krt-muse-artificer
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The script
scripts/run_muse_implementer.pydynamically modifies the Python system path using the--seneschal-skill-dirargument to import integration modules (render_worker_envelope,worker_contract). This is an expected modular design pattern for components within the vendor's ecosystem.\n- [COMMAND_EXECUTION]: The runner script invokes external binaries, specifically themuseCLI and a Python-based validation script, usingsubprocess.runto perform code implementation and verify results. It includes a check to ensure the CLI supports required headless flags before execution.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes structured JSON data from worker contracts which are used to generate the model's prompt, creating an injection surface.\n - Ingestion points: The
--contractJSON file processed inscripts/run_muse_implementer.py.\n - Boundary markers: The generated prompt includes explicit restrictive instructions (e.g., 'Do not stage, commit, switch branches') and a JSON schema for output to constrain the worker model's actions.\n
- Capability inventory: Shell command execution via
muse execand file system access within the Git workspace.\n - Sanitization: The script performs validation on the contract profile and lane, and enforces that task metadata and log paths remain outside the active Git worktree.
Audit Metadata