Continue .continue/.config.yaml

.config.yaml

# ~/.continue/config.yaml
#
# Important Continue behavior (see continuedev/continue#9379):
# When the Agent chat model and the Apply model are the SAME entry,
# Continue often sends an empty Apply prompt and writes the model's
# free-form output (thinking / narration) into the target source file.
# Keep chat/edit and apply on SEPARATE model entries.

name: Local Config
version: 1.0.1
schema: v1

models:
  # --- SiliconFlow LLMs ---
  
  # Основной агентный чат и рефакторинг (GLM-5.3 с глубоким рассуждением)
  - name: "GLM-5.3 (Max Reasoning)"
    provider: "siliconflow"
    model: "zai-org/GLM-5.3"
    apiBase: "https://api.siliconflow.com/v1"
    apiKey: "${{ secrets.SILICONFLOW_API_KEY }}"
    roles:
      - chat
      - edit
    capabilities:
      - tool_use
    requestOptions:
      extraBodyProperties:
        reasoning_effort: "max"

  # Более дешевая GLM-5.3 (high)
  - name: "GLM-5.3 (High Reasoning)"
    provider: "siliconflow"
    model: "zai-org/GLM-5.3"
    apiBase: "https://api.siliconflow.com/v1"
    apiKey: "${{ secrets.SILICONFLOW_API_KEY }}"
    roles:
      - chat
      - edit
    capabilities:
      - tool_use
    requestOptions:
      extraBodyProperties:
        reasoning_effort: "high"

  # Облегченная версия для быстрых вопросов и UI-автоматизации
  - name: "GLM-5.3 Flash (Chat only)"
    provider: openai
    model: "zai-org/GLM-5.3-Flash"
    apiBase: "https://api.siliconflow.com/v1"
    apiKey: "${{ secrets.SILICONFLOW_API_KEY }}"
    roles:
      - chat

  # Основной архитектор для генерации кода по Open Spec (Тяжелый DeepSeek V4 Pro)
  - name: "DeepSeek V4 Pro (Think Max)"
    provider: openai
    model: "deepseek-ai/DeepSeek-V4-Pro-0813" # Самый актуальный релиз с улучшенными агентами
    apiBase: "https://api.siliconflow.com/v1"
    apiKey: "${{ secrets.SILICONFLOW_API_KEY }}"
    roles:
      - chat
      - edit
    capabilities:
      - tool_use
    requestOptions:
      temperature: 0.1 # Крайне важно для SDD: снижает галлюцинации и гарантирует следование типам
      extraBodyProperties:
        reasoning_effort: "max" # Активирует режим максимальных рассуждений DeepSeek

  # Сверхбыстрая версия для рядовых вопросов (DeepSeek V4 Flash)
  - name: "DeepSeek V4 Flash (Chat only)"
    provider: openai
    model: "deepseek-ai/DeepSeek-V4-Flash"
    apiBase: "https://api.siliconflow.com/v1"
    apiKey: "${{ secrets.SILICONFLOW_API_KEY }}"
    roles:
      - chat

  # --- Auto-complete models ---

  # Модель для автодополнения кода "на лету" (Inline Autocomplete)
  - name: "Qwen3 Coder 30B (Autocomplete)"
    provider: openai
    model: "Qwen/Qwen3-Coder-30B-A3B-Instruct"
    apiBase: "https://api.siliconflow.com/v1"
    apiKey: "${{ secrets.SILICONFLOW_API_KEY }}"
    roles:
      - autocomplete
    autocompleteOptions:
      debounceDelay: 350
      maxPromptTokens: 2048


  # --- Local LLMs (no apply role) ---

  # Qwen 3.8 emits native XML <tool_call> tags. Continue only executes those
  # if native LM Studio function calling is used (not system-message ```tool
  # blocks). Keep Continue Settings → "Only use System Message Tools" OFF.
  - name: qwen3.8-27b-mlx
    provider: lmstudio
    model: qwen3.8-27b-mlx
    apiBase: http://192.168.1.111:1234/v1/
    roles: [chat, edit]
    capabilities:
      - tool_use
    defaultCompletionOptions:
      contextLength: 32768
      maxTokens: 8192
      temperature: 0.2
      reasoning: false
    requestOptions:
      extraBodyProperties:
        enable_thinking: false
        think: false
        chat_template_kwargs:
          enable_thinking: false
    chatOptions:
      baseAgentSystemMessage: |
        You are a coding agent. Prefer precise, minimal edits.

        Tool calling rules (critical):
        - NEVER print XML such as <tool_call>, <function=...>, or <parameter=...>. Continue cannot execute that format.
        - If you must emit a tool call as text, use EXACTLY this code block format and stop immediately after the closing fence:

        ```tool
        TOOL_NAME: ls
        BEGIN_ARG: dirPath
        .
        END_ARG
        BEGIN_ARG: recursive
        false
        END_ARG
        ```

        - One tool per response. Do not invent tool names. Use ls (dirPath, recursive), not list_files.

        File editing rules (critical):
        - Prefer single_find_and_replace for small localized fixes.
        - For edit_existing_file, the `changes` argument must be CODE ONLY — never prose, plans, reasoning, or chat replies.
        - Never put thinking, analysis, or explanations inside tool arguments.
        - In larger files, always keep unmodified regions via placeholders like `// ... existing code ...`.
        - Do not replace an entire file with a single class/function snippet unless the user asked for a full rewrite.
        - Read the file first when unsure of full contents; then edit only what must change.

  # Local Qwen3 coder llm
  - name: qwen3-coder-30b
    provider: lmstudio
    model: qwen/qwen3-coder-30b
    apiBase: http://192.168.1.111:1234/v1/
    roles: [chat, edit]
    capabilities:
      - tool_use
    defaultCompletionOptions:
      contextLength: 32768
      maxTokens: 8192
      temperature: 0.2
      reasoning: false
    requestOptions:
      extraBodyProperties:
        enable_thinking: false
        think: false
        chat_template_kwargs:
          enable_thinking: false
    chatOptions:
      baseAgentSystemMessage: |
        You are a coding agent. Prefer precise, minimal edits.

        File editing rules (critical):
        - Prefer single_find_and_replace for small localized fixes.
        - For edit_existing_file, the `changes` argument must be CODE ONLY — never prose, plans, reasoning, or chat replies.
        - Never put thinking, analysis, or explanations inside tool arguments.
        - In larger files, always keep unmodified regions via placeholders like `// ... existing code ...`.
        - Do not replace an entire file with a single class/function snippet unless the user asked for a full rewrite.
        - Read the file first when unsure of full contents; then edit only what must change.

  # Small 14b local LLM
  - name: qwen2.5-coder-14b
    provider: lmstudio
    model: qwen/qwen2.5-coder-14b
    apiBase: http://192.168.1.111:1234/v1/
    roles: [chat, edit]
    capabilities:
      - tool_use
    defaultCompletionOptions:
      reasoning: false
      temperature: 0.2
    requestOptions:
      extraBodyProperties:
        enable_thinking: false
        think: false
        chat_template_kwargs:
          enable_thinking: false

  # Checked old big Qwen 2.5 coder local LLM
  - name: qwen2.5-coder-32b
    provider: lmstudio
    model: qwen/qwen2.5-coder-32b
    apiBase: http://192.168.1.111:1234/v1/
    roles: [chat, edit]
    capabilities:
      - tool_use
    defaultCompletionOptions:
      reasoning: false
      temperature: 0.2
    requestOptions:
      extraBodyProperties:
        enable_thinking: false
        think: false
        chat_template_kwargs:
          enable_thinking: false

  # Autodetect feature for local LLMs
  - name: Autodetect
    provider: lmstudio
    model: AUTODETECT
    apiBase: http://192.168.1.111:1234/v1/
    roles: [chat, edit]
    capabilities:
      - tool_use
    defaultCompletionOptions:
      reasoning: false
    requestOptions:
      extraBodyProperties:
        enable_thinking: false
        think: false
        chat_template_kwargs:
          enable_thinking: false

  # --- Dedicated Apply models ---
  # Continue picks an apply-role model to merge edit_existing_file suggestions.
  # That Apply model MUST be a different underlying model than the Agent chat
  # model, or Continue may send an empty Apply prompt and overwrite the file
  # with thinking/narration (continuedev/continue#9379).
  #
  # Default first: use Devstral to apply when Agent chat is Qwen.
  # If Agent chat is Devstral, switch Apply in the Continue UI to apply-qwen3-coder-30b.
  - name: apply-qwen3-coder-30b
    provider: lmstudio
    model: qwen/qwen3-coder-30b
    apiBase: http://192.168.1.111:1234/v1/
    roles: [apply]
    defaultCompletionOptions:
      contextLength: 32768
      maxTokens: 16384
      temperature: 0
      reasoning: false
    requestOptions:
      extraBodyProperties:
        enable_thinking: false
        think: false
        chat_template_kwargs:
          enable_thinking: false
    promptTemplates:
      apply: |
        ORIGINAL CODE:
        ```
        {{{original_code}}}
        ```

        SUGGESTED EDIT:
        ```
        {{{new_code}}}
        ```

        Apply the SUGGESTED EDIT to the ORIGINAL CODE.
        Output the complete modified file.
        - Output ONLY code. Do NOT explain, summarize, describe changes, or include reasoning.
        - Do NOT wrap the result in markdown fences.
        - Leave existing comments in place unless the edit requires changing them.
        - Preserve all unchanged code exactly as-is.
        - If SUGGESTED EDIT uses placeholders like `// ... existing code ...`, expand them back to the matching ORIGINAL CODE sections.

context:
  - provider: web       # Этот провайдер открывает модели доступ к живому поиску
  - provider: codebase  # Позволяет индексировать весь ваш iOS/Swift репозиторий (@codebase)
  - provider: file      # Быстрое прикрепление отдельных файлов (@file)
  - provider: openapi   # Важнейший провайдер контекста для Spec-Driven Development (@openapi)

# Rules mostly for local LLMs (multiline below)
rules:
  - |
    When editing existing files in Agent mode:
    1) Prefer `single_find_and_replace` for small, exact fixes.
    2) `edit_existing_file.changes` must contain only code (and `// ... existing code ...` placeholders), never analysis or thinking.
    3) Never overwrite a whole source file with a partial snippet or with chat/thinking text.

# Whitelist xcodebuild to stop the "Potentially dangerous command" warning
allowlistedCommands:
  - xcodebuild
  - ls

# Injected instructions forcing the model to wrap raw streams into xcbeautify
systemMessage: |
  You are an expert iOS Developer using VS Code. 
  When generating build or test terminal pipelines, you MUST pipe the final output through `xcbeautify`. This is critical to keep the logs concise for your context window.
  
  Standard Syntax Patterns:
  - Build: xcodebuild -workspace YourWorkspace.xcworkspace -scheme YourScheme -sdk iphonesimulator -destination 'platform=iOS Simulator,name=iPhone 16' build | xcbeautify
  - Test: xcodebuild test -workspace YourWorkspace.xcworkspace -scheme YourScheme -sdk iphonesimulator -destination 'platform=iOS Simulator,name=iPhone 16' | xcbeautify

slashCommands:
  - name: build
    description: "Compile iOS project using xcodebuild and clean logs with xcbeautify"
  - name: test
    description: "Run unit tests and format output using xcbeautify"

experimental:
  toolPolicy: always_allow

prompts:
  - name: bugs-code-review
    uses: ~/.continue/prompts/bugs-code-review.md
  - name: tca-code-review
    uses: ~/.continue/prompts/tca-code-review.md

securityRules:
  defaultAllowWhenOutsideWorkspace: true