Artifact collection

Preserve files produced during a trial.

Harbor collects artifacts after the agent phase for two purposes: 1. saving files for local inspection 2. transferring agent-produced files into a separate verifier sandbox.

Automatic artifact collection

Automatic collection from /logs/artifacts/ will be removed soon. Configure artifacts explicitly for new tasks.

Files written to /logs/artifacts/ in the main sandbox are collected automatically. Harbor does not tell the agent to use this directory; specify it in the task instruction when the agent should produce the artifact.

instruction.md

Save the final report to `/logs/artifacts/report.json`.

They are saved under <trial-dir>/artifacts/logs/artifacts/ and shown in the viewer’s Artifacts tab.

Configure artifact collection

Use --artifact for a run, artifacts in config.json, or artifacts in the task’s task.toml.

Both forms below accept a file or directory. And you can specify multiple artifacts.

Basic

CLI

--artifact is repeatable.

export OPENAI_API_KEY="..."
harbor run \
  -t hello-world/hello-world \
  -a codex -m openai/gpt-5.6-sol \
  --artifact /app/hello.txt

Run config

config.json

{
  "artifacts": ["/app/hello.txt"]
}

Task config

task.toml

artifacts = ["/app/hello.txt"]

/app/hello.txt is saved as <trial-dir>/artifacts/app/hello.txt on the host.

Run-level artifacts are added to task-level artifacts.

Advanced options

Use the object form to change the saved path or exclude files from a directory. These options require a run or task config. Add multiple entries to collect files and directories together.

Run config

config.json

{
  "artifacts": [
    {
      "source": "/workspace/output",
      "destination": "output",
      "exclude": ["*.tmp"]
    },
    {
      "source": "/app/report.json",
      "destination": "report.json"
    }
  ]
}

Task config

task.toml

artifacts = [
  { source = "/workspace/output", destination = "output", exclude = ["*.tmp"] },
  { source = "/app/report.json", destination = "report.json" },
]

On the host, these are saved as <trial-dir>/artifacts/output/ and <trial-dir>/artifacts/report.json.

Field	Description
source	File or directory inside the sandbox.
destination	Optional relative path under <trial-dir>/artifacts/.
exclude	Patterns excluded when collecting a directory.
service	Compose service to collect from. Defaults to main.

Separate verifier

The verifier runs in a new sandbox. From the agent sandbox, Harbor transfers only the declared artifacts and the automatic /logs/artifacts/ directory. Use the same artifacts field; there is no separate artifact configuration.

task.toml

artifacts = ["/app/report.json"]

[verifier]
environment_mode = "separate"

Harbor collects /app/report.json from the agent sandbox and places it at /app/report.json in the verifier sandbox. destination controls only where the artifact is saved on the host.

Learn more about separate verifiers.

Collection manifest

Harbor writes <trial-dir>/artifacts/manifest.json as a collection report. Each entry records the artifact’s source, host destination, service, file type, and status: ok, empty, failed, or skipped.

manifest.json

A collection failure is recorded in the manifest without failing the trial.

If two entries would write to overlapping host paths, Harbor keeps the first and records the other as skipped instead of overwriting it.

Sidecar artifacts

For multi-container tasks, set service to a Compose service. A verifier.collect hook can first write runtime state to a file for collection.

task.toml

artifacts = [
  { source = "/var/log/api/orders.log", service = "api" },
  { source = "/tmp/stats.json", service = "api" },
]

[[verifier.collect]]
service = "api"
command = "curl -s http://localhost:8000/stats > /tmp/stats.json"
timeout_sec = 30

Sidecar artifacts and collection hooks require a Compose-capable sandbox. See the working example and provider capabilities.

Multi-step tasks

For multi-step tasks, step-level artifacts are added to the task- and run-level artifacts. Harbor collects them before each step’s verifier and stores them under steps/<step-name>/artifacts/.

View artifacts

Open a trial’s Artifacts tab in the results viewer.

Collected artifacts in viewer