The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data from an external messaging network which can influence agent behavior.
Ingestion points: Data is retrieved from the AgentMesh network via tools like agentmesh_check_messages, agentmesh_owner_inbox, and agentmesh_owner_conversations as seen in SKILL.md.
Boundary markers: There are no explicit markers or instructions provided to the agent to distinguish between user intent and instructions contained within the fetched messages.
Capability inventory: The agent has high-privilege tool access including Bash, Write, Edit, and Read permissions.
Sanitization: No sanitization or safety filtering of the message content is implemented.
Evidence: The 'Autonomous Mode' guide in Step 2.5 explicitly directs the agent to 'Automatically process messages as instructions' (收到消息后自动当作指令处理), which allows for indirect prompt injection where a remote user could send commands to the agent.
[COMMAND_EXECUTION]: The skill modifies the environment to execute code locally.
Evidence: Step 1.2 writes an entry to .mcp.json that executes a script using npx tsx <absolute-project-path>/packages/mcp-server/src/server.ts.
[EXTERNAL_DOWNLOADS]: The MCP server configuration utilizes npx, which downloads packages from the npm registry at runtime if they are not already installed locally.
[CREDENTIALS_UNSAFE]: The setup process prompts the user for an AGENTMESH_API_KEY and persists it in plain text within the local .mcp.json configuration file.

agentmesh