Env vars applied to ~/.claude/settings.json are read by Claude Code at startup and used to configure the main request loop. That's enough to close most of the regression gap — but not all of it.
Per @Frisch12's binary analysis on issue #42796:
CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1is checked in 1 of 5 adaptive code paths.- The remaining 4 paths — notably subagent spawning through the
V9H()helper and some tool-use fallbacks — do not check the env var. They apply adaptive thinking unconditionally.
Practical consequence: if you rely on subagents (Agent tool invocations, Task-based parallel work), they will still occasionally produce zero-thinking turns even with this tool installed. The main loop, however, is fixed.
It would be trivial to write a patcher that edits cli.js to force-disable adaptive thinking everywhere. We deliberately do not, for three reasons:
-
Terms of Service. Anthropic's ToS forbid reverse engineering or modifying their software. A patcher walks squarely into that grey area. Even if individual patches are low-risk in practice, publishing a tool that automates binary modification normalises it and puts users in a position where Anthropic support may decline to help them.
-
Fragility.
cli.jsis re-minified on every Claude Code release. Variable names likeV9Hchange from one build to the next. A patcher would need per-version signatures, maintained manually. With the 2.1.104+ Bun-compiled binary, the patch surface becomes a native executable — substantially harder and riskier to modify. -
It's a workaround, not a fix. The right remedy is upstream. Anthropic has acknowledged the regression in the issue tracker and on Hacker News. The correct path is: apply env vars as a partial mitigation, track the upstream fix, remove the workaround when it lands. A binary patcher would delay that cycle by making the symptom tolerable.
For users of opinionated workflow frameworks (SuperFlow and similar), the answer is to define agents explicitly in ~/.claude/agents/*.md with effort: max or effort: high frontmatter. That closes the subagent gap for agents you define and spawn explicitly. Built-in subagents (those defined inside the Claude Code binary itself) remain out of reach until Anthropic ships the fix.
Pattern-based agent-file patching is intentionally out of scope for v0.1 of this tool (see v0.2 roadmap). For now: if you use SuperFlow, run superflow update. If you have a custom setup, edit your agent files manually.
Watch issue #42796 for the upstream fix. When Anthropic raises the default effort back to high (or exposes a proper config knob), cc-anti-regression uninstall returns your settings.json to its pre-install state. Backups are kept, so there's no lossy path.