Activity monitor should detect and recover from Claude Code whirlpool hang

[Nanyan]

Problem

Claude Code's Whirlpool (context compaction) can hang indefinitely, leaving the agent completely unresponsive. The current activity monitor does not detect this state — the Claude process is still running, tmux session exists, so all liveness checks pass. But the agent processes zero messages.

Incident: zylos305 (2026-04-14 ~13:16 UTC)

Timeline:

• Agent was actively working (code edits, test runs, GitHub issue creation on coco-dashboard)
• Context reached 82% (control_queue warnings at 71% and 82%)
• Claude entered Whirlpool compaction at ~13:16 UTC
• Whirlpool hung — over 14 minutes with no progress (Whirlpooling… (13m 57s · ↓ 2.6k tokens · thought for 2s))
• 3 messages queued (2 scheduled tasks + 1 Lark group message), none processed
• Agent remained in this state until manual intervention

Environment snapshot:

• Claude Code 2.1.107
• Claude process PID 2298112, RSS 634MB, running for 8h16m
• VM: cocoai-zylos305-1b (7.8Gi RAM, 73% disk)
• Last clean exit was 2026-04-04 (10 days of continuous operation with /clear-based session resets)
• 14 IN vs 5 OUT messages in last hour at time of detection
• state.md was large with extensive test environment context

Tmux pane at time of detection:

● Bash(cd ~/zylos/workspace/coco-dashboard && gh issue create ...)
  ⎿  https://github.com/coco-xyz/coco-dashboard/issues/1329

✶ Whirlpooling… (13m 57s · ↓ 2.6k tokens · thought for 2s)
  ⎿  Tip: Use /btw to ask a quick side question...

  ❯ [Scheduled Task: task-mnv2xy40-p20fiz] ...
  ❯ [Lark GROUP:coco-dashboard研发] 3deca688 said: ...
  ❯ [Scheduled Task: task-mn31ohyb-hiq20f] ...

Expected Behavior

The activity monitor should detect that Claude is stuck (not producing output despite having queued input) and automatically recover — either by aborting the compaction or by killing and restarting the Claude process.

Proposed Detection Approaches

Option A: Tmux screen scrape (targeted)

• Periodically capture tmux pane text
• Detect Whirlpooling or Compacting keywords
• If the text persists for >5 minutes, consider it hung
• Recovery: kill Claude process → activity monitor auto-restarts

Option B: Message flow monitoring (general)

• Monitor c4.db conversations table
• If last IN is newer than last OUT by >N minutes while Claude process is running, flag as stuck
• This would catch whirlpool hangs AND other hang scenarios (API timeouts, etc.)
• More robust but may need tuning to avoid false positives during legitimate long-running tasks

Option C: Combine both

• Use Option B as the general detector with a longer threshold (e.g., 15-20 min)
• Use Option A as a fast-path for known hang patterns (e.g., 5 min for whirlpool)

Impact

• Agent becomes completely unresponsive during a whirlpool hang
• All queued messages (user messages + scheduled tasks) pile up
• No automatic recovery — requires manual SSH intervention
• Users perceive the bot as "dead" with no explanation

Workaround

Manual: SSH into VM → kill Claude PID → activity monitor auto-restarts → queued messages redelivered in new session.