Hermes Labyrinth is a dashboard plugin for Hermes Agent that records and visualises what an agent did during autonomous work. It reads local Hermes state (the sessions and messages in state.db, skills directories, and cron configuration) and presents each journey as an ordered series of crossings: prompts, tool calls, tool results, failures, model switches, subagents, approvals, memory hits, redactions, context compression, and cron runs. It is not a chat interface; it never starts, stops, resumes, or mutates a session.
The plugin installs into ~/.hermes/plugins and appears as a Labyrinth tab in hermes dashboard, with a Python API mounted at dashboard startup and a built static frontend. Secret redaction is applied to summaries, previews, and reports through the Hermes core redactor; if that redactor cannot load, the plugin fails closed and shows [redaction unavailable] instead of raw trace text. A public GitHub Pages demo runs on mocked state. Version 0.1.3 is a hackathon preview.
Features
- Journey index: recent CLI, dashboard, gateway, cron, and delegated work
- Labyrinth map: ordered crossings through a selected journey
- Inspector: input, output, duration, status, evidence, and guideposts for a crossing
- Guideposts: generated observations backed by local evidence
- Skill atlas: effective skills, shadowed overrides, true duplicates, and scan errors reported separately
- Cron gate: scheduled autonomy with next runs, last failures, and working directories
- Model ferry: model and provider transitions across sessions
- Reports: redacted Markdown and JSON exports per journey
- Read-only API: GET endpoints for health, journeys, crossings, skills, cron, guideposts, and reports
- Verification: build checks, fixture tests for API normalisation, and headless Chrome smoke tests
