Documentation
¶
Overview ¶
Command ax-agent-entry runs a one-shot core-agent turn as an AX task command.
AX's runner starts the task command on every boot of the sandbox, and Agent Substrate also runs it once in the template's golden boot, where egress is denied and /workspace is empty. This wrapper makes a core-agent turn safe under those rules:
- With --wait-for-egress it blocks until a TLS handshake to that address succeeds, so no turn starts inside the golden boot. The wait is unbounded by default because the golden boot's length is Substrate's to decide; --egress-timeout bounds it for a real boot where a wrong host would otherwise spin until AX kills the pod. SIGTERM and SIGINT end the wait (#1419).
- It skips the run when --state-dir holds a done marker from an earlier boot, so a resumed task does not repeat a finished turn. The marker is checked after the egress wait, because /workspace is only populated once the real actor is running.
- It copies core-agent's stdout to result.txt in --state-dir, since AX has no channel for reading a command's output back.
Everything after "--" is the core-agent command line, run as given.
ax-agent-entry --wait-for-egress=generativelanguage.googleapis.com:443 -- \ core-agent --task=review --session-db --session-db-path=/workspace/.core-agent/sessions.db -p "..."
Click to show internal directories.
Click to hide internal directories.