The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a Python script scripts/digest_state.py to maintain local execution records in the .octave/digests/ directory. The script employs security best practices, such as checking for symlinks using unsafe_symlink and sanitizing path components with a component() helper to prevent path traversal vulnerabilities.
[DATA_EXFILTRATION]: The skill handles sensitive business data including CRM records and call transcripts. It implements a strict multi-stage "Privacy Gate" and "Evidence-Worthiness Gate" (detailed in references/evidence-and-links.md) to prevent unauthorized exposure of private details like company names, deal amounts, or verbatim quotes to external audiences. Links are correctly restricted to the vendor's official domain (app.octavehq.com).
[INDIRECT_PROMPT_INJECTION]: The skill has a defined attack surface as it ingests untrusted external data.
Ingestion points: External report data, CRM records, and call transcripts are ingested via references/content-packet.md and references/investigation-method.md.
Boundary markers: The skill uses explicit "Editorial Review" protocols and "Output Readiness" checks to ensure content is verified.
Capability inventory: Local filesystem writes (scripts/digest_state.py) and read-only analytical tool calls (references/research-tools.md).
Sanitization: Strict attribution rules and a mandatory review protocol in references/editorial-guidance.md ensure the agent distinguishes between observed evidence and instructions.
[SAFE]: All external references are limited to the vendor's own infrastructure and standard analytical operations. The inclusion of defensive code in helper scripts indicates a high security standard in development.

digest