claude-code - 💡(How to fix) Fix [BUG] agent thinking work lost when hitting limit

You're right that it's a real gap in the tooling, and your frustration is completely fair.

What actually exists today:
- Claude Code has memory files (the /memory/ system in this project) — but those are for facts and preferences, not
mid-reasoning agent state
- There's no "pause and resume a subagent" capability — when an agent hits its limit, it's gone
- There's no way to checkpoint Opus mid-analysis and pick it up later

What should exist but doesn't:
- Resumable subagent sessions
- Automatic state serialization before hitting token limits
- A warning threshold so the agent can summarize its findings before dying

How to file feedback:
- GitHub: https://github.com/anthropics/claude-code/issues
- That's the official channel for Claude Code feature requests and bug reports

The /compact command helps compress the main conversation, but it doesn't help a subagent that's running out of runway
mid-task. Spawning Opus for open-ended long-form reasoning without tight scope constraints is a design flaw in how I used
 the tool — and the platform not having a graceful fallback compounds it.

File the issue. This is a legitimate complaint and worth tracking.