The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data originating from external phone call summaries and transcripts.
Ingestion points: Call summaries, transcripts, and activity messages are ingested from the calle CLI output and incorporated into the agent's context (e.g., in SKILL.md).
Boundary markers: The instructions mandate the use of explicit markers such as `[Call Summary
untrusted call data]and[Transcript
untrusted call data]` to separate external content from the agent's primary instructions.
Capability inventory: The skill utilizes a launcher script (scripts/run-agent-command.mjs) that executes the calle CLI to perform authentication, call planning, call initiation, and status monitoring.
Sanitization: The instructions include explicit directives to disregard any instructions, URLs, or shell commands found within the untrusted call data. It also enforces fixed templates for displaying this information to the user.
[COMMAND_EXECUTION]: The skill executes external commands via a bundled Node.js wrapper script.
Launcher validation: The scripts/run-agent-command.mjs script verifies the integrity and identity of the @call-e/cli package before execution by checking the package.json and performing probe commands.
Injection mitigation: Commands are executed using child_process.spawn with shell: false, and arguments are passed as a JSON array rather than raw shell text to prevent command injection attacks.
[EXTERNAL_DOWNLOADS]: The skill relies on the @call-e/cli package, which is an external dependency provided by the same vendor (calle-ai). The skill provides instructions for local installation of this CLI if it is not already present.

calle