Regression ~2PM CST Dec 5: Aggressive credit consumption during background task polling

Summary

Starting approximately 2:00 PM CST on December 5, 2025, Claude Code began exhibiting degraded behavior with aggressive credit consumption during background task monitoring.

Observed Behavior Changes (Post ~2PM CST)

1. Constant BashOutput polling - Every 2-3 seconds during background tasks
2. Each poll consuming credits - Even with no new output
3. Protocol failures - Not following established patterns, forgetting context
4. User reports: "credits are being consumed like you are fucking hand knitting"

Previous Behavior (Pre ~2PM CST)

• Background tasks ran with minimal credit drain
• Polling was efficient
• Protocol adherence was consistent

Technical Context

• Task: Docker buildx cross-platform build (ARM Mac → linux/amd64)
• Image size: 122MB (not large)
• Build completed all 33 steps successfully
• Got stuck polling step #34 "sending tarball" repeatedly
• User had to intervene to stop the polling loop

Impact

• HIGH severity - Breaks normal CI/CD workflows
• Makes long-running builds cost-prohibitive
• Degraded reasoning/protocol adherence observed simultaneously

Environment

• Claude Code CLI
• macOS (darwin)
• Model: claude-opus-4-5-20251101

Timeline

• ~2:00 PM CST: Behavior change observed
• ~3:12 PM CST: User escalated
• Issue duration: 1+ hour before user intervention

Question

Was there a deployment or change around 2:00 PM CST (8:00 PM UTC) on December 5, 2025 that could explain this regression?

---
Reported by user, filed by Claude Code