The agent run you can open back up.

Switch on session capture and your runs through the gateway become traces you can open, export, share and re-ask a model turn at a time on another model — and the open-source CLI records the shell, filesystem and MCP calls a gateway never sees.

With capture on, a run carries the routing and cost truth: every turn, every model call, and what each one actually cost.
Step-level what-if: re-ask one turn on other models and put the answers, the latency and the bill side by side.
Export any run as an orca-trace v0 archive, or hand it to someone on a share link that expires and can be revoked.

Two halves. One trace format.

The local orca record command captures what a gateway cannot see: the shell, the filesystem, the MCP calls. The gateway captures what local can never reach — every key, every teammate, every CI job — with the routing decision and the real cost attached. Both write the same orca-trace archive, so a run from either side opens the same way.

Ask the question you only thought of afterwards.

Find the model turn where it went wrong and re-ask that one turn on a different model. The comparison runs through your own gateway on your own keys, so what comes back is what you would actually have shipped — with the token bill next to it.