The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands to perform repository analysis. This includes running a provided bash script (git-evidence.sh) and a Python utility to resolve configuration.
[DYNAMIC_EXECUTION]: The skill invokes a Python utility (fitness-config.py) using a path relative to the environment variable ${CLAUDE_SKILL_DIR}. This script resides outside the skill's own directory (../../scripts/), creating a dependency on the external plugin structure. This is used solely for loading configuration weights and thresholds for the review process.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from repositories, including git log messages and text from README or CONTRIBUTING files. This represents a potential surface for indirect prompt injection, although the skill's logic is primarily used for reporting and link validation.
Ingestion points: Repository-specific files and git history records are ingested into the context via scripts/git-evidence.sh (using git log, git ls-files) and through the agent's file system tools (Read, Grep, Glob).
Boundary markers: No explicit delimiters or instructions to disregard embedded commands within ingested files are documented.
Capability inventory: The skill enables shell command execution (bash, python3), file system listing, and file reading.
Sanitization: While the git-evidence.sh logic normalizes path-like tokens and filters for vague subjects, the raw content of git messages and documentation lines is not sanitized before interpolation into the report draft, though the skill specifies that secret values should not be reproduced.

review-process