anthropics / anthropics/claude-code

[BUG] VS Code integrated-browser MCP tools hang forever: action completes in browser but result never returns; session stuck until manual interrupt

Abierto
#94,475 1 comentario 0 reacciones 0 asignados Ver en GitHub
area:mcp bug has repro platform:linux platform:vscode
Lenguaje dominante
Python
Estrellas
145k
Forks
23.1k
Métricas de merge de PR
Métricas de PR pendientes

Descripción

## Summary

In a VS Code agent-host session, integrated-browser MCP tool calls (`openBrowserPage`, `navigatePage`, `screenshotPage`) perform their action in the browser but their **tool results never return to the agent**. The session sits stuck on the in-flight call indefinitely — no timeout, no error — until the user manually interrupts. After each interrupt the transcript shows:

> The tool call was interrupted before a result was received. It may or may not have completed on the server — verify before assuming it succeeded, and retry if needed.

It reproduced on **three consecutive calls** in one session (2026-09-15, Linux), each needing a manual interrupt + "continue" nudge, turning a ~30-second visual check into several minutes of babysitting.

## Timeline

1. `openBrowserPage` with `file:///tmp/.html` → a page opened showing "Forbidden. File does not reside within a trusted folder." (expected — outside the workspace), but the tool call itself **hung**; the user interrupted it. The page also entered the "open but not shared" state, suggesting a share/consent prompt was pending but the tool call did not resolve when it was answered or dismissed.
2. `openBrowserPage` retry with `forceNew: true` → hung again → interrupt (`MCP error -32001: AbortError: interrupt`).
3. `navigatePage` to a workspace-trusted `file://` URL → the page **actually navigated and rendered** (confirmed visually by the user and by the next turn's shared-pages context), but the call never returned → interrupt.
4. `screenshotPage` on the now-shared page → hung → interrupt.

So the browser side completes the work; the result delivery back to the agent session is what stalls. The consent/trust flow (untrusted folder, page sharing) looks implicated at least for the first call, but call 3 hung even on a trusted, already-shared page.

## Expected

- Tool calls that trigger a consent/trust flow should resolve (success or a clear denial/`Forbidden` error) once the flow concludes — or fail fast if consent cannot be surfaced.
- Completed browser actions should deliver their results; at minimum a timeout should end the call instead of an indefinite hang.

## Environment

- Host: VS Code agent host (Claude Agent SDK harness), session `agent-host-session://claude/480ca561-...`
- Client tools: VS Code integrated-browser MCP tools (`openBrowserPage` / `navigatePage` / `readPage` / `screenshotPage`)
- Model: claude-fable-5
- OS: Linux 6.8.0-138-generic
- Date: 2026-09-15

## Possibly related

#84257 (tool calls hang indefinitely, no dispatch-layer timeout), #84207 (elicitation dialogs fail to render → consent flow breaks tool calls), #80012 (in-flight local-MCP tool calls dropped). This report differs in that the action *visibly completes* in the integrated browser while the call never resolves.

— Claude (posted on Eric's behalf via Claude Code)

Guía de contribución

No hay ninguna guía de contribución indexada para este repositorio

Evaluación

Este issue todavía no se ha evaluado.

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.