Thinking blocks missing when using Opus 4.6 with --max-thinking-tokens

[dolphin-wood]

Description

After switching from claude-opus-4-5-20251101 to claude-opus-4-6, thinking blocks are no longer returned in the response stream. The agent previously produced ThinkingBlock content with Opus 4.5, but with Opus 4.6 it goes straight to text/tool_use without any thinking.

Environment

• claude-agent-sdk: 0.1.31
• Bundled CLI: 2.1.33
• Model: claude-opus-4-6
• Platform: macOS arm64 / Linux x86_64

Configuration

options = ClaudeAgentOptions(
    max_thinking_tokens=10000,
    # ... other options
)

The model is set at runtime via client.set_model("claude-opus-4-6").

Root Cause (Suspected)

Opus 4.6 deprecated thinking: {type: "enabled", budget_tokens: N} in favor of adaptive thinking (thinking: {type: "adaptive"}). The --max-thinking-tokens CLI flag likely maps to the deprecated budget_tokens parameter, which may cause Opus 4.6 to skip thinking entirely.

From the Opus 4.6 docs:

Adaptive thinking (thinking: {type: "adaptive"}) is the recommended thinking mode for Opus 4.6. thinking: {type: "enabled"} and budget_tokens are deprecated on Opus 4.6.

Expected Behavior

Thinking blocks should still be returned when using Opus 4.6, either via the deprecated budget_tokens path or (preferably) via adaptive thinking.

Feature Request

Please add support for adaptive thinking in the SDK:

1. A new parameter (e.g., thinking_mode="adaptive") in ClaudeAgentOptions
2. A new CLI flag (e.g., --thinking-mode adaptive) mapping to thinking: {type: "adaptive"}
3. An optional effort parameter (low / medium / high / max) to control thinking depth

This would allow SDK users to migrate from the deprecated budget_tokens to adaptive thinking as recommended by the Opus 4.6 docs.

Workaround

Currently reverting to claude-opus-4-5-20251101 to retain thinking blocks.