The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (such as investigation reports, source code, or issue summaries) to generate HTML documents. While it enforces strict sanitization—prohibiting JavaScript, event handlers, and external CSS dependencies—the ingestion of external content is a known surface for indirect prompt injection.
Ingestion points: The Objective, Scope, and Context fields in the request, which reference external files or URLs as source material.
Boundary markers: The skill uses structured Markdown for requests and instructions to sub-agents, encouraging the use of file paths over direct content injection.
Capability inventory: Utilizes the Bash tool to execute internal scripts for managing the delegation process.
Sanitization: Explicitly forbids script elements, event handler attributes, and javascript: URLs in the generated output, and uses a locked CSS template.
[EXTERNAL_DOWNLOADS]: The model-token-prices.json file contains reference URLs for pricing information from well-known AI service providers, including Anthropic, OpenAI, and Cognition. These are static references for cost analysis and do not trigger runtime downloads or code execution.
[COMMAND_EXECUTION]: The skill includes multiple shell scripts in the scripts/ directory that serve as wrappers for a local Node.js CLI tool (delegate-cli.mjs). These scripts are invoked via the Bash tool to handle task preparation, dispatching to sub-agents, and processing responses.

delegate-htmldoc