[bot] Groq streaming drops `reasoning` content from DeepSeek R1 and Qwen 3 reasoning models

[braintrust-bot]

Summary

The Groq API returns a reasoning field in chat completion responses when using reasoning-capable models (DeepSeek R1, Qwen 3 with reasoning_format: "parsed"), but the Groq plugin's streaming aggregation silently drops all reasoning content. The aggregateGroqChatCompletionChunks function delegates to aggregateChatCompletionChunks from the OpenAI plugin, which only accumulates delta.content, delta.refusal, and delta.tool_calls; the delta.reasoning field falls through and is lost. This is a direct parity gap with Anthropic, Google GenAI, AI SDK, and Cohere reasoning instrumentation in this repo, and the same class of bug as OpenRouter #1883.

What instrumentation is missing

Streaming aggregation: reasoning content silently dropped

In js/src/instrumentation/plugins/groq-plugin.ts, aggregateGroqChatCompletionChunks (line 110) delegates to aggregateChatCompletionChunks from openai-plugin.ts (lines 475–570). That function only captures three delta fields:

if (delta.content) {
  content = (content || "") + delta.content;
}
if (delta.refusal) {
  refusal = (refusal || "") + delta.refusal;
}
if (delta.tool_calls) { /* ... */ }
// delta.reasoning — silently dropped

The final aggregated output (lines 554–569) only includes role, content, refusal, and tool_calls — no reasoning field.

Vendor types: no reasoning field on delta

The Groq chunk type (js/src/vendor-sdk-types/groq.ts, line 34) extends OpenAIChatCompletionChunk, whose OpenAIChatDelta (in openai-common.ts, lines 110–117) declares role, content, refusal, tool_calls, and finish_reason. While [key: string]: unknown allows extra fields at runtime, the aggregation code never reads delta.reasoning.

Non-streaming: reasoning passes through in raw output

For non-streaming responses, the Groq plugin returns result.choices directly, so message.reasoning would be present in the raw output. The gap is specifically in streaming aggregation.

Upstream API format

Groq surfaces reasoning content through the reasoning field when reasoning_format: "parsed":

• Non-GPT-OSS models (Qwen 3 32B): reasoning_format parameter with "raw" (think tags in content), "parsed" (separate reasoning field), or "hidden" modes
• GPT-OSS models (DeepSeek R1 distills): include_reasoning: true returns reasoning field in messages

In streaming, delta.reasoning contains the reasoning text chunks. In "raw" mode, reasoning comes through delta.content with <think> tags (captured correctly). The gap only affects "parsed" mode where delta.reasoning is a separate field.

Comparison with other providers in this repo

Provider	Reasoning content captured in streaming	Issue
Anthropic	thinking_delta aggregated	✅ Working
Google GenAI	thought parts handled	✅ Working
AI SDK	reasoning-delta chunks aggregated	✅ Working
Cohere	thinking content blocks aggregated	✅ Working
Mistral	thinking chunks aggregated	✅ Fixed in #1857
OpenRouter	Silently dropped	❌ Open #1883
Groq	Silently dropped	❌ This issue

Braintrust docs status

not_found — The Braintrust Groq integration page at https://www.braintrust.dev/docs/integrations/ai-providers/groq does not mention reasoning content handling.

Upstream references

• Groq Reasoning docs: https://console.groq.com/docs/reasoning
• Groq API reference: https://console.groq.com/docs/api-reference
• Affected models: DeepSeek-R1-Distill-Llama-70B, DeepSeek-R1-Distill-Qwen-32B, Qwen 3 32B (with reasoning_format: "parsed")
• reasoning_format parameter: "raw" | "parsed" | "hidden"
• include_reasoning parameter: boolean (for GPT-OSS models)

Local files inspected

• js/src/instrumentation/plugins/groq-plugin.ts (line 110: aggregateGroqChatCompletionChunks delegates to OpenAI's aggregation)
• js/src/instrumentation/plugins/openai-plugin.ts (lines 475–570: aggregateChatCompletionChunks only captures content, refusal, tool_calls)
• js/src/vendor-sdk-types/groq.ts (line 34: GroqChatCompletionChunk extends OpenAIChatCompletionChunk)
• js/src/vendor-sdk-types/openai-common.ts (lines 110–117: OpenAIChatDelta has no reasoning field)
• e2e/scenarios/groq-instrumentation/ (no reasoning test scenarios)