GitHub Issue #42796
- Entity ID:
ent-20260419-1c99e6d20827 - Type:
issue - Scope:
shared - Status:
active - Aliases: #42796, Laurenzo thinking regression issue
Description
Stella Laurenzo's filed issue on April 1, 2026 documenting the thinking-depth regression with proxy-measurement evidence. Reached 790 HN points. Boris Cherny's official response (that redaction only hides thinking from UI) was rejected by Laurenzo using signature-field proxy evidence showing decline predated redaction.
Key claims
- Default effort elevated to HIGH as response to regression report
- Adaptive Thinking Zero-Budget Bug is Mechanically Explained by Design
- DISABLE_BACKGROUND_TASKS=1 is B4 Mitigation, DISABLE_ADAPTIVE_THINKING=1 is Zero-Budget Workaround
- Anthropic internal users receive person-oriented prompt; external users receive brevity directive
- Adaptive thinking allocated zero reasoning tokens on fabrication turns
- Shallow thinking drives 80x API request count and 122x billing despite stable prompt count
Relations
- Default Effort Elevation to HIGH --[informed_by]--> GitHub Issue #42796
- GitHub Issue #42796 --[summarizes]--> Thinking Depth Regression (pre-redaction)
- Adaptive Thinking Official Documentation --[summarizes]--> GitHub Issue #42796
- CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING --[fixed]--> GitHub Issue #42796
- Claude Self-Observation Passage --[derived_from]--> GitHub Issue #42796
- Two-Version Output Efficiency Directive --[caused]--> GitHub Issue #42796
- Q1 2026 Feature Release Cluster --[caused]--> GitHub Issue #42796