tools: pydantic-ai agent sandbox driving loseit CLI via homelab Ollama

[phitoduck]

Summary

• Adds tools/agent-sandbox/: single-file pydantic-ai agent (agent.py) that drives the loseit CLI via 5 tools — one per non-destructive subcommand. delete and login are deliberately omitted.
• Control flow is handled entirely by the homelab-hosted qwen3:8b Ollama model at https://ollama.priv.mlops-club.org/v1 (Tailscale-only).
• Docker image based on ghcr.io/astral-sh/uv:python3.12-bookworm-slim, runs as non-root, only mount is ~/.config/loseit/ read-only.
• Per-run logging: one file at runs/run-{UTC}.log capturing the prompt, system prompt, every tool args+return, the full final message history (every model request/response with parsed tool_call args), and token usage. Mountable to the host.

End-to-end run on the goal prompt (255g asparagus/lentils/red-potato mix + 100g guacamole) lands all 4 entries correctly in the diary, verified via readback.

Snacks (2026-06-12):
  Asparagus, Spears, Cooked    85g    18.7 cal
  Beans, Lentils, Cooked       85g    98.6 cal
  Potatoes, Red, Baby, Cooked  85g    75.6 cal
  Guacamole                   100g   202.8 cal
                       Total:        395.7 cal

Notes

• lose-it itself is unpublished, so the image installs it from git+https://github.com/phitoduck/lose-it@${LOSEIT_REF} (default main, override with --build-arg).
• A small tool-side guardrail in log_food refuses absurd-magnitude tbsp/tsp/fl_oz logs — the small model otherwise falls back to grams ÷ g-per-tbsp math instead of re-searching for a gram-supporting food entry.

Test plan

• Build the image: docker build -t loseit-agent tools/agent-sandbox
• Smoke test auth: docker run --rm -v ~/.config/loseit:/home/agent/.config/loseit:ro --entrypoint /home/agent/.venv/bin/loseit loseit-agent whoami
• Run end-to-end with a real prompt (Tailscale required) and confirm entries land via loseit diary
• Inspect runs/run-*.log to verify all LLM + tool I/O is captured