[BUG] Tool calls hang indefinitely with flat token usage — no timeout at dispatch layer (Windows desktop app, "New window" sessions

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Bug report: tool calls intermittently never resolve — session shows "thinking" for hours with flat token usage

Product: Claude Desktop (Cowork) on Windows 11 Home (build 10.0.26200), Claude Code harness
User: dax@foundsm.com · Dates observed: 2026-08-03 through 2026-08-05

Summary

Tool calls occasionally dispatch and never resolve. No timeout fires at any layer,
so the session UI shows "thinking…" with an ever-growing wall-clock timer while
token usage stays flat (i.e., no compute is happening). Only a manual user
interrupt unblocks the session; the retried call then succeeds immediately.

Key detail — window type

Every hang occurred in a session opened via "New window." After returning to
the main-window session, an identical battery of probes across all tool channels
(Bash, PowerShell, file search, WebFetch, remote MCP connector, Chrome-extension
MCP) passed instantly — including the exact calls that had hung. This suggests
the wedge may be specific to secondary-window session plumbing.

Incidents

1. Chrome-extension MCP find call: open 17h 10m 40s, 470 tokens total (screenshot available). Overnight; machine awake.
2. PowerShell tool call: open 37m 26s, 323 tokens — a 2-second local command (file reads only). The harness's own 2-minute Bash/PowerShell timeout never fired, implying the dispatch-layer await was lost rather than the tool running long.
3. ~2h "thinking" episode — user reports no credit/token movement after the first ~25 minutes.
4. Scheduled tasks (weekly, Mondays 08:00): four consecutive runs (Jul 13, 20, 27, Aug 3) produced zero assistant output — sessions show lastActivity within seconds of dispatch. Likely the Chrome extension was not yet connected at fire time; failure was silent (no notification of the wedge).

Reliable observations

• Token counter flat during all hangs → model idle, not looping.
• User interrupt always unblocks; immediate retry always succeeds.
• Machine confirmed awake and actively monitored during hangs (not OS sleep).
• Not tool-specific: both Chrome-extension MCP and built-in PowerShell affected.

Impact

• Sessions appear stuck for hours; users can't distinguish "model thinking" from "awaiting a lost tool reply."
• Unattended scheduled automation fails silently for weeks.

Asks

1. Per-tool-call timeout/heartbeat at the dispatch layer with a loud failure.
2. UI distinction between "model generating" and "waiting on tool result" (+ elapsed).
3. Scheduled-task runs should notify on abnormal termination, not just completion.

<img width="588" height="102" alt="Image" src="https://github.com/user-attachments/assets/3a7142f5-8b32-4ee3-942a-dd1654873934" />
<img width="347" height="154" alt="Image" src="https://github.com/user-attachments/assets/6b2b2cef-8844-4482-a57f-d0243cbf2742" />

What Should Happen?

Tool calls resolve or fail loudly — a per-call timeout/heartbeat should surface an error in minutes, never an indefinite "thinking…" state.

Error Messages/Logs

None — that's the bug itself. Only signals: growing wall-clock timer + flat token count (screenshots: 17h/470 tokens, 37min/323 tokens). Note the contrast: CDP screenshot calls did produce proper 30s-timeout errors, proving some layers have timeouts while tool dispatch doesn't.

Steps to Reproduce

Not deterministic — pattern: open a session via "New window" on the Windows desktop app, do normal tool work, and intermittently a call never resolves; interrupt + retry always succeeds. Scheduled-task variant: fire a browser-dependent task before the Chrome extension connects → silent seconds-long death.

Claude Model

Other

Is this a regression?

I don't know

Last Working Version

First time I used this function to open code in a new window.

Claude Code Version

2.1.107

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

_No response_