call-capture

Pass

Audited by Gen Agent Trust Hub on Oct 5, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes npx tsx to run its internal data collection scripts as part of its scheduled and manual workflows.
  • [EXTERNAL_DOWNLOADS]: The skill communicates with official APIs of well-known call recording services (such as Avoma, Gong, Fireflies, and Fathom) to retrieve meeting transcripts and notes.
  • [DYNAMIC_EXECUTION]: The skill uses npx tsx at runtime to execute its own TypeScript logic for processing recorder data, which is a standard pattern for this type of agent skill.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted call transcript data which is then summarized and stored in the repository.
  • Ingestion points: External transcripts are retrieved in scripts/collect/recorder.ts through the various recorder adapters.
  • Boundary markers: The system prompt in infra/agents/call-scribe.prompt.ts instructs the agent to use a specific structured markdown format and to quote transcripts verbatim for contentious points.
  • Capability inventory: The skill possesses repository write access via the GitHub connector and network access to specific recorder service domains.
  • Sanitization: The skill mitigates risks from untrusted content through a mandatory human review gate (Pull Requests) and a 'repetition bar' logic, which requires a claim to be heard across multiple independent calls before it is promoted to the core knowledge base.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 5, 2026, 08:46 PM