review-fix-signoff-loop

Pass

Audited by Gen Agent Trust Hub on Aug 15, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill implements a structured workflow pattern for high-stakes code implementations. It includes deterministic preflight checks and reconciliation steps to ensure repository state consistency.
  • [COMMAND_EXECUTION]: The skill uses standard development CLIs including git, npm, and gh for repository management, package inspection, and pull request interaction. These tools are used appropriately for their intended purposes in a development environment.
  • [DATA_EXFILTRATION]: No evidence of sensitive data harvesting or unauthorized network transmission was found. Network interactions are limited to well-known services like GitHub and NPM registries.
  • [PROMPT_INJECTION]: The instructions do not contain any patterns typical of prompt injection, such as attempts to bypass safety filters, extract system prompts, or override agent constraints.
  • [INDIRECT_PROMPT_INJECTION]: While the skill processes external data (such as pull request bodies and validation logs), it implements a strict 'Verdict Contract' and deterministic dual-signoff gates to manage the risk of untrusted input influencing the workflow's state transitions.
  • Ingestion points: Reads repository files, diffs, validation logs, and GitHub PR metadata via gh pr view.
  • Boundary markers: Utilizes a strict text contract for verdicts (e.g., VERDICT: COMPREHENSIVELY_SATISFIED) to ensure machine-parsable outcomes.
  • Capability inventory: Includes capabilities for repository modification (git), pull request commenting (gh), and tool execution (codex exec).
  • Sanitization: Relies on deterministic dual-signoff verification and file_exists checks for durable artifacts before proceeding.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 15, 2026, 10:32 PM
Security Audit — agent-trust-hub — review-fix-signoff-loop