The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the gh CLI via Python's subprocess module to interact with the GitHub API. This is the intended mechanism for data collection and reporting. Evidence is found in:
scripts/collect_ci_data.py for fetching workflow and job metadata.
scripts/publish_ci_findings.py for creating issues and posting comments.
[INDIRECT_PROMPT_INJECTION]: The skill processes workflow logs, which could contain untrusted text. Several mitigation measures are implemented:
Ingestion points: scripts/collect_ci_data.py reads data from UKGovernmentBEIS/inspect_ai and meridianlabs-ai/inspect_ai.
Boundary markers: The TRUSTED_HEAD_REPOS constant in scripts/collect_ci_data.py ensures that only data from specific repositories is processed.
Capability inventory: The skill can perform file writes to a local directory and authenticated writes to GitHub issues.
Sanitization: scripts/publish_ci_findings.py includes validate_findings and validate_report functions that check for and reject automation triggers like @auto or @review in the final output, preventing malicious content in logs from influencing downstream automated systems.
[EXTERNAL_DOWNLOADS]: The skill fetches metadata and job logs from the GitHub API. These network operations target a well-known service and specific repositories associated with the project's development, which is consistent with the skill's primary purpose.

ci-perf