The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes and interprets external .tape files which contain shell commands. This creates a surface where malicious instructions could be embedded in files ingested from untrusted sources.\n
Ingestion points: The source: and tape: fields in the type: cast workflow step, and the atmos cast record <input.tape> command-line entry point.\n
Boundary markers: The skill notes that tape syntax is interpreted directly in memory and that double-quoted strings are literal (no backslash escapes), which provides some distinction from standard templating but does not prevent adversarial injection.\n
Capability inventory: The interpreter can execute shell commands (via Type directives), write files (Output, Screenshot), and modify environment variables (export, unset).\n
Sanitization: Documentation specifies that Atmos-native templating is not applied to tape-derived steps, though commands are eventually executed in a PTY-backed shell environment.\n- [DYNAMIC_EXECUTION]: The skill implements a runtime interpreter for the VHS tape DSL, which includes capabilities for dynamic file inclusion and command generation.\n
The Source <file> directive allows for recursive inlining of external tape files at runtime, resolved relative to the working directory.\n
The implementation involves dynamic translation of DSL directives into shell actions across both mode: session and mode: steps execution paths.

atmos-vhs