# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Commands

```bash
uv run uvicorn app:app --reload                          # start dev server at http://localhost:8000
uv run --env-file .env uvicorn app:app --reload          # load env vars from .env file
uv sync                                                  # install/sync dependencies
uv add <package>                                         # add a new dependency
```

Set `LLAMACPP_BASE_URL=http://host:port/v1` to point at a local LLM (inline, exported, or via `.env`).

## Architecture

Two-file app: a FastAPI backend (`app.py`) and a single-page frontend (`static/index.html`).

**`app.py`** parses all GPX files in `garmin_gpx_exports/` at startup into an in-memory dict (`_rides`), keyed by filename stem. Stats (haversine distance, elevation gain/loss, HR zones) are computed once at load time. API routes:
- `GET /api/rides` — summary list (no track points)
- `GET /api/rides/{id}` — full ride metadata + cached insight
- `GET /api/rides/{id}/points` — full lat/lon/ele/hr array
- `POST /api/rides/{id}/insights` — streams AI coaching response
- `POST /api/insights/generate-all` — SSE stream, generates insights for all rides in chronological order
- `POST /api/fetch-rides` — SSE stream, fetches new rides from Garmin Connect (supports MFA)
- `POST /api/mfa` — submits MFA code for an in-progress Garmin session

**AI integration** uses the OpenAI Python SDK pointed at a local LLM endpoint (`LLAMACPP_BASE_URL` env var, defaults to `http://localhost:8080/v1`). The model name is auto-discovered from `GET /v1/models` at startup (falls back to `"local-model"`). If the endpoint is unreachable at startup a warning is logged and the app continues — only insight generation is affected. The chunk reader handles both `content` and `reasoning_content` delta fields for reasoning models.

**Insights cache** is persisted to `insights.json` (gitignored) and loaded at startup. The prompt includes the previous ride's cached insight so the coach can reference its own prior advice.

**`static/index.html`** is a self-contained SPA (no build step). Leaflet.js draws the route as per-segment polylines colored by HR intensity zone. Chart.js renders elevation and HR profiles against cumulative distance. AI insights stream via `fetch` + `ReadableStream`. Units (km/mi) toggle and resting HR are persisted in `localStorage`.

**`main.py`** is a standalone CLI script for bulk-downloading Garmin Connect GPX files without running the web server.

## GPX data notes

- Garmin extension namespace for HR: `http://www.garmin.com/xmlschemas/TrackPointExtension/v1` (`ns3:hr`)
- Activities with an empty `<trkseg/>` (e.g. indoor rides) get `has_gps: false` — frontend shows no map or charts
- HR zone boundaries (fixed, used for API summary): Z1 <120, Z2 120–140, Z3 140–160, Z4 160–180, Z5 >180 bpm
- Karvonen zones (used in the UI and AI prompt) are computed from resting HR input + lifetime max HR observed across all rides

## Logging

httpx, httpcore, and openai loggers are set to WARNING to suppress connection-level debug noise. The `bikeslop` logger runs at DEBUG.