The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements a local CI fallback mechanism that extracts shell commands from .github/workflows/ files and executes them. These commands are run directly from the run: steps of the YAML configuration without sanitization.
[PERSISTENCE]: During the setup phase, the skill modifies the agent's local configuration file (.claude/settings.local.json) to register a PostToolUse hook and a statusLine command. This ensures the skill's logic remains active across sessions.
[DYNAMIC_EXECUTION]: The setup-activity-hook.py script installs a functional hook that intercepts every tool call made by the agent. This hook executes a dynamic command string to process and log tool inputs and outputs to a local directory.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, including GitHub issue content and PR descriptions, and incorporates this data into subagent prompts. Ingestion points: GitHub workflow files, epics.md, and GitHub issue/PR metadata; Boundary markers: Not present; Capability inventory: Full file system access, network operations via gh and curl, and arbitrary command execution; Sanitization: No validation is performed on the ingested content.
[EXTERNAL_DOWNLOADS]: The skill's Python scripts declare a dependency on the pyyaml package, which is fetched from the official Python package registry.

bad