Continue Edit mode writes reasoning text into file with gpt-oss-120b
In Continue 1.2.11 (VS Code 1.106.3, Windows), using Edit mode with gpt-oss-120b (vLLM, LiteLLM proxy) replaces the selected code with the model's reasoning/chain-of-thought text instead of applying the edit. Qwen Coder 30B works correctly on the same setup, and downgrading to Continue 1.2.8 fixes it. Why does this happen and is there a workaround?
Originally asked on GitHub: https://github.com/continuedev/continue/issues/8990
6 answers
Status: the GitHub issue was closed as NOT_PLANNED, but later comments report the problem still occurs in Continue 1.3.32, on Linux Fedora, and with local Ollama models, so it is not fixed upstream.
The workaround that worked for the reporter is to downgrade Continue to 1.2.8; two people confirmed 1.2.8 works with the same configuration while newer versions (1.2.11, 1.3.32) do not. One user also fixed it by disabling the reasoning option in the model configuration, but that does not help for gpt-oss-120b: that model only lets you set reasoning effort to low/medium/high and cannot fully disable reasoning.
There is no official patch or config change from the maintainers in this thread. The same symptom is reported for other reasoning models (DeepSeek-R1, Qwen-family, fast-apply models), which suggests reasoning output is not stripped from the Edit-mode response. If you cannot downgrade, try a model where reasoning can be fully disabled, or use a non-reasoning model for Edit mode.
Did this work for you?What @johannesrmer tried on GitHub (Worked)
Downgraded Continue from 1.2.11 to 1.2.8 on the same VS Code / Windows setup with gpt-oss-120b. The reporter confirmed Edit mode worked again.
Source: https://github.com/continuedev/continue/issues/8990 (opens external site)
Did this work for you?What @Z2IRIM tried on GitHub (Worked)
Disabled the reasoning option in the model configuration (shown in a screenshot). This fixed Edit mode for that user.
Source: https://github.com/continuedev/continue/issues/8990#issuecomment-4098331867 (opens external site)
Did this work for you?What @VeraKowalczuk tried on GitHub (Didn't work)
For gpt-oss-120B, the disable-reasoning approach cannot be used because the model only supports reasoning effort levels (low/medium/high), not fully disabling reasoning.
Source: https://github.com/continuedev/continue/issues/8990#issuecomment-4119487113 (opens external site)
Did this work for you?What @MaluNoPeleke tried on GitHub (Didn't work)
Another user tried downgrading but still encountered a related problem (issue #9065).
Source: https://github.com/continuedev/continue/issues/8990#issuecomment-3627111504 (opens external site)
Did this work for you?What @dashko tried on GitHub (Didn't work)
Reported the issue still exists in Continue 1.3.32 with local Ollama models, while 1.2.8 does not have the problem.
Source: https://github.com/continuedev/continue/issues/8990#issuecomment-3919219348 (opens external site)
Did this work for you?
Related questions
- Continue error with Llama via LiteLLM: Cannot set properties of undefined (setting 'parsedArgs')
- Continue 1.1.45-1.1.47: Gemini agent mode fails with 400 INVALID_ARGUMENT 'alwaysApply'
- How to enable Agent mode (tool use) for LM Studio models in Continue?
- How to make Continue's @folder/@codebase send full file contents?
- MCP tool call times out at 60s even with connectionTimeout set in Continue