# Claude Code compaction mid-task: what do you do so you don't lose the thread? — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/f490b151e622f59c

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-10-21 13:18 ingested 2026-10-22 02:37

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xbmxo7/claude_code_compaction_midtask_what_do_you_do_so/)

## retained text



```
In the claude-code GitHub issues, compaction complaints seem to fall into three groups:

1. It fires earlier than people expect. One issue reports a session where it triggered at roughly 35% of context, which is one person's measurement, not a rule.
2. It fails to run in long conversations. Anthropic's changelog lists a fix for one case in v2.1.288.
3. The summary drops details like earlier decisions, the reasoning behind them, and file paths.

I can't tell how widespread any of these are. The habits people describe are still worth comparing:

* Put decisions and constraints in [CLAUDE.md](http://CLAUDE.md) before a long task
* Compact manually at a clean break instead of mid-task
* Start a fresh session with a short handoff note
* Keep tasks small enough that no session gets that long

Which of these have worked for you, and have you found anything better?
```



document fetched 2026-10-22 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xbmxo7/claude_code_compaction_midtask_what_do_you_do_so
