openclaw - 💡(How to fix) Fix openai-responses API: reasoning items break history replay with 400 error [1 participants]

Bug Description

When using the openai-responses API with reasoning: true on a model (Azure OpenAI GPT-5.3), the first message works correctly and generates reasoning tokens. However, every subsequent message in the same session fails with:

400 Item 'rs_XXXX' of type 'reasoning' was provided without its required following item

The Azure OpenAI Responses API rejects the request because the conversation history contains a reasoning item without its paired output item.

Steps to Reproduce

1. Configure a provider with "api": "openai-responses" and a reasoning-capable model:

{
  "models": {
    "providers": {
      "azure-openai-responses": {
        "baseUrl": "https://<resource>.cognitiveservices.azure.com/openai/v1",
        "api": "openai-responses",
        "authHeader": false,
        "headers": { "api-key": "<key>" },
        "models": [{
          "id": "gpt-5.3-chat",
          "reasoning": true,
          "compat": { "supportsStore": false }
        }]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": { "primary": "azure-openai-responses/gpt-5.3-chat" },
      "thinkingDefault": "xhigh"
    }
  }
}

1. Send any message to the agent → first response works, includes reasoning tokens
2. Send a follow-up message → 400 error every time

Expected Behavior

Follow-up messages should work. OpenClaw should either:

• Strip reasoning items from the input array when replaying history (since they're internal model artifacts, not user-facing content)
• Or ensure reasoning items are always paired with their following output item when constructing the history replay

Actual Behavior

OpenClaw stores the full model response (including reasoning items with rs_* IDs) in the session JSONL file. On the next turn, it replays these items in the input array. The Responses API rejects this because:

• The reasoning item is present without its required following item
• Or the id fields from the original response are being sent back (known to cause issues per OpenAI docs)

Environment

• OpenClaw version: 2026.3.31
• Provider: Azure AI Foundry (cognitiveservices.azure.com)
• API: openai-responses
• Model: gpt-5.3-chat with reasoning: true
• OS: Ubuntu 24.04 (Hetzner VPS)
• Compaction mode: safeguard

What I've Tried

Suggested Fix

When constructing the input array for the Responses API from session history:

1. Strip all reasoning type items from the input (they're model-internal, not meant to be replayed)
2. Or strip id fields from replayed items (per OpenAI guidance)
3. Or add a compat.stripReasoning: true flag to control this behavior

Workaround

Setting "reasoning": false on the model prevents reasoning tokens from being generated, avoiding the issue. The model still works but without extended thinking capabilities.

Alternatively, using "api": "openai-completions" avoids the issue but doesn't pass reasoning_effort to the Azure endpoint, so reasoning doesn't activate.