openclaw - 💡(How to fix) Fix MiniMax-M2.7-highspeed returns no text / incomplete terminal response in OpenClaw 2026.5.7

Bug type

Regression / provider compatibility

Summary

MiniMax-M2.7-highspeed works with the MiniMax API directly, but OpenClaw 2026.5.7 fails to surface any assistant text. The local inference path reports No text output returned, and the gateway path reports incomplete terminal response.

This looks like a MiniMax M2.7 response parsing / terminal response compatibility issue rather than an auth, quota, or network problem.

Steps to reproduce

1. Configure MiniMax API-key auth for minimax/MiniMax-M2.7-highspeed.
2. Verify the model is present / allowed:

openclaw models status --json | jq '{defaultModel, fallbacks, aliases}'

Relevant output:

{
  "fallbacks": [
    "deepseek/deepseek-v4-flash",
    "deepseek/deepseek-v4-pro",
    "minimax/MiniMax-M2.7-highspeed"
  ],
  "aliases": {
    "minimax-m2.7-highspeed": "minimax/MiniMax-M2.7-highspeed"
  }
}

1. Run local inference:

openclaw infer model run --local --json \
  --model minimax/MiniMax-M2.7-highspeed \
  --prompt 'Reply exactly: OK_OPENCLAW_REPRO_42'

1. Run gateway inference:

openclaw infer model run --gateway --json \
  --model minimax/MiniMax-M2.7-highspeed \
  --prompt 'Reply exactly: OK_OPENCLAW_GATEWAY_REPRO_42'

Expected behavior

OpenClaw should return the visible assistant text from MiniMax, e.g. OK_OPENCLAW_REPRO_42.

Actual behavior

Local inference returns:

Error: No text output returned for provider "minimax" model "MiniMax-M2.7-highspeed".

Gateway inference returns:

GatewayClientRequestError: FailoverError: minimax/MiniMax-M2.7-highspeed ended with an incomplete terminal response

Direct API sanity check

Calling MiniMax directly with the same API key succeeds with HTTP 200 and returns a valid choices[0].message.content. The content may include a <think>...</think> section before the visible final answer, for example with enough max_tokens the visible answer is returned after the reasoning section.

This suggests OpenClaw may be treating MiniMax's reasoning/thinking output or terminal event sequence as incomplete, or may be dropping the final text when parsing the provider response.

Environment

• OpenClaw: 2026.5.7 (eeef486)
• Install method: npm
• OS: macOS 26.4.1, arm64
• Node: v26.0.0
• Model: minimax/MiniMax-M2.7-highspeed
• API path: MiniMax API-key setup, OpenAI-compatible provider config

Notes

I also checked related MiniMax issues:

• #50234 added/requested M2.7 support and is closed.
• #34487 is similar in that direct MiniMax API worked while OpenClaw failed, but that issue was a timeout on an older version.
• #21872 was about MiniMax streaming behavior, which may be adjacent, but this repro fails even through the current model run paths.

No API key or private config is included here.