handyman
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process and summarize potentially untrusted content such as project source code, web pages, and previous agent reports. This creates an inherent attack surface for indirect prompt injection where malicious instructions could be embedded in data files.
- Ingestion points: The agent reads from
feature_list.json,memory/*,progress/*,backlog/*, and external web content (via the leader role). - Boundary markers: The skill includes a dedicated
references/security.mdfile defining a "data-not-instructions" boundary and role templates (e.g.,role-leader.template.md) that explicitly instruct agents to ignore directives embedded in ingested data. - Capability inventory: High-privilege tools are available, including file modification (
edit), shell execution (execute), and web browsing (web,browser). - Sanitization: Roles are instructed to report content as quoted observations rather than executing instructions found within them.
- [COMMAND_EXECUTION]: The skill's primary methodology involves agents executing a project-local verifier script (
./init.sh) and other CLI tools to validate environment state and test results. - Evidence: Found in
SKILL.md,references/workflow.md, and thefeature.tsimplementation. - Context: This behavior is fundamental to the skill's purpose of providing an auditable and verified workflow.
- [DYNAMIC_EXECUTION]: During setup, the skill generates executable shell scripts from local templates.
- Evidence:
scripts/scaffold.shcreates aninit.shverifier script from a bundled template and applies executable permissions. - Context: This is a routine setup action used to establish the project's testing and verification harness.
- [EXTERNAL_DOWNLOADS]: The skill makes use of external software packages hosted on public registries.
- Evidence: Instructions direct the use of
npx handyman-harness@3to invoke the toolchain. - Context: The package is provided by the skill author and is hosted on the public NPM registry, a well-known service.
Audit Metadata