Context window blowouts started with v2.1.32 — sessions die mid-task with no recovery

Summary

Since upgrading to v2.1.32, long-running sessions consistently blow out the context window mid-task with no warning or recovery. This did not happen on prior versions. The problem persists on v2.1.34 and v2.1.37.

Suspected cause

v2.1.32 changed the context window blocking limit from ~65% to ~98%. This appears to let sessions run too close to the edge before auto-compaction kicks in. Compaction either fails silently or doesn't reclaim enough context to continue. The new auto-memory feature (also introduced in 2.1.32) likely adds additional context overhead that compounds the problem.

Reproduction

• Workflow: Sequential multi-step code analysis and editing (test robustness review across several files). No parallel subagents — just reading files, analyzing, and writing edits.
• What happens: Session dies mid-operation. No warning, no graceful degradation. All accumulated context and in-progress analysis is lost.
• Frequency: Consistent. Every session involving non-trivial multi-step work hits this.
• Workaround: Manually checkpointing state to memory files as work progresses. This preserves what was being done but not the analysis context, so the work has to be partially re-done (at additional token cost).

Environment

• Versions affected: v2.1.32, v2.1.34, v2.1.37
• Versions that worked: Pre-2.1.32
• Platform: macOS Darwin 25.2.0
• Model: Claude Opus 4.6

Impact

This makes Claude Code unusable for real multi-step engineering tasks. Any operation that takes more than a few steps risks losing all progress. Users end up paying for the same Opus-tier tokens twice (or more) to redo lost work.