Plan: pal-e-docs MCP + Seeding + Frontend + Dogfood
Vision
pal-e is the foundation for an AI agency. Three pillars: pal-e-platform (done), pal-e-services (nearly done), pal-e-docs (this plan). Once all three are done, agents have context — they query for what they need and know the conventions.
User story: "I tell an agent what to do, and it already knows my platform."
Phases
- Phase 1: MCP Server — DONE. pal-e-docs-mcp on GitHub, 11 tools, registered in ~/.mcp.json.
- Phase 2: Seed Platform Fundamentals — DONE. 9 notes, 3 projects, cross-linked. Needs user review.
- Phase 3: Browser Frontend — DONE. PR #6 merged. Jinja2 SSR at /browse/.
- Phase 4: Dogfood — IN PROGRESS. /plan skill updated. CLAUDE.md points to pal-e-docs. This plan stored in pal-e-docs.
Decisions Made
- No auth for now — behind Tailscale, issue #2 tracks
- Platform fundamentals first — platform knows itself before supporting projects
- Small atomic notes, heavy tagging — Zettelkasten-like, query by tag intersection
- MCP repo on GitHub — dev tool, same pattern as notion-mcp (user flagged: should this default to Forgejo?)
- Jinja2 server-side frontend — no client JS, agent can open URLs for user review
Next Plan Seeds
- Session-start hook — auto-query active SOPs on session start, inject into context
- Review seeded content — 9 notes written from agent memory, need user verification
- Repo placement SOP — when does a repo go on GitHub vs Forgejo? Needs a documented convention.
- Full-text search — SQLite FTS5 for content search beyond tags
- Litestream backup — continuous SQLite replication
- basketball-api seeding — project-specific knowledge after platform fundamentals verified
- Hook false positive fix — remind-review-loop.sh fires on PR merges, not just PR creations