Hook inputs: expose the active chat model

未关闭
  • 哪里不对？
#2,851 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

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
35/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
typescript
领域
developer-experience

调研方向

Start from the shared hook input type (BaseHookInput) and the event-specific inputs the issue names: UserPromptSubmit, PreToolUse and PostToolUse. Find where the session and timestamp fields are populated, and check whether the host already supplies any model data before adding an optional model object shaped like { id, vendor, family? }. Done means every hook event carries the field, or the issue records that the host cannot provide it.

由索引模型根据 Issue 内容生成。

描述

Context
This continues a request originally filed against VS Code - microsoft/vscode#301152 ("Add model field to Copilot agent hook inputs") — which was closed as the hooks surface moved to the Copilot SDK. The underlying gap wasn't resolved, only relocated. A related sibling request, #2810 ("expose the active agent and the approval mode"), covers agent identity and approval mode but not the model — this issue is specifically about the LLM model.

Problem
Hook inputs (BaseHookInput and the event-specific types, e.g. UserPromptSubmit, PreToolUse, PostToolUse) carry sessionId, timestamp, workingDirectory, etc., but no identifier for which chat model is active for that call. There is currently no reliable way for an external hook process to determine this:

vscode.lm.selectChatModels() returns models matching a selector, not the one currently selected in the chat UI.
The Chat Participant API's request.model is accurate, but only available inside a chat participant's own request handler, not to hooks, which fire on all chat activity regardless of whether a participant is invoked.
Reading transcript_path / host storage is an undocumented, fragile workaround that can silently break on any host update.

Use case
We build developer tooling on top of UserPromptSubmit that needs to behave differently depending on which chat model is active (e.g., adjusting behavior to a model's known characteristics, or attributing telemetry per model). This is a different use case from the original issue's security/policy-enforcement angle (model allow/denylisting), but hits the identical missing field: there's currently no supported way to know which model is active from a hook.

Proposed solution
Add a model field to the common hook input fields sent to all hook events:

model?: {
  id: string;
  vendor: string;
  family?: string;
};

Matches the shape already returned by vscode.lm's LanguageModelChat objects, so existing model metadata concepts carry over directly.

Why it matters
Without this, any hook-based tool that wants to reason about "which model is this developer using right now" has no supported way to find out - only fragile, undocumented workarounds that can break without notice.

主要语言
Java
星标
10.7k
派生
1.5k
30 天内没有已合并 PR

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

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。

github/copilot-sdk 的其他 Issue

查看 github/copilot-sdk 的全部 Issue

相似的 Issue

更多 Java Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。