The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides documentation and examples for executing shell commands via the Pi agent's bash tool and ! / !! syntax. It explicitly notes that Pi runs without a sandbox by default and inherits the user's permissions.
[EXTERNAL_DOWNLOADS]: The skill documents the installation of the Pi agent via a remote script (https://pi.dev/install.sh) and via npm. It also describes the system for installing additional agent extensions and skills from npm, git, and remote URLs.
[INDIRECT_PROMPT_INJECTION]: The skill describes an indirect injection surface through the pi-web-access extension, which allows the agent to ingest untrusted data from web searches, URLs, and document conversions.
Ingestion points: web_search and fetch_content tools documented in pi-web-access.md.
Boundary markers: serializeConversation() renders tool results with [Tool result]: delimiters as described in compaction.md.
Capability inventory: Access to bash, write, edit, and subagent tools across multiple reference files.
Sanitization: Output truncation (50KB / 2000 lines), SSRF validation, and hostname domain policies are described in pi-web-access.md and pi-mcp-adapter.md.
[DYNAMIC_EXECUTION]: The skill documents the use of runtime-evaluated JavaScript and TypeScript through features such as workflowScript in pi-subagents.md, mcpScript in pi-mcp-adapter.md, and the dynamic loading of TypeScript extensions using jiti in extensions.md.
[DYNAMIC_CONTEXT_INJECTION]: The skill describes the ! command syntax supported in models.json and auth.json configurations, which allows for shell command execution at request time to resolve secrets or headers.

pi-agent