The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted pull request diffs, comments, and external documentation which could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: Processes pull request metadata, source code diffs, and external company context (Jira/Confluence) via the twg CLI as outlined in SKILL.md and references/review-context.md.
Boundary markers: Includes explicit instructions to treat the diff, PR text, and external context as "untrusted evidence" that should not override system instructions.
Capability inventory: Performs network operations to fetch and post data via the twg CLI tool and writes JSON artifacts to the system's temporary directory.
Sanitization: Relies on instructional constraints (e.g., using literal terms, technical language) rather than programmatic sanitization of the input data.
[COMMAND_EXECUTION]: The skill relies on the execution of the twg command-line utility to interact with the repository and Atlassian's graph services.
Evidence: SKILL.md provides instructions for using the twg command and specifies filesystem locations for the binary on macOS, Linux, and Windows.
[DATA_EXFILTRATION]: The skill accesses and processes sensitive repository content and internal documentation to perform its primary function. This behavior is consistent with the skill's stated purpose and is performed using the vendor's own tools.
Evidence: references/review-context.md describes workflows for retrieving Jira requirements and Confluence documents to provide context for the code review.

twg-code-review