{"batch":"read1-070","plugins":[{"pr":0,"name":"done-bell","does":"Rings and records a notification when a Hermes turn ends, with approval-waiting state, mute controls, hooks, a command, and a Desktop chip/pane.","useful":3,"reliable":4,"polish":4,"creative":2,"beyond":3,"total":19,"best":"It is a focused, installable Desktop/CLI quality-of-life integration with explicit state handling and tests.","worst":"The idea is comparatively conventional and the README's Windows-oriented manual pane-copy step makes setup less seamless.","standout":false,"blocked":null},{"pr":0,"name":"homeassistant","does":"Connects Hermes to Home Assistant as a platform plugin, exposing Home Assistant entities and actions through Hermes tools and Desktop integration.","useful":5,"reliable":5,"polish":5,"creative":3,"beyond":5,"total":28,"best":"It makes Hermes operate across a user's own smart-home hardware and app ecosystem, with the depth expected of a platform plugin.","worst":"Its breadth and external Home Assistant setup make the integration heavier to configure and maintain than a local-only plugin.","standout":true,"blocked":null},{"pr":151024,"name":"night-shift","does":"Maintains incident runbooks and metrics context, diagnoses and proposes operational fixes, and requires explicit human approval before executing a fix, including Slack approval flows.","useful":5,"reliable":4,"polish":5,"creative":5,"beyond":5,"total":29,"best":"The approval-code/fingerprint flow, Slack actions, runbook learning, and incident-oriented Desktop/demo material form a compelling agent operations product.","worst":"The implementation depends on substantial external deployment infrastructure and has only one repository test, leaving more integration risk than the documentation suggests.","standout":true,"blocked":null},{"pr":153051,"name":"show-your-work","does":"Builds an acceptance-criteria evidence board from JUnit results, file freshness, linked tests, and human accept/reject sign-off, with agent, CLI, slash, and Desktop surfaces.","useful":5,"reliable":5,"polish":5,"creative":4,"beyond":5,"total":29,"best":"It separates coverage, test outcome, freshness, and human review instead of treating a green command as proof, and backs that model with tests and disclosure rules.","worst":"It is primarily an evidence workflow around existing test runners, so some value still depends on users maintaining meaningful criteria and links.","standout":true,"blocked":null},{"pr":151888,"name":"preferly","does":"Declares a hosted Streamable HTTP MCP server offering Preferly AI-visibility tools plus a skill for crawler, robots, llms.txt, structured-data, and visibility checks.","useful":3,"reliable":2,"polish":3,"creative":3,"beyond":3,"total":17,"best":"The service can expose specialized website visibility checks without requiring the plugin to ship a large local implementation.","worst":"At the pinned commit it is only a manifest, MCP URL, logo, and skill: there are no local handlers or tests, and availability and behavior are delegated to a third-party service.","standout":false,"blocked":null},{"pr":149190,"name":"living-field","does":"Adds a deterministic canvas-based Hermes Desktop pane and status chip whose animation reflects the host's busy, gateway, and model state.","useful":2,"reliable":4,"polish":4,"creative":3,"beyond":5,"total":23,"best":"The visual is unusually honest for a dashboard: it uses host state, avoids fabricated metrics and network access, and has a stable seeded field.","worst":"It is a narrow visualization with no agent tools, tests, or README setup beyond enabling the pane, so relatively few users need it on day one.","standout":false,"blocked":null},{"pr":150873,"name":"skill-court","does":"Replays tasks with and without a skill in isolated Docker profiles, scores outcome differences per model, quarantines harmful skills, and supports evidence-based appeals and automatic trials.","useful":5,"reliable":4,"polish":5,"creative":5,"beyond":5,"total":29,"best":"Paired outcome attribution, model-specific quarantine, blind judging, and reversible appeals directly address the risk of agents writing harmful skills.","worst":"It is operationally demanding: Docker, model calls, replay budgets, and judge uncertainty create substantial setup, cost, and failure modes despite the extensive tests.","standout":true,"blocked":null},{"pr":151252,"name":"proof-run","does":"Records operator-demonstrated tool workflows as contracts, checks fresh read evidence at the tool boundary, and fail-closed blocks mutations that do not match the proven contract.","useful":5,"reliable":4,"polish":4,"creative":5,"beyond":5,"total":28,"best":"It turns demonstrations into an enforceable, versioned action gate rather than a prompt promise, with bounded state, approval-aware capture, and extensive semantic tests.","worst":"The safety logic is subtle and host-API dependent; the catalog folder has a concise README and limited visible integration coverage compared with the complexity of the gate.","standout":true,"blocked":null},{"pr":151242,"name":"study-tree","does":"Branches side questions into separate Hermes sessions, resolves them back into the main thread, and renders requested research or time-change results as interactive widgets.","useful":4,"reliable":4,"polish":5,"creative":4,"beyond":5,"total":27,"best":"It changes conversation topology and Desktop presentation, with explicit prompt policies, multiple tools/hooks, screenshots, and both Python and JavaScript tests.","worst":"The injected system-prompt directives are broad and unusually prescriptive, so prompt interactions and host surface compatibility remain meaningful risks.","standout":true,"blocked":null},{"pr":153802,"name":"leaveby","does":"Uses Google Maps or TomTom to save places, calculate traffic-aware ETAs and leave-by plans, inject route context, and run no-agent background departure alerts.","useful":5,"reliable":4,"polish":4,"creative":4,"beyond":5,"total":27,"best":"It combines real external-world routing, multi-stop planning, a pre-LLM context hook, and token-free cron alerts into a genuinely useful travel assistant.","worst":"Correctness and privacy depend on paid/limited map APIs and cron environment passthrough; the background watcher is an additional deployment surface with only one listed test file.","standout":true,"blocked":null},{"pr":152343,"name":"melange","does":"Records terminal sessions as hash-chained timelines and lets Hermes scrub, explain provenance, fork what-ifs, clamp inputs, recompute dependents, and preserve original recordings.","useful":5,"reliable":4,"polish":5,"creative":5,"beyond":5,"total":29,"best":"The local, append-only replay/what-if engine gives Hermes provenance and counterfactual computation over real terminal work, with a substantial demo and product-quality surfaces.","worst":"It is a very large and ambitious codebase with PTY execution and many moving parts, so the integration and fidelity failure surface is high even with bundled tests.","standout":true,"blocked":null},{"pr":149668,"name":"skillcheck","does":"Observes skill edits, snapshots revisions, generates reviewable regression cases, runs them in narrowed child sessions, reports verdicts, and offers reversible promote/revert commands.","useful":5,"reliable":4,"polish":4,"creative":5,"beyond":5,"total":28,"best":"It creates a practical regression loop for self-modifying skills while explicitly distinguishing first runs, regressions, judge failures, snapshots, and human promotion.","worst":"The child sessions share the real environment and the LLM judge can be unavailable, so its safety and verdict quality depend on Hermes lifecycle behavior and careful review.","standout":true,"blocked":null}]}