Agent Reliability

Provides error monitoring, automatic fallback, confidence evaluation, and multi-agent voting to maintain accuracy in multi-step workflows.

This package appears designed to provide reliability tooling, but it contains multiple coherence issues that make it unreliable rather than malicious. Before installing or enabling it in any agent: - Don't deploy to production yet. Run the tests locally (npm test) in an isolated environment to observe failures. - Manually review the exported API surface: the tests call consensus.getSession and consensus.getStats and the ReportGenerator expects monitor.getStats(), but the codebase does not define these public methods consistently (ReliabilityMonitor provides getReliabilityScore(), VotingConsensus has no getSession/getStats accessor). These mismatches will cause runtime errors. - Inspect ReportGenerator.generateMarkdown: it expects stats.successRate, stats.total, stats.successful, stats.failed, and stats.byStep, but the monitor returns a different object shape — that will produce incorrect reports. - If you plan to use this library, either obtain a fixed release from a trusted source or patch the code to reconcile the API mismatches (add missing accessors or adjust callers). - Because there are no network calls or secret requests in the code, immediate credential exfiltration risk is low; however, the poor code quality could hide other bugs — review all modules before automated or autonomous agent use. If you want, I can list the exact mismatches and propose minimal patches to make the public API consistent with the tests/examples (e.g., add getStats/getSession wrappers or change ReportGenerator to consume getReliabilityScore()).