telexed ~ c / b05ea88a-7c6radar:50 · agent_toolLIVE
← back
NO.
#b05ea88a
Topic
AGENTS & TOOLS
Source
Anthropic Engineering
Published
2026-05-05 06:20:21
Importance
★ 5/10 — radar 50
`Claude Code` quality regressions traced to 3 product-layer changes
FIG-0051:1

`Claude Code` quality regressions traced to 3 product-layer changes

The breakage came from defaults and prompts, not the model API: lower reasoning, broken session memory cleanup, and an anti-verbosity prompt. Fixed by v2.1.116; treat client defaults and prompt tuning as production-critical.

[ KEY POINTS ]
  1. On March 4, default reasoning dropped from high to medium to cut latency; Anthropic reverted it on April 7 after users preferred better output over speed.
  2. A March 26 idle-session cleanup bug kept wiping older thinking every turn after 1 hour idle, making responses forgetful and repetitive until the April 10 fix.
  3. An April 16 system-prompt change to reduce verbosity hurt coding quality and was rolled back on April 20; prompt edits alone were enough to degrade code output.
  4. The API and inference layer were explicitly unaffected. The failures sat in Claude Code, the Agent SDK, and Claude Cowork, which matters if you benchmark via product UX rather than raw API.
  5. Usage limits were reset for subscribers on April 23. Practical takeaway: pin versions, watch default-effort changes, and regression-test prompt/session behavior around coding flows.
Originalwww.anthropic.com/engineering/april-23-postmortemRead original →

// related