Greptile Summary

This PR fixes the missing "disabled" thinking transform path across the Anthropic and ZAI outbound transformers. When a client sends thinking: {type: "disabled"}, the inbound layer now sets ReasoningEffort="none" and the TransformerMetadataKeyThinkingType marker, and the outbound layer correctly maps both signals to Thinking{Type: "disabled"} — including the DeepSeek output_config.effort guard that now skips "none". Additionally, TypeXiaomi is migrated from the generic OpenAI transformer to the ZAI transformer at Version="v1", and TypeCodex/TypeClaudecode are restricted to OAuth-only credentials.

Confidence Score: 5/5

Safe to merge — all findings are P2 or lower, production logic is correct and well-tested.

The core bug fix (disabled thinking not propagated to upstream) is correctly implemented with matching test coverage in both the Anthropic and ZAI transformers. The only finding is a P2 test-quality issue where the ZAI thinking tests assert an inline copy of the logic rather than invoking TransformRequest; this does not affect production behaviour.

llm/transformer/zai/thinking_test.go — tests replicate production logic inline instead of calling TransformRequest.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["Inbound: thinking.type=disabled"] --> B["inbound_convert.go\nSet TransformerMetadata\nthinking_type=disabled\nReasoningEffort=none"]
    B --> C["llm.Request"]
    C --> D{"buildBaseRequest\noutbound_convert.go"}
    D --> E{"TransformerMetadata\nthinking_type=disabled?"}
    E -- yes --> F["req.Thinking =\nThinking{Type:disabled}"]
    E -- no --> G{"ReasoningEffort\n== none?"}
    G -- yes --> F
    G -- no --> H{"ReasoningEffort non-empty\nOR ReasoningBudget set?"}
    H -- yes --> I["buildThinking()\nenabled/budget"]
    H -- no --> J["No thinking"]
    F --> K["Upstream Anthropic API\nthinking: disabled"]
    I --> K
    J --> K
    C --> L{"ZAI outbound.go"}
    L --> M{"ReasoningEffort == none?"}
    M -- yes --> N["Thinking{Type:disabled}"]
    M -- no --> O["Thinking{Type:enabled}"]

_{Reviews (5): Last reviewed commit: "fix(llm): missing anthropic thinking dis..." | Re-trigger Greptile}