Hermes Labyrinth is a read-only observability dashboard plugin for the Hermes Agent. Instead of a chat interface, it acts as a black-box recorder that reconstructs an agent's autonomous work into a navigable map, surfacing prompts, tool calls and results, failures, model switches, subagents, approvals, memory hits, redactions, context compression, and scheduled runs. It is aimed at anyone who needs to inspect and audit what a Hermes agent actually did across a run.
It installs into the Hermes user plugin directory and appears as a Labyrinth tab in the Hermes dashboard, reading local Hermes state (session/message database, skill directories, cron config) through a read-only Python API. Its main surfaces include a journey index of recent CLI, dashboard, gateway, cron, and delegated work; a labyrinth map of the ordered steps in a journey; an inspector showing a step's input, output, duration, status, and evidence; generated guideposts backed by that evidence; a skill atlas that separates effective skills from shadowed overrides and true duplicates; a cron view of scheduled runs; a model-transition view; and redacted Markdown/JSON report exports per journey.
Safety is a stated design constraint: the plugin never starts, stops, resumes, mutates, or creates Hermes sessions. Secret redaction is applied to summaries, previews, and reports, and if the Hermes core redactor cannot load or throws, it fails closed and shows a redaction-unavailable marker rather than raw trace text. The documentation recommends testing against a dedicated Hermes home and verifying that dummy secrets are absent from every surface before pointing it at production data.
The project is a hackathon preview (v0.1.3) described as stable enough to demo and install, with a public GitHub Pages demo running on mocked state. Verification is via an npm test suite covering build reproducibility, frontend/backend parse checks, API-normalization fixtures, and headless-Chrome smoke tests; fuller Hermes dashboard integration tests are noted as still on the roadmap.
