The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from web search results, creating a potential surface for indirect prompt injection where malicious instructions on a website could influence the agent's behavior.
Ingestion points: The subagents in various 'lanes' (Canon, Ecosystem, Commercial, etc.) perform web searches and ingest content from external sources.
Boundary markers: The skill defines an 'Evidence Contract' and a specific 'Candidate row schema' (JSON) which provides structure to the ingested data, acting as a partial boundary against accidental execution of embedded instructions.
Capability inventory: The skill uses web search and file reading capabilities (restricted to project manifests).
Sanitization: The instructions emphasize verification of liveness and source URLs, but do not explicitly mandate sanitization or filtering of the text content for potential injection strings.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses local project manifest files such as package.json, requirements.txt, go.mod, Cargo.toml, and others to determine existing project dependencies.
This access is limited to standard configuration files and is essential for the tool's primary purpose of avoiding redundant library installation.
No access to sensitive files like .env, .ssh, or cloud provider credentials was detected.

melech-scout