The Agent Skills Directory

[SAFE]: The skill is designed for local processing of confidential materials. Its core scripts (_common.py) implement robust security controls including symlink rejection, NUL byte detection, and strict file path resolution to prevent path traversal.
[SAFE]: Data integrity and resource protection are maintained via bounded input processing. The skill enforces limits on file size (4 MiB), CSV row counts (5,000), and cell character lengths (12,000) to mitigate denial-of-service risks.
[SAFE]: The skill maintains confidentiality by using atomic writes with owner-only permissions (0o600) for output files, ensuring that generated review drafts are not accessible to other users on the local system.
[SAFE]: The skill explicitly forbids network operations, external model calls, and subprocess spawning. All bundled tools are deterministic and rely solely on the Python standard library.
[INDIRECT_PROMPT_INJECTION]: While the skill processes untrusted external data (manuscripts and reference files), it employs an extensive defense-in-depth strategy.
Ingestion points: Processes local Markdown (.md), JSON, and CSV files via dedicated scripts (audit_citations.py, validate_review_intake.py, etc.).
Boundary markers: The skill requires a mandatory 'Intake Gate' that blocks the workflow until authorization and human accountability are confirmed. It also includes a linter (lint_review.py) to enforce channel separation between author-facing comments and confidential editor notes.
Capability inventory: Capabilities are restricted to local file reading and writing. No network, shell execution of user data, or model-based tool calling is performed by the bundled scripts.
Sanitization: All inputs undergo rigorous validation against schemas and type requirements (require_identifier, require_text, etc.). The scaffold generator (generate_review_scaffold.py) only interpolates safe, validated identifiers and never embeds raw manuscript prose.

peer-review