{
  "batch": "read1-072",
  "plugins": [
    {
      "pr": 0,
      "name": "done-bell",
      "does": "Notifies users when Hermes finishes a turn or needs approval, with a Desktop status chip, chime and mute controls.",
      "useful": 4,
      "reliable": 3,
      "polish": 3,
      "creative": 2,
      "beyond": 4,
      "total": 20,
      "best": "The Desktop busy-session watcher and approval hooks make a useful everyday feature that a prompt cannot provide.",
      "worst": "The README requires a developer-specific Windows install and manual Desktop copying, while the supplied tests cover only the Python side.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 0,
      "name": "homeassistant",
      "does": "Connects Hermes to Home Assistant state-change events, device queries and service calls, plus notification delivery.",
      "useful": 4,
      "reliable": 4,
      "polish": 4,
      "creative": 3,
      "beyond": 5,
      "total": 25,
      "best": "A real gateway platform combines filtered inbound events, cron delivery and device control with scoped secrets and extensive policy tests.",
      "worst": "Entity restrictions are opt-in; without allow/deny lists, service calls can control any entity reachable by the token.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 152795,
      "name": "weightscope",
      "does": "Inspects and compares safetensors checkpoints using streamed tensor metrics, with optional local GPU or remote worker execution.",
      "useful": 3,
      "reliable": 5,
      "polish": 4,
      "creative": 4,
      "beyond": 3,
      "total": 22,
      "best": "The bounded checkpoint parser, numeric edge-case tests and guarded remote client make the implementation unusually careful.",
      "worst": "It serves a specialist checkpoint-analysis niche and reports numeric weight differences rather than model behavior or quality.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 151264,
      "name": "toolfound",
      "does": "Configures a hosted Toolfound MCP server and bundles a skill for software discovery and product launch submissions.",
      "useful": 2,
      "reliable": 3,
      "polish": 3,
      "creative": 1,
      "beyond": 3,
      "total": 15,
      "best": "The server is explicitly untrusted and the launch workflow asks for confirmation of the exact submission fields.",
      "worst": "The pinned package contains no server implementation or tests; its tool list and behavior are supplied by an unpinned remote service.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 150321,
      "name": "hermes-sentinel",
      "does": "Blocks repeated identical tool calls and credential-shaped new messages, redacts and bounds tool results, and keeps a local audit ledger.",
      "useful": 4,
      "reliable": 4,
      "polish": 4,
      "creative": 2,
      "beyond": 4,
      "total": 22,
      "best": "Four runtime hooks provide deterministic intervention, with redaction before audit previews and tested storage-failure fallbacks.",
      "worst": "The fixed third-call breaker can block legitimate repeated polling, and unconditional output truncation can discard information needed by the agent.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 153821,
      "name": "learnloop",
      "does": "Turns Hermes into a structured tutor with graded questions, practice problems, persistent knowledge graphs and interactive Desktop learning views.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "total": 27,
      "best": "Code enforces lesson phases and grading order while a substantial Desktop UI makes questions, practice and learned facts usable.",
      "worst": "The composer middleware uses the focused chat rather than the sending chat, so an unusual multi-chat layout can submit text as the wrong question's answer.",
      "standout": true,
      "blocked": null
    },
    {
      "pr": 150892,
      "name": "sessionguard",
      "does": "Records local task checkpoints and tool summaries, then briefs later sessions about unfinished work from the same sender and platform.",
      "useful": 4,
      "reliable": 2,
      "polish": 3,
      "creative": 3,
      "beyond": 4,
      "total": 20,
      "best": "Automatic checkpoint evidence and recovery reminders address real continuity problems without network services.",
      "worst": "Although recovery briefings filter by sender, sessionguard_status and /guard read all profile checkpoints without sender filtering, exposing other users' work on a shared gateway.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 151996,
      "name": "expatslist",
      "does": "Connects Hermes to a public ExpatsList MCP endpoint and supplies a skill for city listings and relocation reference lookups.",
      "useful": 3,
      "reliable": 3,
      "polish": 3,
      "creative": 1,
      "beyond": 3,
      "total": 16,
      "best": "The small package clearly discloses remote calls, needs no credentials and marks the server untrusted.",
      "worst": "There is no pinned server implementation or test suite, and the PR explicitly says the package has not been tested in Hermes.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 153324,
      "name": "glance",
      "does": "Captures and compares simulator screenshots around UI edits, nudges Hermes to check failures and displays before/after results.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 4,
      "total": 25,
      "best": "Bounded capture, persistent verification state and image fixtures support a useful feedback loop with an integrated Desktop comparison view.",
      "worst": "The automatic baseline only watches write_file and patch on selected extensions, so terminal edits and UI changes in .ts or .js files escape it.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 153716,
      "name": "circuit",
      "does": "Builds two-layer PCB designs with a part library, placement, routing, design checks, fabrication exports and a Desktop board viewer.",
      "useful": 3,
      "reliable": 3,
      "polish": 5,
      "creative": 5,
      "beyond": 4,
      "total": 24,
      "best": "A substantial local engine produces Gerbers, drills, KiCad and STEP files from a shared board model rather than merely generating design prose.",
      "worst": "In drc.py, trace bounding boxes omit trace width but the clearance prefilter uses those boxes, so it can skip real copper-clearance violations.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 149325,
      "name": "injection-guard",
      "does": "Screens selected tool outputs for prompt-injection patterns, redacts matches and wraps flagged text in an untrusted-data envelope.",
      "useful": 4,
      "reliable": 3,
      "polish": 3,
      "creative": 2,
      "beyond": 4,
      "total": 20,
      "best": "The middleware intervenes before results reach the model, with envelope escaping, report mode and tests for benign-content preservation.",
      "worst": "Enforce mode rebuilds flagged output from only the first 200,000 characters by default, silently dropping the remaining content; later-only attacks are not scanned.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 152438,
      "name": "reflexes",
      "does": "Starts Hermes turns in an existing chat when watched folders, connected drives or appended log lines trigger a saved rule.",
      "useful": 5,
      "reliable": 3,
      "polish": 4,
      "creative": 4,
      "beyond": 5,
      "total": 26,
      "best": "Local event sources, approval for arming rules, queued delivery and Desktop controls make Hermes useful between user messages.",
      "worst": "Injection and delivery acknowledgement are separate operations, so a crash between them can replay an event despite the README's never-fires-twice claim.",
      "standout": false,
      "blocked": null
    }
  ]
}