start-workflow
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from local project files like
workflows.jsonand files specified inpreActions.readFiles. These files could theoretically contain malicious instructions if the repository itself is untrusted. - Ingestion points: Reads from
workflows.json,preActions.readFiles, andpreActions.injectContext(SKILL.md). - Boundary markers: Implements mandatory todo tracking for all steps and an explicit "workflow question" tool to obtain user approval before activating catalog-matched workflows.
- Capability inventory: Can execute local Node.js scripts, perform file system writes in
tmp/, and spawn sub-agents via thespawn_agenttool. - Sanitization: Workflow manifests are resolved through a dedicated JSON-aware script (
read-workflow-entry.mjs) rather than direct file parsing, providing a layer of validation. - [COMMAND_EXECUTION]: The skill executes local Node.js scripts to manage workflow state and manifest resolution.
- Evidence: Specifically calls
node .claude/scripts/codex/read-workflow-entry.mjs,.claude/scripts/lib/workflow-manifest.cjs, and.claude/scripts/lib/workflow-baseline.cjs. - Context: These are project-internal utilities used for infrastructure management; they are executed with arguments passed as data rather than interpolated shell strings, reducing injection risk.
Audit Metadata