Laguna models intermittently inline reasoning as <think> tags in content instead of reasoning_content field

[23wwwtt]

Summary

When calling Laguna S 2.1 through the OpenAI-compatible endpoint (https://inference.poolside.ai/v1) from a third-party client (opencode 1.18.19 using @ai-sdk/openai-compatible), some responses carry the model's chain-of-thought inline in content, delimited by <think>...</think> tags, instead of populating the separate reasoning_content field.

The client-side model definition declares interleaved reasoning via interleaved: { field: "reasoning_content" }, and many requests do honor it. But intermittently, a response arrives as a single content string containing:

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

For those messages the reported reasoning token count is 0 at the message level, even though a full reasoning trace is present in the text.

Expected behavior

Consistent transport of reasoning across all requests — either always via the dedicated reasoning_content field, or (if inline tags are part of the wire format) documented so clients can parse it deterministically.

Steps to reproduce

Use any OpenAI-compatible client against inference.poolside.ai/v1 with model poolside/laguna-s-2.1, send a multi-step coding task requiring tool calls, and inspect raw streamed deltas across several turns. Some turns will show <think> blocks inside content deltas rather than reasoning_content deltas.

Impact

• Third-party UIs render the chain-of-thought as final assistant output (users see raw English monologue text).
• Usage/billing attribution is skewed: reasoning tokens are counted as regular output tokens on affected messages.
• Clients cannot reliably separate "thinking" from "answer" without ad-hoc parsing.

Request

Please either fix the inconsistency (always emit reasoning_content) or document the inline <think> convention officially so client integrations can handle it robustly.

[IgorGanapolsky]

This matches documented behavior in poolsideai/n8n-poolside-node, not only OpenCode / @ai-sdk/openai-compatible.

That package’s README says Laguna thinking is enabled by default on Poolside Platform and “the reasoning text arrives in the message content,” which breaks consumers that expect a structured field. It also notes the node does not echo reasoning_content on follow-up turns.

So a stable wire format — always reasoning_content, or always inline <think> with a parse rule — would let ACP clients, n8n, and OpenAI-compat SDKs handle this deterministically. Intermittent mixing of the two is the painful case.

No affiliation. Happy to re-test a documented format from pool 1.0.15 on macOS.