The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes dynamic context injection (!command syntax) in the SKILL.md file to check for the availability of the ollama executable and to read a configuration file located at ~/.claude/pza-ollama-model during the skill loading phase.
[COMMAND_EXECUTION]: The skill uses node -e to execute inline Javascript code for the purpose of extracting and validating JSON structures from the review results.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface because it processes untrusted data (git diffs and untracked file contents) and interpolates them into a prompt for the Ollama model. While it includes 'Core Constraints' and instructions to stay in 'review-only' mode, it lacks explicit sanitization or boundary markers (beyond simple heredocs) to prevent malicious instructions within the reviewed code from influencing the agent's behavior.
Ingestion points: Git diff output and untracked file contents are read via git diff and head (Step 1).
Boundary markers: The skill uses EOFPROMPT heredocs to wrap the prompt and diff content (Step 2).
Capability inventory: The skill can execute arbitrary shell commands via Bash and interact with the user via AskUserQuestion.
Sanitization: No explicit sanitization or filtering of the code content is performed beyond file type filtering and budget-based truncation.

ollama-review