[BUG] Auto-compact threshold silently regressed — fires at ~73% context used, previously ~83%

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?

I do long agentic sessions where I'm actively building Claude's reasoning context. Compaction destroys that context, so I manage sessions manually — I watch the token percentage and wrap up before compaction fires.

This workflow depends on knowing when compaction will fire. It used to fire consistently at ~83% context used. Today it fired at ~73% with no warning. That's a 10-point shift that invalidated my session boundary habits mid-session.

The problem isn't the number — it's that it changed silently. No changelog, no announcement, no setting I touched. I can't maintain a consistent workflow if this threshold moves without notice.

Expected: threshold stays stable, or breaking changes to session behavior appear in the changelog
Actual: threshold shifted ~10% downward, discovered only when compaction fired earlier than expected

This isn't a request to make it configurable — it's a request to stop changing it without telling people.

What Should Happen?

If the threshold changed intentionally, document it in the changelog
If it changed unintentionally, revert it
Either way, treat session behavior changes as breaking changes that warrant an announcement

Error Messages/Logs

Steps to Reproduce

Start a session
Use context until ~73% consumed
Observe auto-compaction fires — previously did not fire until ~83%

No error message — compaction fires silently, which is the point. The bug is behavioral, not a crash.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

I don't know. I am busy trying to corral Claude.

Claude Code Version

2.1.204 (Claude Code)

Platform

Other

Operating System

Windows

Terminal/Shell

Other

Additional Information

Compaction doesn't end the session — it silently rewrites the context and lets you continue, which makes it more dangerous than a hard stop. You don't always notice it fired. By the time you do, Claude's reasoning context has already been corrupted. That's why I manage session boundaries manually — to avoid compaction entirely.