The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The canvas inspector script reads diagnostic data from the web page's global scope (window.__THREE_APP_DIAGNOSTICS__). If run against a malicious or compromised web page, this data could theoretically contain instructions intended to influence the agent's release decisions.
Ingestion points: scripts/inspect-threejs-canvas.mjs captures data from window.__THREE_APP_DIAGNOSTICS__ and window.__THREE_GAME_DIAGNOSTICS__ on the target page.
Boundary markers: The script outputs structured JSON to the console and to a file, which helps isolate the data, but no explicit instructions are provided to the agent to treat this data as untrusted.
Capability inventory: The skill allows the agent to execute shell commands (running the inspector), write artifacts to the local file system, and interact with web pages via Playwright.
Sanitization: The collected data is serialized using JSON.stringify before being presented to the agent.
[COMMAND_EXECUTION]: The skill provides and instructs the agent to run various shell commands for building projects, running tests, and executing the provided canvas inspector script.
Evidence: Examples include node scripts/inspect-threejs-canvas.mjs, pnpm exec playwright test, and cp tests/bot-playtest.template.ts tests/bot-playtest.spec.ts found in SKILL.md and referenced documentation.
[EXTERNAL_DOWNLOADS]: The use of Playwright for testing and inspection may involve downloading browser binaries if they are not already present in the environment.
Evidence: scripts/inspect-threejs-canvas.mjs imports @playwright/test from the host project's own dependencies to launch a browser instance.

threejs-qa-release