Claude 5.5 effort levels and thinking-off; Copilot uses Anthropic Messages
- Native
xhigh and max effort for Claude Opus/Sonnet/Haiku 5.5 on the Anthropic API, GitHub Copilot and Bedrock. Opus/Sonnet 4.6 get max. A saved xhigh on 4.6 now clamps up to max, which costs more on Sonnet 4.6.
- Haiku 5.5 uses adaptive thinking.
- Each model now turns thinking off its own way (
disabled, between_tools, or no thinking field). You can't choose off on Opus 5.5; it clamps to minimal.
- One shared helper now clamps thinking levels everywhere: session, CLI, agent loop, subagents and arbiter. Clamping no longer overwrites your saved default level.
- GitHub Copilot: Claude models now go through Copilot's
/v1/messages API. Before, the thinking level you picked had no effect on Claude 5.x there. dreb now discovers the models your account can use from Copilot /models at login and token refresh, and hides the rest.
Resolves issue 545 via PR 546.
