Harden ActionManager timeout and error-state handling

[xdCloudy]

Summary

Fix several ActionManager lifecycle bugs that can leak state between actions and lose useful error information.

Problem

Action execution currently keeps timeout/interruption state on the manager instance. After a timeout, that state can affect later actions if it is not reset correctly.

The error path also converts an Error into a string before attempting to record its stack trace, which causes the original stack information to be lost.

These issues make failures harder to diagnose and can make subsequent actions behave as though an earlier timeout is still active.

Changes

• Reset action-local timeout state for each new action
• Preserve the original Error object through failure handling
• Keep stack traces available for logs/debugging
• Tighten resume/action lifecycle handling so stale state does not bleed into the next action
• Keep the external ActionManager API unchanged

Why

Action execution is a core runtime boundary. It should be deterministic per action and should preserve enough diagnostic context to understand failures.

This is particularly important for long-running autonomous agents where one transient timeout should not poison later actions.

Compatibility

No command syntax or agent profile changes are required.

Validation

Validated on the fork CI matrix with Node 20 and Node 22.

[Sweaterdog]

quit spamming our repo with slop