The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the GitHub CLI (gh) and Git to perform repository analysis, fetch user metadata, and collect review activity. These commands are executed locally using standard tools and are limited to fetching data the user is authorized to access.\n
Evidence: The script scripts/collect-review-activity.py calls subprocess.run with specific gh subcommands such as gh api and gh search.\n
Evidence: The workflow.md instructions guide the agent to use git log, git blame, and git shortlog to determine code ownership and activity.\n- [DYNAMIC_EXECUTION]: The trainer agent is instructed to generate a new SKILL.md file that defines the behavior of the 'Clone' agent. This represents a legitimate use of instruction generation to create a functional extension based on analyzed review patterns.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external GitHub review comments which are ingested as data for the learning model. While this creates a surface for untrusted text to enter the agent's context, the skill's purpose is restricted to style analysis rather than executing commands embedded in the comments.\n
Ingestion points: Comments are fetched via the GitHub API and stored in local batch files (scratch/comments/batch-*.json).\n
Boundary markers: Comments are processed in isolated batches to prevent context pollution.\n
Capability inventory: The agent has permissions to read/write to its own skill directory and execute local Git/GitHub commands.\n
Sanitization: The skill performs standard JSON parsing; it does not explicitly sanitize comment bodies for instructions, but its operational logic is focused on statistical and stylistic modeling rather than instruction following.

code-review-clone