The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The primary SKILL.md configures a DevEco MCP server using the command npx -y @deveco-codegenie/mcp. This instruction forces the agent to download and execute the latest version of the specified package from the npm registry at runtime, introducing an external code dependency.
[COMMAND_EXECUTION]: The initialization scripts (scripts/init.sh and scripts/init.ps1) are designed to run automatically when the skill is first loaded. These scripts perform filesystem operations, create directories, and modify local configuration files. On Windows, the initialization is performed using powershell -ExecutionPolicy Bypass, which circumvents local script execution restrictions.
[PROMPT_INJECTION]: The skill setup process automatically modifies project-level configuration files for various AI tools, including .claude/CLAUDE.md, .cursor/rules/deveco-flow.mdc, and .windsurfrules. It injects "MANDATORY" instructions and self-check rules into these files to ensure the agent follows specific HarmonyOS development constraints, which effectively overrides or biases the agent's governing instruction set for the workspace.
[PROMPT_INJECTION]: The skill exhibits a significant surface for indirect prompt injection (Category 8). It is designed to read and analyze untrusted source code and project files (such as Index.ets and contents of the .knowledge/ directory) and interpolate them into prompts for analysis, planning, and coding phases. This ingestion of external data combined with powerful capabilities like command execution (hvigorw, hdc, xcodebuild) and file writing creates an exploitable trust chain.
[COMMAND_EXECUTION]: The harmony-verify sub-skill utilizes wrapper scripts (hdc.sh, hdc.ps1) to execute powerful device management commands. These scripts allow the agent to manage applications, extract logs, and capture screenshots from connected HarmonyOS devices or emulators.

deveco-native-flow