The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents the use of LocalWorkspace and CodeExecutionTool for running shell commands and Python code. It provides clear architectural guidance and explicit security warnings, noting that LocalWorkspace lacks isolation and is intended only for trusted workloads, and advising against exposing sensitive environment variables to the model.
[EXTERNAL_DOWNLOADS]: The skill references standard Pydantic ecosystem tools, documentation at pydantic.dev, and official model providers. It includes instructions for setting up observability using logfire through official CLI tools and package managers.
[INDIRECT_PROMPT_INJECTION]: The skill describes a framework for building agents that ingest untrusted data through prompts, tool results, and enqueued messages.
Ingestion points: Data enters the agent context via agent.run() arguments, ToolReturn content, and ctx.enqueue() for mid-run steering.
Boundary markers: The skill mentions provider-native mechanisms like <system> tags for enqueued instructions to distinguish them from user content.
Capability inventory: The framework supports powerful capabilities including LocalWorkspace for shell/file access, CodeExecutionTool for Python execution, and WebSearch for network operations.
Sanitization: The skill promotes the use of Pydantic BaseModel and output_type validation to ensure structured outputs adhere to a strict schema, and warns users not to execute commands found in telemetry data.
[CREDENTIALS_UNSAFE]: The documentation covers the management of API keys for various providers and the Pydantic AI Gateway. It correctly identifies the risks associated with hardcoding secrets and provides best practices for secure configuration using environment variables and local project directories.

building-pydantic-ai-agents