{ "batch": "read2-019", "plugins": [ { "pr": 0, "name": "done-bell", "does": "Notifies the user when Hermes finishes a turn or needs approval, with OS alerts, a chime, and a Desktop status panel.", "useful": 4, "reliable": 3, "polish": 3, "creative": 2, "beyond": 4, "total": 20, "best": "Uses actual lifecycle hooks and Desktop busy state, with bounded event history and tests that verify approval and mute behavior.", "worst": "Setup still includes an author's local Windows install path and manual Desktop file copying, while tests suppress OS notifications rather than verifying delivery.", "standout": false, "blocked": null }, { "pr": 0, "name": "homeassistant", "does": "Connects Hermes to Home Assistant through a filtered WebSocket event gateway, device-control tools, and notification delivery.", "useful": 4, "reliable": 5, "polish": 4, "creative": 3, "beyond": 5, "total": 26, "best": "Combines scoped secrets, reconnects, timeouts, validated service names, dangerous-domain blocking, operator entity restrictions, and targeted policy and adapter tests.", "worst": "Entity restrictions are opt-in and the default token-backed toolset allows broad device control; users must configure the lists and event subscriptions carefully.", "standout": false, "blocked": null }, { "pr": 153861, "name": "summon", "does": "Lets an owner summon their Hermes in X threads, read thread media, post replies, and request skills through Hermes's approval workflow.", "useful": 4, "reliable": 4, "polish": 4, "creative": 3, "beyond": 5, "total": 25, "best": "A genuine platform adapter includes user-ID filtering, scoped credentials, secret/contact-text guards, safe-tool setup, persistence, and fake-X tests.", "worst": "The budget check is before an answer, not each charged post or contextual read, so a multipart answer or thread expansion can exceed the nominal daily cap.", "standout": false, "blocked": null }, { "pr": 150556, "name": "hermes-ui-cards", "does": "Adds validated interactive cards to Hermes Desktop and Webapp chat, returning user choices as conversation messages with text fallbacks elsewhere.", "useful": 5, "reliable": 4, "polish": 5, "creative": 4, "beyond": 5, "total": 28, "best": "Seventeen card types, readable UI source, shared schema fixtures, receipts, undo, compaction tracking, and extensive Python and Desktop regression tests make this a broad improvement to daily Hermes use.", "worst": "Cross-device receipts are eventual and stale or duplicate confirmation protection relies on model warnings, so simultaneous clients can still submit twice.", "standout": true, "blocked": null }, { "pr": 153724, "name": "roomsmith", "does": "Provides a shared 3D room editor and agent tools for furniture placement, geometric fit checks, product-page dimensions, budgets, revisions, and export.", "useful": 3, "reliable": 4, "polish": 5, "creative": 4, "beyond": 5, "total": 26, "best": "The user and agent edit the same revision-checked SQLite scene, backed by real geometry, price calculations, UI source, and focused hardening tests.", "worst": "Product-fetch SSRF protection resolves hosts for checking and then lets urllib resolve again for the connection, leaving a DNS-rebinding gap despite the documented public-address restriction.", "standout": false, "blocked": null }, { "pr": 149202, "name": "3ds-platform", "does": "Adds an encrypted Nintendo 3DS gateway and homebrew client with streaming chat, media, voice, sessions, and interactive Hermes approvals.", "useful": 3, "reliable": 4, "polish": 5, "creative": 5, "beyond": 5, "total": 27, "best": "A complete hardware-facing platform includes readable console source, Noise transport, expiring pairing, revocation, native approvals, and protocol regression tests.", "worst": "Requiring a homebrew-enabled Nintendo 3DS and a separate console installation sharply limits its audience and adds substantial setup beyond installing the gateway plugin.", "standout": false, "blocked": null }, { "pr": 151887, "name": "cmi", "does": "Replaces Hermes context selection with deterministic shortening of stale tool output, recoverable originals, cache-aware selection, and experimental project memory.", "useful": 5, "reliable": 4, "polish": 4, "creative": 4, "beyond": 5, "total": 27, "best": "A native context engine improves the agent itself without extra model calls, with private bounded recall storage and extensive integrity, cache, and project-boundary tests.", "worst": "Private core imports and direct project/session database reads create compatibility risk, and heuristic shortening is not evidence that answer quality stays unchanged.", "standout": false, "blocked": null }, { "pr": 153491, "name": "atomicmail", "does": "Gives Hermes an Atomic Mail inbox, an allowlisted DKIM-checked email gateway, and email and calendar-invitation tools.", "useful": 4, "reliable": 4, "polish": 4, "creative": 3, "beyond": 5, "total": 25, "best": "Implements the channel as well as tools, including aligned-DKIM verification, loop rejection, scoped secrets, real signed-message tests, and invitation workflows.", "worst": "Sign-ins repeatedly require CPU-intensive proof-of-work and initially delay listening, while unsigned senders and other Atomic Mail senders cannot use the secure default channel.", "standout": false, "blocked": null }, { "pr": 153955, "name": "shadow", "does": "Replays saved Hermes turns on two models, compares answers with blinded order-swapped judging, and manages model switching with rollback.", "useful": 5, "reliable": 4, "polish": 5, "creative": 4, "beyond": 5, "total": 28, "best": "Improves the owner's model choice using their own workload, with two replay-blocking layers, verified isolation, priced usage, statistical verdicts, and a substantial Desktop interface and test suite.", "worst": "Replays block real actions and depend on current workspace files, so their results are imperfect proxies for actual task completion; private core dependencies and hook reload gaps also limit assurance.", "standout": true, "blocked": null }, { "pr": 151281, "name": "tironian", "does": "Learns recurring Hermes work from session traces, trains small local models, evaluates them, and routes approved task families with fallback and drift demotion.", "useful": 4, "reliable": 4, "polish": 5, "creative": 5, "beyond": 5, "total": 28, "best": "Connects capture, redaction, LoRA training, held-out evaluation, shadow trials, schema validation, and live fallback through Hermes middleware, with UI and regression tests.", "worst": "The pooled-pass branch of gated_scorecard allows an undersampled step kind to serve when its raw pass rate is high, weakening the advertised separate minimum-sample and Wilson-bound guarantee.", "standout": true, "blocked": null }, { "pr": 152869, "name": "screen-overlay", "does": "Lets Hermes draw pointers, annotations, interactive guides, and shared sticky notes over Windows apps using UI Automation and a local overlay daemon.", "useful": 4, "reliable": 2, "polish": 5, "creative": 4, "beyond": 5, "total": 25, "best": "Adds a tangible screen-level teaching capability with real UI targets, DPI handling, user-editable notes, expiry, an emergency clear shortcut, and focused client and shape tests.", "worst": "At the pinned commit daemon_env copies the full Hermes environment into the PySide6/Qt child, unnecessarily exposing provider keys to that process.", "standout": false, "blocked": null }, { "pr": 152914, "name": "culprit", "does": "Snapshots file-changing Hermes tool calls, bisects recorded states with a test command, and can reverse only the breaking step.", "useful": 5, "reliable": 4, "polish": 4, "creative": 4, "beyond": 5, "total": 27, "best": "Makes agent mistakes diagnosable using actual test execution, durable private Git snapshots, precise approval prompts, minimal child environments, and tested conflict-safe reversal.", "worst": "Parallel calls can collapse into one recorded step, the search assumes a stable pass-to-fail transition, and snapshots have no automatic retention cap, limiting exact attribution and long-term storage safety.", "standout": false, "blocked": null } ] }