The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The scripts/probe-lib.mjs library uses new Function(step.code) to execute arbitrary JavaScript code within the mini-program's automation context. This code is sourced from plan.json, which is generated by the agent based on potentially untrusted mini-program source code, creating a path for code execution based on external input.
[COMMAND_EXECUTION]: The skill instructions in SKILL.md require the execution of Node.js scripts (scripts/check-artifacts.mjs and scripts/probe.mjs) to perform project validation and runtime analysis. This grants the skill the ability to run local scripts with the user's permissions.
[DATA_EXFILTRATION]: The runtime probing mechanism (scripts/probe.mjs) captures full network requests and responses from the mini-program, including potentially sensitive authentication headers (tokens) and business data payloads. This data is extracted from the mini-program's runtime and saved into local files within the .ai-mode-skills/ directory.
[PROMPT_INJECTION]: The skill is designed to ingest and analyze entire mini-program source codebases. This serves as an ingestion point for untrusted data, where maliciously crafted code files could influence the agent's logic during the analysis or code generation phases (e.g., forcing the generation of malicious triggers in the probing plan).

wxa-skills-generate