losing track of and faking execution of simple tasks: release 0.27.0, 0.28.0, and 0.29.0 with gpt-oss and other models

I consistently have this type of issue especially as context usage increases - Codex will repeatedly write plans and claim to have completed the work, while no work was actually performed.

Hey @neurostream this might be a collection of changes that might result in the wrong symptoms.

0.29 fixed the behavior where the CoT wasn't passed back to the model provider resulting in occasional tool calling amnesia. Passing the CoT back in should overall increase the performance but can increase the likelihood of running out of context with providers like Ollama that default to very small context windows.

The Ollama team is aware of this and is trying to improve context window management but one thing you could try that should improve things is setting a higher context window.

For that create a file called Modelfile and add the following content:

FROM gpt-oss:20b
PARAMETER num_ctx 32000

(Or increase beyond 32k. The default is I think 4k or 8k)

Then run

ollama create -f Modelfile gpt-oss:20b-32k

This should give you a new gpt-oss:20b-32k that you can try instead. Hope that helps

Hey @neurostream this might be a collection of changes that might result in the wrong symptoms. 0.29 fixed the behavior where the CoT wasn't passed back to the model provider resulting in occasional tool calling amnesia. Passing the CoT back in should overall increase the performance but can increase the likelihood of running out of context with providers like Ollama that default to very small context windows. The Ollama team is aware of this and is trying to improve context window management but one thing you could try that should improve things is setting a higher context window. For that create a file called Modelfile and add the following content: `` FROM gpt-oss:20b PARAMETER num_ctx 32000 ` (Or increase beyond 32k. The default is I think 4k or 8k) Then run ` ollama create -f Modelfile gpt-oss:20b-32k `` This should give you a new gpt-oss:20b-32k that you can try instead. Hope that helps

btw, i continued the next iteration of the test above with 0.30.0 and it seems to be a lot better.

I'll try to re-cast the model in ollama (i'm using gpt-oss:120b and other models with codex) with the num_ctx adjustment - i just assumed what was in the metadeta about the context size was what it would be using, like:

worker@msm3u1 local % ollama show gpt-oss:120b
  Model
    architecture        gptoss    
    parameters          116.8B    
    context length      131072    
    embedding length    2880      
    quantization        MXFP4     

  Capabilities
    completion    
    tools         
    thinking      

  Parameters
    temperature    1    

  License
    Apache License               
    Version 2.0, January 2004    
    ...                          

in any case, thank you! very insightful

@dkundel-openai things are looking good. closing this issue, thanks agian.

With default run of ollama codex quickly forgets the task, even though ollama show gpt-oss-20b shows context length 131072. The problem is ollama has a hidden additional context limit of 8192 tokens:

https://github.com/ollama/ollama/blob/main/docs/faq.md#how-can-i-specify-the-context-window-size

With default run of ollama codex quickly forgets the task, even though ollama show gpt-oss-20b shows context length 131072. The problem is ollama has a hidden additional context limit of 8192 tokens: https://github.com/ollama/ollama/blob/main/docs/faq.md#how-can-i-specify-the-context-window-size

right... i've recently started to use "ollama create" to set the ctx size for each local model to more deliberate sizes for my inference hardware; and setting the global OLLAMA_CONTEXT_LENGTH for "ollama serve" to the context size of the largest model i use locally (since the default of like 4k or 8k caps the size of all models regardless of the the models trained context size).