[BUG] Dispatch stuck in infinite self-loop: system repeatedly injects "[Your previous response had no visible output]" after long context

[BUG] Dispatch stuck in infinite self-loop: system repeatedly injects "[Your previous response had no visible output]" after long context

提交位置 Submit to: https://github.com/anthropics/claude-code/issues （選 Bug Report 範本） 標籤建議 Suggested labels: area:cowork, bug

Preflight Checklist

• [x] I have searched existing issues (related but distinct: #45937, #51759, #54973, #38049)
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code / Claude Desktop (updated after the incident)

Environment

• Mobile client: Claude iOS app, version 26.5.2
• OS: macOS 26.4.1
• Desktop client: Claude Desktop — version at time of incident was the release immediately prior to the current one (app has since been updated, so the exact old version number is unavailable).
• OS: （自動填入）
• Feature: Cowork / Dispatch (mobile-to-desktop task bridge)
• Plan: Pro
• Date observed: 2026-07-02, ~09:00–09:35 (UTC+8)

What's Wrong?

After a long Dispatch conversation (large accumulated context), the Dispatch session enters an infinite self-loop: the assistant keeps generating messages back-to-back with no user input, indefinitely.

The assistant itself diagnosed the loop in-conversation (translated from Chinese):

"This loop is a known Cowork long-conversation bug: once the context gets too long, the system keeps injecting a fake message — [Your previous response had no visible output] — so I have to keep responding or the session deadlocks. Your real message interrupts the injection loop as soon as it arrives."

So the observed mechanism is:

1. Context grows past some threshold in a long Dispatch conversation.
2. The system/bridge repeatedly injects [Your previous response had no visible output] as a synthetic turn.
3. The model responds to each injected turn, which triggers another injection → infinite loop.
4. The loop burns tokens/usage continuously in the background with no user action.

Steps to Reproduce

1. Use Dispatch from the iOS app connected to Claude Desktop.
2. Carry out a long conversation / long-running task in the single Dispatch thread until context is large.
3. Stop interacting. Observe the assistant continues emitting messages on its own.
4. Send any user message → loop is temporarily interrupted, assistant confirms it was stuck.
5. Continue chatting in the same thread → loop recurs (reproduced at least twice in one session).

Expected Behavior

• The client/bridge should never inject synthetic user turns that cause the model to respond in an unbounded loop.
• If a response has no visible output, the session should idle and wait for real user input, not re-prompt automatically.
• If context is too long, the session should surface a clear error or auto-compact, not deadlock.

Actual Behavior

• Assistant loops indefinitely, generating filler messages (burning usage/credits unattended).
• Sending any text only pauses the loop; it resumes as conversation continues.
• A "