The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted data in the form of review comments, creating a potential surface for indirect prompt injection. This is mitigated by explicit instructions to separate raw input from analysis.
Ingestion points: Review comments (excerpts and suggested fixes) are ingested from external sources as described in references/context-pack.md.
Boundary markers: SKILL.md contains a hard rule stating that "Raw review prose must not be handed to an implementer," creating a logic boundary between untrusted input and actionable output.
Capability inventory: The skill generates a markdown-based adjudication kernel and warrants. It uses a local Python tool tools/review_adjudication_gate.py to validate the structural consistency of these outputs. No high-privilege operations or network calls are triggered by the analysis.
Sanitization: There are no explicit instructions for escaping or sanitizing the raw review text before processing.
[SAFE]: The Python tool tools/review_adjudication_gate.py is a static analysis script designed to verify markdown tables. It uses standard Python libraries and does not perform any network operations, dynamic code execution, or sensitive file access.

review-adjudication