openclaw - 💡(How to fix) Fix [Bug] MiniMax-M2.7 streaming not working - content field contains thinking tags mixed with response [3 comments, 2 participants]

Bug Description

Problem: Streaming output does not work with minimax-cn/MiniMax-M2.7 model in Feishu channel. The response appears all at once instead of streaming token by token.

Expected: Streaming should work the same as GLM-5 (dashscope provider) which streams correctly.

Actual: minimax-cn/MiniMax-M2.7 shows no streaming, waits for complete response before displaying.

Environment

• OpenClaw version: 2026.4.1 (da64a97)
• Channel: Feishu
• Model: minimax-cn/MiniMax-M2.7

Root Cause Analysis

Compared SSE responses between GLM-5 (working) and MiniMax-M2.7 (broken):

GLM-5 (dashscope) - Streaming works correctly ✅

data: {"choices":[{"delta":{"content":null,"reasoning_content":"Let","role":"assistant"}}]}
data: {"choices":[{"delta":{"content":null,"reasoning_content":" me think about...","role":"assistant"}}]}

• delta.content is null
• Thinking is in a separate reasoning_content field
• Clean separation between thinking and response

MiniMax-M2.7 (minimax-cn) - Streaming broken ❌

data: {"choices":[{"delta":{"content":"(think)\nThe user said hi, we should respond...","role":"assistant","name":"MiniMax AI","audio_content":""}}]}
data: {"choices":[{"delta":{"content":" Okay, greeting back...","role":"assistant","name":"MiniMax AI","audio_content":""}}]}

• delta.content contains BOTH thinking tags and actual response text mixed together
• Content example: (think)\nThe user says "hi". We need to respond.
• Additional fields present: name, audio_content, input_sensitive, output_sensitive, etc.

Hypothesis

OpenClaw's SSE streaming parser likely:

1. Reads delta.content and displays it when non-null
2. Does not understand the (think) thinking tags embedded within the content string
3. Treats the entire (think)\n... sequence as plain text, causing incorrect streaming display behavior

The thinking tags in the content field confuse the streaming display logic.

Workaround

Use dashscope/MiniMax-M2.5 (alias: MiniMax-Dash) instead, which does not have this issue and streams correctly.

Config Changes Tried

1. Changed minimax-cn baseUrl from /anthropic to /v1
2. Changed minimax-cn api from anthropic-messages to openai-completions
3. Added no_proxy for api.minimaxi.com

None of these fixed the streaming issue. The root cause is in how OpenClaw parses the SSE content field when it contains mixed thinking tags.

Expected Behavior

MiniMax-M2.7 streaming should work like GLM-5, with tokens appearing progressively in the chat interface.

Reporter: 骆伟雄 (ou_97861c3cd1c00cff9d8002f2e68bdb94) Reported via: OpenClaw ops agent