claude-code-loop-harness
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of multi-agent harnesses that ingest and process external data, such as repository source code for auditing or PR feedback for automated fixes.
- Ingestion points: Workflow workers are instructed to gather context from source files (e.g.,
src/routes) and external PR content as described inSKILL.mdandreferences/architecture-and-routing.md. - Boundary markers: The skill requires the creation of a 'frozen task contract' before execution, which explicitly defines the objective, allowed scope, and restricted network/write paths.
- Capability inventory: The orchestration scripts and workers have capabilities for file system access (read/write), worktree management for isolation, and shell command execution for planning and validation.
- Sanitization: The skill mandates 'independent verifiers' that check worker output against the frozen contract and requires a 'human gate' for all consequential actions like merging, pushing, or deploying.
- [COMMAND_EXECUTION]: The instructions direct the agent to execute a local Node.js script (
scripts/plan-harness.mjs) to validate harness configurations and generate deterministic plans. The script uses standard Node.js built-ins (node:fs,node:crypto) to perform validation and integrity hashing without external network dependencies.
Audit Metadata