meal-cam-logger

name	meal-cam-logger
description	Log meals in real time from a webcam or IP camera. Opens a capture loop, asks Claude vision what food is visible in each frame, dedupes, and writes each eating session to a standalone Markdown file. Use when the user says start meal cam, log my meal, watch me eat, 记录我吃的, or /meal-cam-logger.
metadata	{"version":"0.1.0","last_updated":"2026-05-31","summary":"Log meals in real time from a webcam or IP camera."}

Skill: meal-cam-logger

Live meal logging. The user turns on a camera, starts eating, and this skill watches the video stream, identifies food as it appears, and writes each detection to a per-session Markdown file under _WorkSpace/6-EndpointStore/ mealcam_logs/.

Trigger phrases: /meal-cam-logger "start meal cam" "log my meal" "watch me eat" "记录我吃的"

---

Overview

Three phases:

1. Start — confirm camera source (built-in webcam / RTSP URL), capture interval (default 5s), and output directory. Launch the capture loop as a background process.
2. Monitor — loop grabs one frame per interval, asks Claude vision what food is visible, dedupes against the last detection, writes / rewrites the session's Markdown file on each new detection.
3. Stop — user says "stop meal cam" → kill the loop → read the session Markdown file and summarise.

---

Phase 1 — Start

Ask the user once:

"Which camera? [1] built-in webcam [2] IP camera (paste RTSP URL)"

"Capture interval? (default 5s — lower = more accurate but more API calls)"

Defaults are fine if they just say "go" — use built-in webcam, 5s interval.

Then launch the loop in the background:

cd /home/jluo41/WellDoc-SPACE
source .venv/bin/activate && source env.sh
mkdir -p _WorkSpace/6-EndpointStore/mealcam_logs
nohup python3 Tools/plugins/health/skills/meal-cam-logger/scripts/meal_cam_loop.py \
    --source 0 \
    --interval 5 \
    --output-dir _WorkSpace/6-EndpointStore/mealcam_logs \
    > /tmp/meal_cam.log 2>&1 &
echo $! > /tmp/meal_cam.pid

Confirm it's running:

sleep 2 && cat /tmp/meal_cam.log | tail -5

Tell the user: "📷 Meal cam is on. Just eat — I'll write down what I see."

---

Phase 2 — Monitor

While the loop runs, the user may ask mid-session:

• "what have I eaten so far?" → tail -20 /tmp/meal_cam.log or cat the current session file, then summarise.
• "skip that last one" → pop the last entry from the in-memory list and rewrite the session file. (Currently requires manual edit of the .md — noted as a TODO.)
• "pause" → kill -STOP $(cat /tmp/meal_cam.pid); "resume" → kill -CONT.

The loop itself handles:

• Frame capture (OpenCV VideoCapture, supports int index or RTSP URL).
• Vision call to Claude (Anthropic SDK, claude-sonnet-4-6 default).
• Dedupe: skip frame if detected food is the same as the previous detection (string match on normalized food label).
• Rewrite mealcam_logs/meal-{YYYY-MM-DD}-{HHMM}.md on each new detection (full-file rewrite keeps the logic simple and avoids partial-write bugs).

---

Phase 3 — Stop

When user says "stop meal cam" / "end meal" / "done eating":

kill $(cat /tmp/meal_cam.pid) && rm /tmp/meal_cam.pid

Then find the current session file and summarise:

ls -t _WorkSpace/6-EndpointStore/mealcam_logs/*.md | head -1 | xargs cat

Present to the user:

"Logged this session (saved to meal-2026-04-19-1204.md): • 12:04 salad with grilled chicken • 12:09 sourdough bread • 12:18 black coffee Duration: 24 min. 3 unique foods from 6 detections."

---

Markdown output format

Written to _WorkSpace/6-EndpointStore/mealcam_logs/meal-{YYYY-MM-DD}-{HHMM}.md. One file per eating session; HHMM is the session start time.

# 2026-04-19

- **Start:**    12:04
- **End:**      12:28
- **Duration:** 24 min

## Foods

- 12:04  salad with grilled chicken
- 12:09  sourdough bread
- 12:18  black coffee
- 12:22  apple

The file is rewritten in full on each new detection (start/end/duration stay in sync without append-tracking).

---

Error Handling

• Camera not found — OpenCV returns no frames. Tell the user to check that no other app is holding the webcam (e.g. Zoom, browser).
• API rate limit — back off: double the interval and warn the user.
• No food in frame — vision returns none; do not write a bullet, just continue.
• Output dir missing — script auto-creates with mkdir -p; no manual setup needed.

---

Privacy note

Frames are sent to the Anthropic API. Tell the user this up front on first run. If they want fully local inference, point them at DeepCamera + a local VLM (Qwen-VL / LLaVA) — not in scope for this skeleton.