bootstrap-agentic-system
Pass
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill defines agents (Implementor, Tester) that utilize terminal execution tools (execute/runInTerminal) and file editing capabilities. These operations are restricted to the target repository environment and are governed by a multi-stage approval process, where implementation must follow a plan containing explicit approval metadata.
- [DATA_EXFILTRATION]: During the discovery phase (Phase B), the system reads local repository metadata (git remotes, issue structures, MCP configurations) to configure the generated agents. The discovered information is stored locally in the agentic-system.answers.yaml and manifest files to ensure local provenance and re-rendering capability.
- [SAFE]: The skill utilizes a 'Copy-First' generation mechanism to create agent contracts and repository instructions from internal mirrors located in the templates/ directory. This process is strictly controlled by user-approved file plans and batch write checkpoints, preventing arbitrary code generation.
- [SAFE]: Ingestion points for potentially untrusted data include the work item retrieval process (Planner agent) and image analysis (Vision agent). The skill mitigates risks by requiring agents to follow a 'Knowledge Authority' rule, which forces adherence to normative rules in local knowledge files over instructions found in external data, and by using structured schemas to enforce artifact consistency.
Audit Metadata