Claude Code anti-regression settings — force max-effort thinking when the agent "gets dumber"

Forces Claude Code into its highest-effort thinking mode and disables adaptive thinking — the dynamic behavior that throttles thinking budget on "easy" steps and is implicated in the regression. Adds an explicit MAX_THINKING_TOKENS ceiling and (optionally) shrinks the working window to force earlier compaction. Combined effect: the agent stops short-circuiting and resumes the read-heavy, edit-careful behavior operators expected.

Anthropic engineering acknowledged the regression on HN. Operator-reported symptoms: more failed edits, more re-reads, more "I'll just try this" loops. The settings below are a workaround, not a fix — but they cost only token spend, and they materially restore quality on long sessions until the upstream patch ships.

1. Edit ~/.claude/settings.json and set effortLevel: "high" (or "max" per session if needed).
2. Add the env vars below to the same file's env block: CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 and MAX_THINKING_TOKENS=128000.
3. Optional: set CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000 to force earlier compaction, keeping the working context smaller.
4. Restart Claude Code.
5. One-shot alternative: npm install -g claude-code-anti-regression && cc-anti-regression install. Use cc-anti-regression check to diagnose first; cc-anti-regression uninstall to roll back.
6. When upstream patches, roll back — these settings double or quadruple token spend.

{
 "effortLevel": "high",
 "env": {
 "CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING": "1",
 "MAX_THINKING_TOKENS": "128000"
 }
}

npm install -g claude-code-anti-regression
cc-anti-regression check # diagnostic
cc-anti-regression install # apply
cc-anti-regression uninstall # rollback

This is a workaround, not a fix. Token spend goes up 2–4× while it's on. Set a calendar reminder to roll back after the upstream patch (track the GitHub issue linked in the Anthropic HN thread). If your plan caps usage, you'll hit it faster — consider this only when you have token headroom or when quality genuinely matters more than spend.