Bug type

Behavior bug (incorrect outbound payload)

Beta release blocker

No

Summary

When using Gemma 4 models (google/gemma-4-26b-a4b, google/gemma-4-31b, etc.) via an OpenAI-compatible endpoint such as LM Studio, OpenClaw re-sends the model's prior-turn reasoning_content on subsequent turns. Google's Gemma 4 documentation explicitly states this should NOT be done:

"You must remove (strip) the model's generated thoughts from the previous turn before passing the conversation history back to the model for the next turn." — Gemma 4 Thinking Mode docs

The exception (function-calling within a single turn, where thoughts must be preserved between tool calls in the same turn) is correctly handled for Anthropic's signed-thinking contract, but it is not the right default for Gemma on OpenAI-compatible transports.

Steps to reproduce

1. Configure LM Studio (or any llama.cpp / vLLM / OpenAI-compat endpoint) serving a Gemma 4 model with thinking enabled at the chat-template level.
2. In openclaw.json, add a provider entry pointing at that endpoint:

   "lmstudio": {
     "baseUrl": "http://<host>:1234/v1",
     "apiKey": "lmstudio",
     "api": "openai-responses",
     "models": [
       { "id": "google/gemma-4-26b-a4b", "name": "Gemma 4 26B A4B",
         "contextWindow": 256000, "maxTokens": 4096,
         "reasoning": true, "input": ["text", "image"],
         "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0} }
     ]
   }

3. Wire the model into an agent and have a multi-turn conversation without tool calls.
4. Capture the second-turn outbound request body (e.g. LM Studio's request log, or an mitm proxy).

Expected behavior

The assistant message from turn 1 replayed in turn 2's history should contain only the visible content, not the prior reasoning, for Gemma-family models on OpenAI-compat transports.

Actual behavior

The outbound assistant message in history includes reasoning_content: "<previous thinking>". Every subsequent turn accumulates more historical thinking in the payload.

Code reference

The re-attachment happens in the bundled @mariozechner/pi-ai library:

node_modules/@mariozechner/pi-ai/dist/providers/openai-completions.js — convertMessages, approximately:

const thinkingBlocks = msg.content.filter((b) => b.type === "thinking");
const nonEmptyThinkingBlocks = thinkingBlocks.filter((b) => b.thinking && b.thinking.trim().length > 0);
if (nonEmptyThinkingBlocks.length > 0) {
    ...
    const signature = nonEmptyThinkingBlocks[0].thinkingSignature;  // "reasoning_content" for Gemma/llama.cpp
    if (signature && signature.length > 0) {
        assistantMsg[signature] = nonEmptyThinkingBlocks.map((b) => b.thinking).join("\n");
    }
}

The signature round-trip is correct for providers that require reasoning-in-history (Anthropic signed thinking, some gpt-oss models). It is incorrect for Gemma 4, which explicitly forbids it.

Suggested fix

Provider-aware stripping: when the target model is a Gemma 4 family model (id prefix gemma-4, or with a flag on the model definition such as stripReasoningInHistory: true), do not re-attach the thinking block to history. Preserving thought blocks between tool calls within a single turn should still be allowed, because Google's guidance makes that explicit exception.

A conservative version would add a model-level opt-in (stripReasoningInHistory) that users can set in models.providers.<p>.models[].stripReasoningInHistory = true for Gemma. A more complete version would default it on based on model-id pattern match.

Impact

• Context bloat that grows every turn (Gemma 4 thought channels routinely run 500-2000 tokens).
• Violates Gemma's documented contract; may subtly degrade multi-turn quality as the model re-sees its own stale reasoning.
• Affects anyone running Gemma 4 via LM Studio, llama.cpp server, Ollama, vLLM, or any OpenAI-compat gateway.

OpenClaw version

2026.4.15 (041266a)

Operating system

macOS 15.x (Darwin 25.3.0)

Install method

curl -fsSL https://openclaw.ai/install.sh | bash

Model

google/gemma-4-26b-a4b (also reproducible on gemma-4-31b)

Provider / routing chain

openclaw → lmstudio provider (openai-responses api) → LM Studio server → Gemma 4

Additional notes

• Google's official guidance: https://ai.google.dev/gemma/docs/capabilities/thinking — "Pass back the conversation history... without the thought channel content"
• Related but distinct: #65533 (MiniMax, which actually needs reasoning preserved — inverse case). This highlights that the right behavior is provider-specific.
• Related: #61995 (MiniMax thinking suppression), #62127/#62411 (Gemma 4 native Google provider thinking-off semantics).