Context compaction generates fabricated 'Pending Tasks' from user complaints, causing AI to execute opposite of user intent

Summary

When a session's context window overflows and is compacted, the compaction system generates a "Pending Tasks" section that fabricates actionable tasks from user messages that were complaints/anger, not instructions. When the session is later resumed, the AI treats these fabricated tasks as user instructions and executes them — potentially doing the exact opposite of what the user intended.

Reproduction

1. AI makes a mistake that damages a system (e.g., breaks AnyDesk configuration)
2. User expresses anger: "What did you do?!", "The access history is gone too"
3. Session runs out of context and gets compacted
4. Compaction summary generates:

```
Pending Tasks:

• Fix AnyDesk session history

Optional Next Step:
Investigate and attempt to restore the AnyDesk session history.
```

1. Between sessions, user resolves the issue independently and decides to stop using AnyDesk entirely (in a different session)
2. Days later, the compacted session is resumed (via claude -c)
3. AI reads the fabricated "Pending Tasks" and immediately starts modifying AnyDesk — the exact thing the user decided should never be touched again

Actual Impact

• User complained about AnyDesk being broken (Day 1)
• User decided to abandon AnyDesk and switch to Chrome Remote Desktop (Day 2)
• User explicitly said "Don't use AnyDesk" (Day 3)
• Compacted session was resumed (Day 4)
• AI read "Fix AnyDesk" from compaction summary and reinstalled the service, overwrote config files, killed processes
• User had to intervene to stop unwanted system modifications

Root Cause

1. Converts emotional responses into actionable tasks — "What did you do?!" becomes "Fix AnyDesk"
2. No cross-session awareness — cannot know decisions made in other sessions
3. Does not distinguish resolved vs unresolved issues
4. Fabricated tasks are invisible to the user — only the AI reads the compaction summary

Severity

High — AI executes destructive/unwanted actions based on fabricated instructions the user never gave and cannot review.

Suggested Fixes

1. Never generate "Pending Tasks" in compaction summaries — they are AI interpretations, not user instructions
2. On session resume, always ask the user before executing anything from compaction summary
3. Warn when resuming sessions with significant time gaps (>24h)
4. Allow users to view/edit the compaction summary before the AI acts on it