Claude in Chrome: extension permanently disconnects after screenshot/scroll operations

Description

The Claude in Chrome browser extension frequently loses its connection during screenshot and scroll operations, entering an unrecoverable state that requires a full Chrome restart.

Environment

• OS: macOS (Darwin 25.1.0, Apple Silicon)
• Chrome: Latest stable
• Claude Code: Latest (installed via Homebrew)
• Node: v25.6.1 (native host runtime)

Steps to Reproduce

1. Start a browser automation session via tabs_context_mcp
2. Navigate to a page (navigate tool) — works fine
3. Take a screenshot (computer tool, action: screenshot) — usually works
4. Scroll the page (computer tool, action: scroll)
5. Take another screenshot immediately after

Result: "Browser extension is not connected" or "Error capturing screenshot: Detached while handling command."

The extension then enters an infinite exponential backoff reconnection loop (100+ attempts over ~2.5 hours) and never recovers. Only a full Chrome restart resolves the issue.

Log Evidence

From ~/.claude/logs/:

1. Timeout triggers the disconnect

[WARN] [Claude in Chrome] Tool call timeout: computer (1c991a7a) after 120000ms, pending calls: 0

2. Delayed response arrives to dead context

[DEBUG] [Claude in Chrome] Bridge received: {"type":"tool_result","tool_use_id":"1c991a7a-...","content":"Error capturing screenshot: Detached while handling command.","is_error":true}
[Claude in Chrome] Received tool_result for unknown call: 1c991a7a

3. Infinite reconnection loop (never recovers)

[INFO] [Claude in Chrome] Bridge reconnecting in 2000ms (attempt 1)
[INFO] [Claude in Chrome] Bridge reconnecting in 15188ms (attempt 6)
...
[INFO] [Claude in Chrome] Bridge reconnecting in 30000ms (attempt 100)

Operation Timing Analysis

| Operation | Avg Time | Reliability |
|-----------|----------|-------------|
| tabs_context_mcp | ~50ms | Always works |
| navigate | ~350ms | Always works |
| 1st screenshot | ~2-5s | Usually works |
| scroll + screenshot | ~20s+ | Frequently disconnects |

Root Cause Analysis

1. CDP connection fragility: The Chromium DevTools Protocol connection drops during slow operations (screenshots are ~20s)
2. No graceful recovery: Once the CDP session detaches, there is no mechanism to re-establish it
3. Orphaned tool calls: Timed-out calls leave no cleanup, and late responses arrive to dead contexts
4. Exponential backoff never succeeds: The reconnection strategy maxes out at 30s intervals but never actually reconnects — only a Chrome restart works

Configuration

• ~/.claude/.mcp.json has empty mcpServers: {} — no timeout or keepalive settings available
• Native host script (~/.claude/chrome/chrome-native-host) is a simple exec node wrapper with no error handling

Suggested Improvements

1. CDP session recovery: Detect CDP detachment and re-establish the debugging session without requiring Chrome restart
2. Reconnection cap: After N failed attempts, surface a clear error to the user instead of silent infinite retry
3. Operation-specific timeouts: Screenshots need longer timeouts than navigation; consider per-operation timeout configuration
4. Keepalive mechanism: Periodic health check pings on the CDP connection to detect disconnects early
5. Serialized screenshot queue: Prevent overlapping screenshot operations that can race-condition the CDP session