coding-agent-harness
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze session transcripts, history logs, and execution records which could contain instructions from untrusted sources or previous agent turns. Ingestion points: SESSION-DIAGNOSIS.md and OPERATOR-RUNS.md establish workflows for reading session history, worker transcripts, and recorded evidence. Boundary markers: The skill requires distinguishing between human prompts, injected instructions, and tool results, and recommends using a structured TSV format for logging decisions to maintain a clear audit trail. Capability inventory: The harness enables the execution of project-level build/test commands (INSTRUCTIONS.md) and the spawning of subagents via dynamic JavaScript-based workflows (CLAUDE-WORKFLOWS.md). Sanitization: SESSION-DIAGNOSIS.md specifies a comprehensive scrubbing and redaction policy to replace secrets, account identifiers, and proprietary terms with placeholders before sharing session data.
- [COMMAND_EXECUTION]: Instructions in INSTRUCTIONS.md direct the agent to execute project-specific commands such as tests, builds, and type-checkers to verify the correctness of code changes before reporting them as complete.
- [EXTERNAL_DOWNLOADS]: The INSTRUCTIONS.md file provides guidance for the agent to install missing project-level dependencies using standard tools such as npm install or pip install when necessary for building or testing.
Audit Metadata