[rekram1-node]

Problem

OpenRouter lowered assistant reasoning details in a post-processing pass over body.messages using sourceAssistants[assistantIndex++], and cast the returned payload as OpenRouterBody (whose rest record used Schema.Any). When OpenAIChat.fromRequest skipped an empty assistant message during lowering, assistantIndex fell out of alignment with request.messages.

Change

• Project assistant reasoning and reasoning_details inside OpenAIChat.fromRequest's assistant lowering hook so each lowered assistant message is paired with its exact source Message.
• Decode known OpenRouter options with lenient(...) schemas and type OpenRouterBody with Options.fields + JsonObject so protocol.body.from returns a strongly-typed OpenRouterBody without as OpenRouterBody.

 protocol.body.from(request)
   OpenAIChat.fromRequest(request, {
     cacheControl: cacheControl(),
+    assistant: (source, message) => attachReasoningDetails(source, message),
   })
-  zip body.messages with sourceAssistants[assistantIndex++]
-  return { ...body, messages, ...bodyOptions(...) } as OpenRouterBody
+  bodyOptions(body, request.providerOptions, request.generation?.maxTokens)
+    decodeOptions(openrouter)
+    return { ...options, ...body, ...knownOptions }

Out of scope

• Streaming parser behavior (OpenAIChat.protocol.stream is unchanged).

[opencode-agent]

Thanks, the reasoning-alignment fix works. I ran the updated OpenRouter test against the base and it fails there, with reasoning coming back undefined. With this PR all OpenRouter tests pass, and packages/ai type-checks.

Public API is unchanged; this is internal to the OpenRouter provider body builder.

One regression to fix before merging. Unknown pass-through options in providerOptions used to override the lowered body, and now the body overrides them. OpenAIChat.fromRequest emits keys like temperature even when their value is undefined, so options such as providerOptions: { temperature: 0.1 } are now silently dropped even when the request doesn't set them (details inline).

[opencode-agent]

This flips the precedence between pass-through options and the lowered body. Before, the result was { ...body, messages, ...bodyOptions(...) }, so unknown providerOptions could override body fields. Now ...body comes after ...options. OpenAIChat.fromRequest emits keys such as temperature, top_p and max_tokens even when their value is undefined, so providerOptions: { temperature: 0.1 } now vanishes from the request even though the request never set a temperature. I checked: base sends "temperature": 0.1, and this PR omits it. Spreading ...options after ...body, as before, would keep the old behaviour.

[rekram1-node]

body should win