You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[MiniMax M2] Agent stop at the middle of the work #4112
The title/header of the right-hand side panel, which appears to contain session context, metadata, and task lists, is displaying an outdated or generic title.
Problem: The header currently reads "New session - 2025-11-07T07:33:05.124Z".
Context: The title should be updated to reflect the actual status or the user's current task within the agent's environment.
2. Agent Stops in the Middle of the Work (Incomplete Execution)
The interaction flow suggests the AI agent (MiniMax) failed to complete its intended action after a code modification step and is now waiting for user input, indicating a premature halt or failure in the execution pipeline.
This issue might be a duplicate of existing issues. Please check:
Hanging During Agent Tasks #4073: Both issues describe agents stopping/hanging during execution and not completing their intended work. Issue Hanging During Agent Tasks #4073 specifically mentions agents hanging during tasks with ESC becoming unresponsive, which seems very similar to your MiniMax M2 agent stopping in the middle of work.
Feel free to ignore if none of these address your specific case.
@lst97 do u not have any openrouter credits? When u select a model we use a tiny cheap one for title gen that your provider has if u are using openrouter this would be haiku 4.5
For the Issue 2, it may not be the issue with opencode.
According to the official announcement: After 24:00 UTC on November 7, the RPM for free users will be reduced. If you are running high-concurrency task, we recommend switching to the MiniMax-M2-Stable in advance or adjusting your task concurrency to ensure uninterrupted execution.
However it's better to display a message if it's the case.
I'm currently on the new MiniMax Coding Plan and it keeps happening all the time. Execution stops and I need to say something like 'resume' to get it going again.
@bennyzen if the model stops working, that's a model / prompting thing typically unless u are seeing an error? Otherwise the model is saying it is done
changed the title [-][MiniMax M2 via OpenRouter] Incorrect side panel title & agent stop at the middle of the work[/-][+][MiniMax M2 via OpenRouter] Agent stop at the middle of the work[/+]on Nov 12, 2025
I’m encountering the same issue with Gemini 3 Pro (from Google APIs directly). #4953
I think it’s more likely an issue within Opencode and not much to do with the specific model.
This issue has more traction so I figured we’d consolidate here. Is there anything in particular I can do to debug?
@0dragosh it has to do with interleavened thinking models, but for anything regarding OpenRouter then it is an issue w/ their API or their sdk. I have been talking with them about several issues and they are still addressing some.
As for minimax outside of openrouter, that's separate and ill have to make an account to check it out
@0dragosh it has to do with interleavened thinking models, but for anything regarding OpenRouter then it is an issue w/ their API or their sdk. I have been talking with them about several issues and they are still addressing some.
As for minimax outside of openrouter, that's separate and ill have to make an account to check it out
Appreciate the insight! For me it happens with Gemini 3 Pro from Google APIs directly (not OpenRouter).
Can anyone send me example sessions of this happening? I have a minimax coding plan setup. Trying to get this stop behavior mentioned but minimax through the minimax provider seems to work great?
Maybe if anyone can replicate the issue they are describing, try sending me your session: opencode export > session.json
OR send me steps to replicate this issue?
Also please tell me:
are you on minimax or minimax china
did you set the provider in your opencode.json or are you using the default (using the default works for me you shouldn't need custom config here)
Okay I managed to replicate, setup a proxy to validate and in fact the minimax provider IS telling us the response is done. I still don't think this is an opencode bug, we are doing exactly what their api / model is saying...
Okay they said they fixed it, doing some testing myself to see if the issue persists. If I cannot replicate I will be closing this but please comment if you continue having issues & we will reopen and continue discussion w/ their team.
Activity
github-actions commentedon Nov 9, 2025
This issue might be a duplicate of existing issues. Please check:
Feel free to ignore if none of these address your specific case.
rekram1-node commentedon Nov 9, 2025
@lst97 do u not have any openrouter credits? When u select a model we use a tiny cheap one for title gen that your provider has if u are using openrouter this would be haiku 4.5
seems like it is failing for u tho
lst97 commentedon Nov 9, 2025
@rekram1-node Ah i see, thank you for your response :)
rekram1-node commentedon Nov 9, 2025
@lst97 for now id recommend setting “small_model” in your global opencode.json
choose any model from the following output: “opencode models”
ofc choose a model u have access to, it can even be a free one like:
opencode/big-pickle
lst97 commentedon Nov 10, 2025
For the Issue 2, it may not be the issue with opencode.
According to the official announcement: After 24:00 UTC on November 7, the RPM for free users will be reduced. If you are running high-concurrency task, we recommend switching to the MiniMax-M2-Stable in advance or adjusting your task concurrency to ensure uninterrupted execution.
However it's better to display a message if it's the case.
ref: https://x.com/MiniMax__AI/status/1986815058249408541
bennyzen commentedon Nov 10, 2025
I'm currently on the new MiniMax Coding Plan and it keeps happening all the time. Execution stops and I need to say something like 'resume' to get it going again.
bennyzen commentedon Nov 11, 2025
Needless to say, but in its current state v1.0.55, opencode is almost unusable with MiniMax M2. it halts after nearly every turn.
rekram1-node commentedon Nov 12, 2025
@bennyzen if the model stops working, that's a model / prompting thing typically unless u are seeing an error? Otherwise the model is saying it is done
lst97 commentedon Nov 12, 2025
MiniMax-AI/MiniMax-M2#37
[-][MiniMax M2 via OpenRouter] Incorrect side panel title & agent stop at the middle of the work[/-][+][MiniMax M2 via OpenRouter] Agent stop at the middle of the work[/+]bennyzen commentedon Nov 12, 2025
@rekram1-node not that sure about that, as other coding agents don't show that behavior with MiniMax M2.
8 remaining items
legout commentedon Dec 3, 2025
Same here. Hope this could be fixed soon.
Thanks!
0dragosh commentedon Dec 4, 2025
I’m encountering the same issue with Gemini 3 Pro (from Google APIs directly). #4953
I think it’s more likely an issue within Opencode and not much to do with the specific model.
This issue has more traction so I figured we’d consolidate here. Is there anything in particular I can do to debug?
rekram1-node commentedon Dec 4, 2025
@0dragosh it has to do with interleavened thinking models, but for anything regarding OpenRouter then it is an issue w/ their API or their sdk. I have been talking with them about several issues and they are still addressing some.
As for minimax outside of openrouter, that's separate and ill have to make an account to check it out
0dragosh commentedon Dec 4, 2025
Appreciate the insight! For me it happens with Gemini 3 Pro from Google APIs directly (not OpenRouter).
hicdercpl commentedon Dec 4, 2025
I think this PR #4933 seems to fix this issue, but not sure why it's closed.
rekram1-node commentedon Dec 10, 2025
Can anyone send me example sessions of this happening? I have a minimax coding plan setup. Trying to get this stop behavior mentioned but minimax through the minimax provider seems to work great?
Maybe if anyone can replicate the issue they are describing, try sending me your session:
opencode export > session.jsonOR send me steps to replicate this issue?
Also please tell me:
rekram1-node commentedon Dec 10, 2025
Okay I managed to replicate, setup a proxy to validate and in fact the minimax provider IS telling us the response is done. I still don't think this is an opencode bug, we are doing exactly what their api / model is saying...
Artifacts for evidence:
Share link of the conversation (note the model says it's about to do something, then stops after a read):
https://docs.dev.opencode.ai/s/djgvEIGD
Here is a gist containing a conversation where the model stops (both request & response) as well as a basic conversation where I just send "hello"
In both cases their api is responding with the same stop reason, it seems like they aren't correctly doing a tool-calls stop reason?
Stopped randomly excerpt:
Hello excerpt:
Here is my gist w/ full payloads for each:
https://gist.github.com/rekram1-node/96907df65c6385d534151f6d35ae8fcd
rekram1-node commentedon Dec 10, 2025
Okay I presented it to their team and in fact the issue is on their end, they said they will be addressing it soon.
So not an opencode bug
rekram1-node commentedon Dec 11, 2025
Okay they said they fixed it, doing some testing myself to see if the issue persists. If I cannot replicate I will be closing this but please comment if you continue having issues & we will reopen and continue discussion w/ their team.
rekram1-node commentedon Dec 11, 2025
Tested quite a bit, couldn't get it to stop randomly!
Now anyone who is using Minimax through other providers, lmk if you hit snags but it should function properly through openrouter & minimax providers
legout commentedon Dec 11, 2025
You rock!! Thanks
Capevace commentedon Feb 23, 2026
have this same issue with Kimi K2.5 and GLM 5.
Looks like it fails during tool calls?
From memory I haven't had this issue before updating to OpenCode Desktop 1.2 (currently 1.2.10)
JoshZA commentedon Feb 27, 2026
Getting this on minimax2.5 coding plan too. Opencode cli just seems to hang randonly, no timeouts or anything
khaled4vokalz commentedon Mar 5, 2026
I'm getting this with GLM 5 as well... I need to make it alive again using some dummy prompts like
.in this image e.g.