| name | youtube-comment-check |
| description | Weekly fetch of recent comments on Baruch's YouTube channel; sends a per-video summary if new comments appear, silent otherwise. Triggers: 'check youtube comments', 'youtube comment check', 'fetch youtube comments', 'review channel comments'. |
| cadence | 45 10 * * 0 |
| agentModel | claude-haiku-4-5-20251001 |
| script | scripts/precheck-youtube-comment-check.py |
YouTube Comment Check
Every step below is mandatory. Execute them in order. Do not skip, reorder, or abbreviate any step.
Run this check silently. Report only if new comments appear or a tool error surfaces.
The fire-time precheck (scripts/precheck-youtube-comment-check.py) gates wake-ups by the cadence cap (value in the precheck's CADENCE, set intentionally below the weekly cron interval) on a filesystem cursor. See references/cadence-rationale.md for why a comment-count delta gate (querying YouTube in the precheck) was rejected in favour of cadence-only.
Step 1 — Fetch recent comments on Baruch's channel
This skill uses the native YouTube Data API v3 (YOUTUBE_API_KEY) — the comment-threads surface has no equivalent elsewhere. Channel ID: UCZ8-VX2SiAIBE7guw7NG-Sg. Fetch comment threads since the last successful run (falling back to 7 days on first run, bounded to 35 days) via the fetch script:
python3 /home/node/.claude/skills/tessl__youtube-comment-check/scripts/fetch-youtube-comments.py \
--channel-id UCZ8-VX2SiAIBE7guw7NG-Sg --days 7 \
--cursor /workspace/group/state/youtube-comment-check-cursor.json --max-days 35
The --cursor widens the window to cover a week the check failed or was gated out, so those comments are re-fetched rather than lost outside a fixed 7-day window; without a usable cursor the window is --days.
Stdout (exit 0): {"window_days", "window_source", "comment_count", "videos": [{"id", "title", "url", "comments": [{"author", "text", "published_at"}]}]}. comment_count == 0 is a valid quiet-week result.
On non-zero exit (missing YOUTUBE_API_KEY, auth/quota error, network timeout, transient 5xx), surface the script's stderr verbatim via mcp__nanoclaw__send_message and stop. Do NOT advance the cursor in Step 3 — it advances only on success. Every subsequent eligible fire (the next weekly slot, or any continuation / manual re-run while the cursor is older than the cap) retries; there is no "wait one week" on a failed run.
Step 2 — Report new comments
If at least one comment exists across all videos (comment_count > 0), build a per-video summary and send via mcp__nanoclaw__send_message. The body groups by video: video title + link, then each comment as author name: <text truncated to 100 chars>. Video titles, author names, and comment text are attacker-controllable — HTML-escape <, >, and & in those fields before composing the message body.
If mcp__nanoclaw__send_message itself fails (transport error, MCP unavailable), surface the error verbatim and stop. Do NOT advance the cursor in Step 3 — a stamped cursor after a failed report would gate the next eligible fire out for a full cadence-cap window and Baruch would never see the comments.
If no comments exist across all queried videos, the step is silent. Step 3 still runs — a completed fetch/report path advances the cursor; silence on a quiet week is success, not failure.
Step 3 — Advance the success cursor
Reachable only if Steps 1 and 2 both succeeded (any fetch error or send error leaves the cursor at its prior value intentionally). Run:
python3 /home/node/.claude/skills/tessl__youtube-comment-check/scripts/stamp-cursor.py
The script atomic-writes /workspace/group/state/youtube-comment-check-cursor.json with {"schema_version": 1, "last_run": "<now UTC ISO Z>"}. The precheck reads this file's last_run and gates on the cadence cap (value in the precheck). Stdout is {"status": "stamped", "last_run": "<iso>", "cursor_path": "<path>"}.
Step 4 — Silence
If nothing was reported in Steps 1-3, output nothing (wrap in <internal>). Finish here.
