The Agent Skills Directory

[PROMPT_INJECTION]: The file references/model-flags.md contains the phrase CRITICAL INSTRUCTION FOR AGENT, and references/provider-cli.md contains a section titled 给 Agent 的硬性要求 (Hard requirements for Agent). These are patterns used to direct or override the agent's default processing behavior.
[COMMAND_EXECUTION]: The script scripts/gen.mjs and the provider library scripts/lib/providers.mjs use child_process.spawn to run external CLI tools. The specific binary executed can be controlled via the DLAZY_BIN environment variable, which presents a risk if the environment is compromised.
[CREDENTIALS_UNSAFE]: The skill relies on sensitive API keys (e.g., DLAZY_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY, FAL_KEY) retrieved from environment variables. Documentation in references/provider-cli.md also indicates that credentials may be stored locally in ~/.dlazy/config.json.
[EXTERNAL_DOWNLOADS]: The persist function in scripts/gen.mjs uses fetch to download data from arbitrary URLs provided as image inputs and writes the resulting buffers to the local file system.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data, such as user prompts and image assets, which are interpolated into command arguments and API requests without explicit sanitization or boundary markers, creating a surface for indirect injection attacks.
[DYNAMIC_EXECUTION]: The code dynamically constructs command-line arguments for the dlazy CLI tool based on user input and configuration files, executing them at runtime.

item-change-background