Auto Compact triggering too early
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
Recently I’m seeing auto-compact trigger much earlier than expected, well before the context window is anywhere near full. At first I assumed this was a bug, but after observing it a few times I’m wondering if this is an intentional behavior change? possibly triggered when a large amount of content or multiple documents are ingested early in the session.
Observed behavior:
Auto-compact runs even though the context usage appears low
This happens early in the session, not after long conversations
Leads to unexpected loss of context / summaries kicking in too soon
Is this: a known bug?
an intentional change in auto-compact heuristics?
expected behavior when ingesting many documents at the start of a session?
<img width="824" height="987" alt="Image" src="https://github.com/user-attachments/assets/4c70f94e-7a2c-4393-ab3b-13aef6f77345" />
What Should Happen?
My expectations from the past year working with Claude is that Auto-compact should only trigger when the context window is actually close to being full, or at least much later than what I’m seeing now.
Error Messages/Logs
Steps to Reproduce
Steps to Reproduce
Start a fresh Claude Code session (tested on Claude Code v2.1.12, Opus 4.5 / Claude Max).
Open a medium–large real-world codebase (TypeScript / React in my case).
Ask Claude to review or verify implementation across multiple files (e.g. “check if everything is implemented correctly before testing”).
Allow Claude to sequentially read multiple files,
L
Example observed reads:
~350–450 lines markdown specs
~1k–2k line TS / TSX files
Multiple files in a row, no back-and-forth conversation
Observe that:
Claude successfully reads many files in sequence
Context does not appear close to full (checked multiple times!)
No warning about imminent context exhaustion
While continuing the same task (still mid-review, not idle), auto-compact triggers:
- Compacting conversation…
This happens mid-task, before the review completes, and results in:
Loss of earlier file context
Task continuation based on a summarized / compacted state
Claude Model
Opus
Is this a regression?
Yes, this worked in a previous version
Last Working Version
_No response_
Claude Code Version
2.1.12
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
Cursor
Additional Information
_No response_
This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗