Integrate datalib as the default personal-data store for agents

[qi-imbue]

What

Makes datalib (the frankweiler tools) the default way Minds agents retrieve, search, and import the user's own personal data -- past chat conversations, Slack, email, GitHub, Notion, contacts -- into a local searchable mirror. Mirrors the mechanism that makes latchkey the default for outbound authenticated requests: a Claude Code skill + environment wiring + a binary installed into the image.

Changes

• .agents/skills/datalib/SKILL.md (new) -- the core deliverable. Its description makes the agent reach for it whenever the user asks about their own history. Teaches the agent to search the local store over its HTTP API (127.0.0.1:8731/api/search) or the qmd index, and to import/refresh sources with frankweiler-sync. Credentials flow through the existing latchkey gateway; the skill defers to the latchkey skill for the permission handshake.
• scripts/setup_system.sh -- installs the fully-static musl frankweiler-sync / frankweiler-http binaries (v0.16.0) via datalib's install.sh with FRANKWEILER_LIBC=musl, so they run as-is regardless of the base image's libc. Version pinned; install.sh fetched at the pinned tag for a reproducible build.
• .mngr/settings.toml -- sets FRANKWEILER_ROOT=/mngr/datalib: on the persistent /mngr volume so it survives restarts, but deliberately outside /mngr/code (git workspace) and runtime/ so the large doltlite + qmd store does not ride the runtime-backup git branch.
• CLAUDE.md -- steers agents to the skill for questions about the user's history.

Auth: no new plumbing

frankweiler-sync shells out to its own latchkey, which reads the LATCHKEY_GATEWAY* env vars Minds already injects -- so web-API syncs route through the same gateway and permission dialog the user already approves for Slack/GitHub/etc. No LATCHKEY_CURL wiring needed (the binary auto-locates its shim as a sibling).

MVP scope / known limitation

Reliable through the Minds gateway: Slack, GitHub, Notion, email (Google Takeout .mbox). Not reliable: Cloudflare-walled sources (claude_api, chatgpt_api) -- inside Minds, latchkey routes through its gateway and bypasses datalib's Chrome-impersonating curl shim, so Cloudflare challenges them. The skill tells the agent this and to prefer an on-disk export for that data.

Open questions for review

• data_root location. /mngr/datalib persists per-workspace and is not part of the runtime git backup (too large). A future iteration could add targeted backup of just the precious raw/ stores, or move to the "local datalib over a gateway" model.
• Retrieval server lifecycle. The skill starts frankweiler-http on demand. A services.toml-supervised background service would be cleaner but needs to tolerate an empty data root -- deferred.
• Sensitivity. datalib accumulates high-value private data; the skill flags it as private + untrusted, but making it a default for autonomous agents warrants a policy pass (lethal-trifecta exposure).

Testing

Not yet exercised end-to-end. Plan: just minds-start -> create a Docker agent (rebuilds the image, exercising the musl binary install) -> ask it to mirror + search a Slack channel and confirm the skill auto-invokes, the sync runs through the latchkey permission dialog, and a query returns.

🤖 Generated with Claude Code