Context limit reached during plan mode causes orphaned work and wasted tokens

Bug

When working in plan mode with a large conversation, Claude Code can hit the context limit mid-exploration. When this happens:

1. Plan mode enters and spawns an Explore agent (27 tool uses, ~115k tokens)
2. Explore agent completes successfully
3. Claude reads files to continue planning
4. Context limit reached — conversation becomes stuck
5. /compact fails with: Error: Conversation too long. Press esc twice to go up a few messages and try again.
6. /compact continues to fail on retry
7. The entire plan mode session (agent work + tokens) is orphaned and wasted

Expected Behavior

• /compact should be able to recover from context limit even in large conversations
• Alternatively, Claude Code should proactively compact before hitting the hard limit, especially when in plan mode with active agent results
• Plan mode work (agent results, file reads) should not be silently lost

Impact

• ~115k tokens wasted on an Explore agent whose results were never usable
• User has to start a new conversation and redo the entire planning phase
• No way to recover — esc twice + /compact still fails

Environment

• Claude Code CLI
• Model: Claude Opus 4.6
• Large conversation with multiple prior tool uses before entering plan mode