Summary

For OpenAI-compatible providers whose models emit reasoning inline in content wrapped in <think>...</think> tags, opencode stores the raw tagged text verbatim as a text part. The UI then renders the model's chain-of-thought as regular assistant output instead of a collapsible thinking block.

Environment

• opencode 1.18.19
• Provider: custom OpenAI-compatible (poolside, endpoint https://inference.poolside.ai/v1)
• Model: poolside/laguna-s-2.1
• Model definition declares interleaved reasoning via interleaved: { field: "reasoning_content" }

Observed behavior

• Many requests correctly populate the separate reasoning channel and are displayed as thinking.
• Intermittently, a response arrives with the entire reasoning trace inline in content:

<long reasoning monologue ...></think><final answer>

opencode saves this whole string as one text part. The affected message reports reasoning: 0 tokens at message level, while the session total still shows accumulated reasoning tokens — so token accounting is also distorted (reasoning counted as output).

Expected behavior

Inline <think> blocks should be extracted into separate reasoning parts so they render as thinking, not as body text.

Suggested fix

Wrap affected provider/model instances with the AI SDK's extraction middleware, e.g.:

import { wrapLanguageModel, extractReasoningMiddleware } from 'ai'

const model = wrapLanguageModel({
  model: openaiCompatibleModel,
  middleware: extractReasoningMiddleware({ tagName: 'think' }),
})

Options:

1. Apply extractReasoningMiddleware automatically for providers/models known to inline think tags (similar to how DeepSeek R1 is handled).
2. Expose a per-model config flag (e.g. reasoning.extract: "think-tags") in the model/provider schema so users can enable it for any OpenAI-compatible endpoint.
3. A safe fallback could also apply startWithReasoning: true semantics for models that omit the opening tag.

Impact

• Users cannot review the agent's reasoning process (it appears as unreadable body text mixed with output).
• Reasoning/output token attribution is wrong on affected messages.