Sandbox log delivery rides the ReliableStream buffer (bounded by AGENTIX_LOG_BUFFER, drops oldest unacked events when full). If the Socket.IO link is down long enough — or the host never reconnects because the rollout driver crashed — buffered records are dropped or die with the worker, and there is no post-mortem artifact inside the sandbox to docker exec / provider-API your way back to.
PR #122's branch appended every captured line (flushed) to $AGENTIX_LOG_DIR/sandbox.log (default /tmp/agentix) as the durable backing for the lossy live stream. That landed nowhere: git grep AGENTIX_LOG_DIR on master returns nothing.
Follow-up: write bridged records (and captured fd-2 output, see companion issue) to an on-disk file inside the sandbox, independent of stream delivery. Bounded rotation is fine; the point is a post-mortem artifact, not an archive.
Related: #122.
🤖 Generated with Claude Code
Sandbox log delivery rides the ReliableStream buffer (bounded by
AGENTIX_LOG_BUFFER, drops oldest unacked events when full). If the Socket.IO link is down long enough — or the host never reconnects because the rollout driver crashed — buffered records are dropped or die with the worker, and there is no post-mortem artifact inside the sandbox todocker exec/ provider-API your way back to.PR #122's branch appended every captured line (flushed) to
$AGENTIX_LOG_DIR/sandbox.log(default/tmp/agentix) as the durable backing for the lossy live stream. That landed nowhere:git grep AGENTIX_LOG_DIRon master returns nothing.Follow-up: write bridged records (and captured fd-2 output, see companion issue) to an on-disk file inside the sandbox, independent of stream delivery. Bounded rotation is fine; the point is a post-mortem artifact, not an archive.
Related: #122.
🤖 Generated with Claude Code