The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The framework uses JSDoc tags to trigger the download and installation of external resources.
The @dependencies tag in references/docblock-tags.md and SKILL.md allows for the automatic installation of NPM packages (e.g., axios, cheerio, lodash).
The @cli tag enables the specification of system tool dependencies that can be downloaded from arbitrary URLs (e.g., `git
https://git-scm.com/downloads`).
The @mcp and @photon tags allow the framework to fetch and load external MCP servers and photons dynamically from registries or repositories.
[REMOTE_CODE_EXECUTION]: The skill facilitates the creation and execution of code that can be updated or extended at runtime.
Photons are executed directly using tsx via the photon runtime, which means any code generated by an agent using this skill can be immediately run on the host system.
The this.schedule.create API and @cron tags allow for persistent execution of code on a schedule, providing a mechanism for long-term persistence of potentially malicious scripts.
The bridge in references/mcp-apps.md injects communication APIs into sandboxed iframes, which, while sandboxed, still allow for allow-same-origin and allow-scripts permissions.
[COMMAND_EXECUTION]: The skill instructions guide the agent to perform high-level CLI operations that modify the system environment.
Instructions include global installation of the runtime (npm install -g @portel/photon) and management of background daemons.
[DATA_EXFILTRATION]: The dependency injection system automatically maps environment variables to constructor parameters.
references/dependency-injection.md describes how private apiKey: string maps to PHOTON_NAME_APIKEY, which could lead to accidental exposure of credentials if the generated photon logic or logging is compromised.
[INDIRECT_PROMPT_INJECTION]: Several examples provided in the skill demonstrate the ingestion of untrusted data from external sources.
Example 2 in references/examples.md is a web scraper that fetches data from user-provided URLs.
Example 3 in references/examples.md fetches GitHub issues and passes them directly to an AI model for categorization.
Ingestion points: WebScraper.run (URL input), IssueTriage.run (GitHub repo input).
Boundary markers: No explicit sanitization or strict boundary markers are shown in the example code snippets to prevent malicious data from influencing the agent's logic.
Capability inventory: The framework has access to the filesystem (via @mcp fs), network (via fetch), and task scheduling (this.schedule).

photon