openclaw - 💡(How to fix) Fix autoglm-browser-agent: mcp_server returns empty execution result, extension works but data lost [1 comments, 1 participants]

Problem

autoglm-browser-agent skill fails to return task results. The browser extension executes the task correctly (user can observe the agent navigating, clicking, taking screenshots in Chrome), but mcp_server returns 'empty execution result' after receiving only an intermediate get_current_state action.

Environment

• OS: Windows 11 (x64)
• OpenClaw runtime: agent=main, model=glm-5-turbo
• Browser: Chrome with AutoGLM extension (jelniggicmclhfgnlapbkgfibmgelfnp)
• Extension is active and connected to relay (ws://127.0.0.1:62031)
• relay.exe running and listening on port 62031
• mcp_server.exe configured with --if_subagent flag

Reproduction

1. Register mcp_server via mcporter config add (standard INSTALL.md steps)
2. Start relay.exe
3. Call browser_subagent with any task
4. Observe: Chrome opens tabs and agent navigates (extension works)
5. mcp_server returns after ~10s with: 'Browser subagent execution error: empty execution result'
6. The result only contains get_current_state at round 2, session marked as finished

Key Log Evidence

mcp_env.log shows the extension sends back:

• msg_type: task_processing (not task_completed)
• action: get_current_state (intermediate step, not final result)
• Session immediately marked as finished by mcp_server

What works

• mcporter list --schema: works (864ms)
• relay ↔ extension connection: works (Chrome connects to relay)
• Extension navigates and performs actions: works (user sees it)
• Second call with same session_id: returns result file, but only contains [summary] without actual data

What does not work

• First call returns empty execution result
• Result file contains only summary, no collected data
• Backend is hardcoded as claude_code in extension, api_url/api_key passed from mcp_server are ignored by extension

Possible causes

1. mcp_server closes WebSocket connection before extension finishes task (race condition)
2. Extension uses claude_code backend internally but has no Claude API key configured
3. Result data is generated but not written to the result file (only summary)
4. subagent_timeout or websocket_timeout too short for multi-page tasks

Suggested fixes

• Allow configuring the extension backend (currently hardcoded claude_code)
• Increase mcp_server wait time for extension results
• Write full task data to result file, not just summary
• Support resuming session to retrieve completed results