The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python scripts to perform local file system analysis of repository instructions. These operations are restricted to deterministic path checks, command manifest parsing (e.g., package.json, Makefile), and generating reports in a local .harness-eval directory.
[REMOTE_CODE_EXECUTION]: While the skill manages subagents to perform dual-judge evaluations, these are orchestrated within the platform's safety boundaries using allowlisted models and do not involve executing external code from untrusted sources.
[DATA_EXPOSURE]: The analysis is performed on the local repository to improve instruction quality. There is no evidence of sensitive data being exfiltrated or unauthorized network access. The script inventory_extract.py specifically identifies manifests to discover legitimate project commands.
[PROMPT_INJECTION]: The skill includes a 'planted traps' protocol to ensure the subagents (judges) are calibrated and following instructions correctly, which is a defensive measure rather than a malicious injection.
[DYNAMIC_EXECUTION]: The skill involves running local Python scripts distributed within the skill folder (e.g., inventory_extract.py, track_a_correctness.py) to automate the evaluation workflow, which is consistent with its stated purpose.

harness-eval