name = "swe-sweep__agda" # The agent's change set, captured by the [[verifier.collect]] hook below after the # agent phase and handed to the separate verifier at the same path. artifacts = ["/logs/artifacts/model.patch"] [metadata] benchmark = "swe-sweep" upstream_task = "agda" version = "2.6.3-dev-20210817" base_ref = "c9c20bf3683f" clone_url = "https://github.com/agda/agda.git" container_repo = "/agda" n_subtasks = 45 upstream_commit = "9f7399e15e6172900e0591deb5a57a45769d8205" category = "software-engineering" tags = ["debugging", "repository-level", "swe-sweep"] [environment] # No prebuilt image: the sandbox is built from environment/Dockerfile. workdir = "/agda" cpus = 16 memory_mb = 65536 storage_mb = 102400 build_timeout_sec = 14400 # Public baseline so the harness can install its agent; the scored phases below are # offline (upstream runs the agent container with --network=none). A hosted model's # endpoint goes through the harness's run-level allowlist (harbor: --extra-allowed-hosts). network_mode = "public" [agent] timeout_sec = 43200 network_mode = "no-network" [verifier] environment_mode = "separate" network_mode = "no-network" timeout_sec = 43200 # Upstream `infer/base.py`'s patch extraction, run in the agent sandbox once the agent is # done: stage everything (new and deleted files included) and diff against the frozen # `baseline` tag; `--binary` keeps the patch appliable when the agent's test runs wrote # binary files into the tree. [[verifier.collect]] command = "mkdir -p /logs/artifacts && cd /agda && git add -A && git diff --cached --binary baseline > /logs/artifacts/model.patch" timeout_sec = 900