ofri-prompts

Prompt template library for the Agentic Coding stack — pi + pi-agenticoding + pi-mcp-adapter + ChunkHound + agent-browser. Type /command-name in the pi editor to invoke a template.

Install

Requires pi ≥ 0.83 and the two packages below; most commands rely on ChunkHound for code research (MCP compatibility via pi-mcp-adapter).

pi install npm:pi-agenticoding
pi install npm:pi-mcp-adapter
pi install git:github.com/agenticoding/ofri-prompts

mkdir -p ~/.pi/agent/pi-agenticoding
cp ~/.pi/agent/git/github.com/agenticoding/ofri-prompts/model-groups/model-groups.json ~/.pi/agent/pi-agenticoding/model-groups.json

Installs 13 /command templates as a pi package and places the model routing policy (model-groups.json) where pi-agenticoding reads it. Upgrade with pi update --extensions.

ChunkHound (code research)

Most /command templates rely on ChunkHound — code_research over the git history (vector_source="diff"), semantic search, architecture research (vector_source="db"), and web search. Install it, configure it, and index your repo before using them — follow ChunkHound's getting-started: uv tool install chunkhound, create .chunkhound.json (embedding + LLM keys — these also gate web search — config docs, gitignore it), then chunkhound index . from the repo root.

Then register the server in your project's .mcp.json. pi-mcp-adapter loads it automatically:

{
  "mcpServers": {
    "ChunkHound": {
      "command": "chunkhound",
      "args": ["mcp"],
      "directTools": true,
      "lifecycle": "keep-alive",
      "idleTimeout": 12000,
      "requestTimeoutMs": 12000000
    }
  },
  "settings": {
    "outputGuard": false
  }
}

Model routing

model-group: frontmatter (e.g. /pr-review, /merge-conflicts, /debug-tests) routes to a group: a set of provider/model entries with a thinking level. At spawn time the router picks whichever provider in the group is authenticated on your machine.

Maintain the policy in the copy you placed during install — swap provider/model entries for your own providers:

nano ~/.pi/agent/pi-agenticoding/model-groups.json

A project-scoped .pi/pi-agenticoding/model-groups.json (relative to the project root) wins over the global file.

Providers

All providers are built into pi — authenticate with pi /login and you're done. Poolside Laguna is served through the built-in Vercel AI Gateway (vercel-ai-gateway), so no custom provider configuration is needed:

1. Run pi /login, select Vercel AI Gateway, and paste your API key (or set AI_GATEWAY_API_KEY).
2. The gateway model catalog ships with pi — poolside/laguna-s-2.1 and the free tier poolside/laguna-s-2.1-free are available as-is.

Custom/self-hosted providers (Ollama, vLLM, proxies) go in ~/.pi/agent/models.json — see pi's docs/models.md and docs/custom-provider.md.

License

MIT — see LICENSE.