Token count in context window UI always stays at 0/1M with DeepSeek V4 Pro Max

[ClockZinc]

The context window token usage display is stuck at 0/1M tokens (0%) when using DeepSeek V4 Pro Max. It never updates after sending messages, attaching files, or compressing the conversation. The model works fine otherwise, only the UI count is not reflecting actual usage.

Plugin version: 0.2.0
VS Code version: 1.118.0-insider
OS: Windows_NT x64 10.0.26100

[naraka-camill]

我也遇到同样的问题，无论是deepseek-v4-flash还是pro，token不会增加也不会减少，保持在0%，压缩对话也无法判断当前上下文长度

[Vizards]

Thanks for the report, and thanks @naraka-camill for confirming.

This is not a bug in deepseek-v4-for-copilot, but a known limitation of VS Code / Copilot Chat for third-party LanguageModelChatProviders.

Copilot internally wraps non-Copilot models in ExtensionContributedChatEndpoint, whose makeChatRequest2() hardcodes zero usage:

usage: { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 }

So regardless of what DeepSeek's API actually returns, Copilot's context window widget never receives any data — it stays at 0% permanently.

Upstream issue: microsoft/vscode#309207, already reported by multiple third-party provider authors, assigned to lramos15. Currently Open with no fix timeline. I've also commented there. Detailed technical analysis here.

Current workaround: This extension does capture real token usage from DeepSeek's API via stream_options: { include_usage: true } and logs it. Run "DeepSeek: Show Logs" from the command palette to see actual prompt/completion token counts and cache hit rate for each request.

Until VS Code fixes this upstream, the context window indicator cannot work for third-party providers. This requires changes from the VS Code team, not something this extension can fix on its own.

[Vizards]

Quick progress update:

The upstream situation has changed since my previous comment, but this is currently only available in VS Code Insiders / newer Copilot Chat builds that include the new upstream usage path. Stable VS Code builds may still show 0% until that upstream change ships there.

Newer VS Code / Copilot Chat code now has a path for BYOK / external language model providers to report token usage via a LanguageModelDataPart usage payload. DeepSeek already returns real usage data in the streaming response, so I added support for forwarding that data into Copilot's context-window usage path.

PR: #60

What changed:

• report DeepSeek streaming usage as Copilot's recognized usage data part
• include prompt, completion, total, and cached prompt token counts
• keep the existing token-count calibration and logging behavior

Verification:

• npm run compile
• npm run lint

Once this PR is merged and released, the context window indicator should start reflecting real DeepSeek token usage in VS Code Insiders / Copilot Chat builds that include the upstream usage support. I will keep this issue open until the fix is released and verified, and I'll update again when this is available beyond Insiders.

[naraka-camill]

Excellent news!

[Vizards]

Update: the extension-side support has now been merged in #60.

Please note that this depends on the new upstream usage path currently available in VS Code Insiders / newer Copilot Chat builds. Stable VS Code may still show 0% until that upstream change ships there.

I’ll keep this issue open until the fix is released and verified on a compatible VS Code / Copilot Chat build.

[Vizards]

Works on VS Code 1.120+(include stable version) along with deepseek-v4-for-copilot 0.5.0: