Process delegate tasks

process-delegate-tasks

Automate the full lifecycle of every URL-based task in the "Delegate" list:

1. Pick up the task from Google Tasks
2. Detect the URL type (Threads post or YouTube video)
3. Fetch or transcribe the content
4. Summarize in Traditional Chinese with the raw content appended verbatim
5. Save a report
6. Mark the task done

Two URL types are handled:

URL Type	Trigger	Processing
Threads post	threads.net or threads.com	fetch-threads-post skill → synchronous
YouTube video	youtube.com or youtu.be	yt2doc Docker → background (async)

DO NOT STOP after processing a single task. This skill clears the entire backlog. The workflow is idempotent — skipping tasks already completed or lacking a supported URL ensures safe re-runs.

Prerequisites

• gws CLI installed and authenticated (gws auth login if needed)
• agent-browser installed (for Threads tasks)
• Docker running (for YouTube tasks) — verify with docker info
• For all Google Tasks API calls, refer to ../gws-tasks/SKILL.md

---

Procedure

Step 1 — Discover the "Delegate" task list

The list ID can change, so always resolve it by name:

gws tasks tasklists list

Parse the JSON and find the item where title == "Delegate". Extract its id. If the list doesn't exist, stop and tell the user.

Step 2 — Fetch incomplete tasks

gws tasks tasks list \
  --params '{"tasklist": "<DELEGATE_LIST_ID>", "showCompleted": false, "maxResults": 100}'

This returns only tasks with status == "needsAction". If items is empty or missing, the list is clear — tell the user.

Step 3 — Classify each task

For each task, scan these fields for a URL: title, links[].link, links[].description, notes.

URL contains	Task type	Action
threads.net or threads.com	Threads task	Queue for Threads flow
youtube.com or youtu.be	YouTube task	Queue for YouTube flow
Neither	Unrecognised	Skip — log: "Skipping '<title>': no supported URL"

Extract the first matching URL from each task. Build two separate queues: threads_queue and youtube_queue.

---

Async Strategy

Because YouTube transcription takes 5–80 minutes, use a fire-and-poll approach:

1. Fire — launch all YouTube background jobs first (Step 4Y)
2. Process — immediately handle all Threads tasks synchronously (Steps 4T–7T)
3. Poll — once Threads tasks are done, poll each YouTube job until complete (Steps 5Y–7Y)

This means a 45-minute video never blocks five Threads tasks from finishing.

---

Threads Flow (Steps 4T–7T)

Step 4T — Fetch the Threads post content

For each task in threads_queue, follow the full procedure defined in fetch-threads-post:

📄 Read and follow .agents/skills/fetch-threads-post/SKILL.md

Use a unique session name per task to avoid state collisions:

agent-browser --session delegate-task-<task-id> open "<THREADS_URL>"
# ... follow fetch-threads-post procedure to expand and extract content ...
agent-browser --session delegate-task-<task-id> close

CRITICAL: Ensure the post is fully expanded (click "Read more") and captured entirely. Truncated raw content is unacceptable.

1. Scroll first: Run agent-browser --session delegate-task-<task-id> scroll down 1000 to trigger lazy-loading and move past overlays.
2. Extract body: Use agent-browser --session delegate-task-<task-id> get text body as the primary extraction method for the 📄 原始內容 section.

Step 5T — Verify and generate a Traditional Chinese summary

Verification first: Compare extracted text length against the visible post. If the raw content appears cut off mid-sentence, repeat extraction with get text body.

Produce a summary in Traditional Chinese (繁體中文):

• 零幻覺（Zero Hallucination）: Only summarize what is in the post. No inference or extrapolation.
• 全面性（Comprehensiveness）: Include all key points — don't drop information for brevity.
• 客觀性（Objectivity）: Neutral tone, no personal commentary.

Step 6T — Save the Threads report

Output directory: reports/Threads_YYYY_MM_DD/ (today's date).

Filename: Derived from author handle + topic. Strip invalid path characters (/, ?, =, &, spaces → _).

@cooljerrett + topic "AI productivity" → cooljerrett_AI_productivity.md

Write the report using the Threads 報告格式 defined in assets/output_template.md.

🎯 產生「AI 分析」區塊前，必須遵照 assets/output_template.md 中的指示讀取 data/goals.md 與 data/user_preferences.md（若存在）。

Confirm the file is written before proceeding.

Step 6Tb — Append suggestion to pending backlog

Append the AI analysis suggestion from this report to data/suggestions_pending.md:

---

### YYYY-MM-DD | Threads | [@{handle} — {topic}]({threads_url})
- 🏷️ {分類} | 💎 {價值評分} | ⚡ {可行動性} | 🎯 {決策建議}
- 📋 建議：{建議下一步}
- 📄 [報告](file:///absolute/path/to/report.md)

If data/suggestions_pending.md doesn't exist, create it with the # 📋 Pending Suggestions heading first.

Step 7T — Mark the Threads task as completed

gws tasks tasks patch \
  --params '{"tasklist": "<DELEGATE_LIST_ID>", "task": "<TASK_ID>"}' \
  --json '{"status": "completed"}'

Log: "✅ Task '<title>' marked as completed. Report saved to reports/Threads_YYYY_MM_DD/<filename>.md"

---

YouTube Flow (Steps 4Y–7Y)

Step 4Y — Launch YouTube background job

For each task in youtube_queue, apply the Video Strategist from yt2doc:

📄 Read the Video Strategist table in .agents/skills/yt2doc/SKILL.md (Step 2)

Video Duration	Whisper Model	Est. Time	Min Docker RAM
< 30 min	medium	5–10 min	4 GB
30–60 min	small	10–20 min	6 GB
1–2 hours	small	35–55 min	8 GB
> 2 hours	base	50–80 min	10 GB

If duration is unknown, look it up via web search or yt-dlp --print duration_string. Default to the conservative path when uncertain.

Create the output directory and launch in the background (use run_command with WaitMsBeforeAsync=5000, then poll with command_status):

mkdir -p reports/YouTube_YYYY_MM_DD

docker run --rm \
  -v "$(pwd)/reports/YouTube_YYYY_MM_DD:/output" \
  ghcr.io/shun-liang/yt2doc \
  --video "<YouTube URL>" \
  --output /output/<video_id>.md \
  --whisper-model <model> \
  --add-table-of-contents

Tell the user what's happening:

"Launching yt2doc for <url> using the <model> model (~X–Y minutes). Running in the background while I process Threads tasks."

Store the job metadata in-memory: { task_id, youtube_url, command_id, output_path, model }.

Docker not running? Skip this task with a warning: "⚠️ Skipping YouTube task '<title>': Docker is not running." Continue with remaining tasks.

Step 5Y — Poll for YouTube job completion

After all Threads tasks are done, poll each YouTube command_id:

command_status(command_id, WaitDurationSeconds=60)  # repeat until DONE

While polling, periodically report elapsed time to the user:

"Still transcribing <title>… (N minutes elapsed)."

On completion, check the exit code:

• Exit code 0 → proceed to Step 6Y
• Exit code 137 (OOM) → report: "⚠️ Task '<title>' FAILED: Docker ran out of memory. Increase Docker RAM: Docker Desktop → Settings → Resources → Advanced → Memory (8 GB minimum, 12 GB recommended)." Do not mark the task as done.
• Other non-zero → show last 20 lines of stderr. Do not mark the task as done.

Step 6Y — Generate the YouTube report

Once the yt2doc output file is confirmed non-empty:

1. Read the full transcript file content
2. Extract: video title (first # heading), chapter count, approximate character count
3. Generate a Traditional Chinese summary from the transcript content following the same quality rules as the Threads flow (Zero Hallucination, Comprehensiveness, Objectivity)
4. Write the report using the YouTube 影片報告格式 defined in assets/output_template.md.

🎯 產生「AI 分析」區塊前，必須遵照 assets/output_template.md 中的指示讀取 data/goals.md 與 data/user_preferences.md（若存在）。

Confirm the file is written before proceeding.

Step 6Yb — Append suggestion to pending backlog

Append the AI analysis suggestion from this report to data/suggestions_pending.md:

---

### YYYY-MM-DD | YouTube | [{Video Title}]({youtube_url})
- 🏷️ {分類} | 💎 {價值評分} | ⚡ {可行動性} | 🎯 {決策建議}
- 📋 建議：{建議下一步}
- 📄 [報告](file:///absolute/path/to/report.md)

If data/suggestions_pending.md doesn't exist, create it with the # 📋 Pending Suggestions heading first.

Step 7Y — Mark the YouTube task as completed and cleanup

# Mark as completed
gws tasks tasks patch \
  --params '{"tasklist": "<DELEGATE_LIST_ID>", "task": "<TASK_ID>"}' \
  --json '{"status": "completed"}'

# Remove the intermediate raw transcription file (already included in the final report)
rm "/Users/allanbian/my-ai-workflow/reports/YouTube_YYYY_MM_DD/<video_id>.md"

Log: "✅ Task '<title>' marked as completed. Report saved to reports/YouTube_YYYY_MM_DD/<filename>.md. Intermediate file removed."

---

Step 8 — Final Summary

After all queues are processed, output a unified summary:

Processed X Threads task(s):
  ✅ @handle1 — topic → reports/Threads_2026_04_29/filename1.md
  ✅ @handle2 — topic → reports/Threads_2026_04_29/filename2.md

Processed Y YouTube task(s):
  ✅ Video Title → reports/YouTube_2026_04_29/video_id.md
  ⚠️ Another Video → FAILED (OOM — increase Docker RAM to 8 GB)

Skipped Z task(s) (no supported URL):
  — "Buy groceries"
  — "Call dentist"

Step 9 — Ephemeral Cleanup

Once the entire backlog is cleared, delete all temporary diagnostic files created during the session:

• Screenshots (*.png)
• Temporary scratch files or ID lists (*.txt, *.json generated for the task)
• Any debug logs outside reports/

Log: "🧹 Session cleanup complete: removed temporary diagnostic files."

---

Troubleshooting

Task list not found: Run gws tasks tasklists list and confirm "Delegate" exists.

Threads URL behind login wall: Follow the authenticated access section in fetch-threads-post/SKILL.md.

gws tasks tasks patch fails: Double-check tasklist and task are IDs (not titles). The task ID comes from the id field in the tasks list response.

YouTube — Exit code 137 (OOM): Increase Docker RAM: Docker Desktop → Settings → Resources → Advanced → Memory (8 GB minimum, 12 GB recommended). Retry with --whisper-model base if RAM is still constrained.

YouTube — LLMModelNotSpecified error: Do NOT use --segment-unchaptered unless you also have a local Ollama running. Remove that flag.

YouTube — ChunkedEncodingError: Network interruption during model download — retry the same command; it resumes from cache.

URL contains & in shell: Always wrap URLs in quotes.

---

Anti-Truncation Standard (Threads)

1. Click "Read more" if visible.
2. Scroll down mandatory: Always run scroll down 1000 before extraction to trigger lazy-loading of long posts and masked content.
3. Use get text body: Always use get text body as the primary extraction method to capture content behind modals or overlays.
4. Manual check: A complete post usually ends with a footer, signature, or engagement metrics. If the text ends abruptly, it is likely truncated.
5. No summarization of Raw Content: The 📄 原始內容 section MUST contain verbatim extraction — never summarize or omit parts.