fix: pass --thinking flag for adaptive/disabled instead of budget tokens

[hackyon-anthropic]

Summary

thinking={"type":"adaptive"} was being lowered to --max-thinking-tokens 32000, which puts the CLI into fixed-budget mode rather than adaptive mode. disabled similarly mapped to --max-thinking-tokens 0.

The CLI exposes an explicit --thinking <adaptive|enabled|disabled> flag that takes precedence over --max-thinking-tokens. The TypeScript SDK already uses it. This change mirrors that behavior:

• adaptive → --thinking adaptive
• disabled → --thinking disabled
• enabled → --max-thinking-tokens <budget_tokens>
• deprecated max_thinking_tokens is only emitted when thinking is unset

Test plan

• pytest tests/test_transport.py (59 passed)
• New parametrized test for all three thinking types
• New precedence test: thinking overrides max_thinking_tokens

[claude]

LGTM — clean fix that correctly maps thinking types to the right CLI flags, with good test coverage.

Extended reasoning...

Overview

The PR modifies in to fix how thinking configuration is translated to CLI flags. It also adds tests in . Two files changed, both in the transport layer.

Security Risks

None. This is purely a CLI argument construction fix with no auth, crypto, or permission-sensitive code involved.

Level of Scrutiny

Low. The change is small, mechanical, and clearly correct. The old code incorrectly mapped adaptive to --max-thinking-tokens 32000 (fixed budget) and disabled to --max-thinking-tokens 0, when both should use the dedicated --thinking flag. The fix mirrors the TypeScript SDK behavior as stated in the description.

Other Factors

The new tests cover all three thinking types parametrically and verify the precedence of thinking over the deprecated max_thinking_tokens. The logic restructure from if/elif/elif + if to if/elif is clean and correct. No outstanding reviewer comments.