deep-interview

Pass

Audited by Gen Agent Trust Hub on Oct 30, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes arbitrary user-supplied ideas and interview responses to generate technical specifications, which represents an attack surface for indirect prompt injection.
  • Ingestion points: The skill accepts untrusted data through initial CLI arguments and subsequent answers captured via the omx question tool in SKILL.md (Phase 2).
  • Boundary markers: It utilizes a structured JSON state model (Phase 1) and distinct execution phases to isolate and validate data at each round of the Socratic loop.
  • Capability inventory: The skill can perform read-only codebase exploration via explore and sparkshell and writes resulting specifications to the .omx/ project directory. It also facilitates handoffs to downstream execution skills.
  • Sanitization: Implements an "oversized initial-context summary gate" (Phase 0) to mitigate prompt-budget exhaustion and requires explicit user opt-in for the creation of any durable documentation or memory updates, ensuring a human-in-the-loop for public-facing artifacts.
  • [SAFE]: The skill's operational logic is focused on clarification and follows the principle of least privilege for its platform-native tool usage.
  • Evidence: Ambiguity thresholds and maximum round caps (Depth Profiles) prevent resource exhaustion and ensure the agent does not loop indefinitely.
  • Evidence: All repository interactions for "brownfield" tasks are specified as read-only fact-gathering, and the skill explicitly forbids direct implementation, routing work to dedicated execution skills instead.
  • Evidence: The handoff bridge (Phase 5) uses structured contracts to preserve intent and non-goals, preventing scope creep or misalignment during transitions.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 30, 2026, 07:52 PM