The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to transmit local code, documentation, and configuration files to external LLM providers such as OpenAI, Anthropic, and OpenRouter for the purpose of technical analysis. This involves the movement of potentially sensitive repository content to third-party cloud services. The skill mitigates exposure by instructing the agent to redact secrets and credentials from any evidence generated during the review process.
[COMMAND_EXECUTION]: The skill can write and modify files on the local filesystem. It creates new review artifacts in the repository's documentation directory and allows for direct edits to documentation, prose, and comments within the codebase. This filesystem write access is restricted to non-executable content but represents a significant capability that could be targeted if the agent processes malicious input.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests and processes untrusted data (source code and documentation) as its primary input. Maliciously crafted instructions within these files could attempt to influence the sub-agent reviewers or the final aggregation logic.
Ingestion points: Source code, configuration files, and documentation files within the local repository repository scope.
Boundary markers: The skill employs a 'counterfactual' requirement and an adversarial hypothesis (treating artifacts as defective until proven otherwise) to distinguish code behavior from embedded instructions.
Capability inventory: The agent can write review reports to the repository filesystem and perform direct edits to documentation and comment sections.
Sanitization: The skill requires the redaction of credentials and secrets before transmission to workers or inclusion in output, but it does not specify automated sanitization for general text inputs.

ts-review