Apply corruption in vscode

[chrisdrumz]

I can't for the life of me get apply to work reliably in medium to large files. Agent or manual apply. Using latest release or pre release in Win 11. Tried with gpt-5.2, and Gemini 3.1 pro through openAI compat. Api through Open WebUI.

All openwebui reverse proxy related settings have been implemented.

Typically lazy blocks or occasionally conversational text gets added to my file, and everything else gets removed... Often VScode hangs during apply.

Is fast apply supposed to work for lazy blocked suggestions in ~2000 line files? It seems to fail, and then llm fallback ends up trying to replace my entire file with just the llm suggestion.

Doesn't seem to matter if the the function has lazy blocks or not.

I system prompted the apply model to output a diff patch but it came through as a thousand code blocks, one for each line of code.. Even with streaming off..

I figured out that adding a system prompt to the apply model forces continue to use chat api VS completion api. But still broken applies.

I tried adding "gpt" to model name to force a different apply path.

I've tried with and without streaming.

I did notice I'm getting "file does not exist in ast tracker" errors in VScode console, but I don't think is related to the fast apply tree sitter stuff.

Any tips or ideas would help. Maybe I need to configure mitmproxy to see the raw data...

[devMiikaK]

Hey, I had a similar issue but with local models. I'm positive that both thinking/reasoning and model context length are the primary cause. Sometimes reasoning tags like would "leak" into the code snippets and cause errors too.

Bumping up the context on the apply model and switching to a non-reasoning apply model fixed this issue for me.

For what it's worth, here's the config.yaml I'm using:

name: llamacpp
version: 0.0.1
schema: v1

models:
  - name: Qwen3.5
    provider: openai
    apiBase: http://localhost:8080/v1
    model: qwen3.5-27b
    roles:
      - chat
      - edit
    capabilities:
      - tool_use
    defaultCompletionOptions:
      temperature: 0.2
      topP: 0.95
      topK: 20
      maxTokens: 32768
      contextLength: 131072
    chatOptions:
      basePlanSystemMessage: >
        You are an expert at programming.
        
        # Coding guidelines
        Try to avoid adding redundant helpers or functions if possible.
        Efficiency is the key - Ensure code is optimized.

        # Bash commands
        If suggesting bash commands, ensure they are linux style commands (not powershell), unless specifically requested.
        Prefer single-row commands, instead of multi-row commands seperated with \.
        If downloading packages / dependencies with links and they do not work, ask the user to download them instead.

        # Chat
        If user's request is ambigious or unclear in any way, ask user for clarification before proceeding.
        Do not include unnecessary conversation filler.

  # apply model
  - name: Qwen2.5
    provider: openai
    apiBase: http://localhost:8081/v1
    model: qwen3.5-4b
    roles:
      - apply
    defaultCompletionOptions:
      temperature: 0
      contextLength: 65536
      maxTokens: 32768
    promptTemplates:
      apply: |
        Original:
        {{{original_code}}}

        New:
        {{{new_code}}}

        Return only the complete new file contents.
        No markdown fences.
        No explanations.
        
context:
  - provider: file
  - provider: code
  - provider: currentFile
  - provider: diff
  - provider: tree
  - provider: search
    params:
      maxResults: 50
  - provider: repo-map
    params:
      includeSignatures: false
  - provider: codebase

[jdcfd]

I had the same solution. You need to use a fast, non-reasoning model for “apply” for it to work well. It is a very weird way of editing files imo, but it works.