Skip to content

fix(sandbox): shell_execution_timeout kills agent turns in Kata VM (WatchdogTimeoutGrace too short) #848

Description

@sabbour

Summary

Agent turns (Boomer/Gaeta in Software Delivery runs) consistently fail with:

\
WorkflowAgentInfrastructureException: Shell execution exceeded its hard deadline of 2 minutes and was terminated.
---> NotSupportedException: Only message, task, task update events are supported from A2A agents. Received: None
\\

This blocks the run from reaching the Human Review gate (beat 2.5 capture).

Root cause

The ~2 minute figure is: \ imeout_ms\ (model-supplied, e.g. 60 s) + \WatchdogTimeoutGrace\ (60 s) = 120 s ≈ 2 min.

In Kata hardware-isolation, SIGTERM relay through the Kata agent can take 10-60+ seconds. The 60-second \WatchdogTimeoutGrace\ is consumed before the process exits, causing the streaming watchdog to fatally abort the entire agent turn.

Additionally: \DefaultTimeoutMs = 30_000\ (30 s) is too short for
pm install, \git clone, \cargo build\ in Kata VMs. When the model doesn't supply \ imeout_ms, commands time out prematurely.

Files

  • \packages/Agentweaver.AgentTools/SandboxToolOptions.cs\ — \WatchdogTimeoutGrace\
  • \packages/Agentweaver.AgentRuntime/CopilotAIAgent.cs\ — \DefaultTimeoutMs\

Fix

PR #847: Increase \WatchdogTimeoutGrace\ 60 s → 5 min and \DefaultTimeoutMs\ 30 s → 5 min.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions