The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The provisioning script downloads a shell script from a remote URL and executes it directly by piping it to the shell.
Evidence: curl -LsSf https://astral.sh/uv/install.sh | sh in scripts/provision-golden-checkpoint.sh.
[EXTERNAL_DOWNLOADS]: The skill downloads code, corpora, and container images from various third-party sources during setup and execution.
Evidence: Clones the Deep-SWE corpus from https://github.com/datacurve-ai/deep-swe in scripts/provision-golden-checkpoint.sh.
Evidence: Downloads task-specific Docker images from various registries during trials.
[COMMAND_EXECUTION]: The skill constructs and executes shell commands dynamically based on user-provided harness names and task definitions.
Evidence: The render() function in scripts/run-trials.sh interpolates variables into shell templates which are then executed via runta exec.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data produced by third-party coding agents, creating a potential path for malicious data to influence the benchmarking process.
Ingestion points: Reads result.json, trial.json, and agent trajectories from harness job directories in scripts/calculate-cost.py and scripts/normalize-results.mjs.
Boundary markers: Absent; the skill does not use specific delimiters or instructions to ignore embedded content when parsing agent outputs.
Capability inventory: The skill has extensive shell execution capabilities within its sandboxed runtimes via runta exec and trial-worker.sh.
Sanitization: While the skill uses jq for structured JSON parsing, it does not explicitly sanitize the text content of agent trajectories against malicious instructions.

frontierharness-eval