中文

问题描述

在 OpenClaw 2026.2.23 中，使用 siliconflow provider 的 Pro/* 模型时，如果会话 thinking 为 off，embedded run 会稳定返回 400 status code (no body)，随后触发 compaction/retry，导致 WebUI 看起来“每发一条就 compact 一次”。

这不是实际上下文溢出：fresh session 下 tokensBefore 很低（例如 121、15）也会触发。

影响模型（已实测）

• Pro/deepseek-ai/DeepSeek-V3.2
• Pro/zai-org/GLM-4.7
• Pro/MiniMaxAI/MiniMax-M2.1
• Pro/moonshotai/Kimi-K2.5

复现步骤

1. 配置 siliconflow provider（api: "openai-completions"）并将主模型设置为任一 Pro/*。
2. 使用默认 thinking off（例如 agents.defaults.thinkingDefault: "off"）。
3. 在 WebUI 或本地 embedded run 发起一次普通对话。
4. 观察 gateway log：embedded run start ... thinking=off 后紧跟 isError=true error=400 status code (no body)，随后 compaction start/retry。

日志特征（精简）

• embedded run start: ... provider=siliconflow model=Pro/... thinking=off
• embedded run agent end: ... isError=true error=400 status code (no body)
• embedded run compaction start ...

API 对照证据（同一 API Key）

对上述 4 个模型均一致：

• 不带 thinking -> 200
• thinking: "off" -> 400（code: 20015, The parameter is invalid. Please check again.）
• thinking: null -> 200

期望行为

当 provider=siliconflow 且模型为 Pro/* 时，thinking=off 不应导致 400；应避免进入误导性的 compaction 重试路径。

建议修复

在 payload 发送前做 provider/model 兼容处理：

• 条件：provider === "siliconflow" && modelId.startsWith("Pro/") && thinkingLevel === "off"
• 行为：将 payload 中 thinking: "off" 归一为 thinking: null

这样可保持“关闭 thinking”的语义，同时兼容 SiliconFlow 参数校验。

English

Description

On OpenClaw 2026.2.23, when using siliconflow Pro/* models, embedded runs fail with 400 status code (no body) whenever thinking is off. The runtime then enters compaction/retry, so WebUI appears to compact on every message.

This is not a real context overflow. Even in fresh sessions with low tokensBefore (e.g. 121, 15), the same loop occurs.

Affected models (verified)

• Pro/deepseek-ai/DeepSeek-V3.2
• Pro/zai-org/GLM-4.7
• Pro/MiniMaxAI/MiniMax-M2.1
• Pro/moonshotai/Kimi-K2.5

Reproduction

1. Configure siliconflow provider (api: "openai-completions") and set any Pro/* model as primary.
2. Keep thinking at off (for example agents.defaults.thinkingDefault: "off").
3. Send a normal message from WebUI (or embedded local run).
4. Observe logs: embedded run start ... thinking=off -> isError=true error=400 status code (no body) -> compaction start/retry.

Log signature (short)

• embedded run start: ... provider=siliconflow model=Pro/... thinking=off
• embedded run agent end: ... isError=true error=400 status code (no body)
• embedded run compaction start ...

API comparison evidence (same API key)

For all four models above:

• Without thinking -> 200
• With thinking: "off" -> 400 (code: 20015, The parameter is invalid. Please check again.)
• With thinking: null -> 200

Expected behavior

For provider=siliconflow and Pro/* models, thinking=off should not cause HTTP 400 and should not trigger compaction retry flow.

Proposed fix

Add a provider/model compatibility shim before payload send:

• Condition: provider === "siliconflow" && modelId.startsWith("Pro/") && thinkingLevel === "off"
• Behavior: normalize payload thinking: "off" to thinking: null

This preserves the "thinking disabled" intent while matching SiliconFlow's payload validation behavior.