light-decide
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes structured data from other skills and external decision backends, creating an injection surface.\n
- Ingestion points: The skill takes input from callers (question, options, and context) and from external sources (decision backends providing answers and probabilities).\n
- Boundary markers: Explicit instructions in SKILL.md require the agent to build a "minimal state" and exclude source code, diffs, credentials, and personal data from the payload sent to sources.\n
- Capability inventory: The skill can write logs to .scratch/, modify the project's docs/agents/light-project.md, update local user configuration in ~/.config/light/, and execute system commands for hardware profiling.\n
- Sanitization: External input is subject to normalization and policy application. Malformed or low-confidence responses trigger a fallback to internal self-judgment rather than immediate execution.\n- [COMMAND_EXECUTION]: The skill uses shell commands for environment discovery and logging timestamps.\n
- Evidence: references/setup.md directs the agent to check the host machine's memory, CPU/GPU, and disk space using commands to recommend model setups. Additionally, references/formats.md specifies using
date -u +%Y-%m-%dT%H:%M:%SZfor log entries. These are benign utility operations used for profiling and auditing.
Audit Metadata