What probably matters to youSection of the release
What
The shared model-capability gate helper (dm, formerly Rm) now distinguishes an explicit false result from an undefined/unset one. Previously only a falsy or undefined answer fell through to the same "no answer" fallback; now an explicit false immediately disables the feature.
This affects checks such as mid_conv_system, adaptive_thinking, rejects_disabled_thinking, context_management, fast_mode, lean_prompt, thrifty_sonic, mid_conv_tool_change, and refusal_fallback, which now short-circuit to "off" as soon as the gate explicitly says false, instead of falling back to matching on model-name strings like opus-4-8 or claude-mythos-5.
The fable_5_mitigations gate similarly now respects an explicit false for the claude-mythos-5 model, where before mitigations were always applied for that model regardless of the gate value.
A new CLAUDE_CODE_MODEL_CAPABILITIES environment variable lets you override these capability flags directly, as a semicolon-separated list of rules in the form modelPattern=flag1,-flag2,..., with a trailing * wildcard supported when matching model IDs.
Why
Makes model-specific behavior more predictable: capability gates can now definitively turn a feature off rather than silently falling back to guesswork based on the model's name, and the new env var gives a direct way to override these flags.
