Known limitation: Selecting DeepSeek does not guarantee all Copilot internal requests avoid Copilot models/credits

[Vizards]

Summary

Selecting DeepSeek V4 Pro or DeepSeek V4 Flash in the Copilot Chat model picker only controls the main chat/agent model request.

VS Code and GitHub Copilot still have several internal routes, such as utility models, inline chat, Plan/Explore agents, title/summary generation, risk assessment, semantic/search helpers, embeddings, and other helper calls. Some of these routes may still use GitHub Copilot built-in models or require Copilot subscription capabilities.

This is a VS Code / Copilot Chat routing limitation. The extension can only handle requests that VS Code sends to the deepseek language model provider.

Important

There are two different costs to think about here.

If you leave chat.utilityModel / chat.utilitySmallModel unset, some Copilot internal helper requests may continue to use Copilot built-in utility models or Copilot credits.

If you set chat.utilityModel / chat.utilitySmallModel to DeepSeek, those helper requests may use your DeepSeek API key instead. These requests can include title generation, progress messages, prompt categorization, todo tracking, commit messages, summaries, search/ranking helpers, and similar small tasks.

Many of these helper requests are independent from each other and usually cannot reuse the same prefix cache as the main agent conversation. This means Flash traffic can still produce a high cache-miss rate and visible API cost during intensive coding sessions, even though deepseek-v4-flash is cheaper than deepseek-v4-pro.

This workaround is therefore a tradeoff, not a free switch. Please monitor both Copilot usage and DeepSeek API usage after changing these settings.

Workaround: reduce common Copilot model routes

If you want the common chat / inline / Plan / Explore / utility routes to use DeepSeek as much as possible, try adding this to your VS Code settings:

{
  "chat.utilityModel": "deepseek/deepseek-v4-flash",
  "chat.utilitySmallModel": "deepseek/deepseek-v4-flash",
  "inlineChat.defaultModel": "DeepSeek V4 Pro (deepseek)",
  "chat.planAgent.defaultModel": "DeepSeek V4 Pro (deepseek)",
  "chat.exploreAgent.defaultModel": "DeepSeek V4 Pro (deepseek)",
  "github.copilot.chat.askAgent.model": "DeepSeek V4 Pro (deepseek)",
  "github.copilot.chat.implementAgent.model": "DeepSeek V4 Pro (deepseek)",
  "github.copilot.chat.exploreAgent.model": "DeepSeek V4 Pro (deepseek)"
}

Notes:

• chat.utilityModel and chat.utilitySmallModel use the vendor/id format, for example deepseek/deepseek-v4-flash.
• Agent/default model settings use the qualified display name format, for example DeepSeek V4 Pro (deepseek).
• This workaround reduces common Copilot-model routes, but it does not guarantee zero Copilot credit usage.
• It may increase DeepSeek API usage because Copilot utility/helper traffic is now routed to DeepSeek.

Alternative: avoid DeepSeek utility/helper cost

If you do not want Copilot internal utility/helper requests to use your DeepSeek API key, remove these settings or set them back to the default value:

{
  "chat.utilityModel": "",
  "chat.utilitySmallModel": ""
}

The tradeoff is that VS Code / Copilot may then use Copilot built-in utility models or Copilot credits for some internal tasks.

Ongoing mitigation

The extension is trying to reduce unnecessary DeepSeek token usage where it can. For example, low-risk helper requests are being routed with reasoning_effort: none where possible. See #147 for the current discussion.

I am also investigating whether these helper requests can be classified in a way that makes prefix-cache reuse more stable, but this is still early and may require changes or cooperation from the VS Code / Copilot side.

If you need to avoid all Copilot model credits

The most conservative option is to sign out of GitHub Copilot in VS Code and continue using the DeepSeek models provided by this extension.

However, some Copilot/VS Code features may degrade or become unavailable when Copilot is signed out, especially features that still depend on Copilot tokens, embeddings, or Copilot-only internal services.

Image input

If image recognition is involved, use:

DeepSeek: Configure Vision Proxy

Then choose API Endpoint and configure your own vision endpoint. In that mode, image recognition uses your configured endpoint instead of Copilot built-in vision models.

Please comment here if you still see Copilot built-in model usage after applying the workaround, or if DeepSeek Flash utility traffic is producing unexpectedly high cost. Logs and screenshots are helpful, especially the exact VS Code version, extension version, selected model, DeepSeek usage breakdown, and relevant settings.

[SilverWhite]

个人经验性补充（如不应直接追加评论或存在问题请删，提前致歉）：
不要使用0.48.1的copilot插件，插件自动更新要关掉，不然会卡在work，copilot找不到deepseek。
目前VS Code 1.116.0 + Copilot Chat 0.44.2 + DeepSeek 插件 0.6.0可以正常使用。
不过0.44.2中会偶发性出现Sorry, no response was returned错误。

[Vizards]

个人经验性补充（如不应直接追加评论或存在问题请删，提前致歉）： 不要使用0.48.1的copilot插件，插件自动更新要关掉，不然会卡在work，copilot找不到deepseek。 目前VS Code 1.116.0 + Copilot Chat 0.44.2 + DeepSeek 插件 0.6.0可以正常使用。 不过0.44.2中会偶发性出现Sorry, no response was returned错误。

@SilverWhite 你是否在 Remote-SSH/WSL/Codespaces 中使用 VSCode + Copilot + 当前拓展？如果是的话，请查看 #136 获得解决方案
建议将 VSCode 升级到 1.122.0+，并使用内置于 VSCode 的 Copilot Chat
1.122.0 之前版本的 VSCode 将导致 DeepSeek 缓存命中异常，给你造成更多的费用消耗。详见：#25

[SilverWhite]

@SilverWhite您是否在 Remote-SSH/WSL/Codespaces 中使用 VSCode + Copilot + 当前拓展？如果是的话，请查看#136获得解决方案 建议将 VSCode 升级到 1.122.0+，并使用内置于 VSCode 的 Copilot Chat 1.122.0 之前版本的 VSCode 将导致 DeepSeek 存储命中异常，给您造成更多的剩余。详情：#25

@Vizards 佬，我装了WSL但是并没有使用，是在普通win11环境下。
目前长任务中DeepSeek缓存命中非常高，昨天用了9392w token，命中9131w，未见＃25问题

[50er]

@Vizards 大佬，内联建议可以切deepseek吗，我copilot额度用完了，设置了这些配置，但似乎不再弹出内联建议虚影了

[Vizards]

@50er 暂时还不支持把自动内联补全的灰色 ghost text 切到 DeepSeek。

这里有两个概念需要区分一下：inlineChat.defaultModel 影响的是 Inline Chat，也就是显式触发的内联聊天；它不会接管编辑器里自动弹出的内联补全建议。

自动内联建议目前仍然属于 GitHub Copilot 自己的 inline completion pipeline。本扩展只是通过 VS Code LanguageModelChatProvider 提供 DeepSeek Chat/Agent 模型，只能处理 VS Code / Copilot Chat 发给 deepseek provider 的请求，暂时不能替换 Copilot 的自动补全模型。

所以如果 Copilot 额度用完，自动补全虚影可能确实不会继续出现。这个方向可以在 #83 继续讨论。

[burningtnt]

如上，已在配置中设置，但依然遇到了 You've reached your monthly credit limit. Upgrade to Copilot Pro or wait until your credits reset on July 1, 2026 at 8:00 AM. 错误