I ran into this while dogfooding a language-model provider for Vercel AI Gateway I'm working on, using a Copilot chat.
While the agent was working, I attached a few source selections and asked whether some verbose validation code would benefit from Zod or another approach. The selections appeared as attachment chips in my message, but the agent asked which file I meant. When I pointed out that the selections were attached, it said it had received only a browser-page attachment.
That was confusing: from my perspective, I’d already supplied exactly the context it was asking for.
What we found
I worked with an agent to trace the selections through the installed VS Code agent host. It forwards their text, file paths and ranges to the SDK, and the SDK persists them in its user.message event. However, they’re absent from the request that reaches our model provider.
We then reproduced the behavior using the installed SDK/runtime directly, with synthetic model responses and no VS Code or Vercel provider involved:
| Same selection and text-blob attachments |
Selection reaches model |
Text blob reaches model |
| Message sent while idle |
Yes |
Yes |
mode: "immediate" during a tool call |
No |
Yes |
Both cases persist the attachments correctly. The comparison reproduced twice.
Reproduction
- Create an SDK session with a tool whose handler can be paused.
- Have a synthetic model response invoke that tool.
- While its handler is paused, send an immediate message containing a
selection attachment with text, file path and range, plus a text-blob attachment as a control.
- Release the tool and inspect the next model request.
The steering message and text blob are present, but the selection is missing. Sending those same attachments while idle includes the selected code and its location.
Tested with the versions bundled in my VS Code installation:
- SDK:
1.0.17-unstable.37397886721.gad270aa
- Runtime:
1.0.92-4.unstable.r37397886721.gad270aa
Expected behavior
I’d expect source selections attached to a steering message to reach the model just as they do when sent while idle.
Reviewing code while an agent continues working feels like a pretty normal use of steering. Ideally, selection rendering would follow the same path for both delivery modes, with a regression test covering attachments sent during a tool call. The visible attachment chips should accurately represent the context the agent receives.
I ran into this while dogfooding a language-model provider for Vercel AI Gateway I'm working on, using a Copilot chat.
While the agent was working, I attached a few source selections and asked whether some verbose validation code would benefit from Zod or another approach. The selections appeared as attachment chips in my message, but the agent asked which file I meant. When I pointed out that the selections were attached, it said it had received only a browser-page attachment.
That was confusing: from my perspective, I’d already supplied exactly the context it was asking for.
What we found
I worked with an agent to trace the selections through the installed VS Code agent host. It forwards their text, file paths and ranges to the SDK, and the SDK persists them in its
user.messageevent. However, they’re absent from the request that reaches our model provider.We then reproduced the behavior using the installed SDK/runtime directly, with synthetic model responses and no VS Code or Vercel provider involved:
mode: "immediate"during a tool callBoth cases persist the attachments correctly. The comparison reproduced twice.
Reproduction
selectionattachment with text, file path and range, plus a text-blob attachment as a control.The steering message and text blob are present, but the selection is missing. Sending those same attachments while idle includes the selected code and its location.
Tested with the versions bundled in my VS Code installation:
1.0.17-unstable.37397886721.gad270aa1.0.92-4.unstable.r37397886721.gad270aaExpected behavior
I’d expect source selections attached to a steering message to reach the model just as they do when sent while idle.
Reviewing code while an agent continues working feels like a pretty normal use of steering. Ideally, selection rendering would follow the same path for both delivery modes, with a regression test covering attachments sent during a tool call. The visible attachment chips should accurately represent the context the agent receives.