The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches a container image from quay.io/brexhq/crabtrap:latest and clones the source code repository from github.com/brexhq/CrabTrap to facilitate local deployment and building.
[COMMAND_EXECUTION]: Instructions involve using make build to compile the Go-based software and docker compose to manage the lifecycle of the proxy and its associated PostgreSQL database.
[INDIRECT_PROMPT_INJECTION]: The skill features a "Policy Builder" and "Eval System" that ingest and analyze historical agent traffic logs to automatically draft or test security policies. If the traffic being analyzed contains adversarial content, it could influence the generated security rules.
Ingestion points: Outbound HTTP request logs including URLs, headers, and payloads stored in the PostgreSQL audit log (SKILL.md).
Boundary markers: None explicitly stated for the automated generation process to distinguish between security policy instructions and logged traffic data.
Capability inventory: The system can block or allow network traffic, write to persistent audit logs, and modify its natural-language security policies based on judge decisions.
Sanitization: Documentation does not specify filtering or escaping mechanisms used when the policy builder agent processes raw traffic logs.

crabtrap-llm-proxy