Copilot Chat Shows "0 Tokens Used" for Third-Party Language Model Providers

[vleeuwenmenno]

Bug Report: Copilot Chat Shows "0 Tokens Used" for Third-Party Language Model Providers

Summary

When using third-party language model providers registered via the VS Code LanguageModelChatProvider API, the Copilot Chat context size indicator always shows 0% / "0 tokens used", regardless of the actual token usage.

Copilot Chat version analyzed: 0.43.0

Steps to Reproduce

1. Install a VS Code extension that registers a LanguageModelChatProvider (e.g., the LLM API Proxy extension)
2. Open Copilot Chat in VS Code
3. Select a model provided by the third-party extension
4. Send a message and wait for the response to complete
5. Observe the context size indicator (bottom-left of the chat input area)

Expected: The indicator shows actual token usage (e.g., "1.2k / 128k tokens, 1%").
Actual: The indicator shows "0 tokens used" / 0%.

Root Cause

The Copilot Chat extension's LN class (which handles all non-Copilot language model providers) hardcodes zero usage in its return value from makeChatRequest2():

// In github.copilot-chat/dist/extension.js (~byte offset 19966000)
return f || g > 0 ? {
    type: "success",
    usage: {
        prompt_tokens: 0,
        completion_tokens: 0,
        total_tokens: 0,
        prompt_tokens_details: { cached_tokens: 0 }
    },
    value: f,
    resolvedModel: this.languageModel.id
} : { type: "unknown", ... };

The model selection logic instantiates LN for any non-Copilot provider:

if (r.vendor !== "copilot") return createInstance(LN, r);

Data Flow

Third-party provider
  → progress.report(LanguageModelTextPart)
    → VS Code extension host
      → $reportResponsePart → main thread stream
        → Copilot LN.sendRequest() reads stream (text/tool-call parts only)
          → LN.makeChatRequest2() returns HARDCODED usage: {prompt_tokens: 0, ...}
            → Copilot agent reports {kind: "usage", promptTokens: 0, completionTokens: 0}
              → chatServiceImpl.setUsage() with zeros
                → Context widget reads response.usage → displays 0%

Impact

• Users cannot see how much of their context window is consumed
• The percentage indicator is completely non-functional for third-party models
• provideTokenCount IS used correctly for prompt budgeting (truncation), but the display is broken
• maxInputTokens and maxOutputTokens are correctly passed through

Suggested Fix

The LN class should estimate usage from the streamed response rather than hardcoding zeros. Options:

Option A: Estimate from streamed content (minimal change)

Count characters from streamed LanguageModelTextParts and LanguageModelToolCallParts, then apply a rough token estimation (e.g., chars / 4). Also use provideTokenCount on the input messages for the prompt side.

Option B: Expose usage via LanguageModelDataPart convention

Define a convention where providers can report actual usage via LanguageModelDataPart with a specific MIME type (e.g., application/vnd.llm.usage). The extension host would process this and Copilot would read it.

Option C: Add usage reporting to the LanguageModelChatProvider API

Extend the VS Code API to allow providers to return usage metadata from provideLanguageModelChatResponse. This is the cleanest long-term solution but requires API changes.

Workarounds

Since this cannot be fixed from the extension side alone (the LanguageModelChatProvider API has no usage reporting mechanism, and the extension host drops unknown response part types), the best current workaround is:

1. Track actual usage from the SSE stream's final chunk (stream_options: { include_usage: true })
2. Display usage in a status bar item or output channel
3. Use an adaptive provideTokenCount that learns from actual usage data for better prompt budgeting

Related

• VS Code LanguageModelChatProvider API
• LanguageModelResponsePart type (text, tool-call, and data parts)
• provideTokenCount method (used for prompt budgeting, NOT for display)

[kadirgun]

In fact, if custom baseURL support were added to the Anthropic and OpenAI providers already available in the Copilot Chat extension, there would be little need for most LanguageModelChatProvider extensions. Most well-known AI providers are compatible with the Anthropic or OpenAI APIs.

Edit: The token count also shows as 0 in the LanguageModelChatProviders provided by Copilot Chat itself.

[okanfidan]

The problem is still ongoing for me as well....

[Vizards]

I‘m maintaining a third-party LanguageModelChatProvider that adds DeepSeek models to the Copilot Chat model picker. I can independently confirm this. the context window indicator permanently shows 0% for all our models, same root cause (ExtensionContributedChatEndpoint hardcoding zero usage).

The cheapest fix that would cover most cases: instead of hardcoding zeros, estimate from the streamed response. Copilot already calls provideTokenCount() for prompt budgeting — reuse that for prompt tokens. For completion tokens, chars / 4 on accumulated LanguageModelTextPart content is a reasonable approximation. maxInputTokens / maxOutputTokens metadata is already available. No API changes needed — just a change in ExtensionContributedChatEndpoint to compute estimates rather than returning zeros.

Even a rough estimate is vastly better than permanently showing 0%. A proper LanguageModelUsagePart could come later.