The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses npx to dynamically fetch and run tools required for its operation.
Evidence: assets/templates/agents-sdk/python/codex_mcp_multiagent_workflow.py uses npx -y codex mcp-server to launch the Codex MCP server.
Evidence: assets/templates/agents-sdk/typescript/mcp-filesystem-example.ts uses npx -y @modelcontextprotocol/server-filesystem to access the filesystem via MCP.
[COMMAND_EXECUTION]: The skill is designed to execute shell commands but includes robust guardrails.
Evidence: references/safety-and-execpolicy.md outlines a threat model and suggests read-only sandboxing by default.
Evidence: assets/templates/execpolicy/default.rules provides a Starlark-based policy to block dangerous commands like sudo and rm -rf.
[PROMPT_INJECTION]: The skill provides patterns to mitigate indirect prompt injection from repository content.
Evidence: references/rag-and-memory.md and references/context-personalization.md discuss treating retrieved memories and RAG chunks as untrusted input and using explicit delimiters to isolate them.

codex-sdk