What happened?
Hello, I have made following request to get some information: "A. Latest up to date documentation and B. The actual state of the current installation"
I have used DeepSeek V4 Flash High to get the information.
Here is the DS response:
Here's the summary in English for the supporter:
Issue: MiniMax M3 thinking content leaks into normal chat response instead of rendering in Kilo's separate thinking/reasoning UI block.
Current config (~/.config/kilo/kilo.jsonc):
- Custom provider
gmi-cloud via @ai-sdk/openai-compatible at https://api.gmi-serving.com/v1
- MiniMax-M3 configured as a custom model with reasoning enabled
- Variant
adaptive-thinking uses "enable_thinking": true and "reasoning_split": true
Config details for the model:
Observed behavior:
- MiniMax's reasoning/thinking appears as plain text inline in the response, not parsed into the collapsible thinking block (as happens with Claude or DeepSeek)
- The model config UI's "Thinking type" dropdown only offers
enabled/disabled/(not set) — MiniMax API rejects enabled with a 400 error and requires adaptive. Workaround is manual config editing.
- Even with
thinking: { "type": "adaptive" } set manually, reasoning content still isn't separated into the thinking block.
Matches known bugs:
Environment:
- Kilo CLI + VS Code extension on Windows
- Config sources:
~/.config/kilo/kilo.json (MCP servers only), ~/.config/kilo/kilo.jsonc (providers, agents, permissions, models)
- MiniMax accessed through custom GMI-Cloud provider, not the built-in MiniMax provider
<@&1349299159445798952>
Area
Mayor / Chat
Town ID (if known)
No response
Steps to reproduce
Simple steps to reproduce.
I have set up GMI Cloud as Provider.
I have set MiniMax M3 as reasoning - yes ; name - adaptive thinking; activate reasoning - yes ; style of reasoning f.e. z.ai - adaptive; split reasoning - yes ; reasoning effort - medium ; output effort - not set; thinking over chat presets - true
I have sent a message where I tried to enforce thinking for the model. I have evaluated it several times.
Reasoning / Thinking is in the content itself in .... parameters but not special sent in a reasoning tag. It is sent in the context tag, as far as I could debug it with chatGPT, then I got bored.
Relevant logs or screenshots
No errors message available :/. Just Feature bug.
What happened?
Hello, I have made following request to get some information: "A. Latest up to date documentation and B. The actual state of the current installation"
I have used DeepSeek V4 Flash High to get the information.
Here is the DS response:
Here's the summary in English for the supporter:
Issue: MiniMax M3 thinking content leaks into normal chat response instead of rendering in Kilo's separate thinking/reasoning UI block.
Current config (
~/.config/kilo/kilo.jsonc):gmi-cloudvia@ai-sdk/openai-compatibleathttps://api.gmi-serving.com/v1adaptive-thinkinguses"enable_thinking": trueand"reasoning_split": trueConfig details for the model:
Observed behavior:
enabled/disabled/(not set)— MiniMax API rejectsenabledwith a 400 error and requiresadaptive. Workaround is manual config editing.thinking: { "type": "adaptive" }set manually, reasoning content still isn't separated into the thinking block.Matches known bugs:
Environment:
~/.config/kilo/kilo.json(MCP servers only),~/.config/kilo/kilo.jsonc(providers, agents, permissions, models)<@&1349299159445798952>
Area
Mayor / Chat
Town ID (if known)
No response
Steps to reproduce
Simple steps to reproduce.
I have set up GMI Cloud as Provider.
I have set MiniMax M3 as reasoning - yes ; name - adaptive thinking; activate reasoning - yes ; style of reasoning f.e. z.ai - adaptive; split reasoning - yes ; reasoning effort - medium ; output effort - not set; thinking over chat presets - true
I have sent a message where I tried to enforce thinking for the model. I have evaluated it several times.
Reasoning / Thinking is in the content itself in .... parameters but not special sent in a reasoning tag. It is sent in the context tag, as far as I could debug it with chatGPT, then I got bored.
Relevant logs or screenshots
No errors message available :/. Just Feature bug.