fix(proxy): Responses→Chat format conversion issues (cache_control, tool_choice, stream truncation)

[thedavidweng]

Problem

Several Responses→Chat format conversion bugs can cause upstream 400 errors or masked stream failures when routing Codex through Chat Completions providers.

1. cache_control passthrough causes 400 errors

During Anthropic→OpenAI format conversion, cache_control fields are preserved on system messages, text content blocks, and tools. Strict OpenAI-compatible endpoints (e.g. MiniMax, some Chinese providers) reject these unknown fields with 400 errors.

Additionally, when a user message has exactly one text content block with cache_control, the content is kept in array-of-objects format instead of being simplified to a plain string. Some providers expect the simpler format.

Location: src-tauri/src/proxy/providers/transform.rs — system message construction (line ~122), text block construction (line ~339), tool construction (line ~187), and single-block simplification (line ~407).

2. tool_choice sent without tools

In the Responses→Chat conversion, tool_choice is copied unconditionally from the input body, but tools is only included when non-empty. If all tools are filtered out during conversion, the request contains tool_choice without a corresponding tools field, which causes errors on providers that enforce this constraint.

Location: src-tauri/src/proxy/providers/transform_codex_chat.rs — lines 308-315. The tool_choice block should be inside the if !tools.is_empty() guard.

3. Truncated chat streams masked as normal completion

When a Codex Chat→Responses stream ends without receiving [DONE] or a finish_reason, the code unconditionally calls finalize() which emits response.completed. This masks truncation as successful completion. The fix should distinguish three cases:

• Normal completion: finish_reason or [DONE] received → response.completed
• Incomplete with output: substantive output exists but no finish_reason → response.incomplete
• Empty truncation: no output and no finish_reason → response.failed with stream_truncated

Location: src-tauri/src/proxy/providers/streaming_codex_chat.rs — finalize() method (line ~525) and response_status_from_finish_reason() in transform_codex_chat.rs (line ~1566).

Expected Behavior

• cache_control fields should be stripped from the OpenAI-format request body to avoid 400 errors on strict endpoints
• Single text content blocks should always be simplified to plain string format
• tool_choice should be omitted when tools is empty
• Truncated streams should be reported as incomplete or failed, not completed