The Agent Skills Directory

[COMMAND_EXECUTION]: The skill scaffolds a shell tool capable of executing arbitrary system commands using the user's default shell.
Evidence: sample/src/tools/shell.ts uses Bun.spawn to execute command strings provided as tool input.
Context: This capability is a primary feature of the generated coding agent and includes timeout and output truncation logic to manage execution risk.
[DYNAMIC_EXECUTION]: Includes an optional JavaScript/TypeScript REPL tool and utilizes dynamic imports for validation logic.
Evidence: references/tools.md defines a js_repl tool that maintains a long-lived Bun evaluation process.
Evidence: sample/src/cli.ts dynamically imports the ajv library for JSON Schema validation.
[INDIRECT_PROMPT_INJECTION]: The generated agent is designed to ingest data from local files and remote web pages, creating a surface for indirect prompt injection.
Ingestion Points: Tools like file_read, grep, and web_fetch (described in references/tools.md).
Sanitization: The skill provides specific implementation requirements for web_fetch to prevent SSRF by validating resolved IP addresses against private and reserved ranges.
Capability Inventory: The agent possesses file-write and shell-execution capabilities, but the skill documentation includes modules for tool approval flows to mitigate autonomous abuse.

create-headless-agent