifest in package.json and auto-discovers extensions/, skills/, and prompts/. TypeScript loads via jiti — edit a .ts and /reload, no compile step.

Use with Claude Code

Open the repo in Claude Code. It reads CLAUDE.md (→ AGENTS.md) and discovers skills via .claude/skills (→ skills/). Slash commands live in .claude/commands/.

Benchmarks

bash benchmarks/scripts/run.sh lru-cache glm-5.2 pi
python3 benchmarks/scripts/summary.py

Tasks are markdown specs; results are JSON with model + metrics. See benchmarks/README.md.

Documentation

Doc	What it covers
docs/setup.md	Fresh-machine install guide
docs/architecture.md	Portability matrix, sync design, why this layout
docs/extensions.md	Pi extension loading + migration from loose files
docs/skills.md	Skill-writing conventions for Pi + Claude Code
TODO.md	Open work, P0/P1/P2 priorities

License

MIT