[BUG] Auto-compaction cannot prevent context overflow from atomic background agent result injection

Open 💬 1 comment Opened Jun 15, 2026 by lg320531124

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?

Auto-compaction triggers when context usage exceeds a threshold (e.g., CLAUDE_AUTOCOMPACT_PCT_OVERRIDE). However, when a background agent completes and its result is injected into the main context, no pre-injection compaction check occurs. The entire agent output (potentially 50-100K tokens) is injected atomically in a single turn, pushing context directly from ~75% to 100%+ — bypassing compaction entirely.

This is related to but distinct from #52390 (AUTOCOMPACT_PCT_OVERRIDE not triggering). Even if autocompact fires correctly at the threshold, it cannot prevent overflow from a single oversized injection because compaction only runs between turns, not during a turn.

The key insight: normal conversation grows context incrementally (small steps between turns), giving compaction a chance to intervene. Background agent results grow context atomically (one large injection within a turn), giving compaction no opportunity.

What Should Happen?

Before injecting any background agent result into the main context, the system should:

  1. Estimate the token count of the incoming result
  2. 2. Check if injection would exceed the autocompact threshold
  3. 3. If yes, run compaction FIRST to free up space
  4. 4. After compaction, check again — if still too large, truncate the result
  5. 5. Only then inject the result

This ensures compaction always runs BEFORE oversized content arrives, not after.

Error Messages/Logs

Steps to Reproduce

  1. Set CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=80
  2. 2. Use context until ~75%
  3. 3. Launch a background agent that returns ~50K tokens of output
  4. 4. Agent completes — its result is injected, pushing context from ~75% to ~100%
  5. 5. Autocompact never ran because the injection happened atomically within a single turn

Claude Model

Not sure / Multiple models

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.177

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

See companion issue #68582 for the broader context overflow problem. This issue focuses specifically on the compaction timing gap.

Related: #52390, #53065, #17208

View original on GitHub ↗

This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗