VS Code Copilot Chat: after `view_image` or the browser screenshot tool succeeds, the next request fails with `400 ... "Error while downloading file. Upstream status code: 404."`

Context

Source(s): https://github.com/microsoft/vscode/issues/337762 Status: reported working by source, not reproduced by us (workaround only; root fix not confirmed).

Problem

The agent calls the browser tool screenshot_page or view_image on a local PNG, the tool reports success (for example "Captured screenshot"), and the very next model request fails with the error above. The chat cannot continue, and it happens again on every attempt. A commenter sometimes got a download timeout instead of the 404. The original poster also saw garbled tool-result text such as "Checked , , , , no problems found."

Error (redacted / normalized)

Sorry, your request failed. Please try again.
Reason: Request Failed: 400 {"error":{"message":"Error while downloading file. Upstream status code: 404.","code":"invalid_request_body"}}

Applies when

VS Code 1.139.0 and 1.139.1 with Copilot Chat 0.66.0 and 0.67.0, on Windows and macOS; GPT-5.6 Sol and GPT-6 Sol; the "Local" agent harness; reported 2026-09-24 to 2026-09-28. A contributor said it only happened on certain models.

Confirm it is this problem

1. Check your user settings for github.copilot.chat.responsesApi.promptCacheBreakpoint.enabled → expect: it is true, or not set (a contributor said it defaults to true from 1.139.0).
2. In a new Local-harness chat, ask the agent to view a small local PNG with view_image → expect: the tool succeeds, then the next request fails with the error above.

Not this when / boundary

You are on the "Copilot" harness; a contributor could not reproduce it there. A 404 for an image you attached yourself earlier in a long chat may be the separate attachment-lifetime issue a commenter linked (#313973) (our inference).

Root cause

A contributor found it only happened when github.copilot.chat.responsesApi.promptCacheBreakpoint.enabled was true, and the same contributor said its default was changed to false in PR #338265 until the service side was fixed. The thread does not explain why the cache breakpoint made the service lose access to the image. On 2026-10-01 a commenter who said they work on the service side said they had deployed a server-side fix and saw no more cases in telemetry; a contributor thanked them for the fix and closed the issue on 2026-10-06.

Variables

none.

Fix

1. Add "github.copilot.chat.responsesApi.promptCacheBreakpoint.enabled": false to your user settings.json, open a new chat and rerun confirm 2 → expect: a commenter's A/B test with the same PNG and model: with false, view_image worked and the model named the image's color; set back to true, the next request failed again; with false again, the chat that had kept failing also worked.
2. Or run the chat on the "Copilot" harness → expect: a contributor's screenshots show no failure there.
3. Update VS Code and Copilot Chat to a build that has PR #338265 (which builds is our inference; the thread names only the PR) → expect: the setting defaults to false.

How to verify

In the Local harness, have the agent view an image and continue the task → expect: no 400 and no Upstream status code: 404.

What did not work

WebSocket fallback and the connectivity checks passed for a commenter, so it is not a general connection problem. Turning off github.copilot.chat.imageUpload.enabled was not needed; the commenter left it on throughout.

Rollback

Remove the promptCacheBreakpoint line from settings.json.

Risks

With the breakpoint off, fewer prompt-cache hits may mean slower or more costly requests (our inference). Commenters posted request IDs in public; keep your own out of public posts unless support asks (our inference).

Verified scope

VS Code 1.139.x with Copilot Chat 0.66 to 0.67, September to October 2026.

Becomes obsolete when

The server-side fix holds with the setting back on true. Open question: nobody in the thread re-tested with the setting set to true after the 2026-10-01 server-side fix.