trash-truck
Pass
Audited by Gen Agent Trust Hub on Oct 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted repository data to nominate code for retirement.\n
- Ingestion points: Repository issues, PRs, commits, runtime logs, analytics (PostHog/Mixpanel), and CloudWatch logs as defined in SKILL.md and references/candidate-packet.md.\n
- Boundary markers: The skill uses stable markers for GitHub issues (e.g., trash-truck:retirement-review:v1) and requires investigation results to be normalized into structured JSON packets.\n
- Capability inventory: Interactive mode allows code deletion, infrastructure decommissioning, and job disabling based on the investigation outcome.\n
- Sanitization: All ranking logic is delegated to a deterministic local Python script (rank_candidates.py) that enforces strict eligibility rubrics, and the workflow requires explicit human selection of a specific manifest before execution.\n- [COMMAND_EXECUTION]: The skill executes a bundled local script (scripts/rank_candidates.py) via the command line to handle its internal ranking and validation logic. These calls use structured JSON inputs to perform decision-making logic in a deterministic environment.
Audit Metadata