Bug Description

The OpenAI Chat Completions streaming parser (StreamingDelta in rig-core/src/providers/openai/completion/streaming.rs) does not include a reasoning_content field. This causes reasoning/thinking content from OpenAI-compatible providers to be silently dropped during streaming.

RawStreamingChoice::ReasoningDelta was added in PR #1395, and PR #1396 implemented reasoning support for 5 specialized providers (OpenAI Responses API, Anthropic, Gemini, xAI, OpenRouter). However, the generic OpenAI Chat Completions streaming parser — used by send_compatible_streaming_request — was not updated.

The DeepSeek provider has its own StreamingDelta with reasoning_content (line 722 of deepseek.rs), so it works correctly. But any provider using the generic OpenAI-compatible path loses reasoning content.

Affected Providers

All models that send reasoning_content via the standard Chat Completions streaming format:

• GLM-4.7 (Zhipu AI) — verified, 193 reasoning deltas silently dropped
• DeepSeek models via OpenAI-compatible endpoints (not the dedicated provider)
• Qwen with thinking mode (Alibaba DashScope)
• vLLM / Ollama serving reasoning models via OpenAI-compatible API
• Any other OpenAI-compatible endpoint returning delta.reasoning_content

Root Cause

StreamingDelta (line 35) only has content and tool_calls:

ust struct StreamingDelta { #[serde(default)] content: Option<String>, #[serde(default, deserialize_with = "json_utils::null_or_vec")] tool_calls: Vec<StreamingToolCall>, }

The reasoning_content field from SSE chunks like:
json {"choices":[{"delta":{"reasoning_content":"Let me think...","content":null}}]}
is silently ignored by serde (#[serde(default)] not present → field not parsed).

Suggested Fix

Two changes in rig-core/src/providers/openai/completion/streaming.rs:

1. Add reasoning_content to StreamingDelta:
   ust struct StreamingDelta { #[serde(default)] content: Option<String>, #[serde(default)] reasoning_content: Option<String>, // ← add this #[serde(default, deserialize_with = "json_utils::null_or_vec")] tool_calls: Vec<StreamingToolCall>, }

2. Yield ReasoningDelta in the stream (before the text content block):
   ust if let Some(reasoning) = &delta.reasoning_content && !reasoning.is_empty() { yield Ok(streaming::RawStreamingChoice::ReasoningDelta { id: None, reasoning: reasoning.clone(), }); }

This matches the existing pattern in the DeepSeek provider (deepseek.rs line 840).

Reproduction

1. Configure an OpenAI-compatible provider with a reasoning model (e.g. GLM-4.7 via https://open.bigmodel.cn/api/coding/paas/v4)
2. Use model.stream() or agent.stream_chat()
3. Observe: no ReasoningDelta / Reasoning events in the stream
4. The model's thinking content is silently discarded

Environment

• rig-core: 0.31.0 (latest main)
• Tested with: GLM-4.7 (Zhipu AI OpenAI-compatible API)