Skip to main content
Do this whenever a command produces something you want after the task ends. The sandbox filesystem is deleted with its pod, so a file that is not copied out is gone — asking for it afterwards is not possible. Do not use artifact collection for large intermediate files. Copy out the deliverable, not the build tree; per-file and per-task ceilings apply, and exceeding them fails the copy rather than truncating it.

​
Prerequisites

  • An agent that can run commands — see run a command
  • A configured object store behind the artifact service
  • Sessions for when the sandbox disk disappears

​
Steps

1

Name the files when running the command

The bash tool takes artifact_paths, a list of relative paths to copy out after the command finishes. The agent supplies them; instruct it to do so:
Paths are relative to the working directory. Absolute paths, .. segments, and NUL bytes are refused before the copy is attempted.
2

Choose how output is captured

stdout and stderr are captured with byte limits rather than being returned whole. Two ways to get them, and they suit different sizes:Read them from the tool result. Fine for ordinary command output. The result carries the captured text along with stdout_truncated and stderr_truncated flags so you can tell a clipped stream from a short one.Read them as stored references. For anything large. A persisted execution record does not carry inline stdout or stderr at all — those are rejected — so the record instead holds stdout_ref and stderr_ref pointing at stored objects. Pick this when output routinely exceeds the capture limit, because the reference is the full stream while the inline copy is the clipped one.
3

Retrieve the results

List everything the task produced:
Each item carries an AgentArea download URL rather than an object-storage URL, so access stays behind the platform’s authorization and audit checks. Download one:

​
Verify

The listing is the check — an artifact that copied out appears in it with a non-zero size and a SHA-256:
Copy-out verifies content on the way through: the executor hashes the file it discovered, and the copy recomputes the digest over the bytes actually read, refusing a mismatch. So an artifact present in the listing with a digest is one whose bytes were confirmed, not merely reported. If a requested path is missing from the listing, check the per-artifact error field in the command result — a failed copy is reported there rather than silently dropped.

​
Troubleshooting

A single artifact may not exceed 16 MiB. This is deliberate: reads above that ceiling fail at the sandbox file API anyway, so the copy is refused loudly rather than half-completed. Compress the file, split it, or write a summary instead. There is no streaming path for larger deliverables.
The task’s whole durable workspace is capped at 1 GiB, so many mid-size artifacts can each pass the per-file check and still exhaust the aggregate. Note that input attachments already occupy part of this quota.
The command did not request the file. artifact_paths is opt-in — writing a file is not enough, the path has to be named in the same tool call. Check the event stream for whether the agent passed artifact_paths at all.
It was. Compare stdout_truncated in the result; if true, the stream exceeded the capture limit and the stored reference holds the rest. Capture limits are bounded at 16 MiB per stream.
Copy-out happens as part of the command that requested it. A file created by an earlier command and never requested is not retroactively collectable once the pod is reclaimed — see sessions for the reclaim timing.

Run a command in a sandbox

Where artifact_paths is passed

Provide input files

The inbound direction

Limits

Every execution and workspace ceiling in one table

Artifacts

The durable store and its history

The file model

Durable versus ephemeral writes
Last modified on September 17, 2026