Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

requiresReasoningContentOnAssistantMessages is used at runtime in detectCompat() / getCompat() to control whether reasoning_content: "" is injected on assistant messages (required by DeepSeek's API). However, this field is not in the Zod ModelCompatSchema, so setting it in user config triggers a .strict() validation error on gateway startup.

This makes it impossible to fully replicate native DeepSeek provider behavior on a custom OpenAI-compatible proxy.

Steps to reproduce

Configure a custom provider pointing to a DeepSeek-compatible proxy:

{
  "models": {
    "providers": {
      "my-proxy": {
        "baseUrl": "https://my-proxy.example.com/v1",
        "models": [
          {
            "id": "deepseek-v4-pro",
            "reasoning": true,
            "compat": {
              "thinkingFormat": "deepseek",
              "requiresReasoningContentOnAssistantMessages": true
            }
          }
        ]
      }
    }
  }
}

Expected: Config accepted, runtime behavior matches native deepseek provider.

Actual: Gateway rejects with Unrecognized key(s) in object: 'requiresReasoningContentOnAssistantMessages' because ModelCompatSchema is .strict() and doesn't include this field.

Setting thinkingFormat: "deepseek" works (it IS in the schema) and correctly triggers the thinking block + reasoning_effort. But the requiresReasoningContentOnAssistantMessages flag, which controls this behavior in convertMessages(), cannot be toggled by users:

// openai-completions.ts

if (compat.requiresReasoningContentOnAssistantMessages && model.reasoning
    && assistantMsg.reasoning_content === undefined) {
    assistantMsg.reasoning_content = "";
}

Currently this flag is only true when detectCompat sees provider === "deepseek" || baseUrl.includes("deepseek.com"). Custom providers never match.

Root Cause

ModelCompatSchema in zod-schema.core.ts doesn't include requiresReasoningContentOnAssistantMessages. The .strict() modifier rejects any unrecognized key.

Suggested Fix

 const ModelCompatSchema = object({
     // ... existing fields ...
+    requiresReasoningContentOnAssistantMessages: boolean().optional(),
 }).strict().optional();

Additional missing fields (minor)

While here, these compat fields also exist in getCompat() but are missing from the schema:

cacheControlFormat (string | undefined)
sendSessionAffinityHeaders (boolean)
supportsLongCacheRetention (boolean)
zaiToolStream (boolean)
openRouterRouting (object)
vercelGatewayRouting (object)

Expected behavior

Configuring a custom OpenAI-compatible provider with:
"compat": { "requiresReasoningContentOnAssistantMessages": true }
should be accepted and cause assistant messages to include reasoning_content: ""

Actual behavior

Gateway startup fails with:
Unrecognized key(s) in object: 'requiresReasoningContentOnAssistantMessages'
because ModelCompatSchema is .strict() and does not include this field.

OpenClaw version

2026.5.28

Operating system

Linux (Ubuntu)

Install method

npm global

Model

deepseek-v4-pro

Provider / routing chain

Custom provider (e.g. Tencent TokenHub / tencentmaas-deepseek) → DeepSeek-compatible API

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response