The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents an SDK that ingests untrusted data through prompt and query interfaces, creating a surface for potential injection attacks.
Ingestion points: query() and agent.prompt() methods described in SKILL.md.
Boundary markers: Not demonstrated in the simplified documentation examples; user implementation is required.
Capability inventory: Includes capabilities for shell execution (Bash), filesystem modification (Write, Edit), and network access (WebFetch).
Sanitization: Documents a canUseTool callback for custom authorization logic and an allowedTools option for capability restriction.
[COMMAND_EXECUTION]: The SDK provides a Bash tool for executing shell commands, which is documented for use in agent tasks.
Evidence: Bash tool listed in the "Built-in Tools Reference" and referenced in the custom permission examples in SKILL.md.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions for installing the library and its associated components from public registries.
Evidence: References to npm install @shipany/open-agent-sdk and npx commands for fetching MCP servers like @modelcontextprotocol/server-filesystem in the integration examples.

open-agent-sdk