URGENT: Auto-Compaction Regression — Context Loss Causing Cascading Hallucination Mid-Session

Summary

The recent Claude Code push changed auto-compaction behavior in a way that drops user-stated intent and corrections from the context window. This causes agents to lose scope mid-session, leading to cascading hallucination and significant degradation in work product quality.

The Problem

When auto-compaction fires, the user's actual words — their stated intent, corrections, and constraints — get dropped from context. This causes:

1. Cascading hallucination — agents forget what they were asked to do and begin acting on assumptions rather than the user's stated intent
2. Repeated re-alignment cycles — users must manually re-inject context ("you keep forgetting context"), burning significant time and breaking flow
3. Compounding scope drift — each compaction loses more signal, and agents cannot self-correct because the reference material (the user's actual instructions) is gone

Steps to Reproduce

1. Start a Claude Code session with a complex, multi-step task
2. Work through enough turns that auto-compaction triggers
3. After compaction, observe that the agent has lost awareness of previously stated constraints, corrections, and scope
4. Attempt to continue — the agent acts on stale or hallucinated context rather than what was explicitly stated earlier in the conversation

What's Needed (Priority Order)

1. Immediate: Provide an opt-out for auto-compaction. Users with their own context management systems need the ability to disable built-in compaction. This is the fastest path to restoring quality for power users.
2. Short-term: Revert the compaction behavior to the previous version until the regression is resolved.
3. The core UX gap: Auto-compaction must re-inject the user's actual words and stated intent into the post-compaction context — with no duplication. The user's explicit instructions are ground truth; dropping them is a data loss event, not a compression optimization.

Impact

• Power users with calibrated skill files, hooks, and alignment systems are seeing significant quality drops
• Users without calibrated skill files are hit even harder — they have no infrastructure to detect or recover from context loss
• Before this change, Claude Code was easily 10X better than alternatives. This regression materially narrows that gap — not because the model is worse, but because the scaffolding is discarding the most important signal: what the user actually said.

Proposed Solution

Compaction should be lossy on tool output and verbose reasoning, never on user-stated intent, corrections, and constraints. Key principle:

• User's actual words → always preserved through compaction boundaries
• Agent reasoning/tool output → compressible
• User corrections mid-session → highest priority retention (these are the most valuable signal)

I have built in-house compaction systems that preserve intent through context boundaries and would be happy to contribute those patterns. The architecture uses structured extraction of user intent before compaction fires, then re-injection into the post-compaction context window.

Environment

• Claude Code CLI (latest as of 2026-03-18)
• macOS
• Long-running sessions with 20+ turns, multiple file reads, and complex multi-step tasks

---

Filed by Jonathan Haber, Founder @ IX Coach / Next AI Labs (founder@ixcoach.com)
Happy to jump on a call or share compaction architecture in detail.