One process per streaming E2B command: owns the Connect stream, translates envd events into the Managoat.Sandbox owner-frame contract, and heartbeats the sandbox TTL while the command lives.

E2B sandboxes always carry a TTL, so a long agent turn must keep extending it — a missed heartbeat degrades to an auto-pause (state preserved), never a kill, because the adapter creates sandboxes with autoPause: true.

Frame mapping (envd StartResponse/ConnectResponse events, JSON codec):

• data.stdout / data.stderr (base64) → {:stdout | :stderr, %{ref: ref}, bin}

• end.exitCode → {:exit, %{ref: ref}, code}
• EndStream (flag 2) with an error → {:error, %{ref: ref}, reason}
• EndStream without a preceding end event → {:exit, %{ref: ref}, 0} — the server finished the stream deliberately; per the contract a close without an exit frame reads as success
• transport failure → {:error, %{ref: ref}, reason} — the process may still be running sandbox-side; reattach exists for exactly this

In :attach mode the stream is a tail-based replayer over the journal files the spawn shim wrote (see Managoat.Sandbox.E2B), so the buffered output replays from byte zero; the real exit code is read from the shim's exit file when the replayer ends.