Close Harbor integration gaps: verifier isolation, artifact collection

[mhrezaei1]

Closes the three open gaps noted in #8.

Changes

Verifier isolation (reward hacking mitigation)

adapter.py now computes the SHA256 of evaluate.py at task-generation time and injects it into test.sh. If the agent modifies evaluate.py to manipulate its score, the verifier detects the hash mismatch and outputs a reward of 0.

The tests/ directory is copied by Harbor separately from the agent's workspace, so the agent cannot alter test.sh itself.

Artifact collection

Two levels now work out of the box:

• Automatic: test.sh copies the workspace to /logs/artifacts/workspace/ at the end of verification, excluding large model weight files (*.safetensors, *.bin, *.pt, *.pth, *.ckpt). Harbor auto-collects /logs/artifacts/ with no extra config.
• Full workspace (opt-in): Each generated task now includes a job.yaml with a commented-out artifacts block. Uncomment to also download model weights:

  artifacts:
    - source: /home/agent/workspace
      destination: full-workspace

  Tasks can then be run as: harbor run -c <task_dir>/job.yaml

Pre-agent hooks / timer

Already handled by the sentinel-file approach in timer.sh — no changes needed.

Testing

• All 28 task combinations generate without errors
• SHA256 hashes are verified to match evaluate.py for each generated task
• Hashes differ correctly across benchmarks (each has its own evaluate.py)

[hrdkbhatnagar]

Thanks for the PR!

Were you able to verify if this worked, by running an actual task on with Harbor (such as with the Modal backend)? We would need to verify if the actual workflow works or not. The task generation is not sufficient

Already handled by the sentinel-file approach in timer.sh — no changes needed.

The sentinel-file approach doesn't work, that's why we way for a pre-agent hook.