Compaction causes agent to resume stale plans and force-approve destructive actions

Summary

After context compaction, Claude Code frequently resurfaces plans that were already completed, shelved, or canceled — then presents them as a yes/no approval with no visible conversation history, forcing the user to approve blindly. ~50% of the time this results in the agent executing work that shouldn't happen.

Steps to Reproduce

1. Have a long session with multiple plans discussed (some completed, some shelved)
2. Continue working until compaction triggers
3. After compaction, Claude immediately presents a stale plan as an actionable checklist
4. User sees only the approval prompt — no prior conversation visible for context
5. User approves (only option without context)
6. Claude begins building/executing the stale plan

Frequency

4-8 times per session. Every compaction event risks this.

Impact

• Agent creates files, writes code, modifies projects based on canceled/shelved plans
• User must catch and manually stop the agent mid-execution
• Work gets broken and must be reverted
• Significant time wasted per occurrence (user is a solo founder with limited runway)

Expected Behavior

• After compaction, the agent should ask what the user is currently working on rather than auto-resuming a plan
• Stale plans should not be presented as actionable
• If a plan must be re-presented, the user should be able to see recent conversation history before approving

Actual Behavior

• Agent treats the most recent plan in compressed context as active and immediately acts on it
• No distinction between "plan we're building" vs "plan we discussed and shelved"
• The approval UI shows the plan but not the conversation that led to shelving/completing it

Environment

• Claude Code CLI on Windows (MINGW64)
• Long sessions (4+ hours) with multiple planning/building cycles
• Model: Claude Opus