The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes code and diffs provided by the user, which presents a surface for indirect prompt injection where malicious instructions embedded in the reviewed code could attempt to influence the reviewer subagents.\n
Ingestion points: Source code and git diffs are read in SKILL.md (Step 1) and passed to reviewers.\n
Boundary markers: The references/reviewer-prompt.md template uses markdown headers to delimit the code under review, but it lacks explicit instructions to ignore any embedded directives within that code.\n
Capability inventory: The agent spawns multiple subagents via the Task tool and synthesizes their findings. The subagents are configured with readonly: true, limiting their ability to modify the environment.\n
Sanitization: The input code is interpolated directly into the prompt template without sanitization or escaping mechanisms.\n- [COMMAND_EXECUTION]: The skill uses shell commands to interact with the version control system and manage its own configuration.\n
Evidence: The skill executes git diff to gather review context and is instructed to autonomously open pull requests to update model configurations if a model slug is found to be invalid.\n- [DATA_EXFILTRATION]: The skill accesses configuration files within the user's home directory to determine model routing.\n
Evidence: It reads model lists from ~/.cursor/rules/pstack-models.mdc to configure the subagent spawning process.

interrogate