claude-codex-consensus
Pass
Audited by Gen Agent Trust Hub on Jun 28, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [SAFE]: The skill implements a structured peer-review process between agents, utilizing specific reconciliation rules and templates to manage disagreements without introducing suspicious logic.\n- [COMMAND_EXECUTION]: The skill invokes the
agentCLI utility (e.g.,agent opus,agent gpt) to facilitate communication with other models. This is the intended mechanism for the consensus workflow.\n- [PROMPT_INJECTION]: The skill exhibits a potential surface for indirect prompt injection because it ingests project data and passes it to another model for review.\n - Ingestion points: The workflow in
SKILL.mdrequires the agent to gather relevant files, diffs, and work orders from the local environment.\n - Boundary markers: The templates provided in
references/review-prompts.mddo not include explicit delimiters or instructions to prevent the reviewing agent from being influenced by instructions embedded within the ingested data.\n - Capability inventory: The
agentCLI allows the secondary agent to write findings to review artifacts on the filesystem.\n - Sanitization: No sanitization of the input content is described before it is interpolated into the prompts for the second agent.
Audit Metadata