Foundry Toolkit Copilot model provider drops token usage and always reports zero context tokens

未关闭
  • 哪里不对？
#940 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
48/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
github, javascript, vscode
领域
developer-experience, observability

调研方向

The report identifies provideTokenCount() and provideLanguageModelResponse in the bundled dist/extension.js, and mentions convertOpenAIChunkToChatChunk as the usage conversion point. Start by locating these in the unminified extension source and tracing how response chunks reach Copilot. Done means actual input, output, and cached token usage is forwarded and token counting no longer always returns zero; the issue names no tests to run.

由索引模型根据 Issue 内容生成。

描述

needs attention

Title: Foundry Toolkit Copilot model provider drops token usage and always reports zero context tokens

Environment
  • VS Code: 1.142.0
  • Integrated Copilot extension: 0.70.0
  • Microsoft Foundry Toolkit: 1.6.19
  • OS: Windows x64
  • Model: Azure Foundry deployment gpt-5.6-sol
  • Integration: Foundry Toolkit language model provider in VS Code Copilot Chat
Problem

When using an Azure Foundry model through Foundry Toolkit in GitHub Copilot Chat, token usage is not reported correctly.

Observed behavior:

  • Context window indicator always displays 0 / 1M tokens.
  • Agent Debug Logs report 0 input and output tokens.
  • Copilot /usage reports AI Units but no Azure API token usage.
  • OpenTelemetry file export is enabled, but no chat token usage metrics are recorded.

Azure Foundry itself processes and bills the model requests normally. A separate direct API test confirms that the service returns valid usage information, including prompt_tokens, completion_tokens, and total_tokens.

Root cause identified through local extension analysis

1. Token counting always returns zero

In the installed Foundry Toolkit bundle:

dist/extension.js, around line 2520:

async provideTokenCount(i,a,o){return 0}

This explains why the Copilot context window indicator always displays zero.

2. API usage is not forwarded to Copilot

The internal response conversion handles the API usage information:

convertOpenAIChunkToChatChunk, around line 4988.

However, the language model provider implementation:

provideLanguageModelResponse, around line 2517

appears to forward text and tool-call content without forwarding the token usage data.

In the integrated Copilot extension, around line 7013, the consumer expects usage information as a LanguageModelDataPart. When it is missing, input, output, and cache token counts are set to zero.

These findings are based on inspection of the installed, bundled JavaScript code and have not been independently validated against the unminified source.

Expected behavior
  • Forward actual API input, output, and cached token usage to VS Code Copilot.
  • Implement meaningful token counting instead of returning zero.
  • Make token usage available to the Agent Debug Logs and supported telemetry consumers.
  • Allow developers using Azure Foundry BYOK deployments to monitor token consumption per model request.
Additional observation

OpenTelemetry file export works and produces telemetry data, but the inspected file contains no model-request token usage metrics. The missing model-turn spans may require separate investigation.

Impact

Without reliable token reporting, developers cannot monitor or optimize API token consumption within Copilot Chat.

In our environment, 296 Azure requests consumed approximately 3.28 million tokens, while the Copilot UI continued displaying zero usage.

Request

Could you confirm whether forwarding token usage from Foundry Toolkit language model providers to Copilot is currently supported?

If supported, please investigate the missing usage forwarding and the provideTokenCount() implementation.

If not yet supported, please consider implementing it so that users of Azure Foundry BYOK models can monitor their actual token consumption.

主要语言
JavaScript
星标
2.1k
派生
260
平均合并
3 小时 22 分钟
30 天内合并 PR
42

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

microsoft/foundry-dev-tools 的其他 Issue

查看 microsoft/foundry-dev-tools 的全部 Issue

相似的 Issue

更多 JavaScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。