Long-running slave command times out at tools/call layer and blocks agent discovery

[yzs15]

Summary

A long-running driver/run_slave_powershell call timed out at the MCP tools/call layer after about 120 seconds even though the request passed timeout_sec=600. The underlying command appeared to continue running on the Windows slave, but the caller received no task_id, so there was no way to poll or cancel it.

While the command was still running, even driver/list_agents timed out. After waiting, list_agents recovered and the command had completed successfully.

I am filing this in both agentserver/agentserver and agentserver/loom because I am not sure which layer owns the behavior.

Expected behavior

• timeout_sec is honored, or the call returns an async task_id before the outer tools/call timeout.
• Long-running slave commands do not block list_agents or unrelated driver queries.
• If a call times out, the response exposes enough state to poll or cancel the underlying task.

Actual behavior

• driver/run_slave_powershell timed out at about 120 seconds despite timeout_sec=600.
• No task_id was returned to the caller for the long-running command.
• driver/list_agents also timed out while the command was still running.
• Later, the slave/driver recovered and the command had completed successfully.

Reproduction

1. Use a Windows slave with PowerShell support.
2. Run a long command through driver/run_slave_powershell with timeout_sec=600, for example:

winget install --id Python.Python.3.14 -e --scope machine --silent --accept-package-agreements --accept-source-agreements

3. Observe the MCP tools/call timing out after about 120 seconds.
4. Immediately call driver/list_agents.
5. Observe list_agents timing out until the long command finishes.

Environment observed

• Slave OS: Microsoft Windows Server 2025 Datacenter
• Slave command interface: PowerShell
• Driver tool: driver/run_slave_powershell
• Related driver query: driver/list_agents

Why this matters

From the orchestrator/user perspective, the workspace looks stuck: the long-running task cannot be polled or cancelled, and agent discovery also becomes unavailable. This makes normal recovery logic hard to implement.