Read what a sandboxed run left behind after its container is gone

Agno's open_run_checkpoint(run_id) looks up the checkpoint reference on the run's row, has the checkpoint store verify the archive, and yields its path for the length of the with block. CheckpointArchive.restore unpacks that path into workspace/ and native/, minus the credential files the capture left out, so a host process can turn an engineer's edits into a patch.

from pathlib import Path
 
from agno.sandbox.checkpoint import CheckpointArchive
 
with engineer.open_run_checkpoint(result.run_id) as archive:
    CheckpointArchive().restore(archive, Path("/tmp/review/turn"))
 
# async code, or an async database:
async with engineer.aopen_run_checkpoint(result.run_id) as archive:
    await CheckpointArchive().arestore(archive, Path("/tmp/review/turn-async"))

engineer.checkpoint_store returns the sandbox's checkpoints= store, or the host-local file default when none is set. Agno refuses that default while a durable queue worker is active, since another replica cannot read this host's disk. A file store now writes its objects read-only, so a reader writing through the yielded path cannot damage the session's recovery point.

We ran a FixtureAgent with Repository("https://github.com/octocat/Hello-World", branch="master") on a real DockerSandbox and Postgres, with the fixture runtime standing in for the model. A shell command in the warm computer appended a line to README, turn 2 committed, and we removed the container. With no container left in the namespace, open_run_checkpoint opened turn 2's 12,156-byte archive, and git diff in the restored checkout printed the edit.

Only a completed run of the same agent opens. On a second session without a repository, an agent with the id reviewer got Run open-26783431-2 belongs to another agent, and an id that was never run got Run open-26783431-never does not exist. A run that failed, was cancelled or is still running is refused the same way, since only a completed turn commits a checkpoint.

See the cookbook, and learn more about deploying harness agents in the documentation.