.pi

Personal Pi coding-agent configuration.

This repo tracks the durable parts of my Pi setup: settings, skills, extensions, prompts, themes, and decisions about how the setup should evolve.

Policy

• GitHub Issues track candidate config changes and setup backlog.
• docs/decisions/ tracks settled decisions and rationale.
• Config files are the executable result.
• Do not commit auth files, sessions, local caches, package installs, or secrets.

Dependencies and local tools

• Pi extension dependencies are tracked with agent/package.json and agent/pnpm-lock.yaml.
• Installed packages and local tool downloads are not committed (agent/node_modules/, agent/npm/, agent/git/, agent/bin/).
• If a helper binary is needed locally, install it outside this repo and keep the requirement documented here. Current local helper: fd (fd-find on some Linux distributions).

Current goals

• Replace heavier agent orchestration with lightweight Pi + tmux worker sessions.
• Use transparent file-backed memory and task tracking where possible.
• Curate a small set of high-value skills instead of importing large packs wholesale.
• Keep the setup understandable, portable, and easy to audit.

Areas under evaluation

• dmmulroy Pi config structure and extensions
• Matt Pocock engineering skills
• Superpowers skills worth adapting
• Karpathy-style LLM wiki memory
• Local todos vs GitHub Issues vs Linear for task tracking
• Tmux-based Pi worker delegation

Implemented extensions

• agent/extensions/git-interceptor.ts: protects agent-driven git commands from editor hangs and blocks --no-verify.
• agent/extensions/web-access/: focused web search/fetch extension with Exa search, Firecrawl clean scraping, direct static fetches, GitHub clone/API handling, and SQLite response/cache storage.

Web access extension

The web access extension is a global personal Pi extension at agent/extensions/web-access/. It intentionally implements only the web capabilities expected to be useful across projects, rather than installing a broad third-party web-access package wholesale.

Available tools:

• web_search: compact Exa-backed web discovery for finding sources.
• code_search: Exa-backed programming/docs/API/example search.
• web_fetch: cleanly reads URLs, with GitHub-aware routing and capped output.
• get_web_content: retrieves stored full or paged content from previous web tool results by responseId.

Design decisions:

• Generic tool names are exposed to the model; provider-specific SDKs stay behind local adapters.
• Startup validation is local-only: it checks .env key presence but does not call Exa, Firecrawl, GitHub, or spend API credits.
• Missing provider keys are non-fatal at startup; tools fail clearly only when a required provider is used.
• Search/fetch results are cached in SQLite at agent/cache/web-access/cache.sqlite and full results are stored behind response IDs.
• Tool output is context-clean by default: search is compact, fetch output is capped/summarized, and uncapped full content requires an explicit get_web_content({ full: true }) call.
• Normal webpages route to Firecrawl for clean markdown; cheap/static/agent-friendly sources can use direct HTTP.
• GitHub Gist pages route to raw gist content to avoid noisy scraped GitHub chrome.
• GitHub repo/tree/blob URLs clone locally; GitHub issue/PR URLs prefer structured gh/REST metadata before any scraping fallback.
• Secrets belong in agent/.env; only agent/.env.example is committed.

See docs/decisions/0003-add-focused-web-access-extension.md for the durable rationale and scope.