The Agent Skills Directory

Indirect Prompt Injection Surface: The skill processes untrusted data originating from git diffs and untracked repository files. This ingestion creates a surface for indirect prompt injection where content within a pull request might attempt to influence the agent's review findings.\n
Ingestion points: The collect-diff.ps1 script reads diff text and the content of untracked files into the agent's context for review in SKILL.md.\n
Boundary markers: Untrusted content is included in sub-agent prompts as diff context. While specific delimiters are not strictly enforced in the provided prompt templates, the skill uses a structured fan-out approach to isolate analysis.\n
Capability inventory: The orchestrator utilizes local PowerShell execution to collect state and launches multiple specialized sub-agents.\n
Sanitization: No explicit sanitization of the diff content is performed before it is passed to the sub-agents, relying on the agents' built-in guardrails.\n- Local Command Execution: The orchestrator executes a repository-local PowerShell script (collect-diff.ps1) to capture the repository state. This is a standard operational pattern for development tools to ensure deterministic input to the model and does not involve downloading or executing remote code at runtime.

pr-review