{
  "batch": "read2-020",
  "plugins": [
    {
      "pr": 0,
      "name": "done-bell",
      "does": "Notifies users when Hermes finishes a turn or awaits approval, with a chime, OS notification and Desktop status pane.",
      "useful": 4,
      "reliable": 3,
      "polish": 3,
      "creative": 1,
      "beyond": 4,
      "best": "Real lifecycle hooks and Desktop busy-state observation make unattended agent work easier to follow.",
      "worst": "Setup instructions use the author's local Windows paths and manual pane copying, while tests cover Python state rather than real Desktop notifications.",
      "standout": false,
      "blocked": null,
      "total": 19
    },
    {
      "pr": 0,
      "name": "homeassistant",
      "does": "Connects Hermes to Home Assistant through a filtered event gateway, notification delivery and four smart-home control tools.",
      "useful": 4,
      "reliable": 5,
      "polish": 4,
      "creative": 2,
      "beyond": 5,
      "best": "A complete gateway platform combines scoped credentials, reconnect handling and tested service/entity restrictions with device tools.",
      "worst": "Entity restrictions are opt-in and a permitted script can still perform actions beyond the entities named in its call.",
      "standout": false,
      "blocked": null,
      "total": 25
    },
    {
      "pr": 151543,
      "name": "seedguard",
      "does": "Redacts recognized secrets before main-loop model requests and restores them only to allowed tools under user-controlled arming.",
      "useful": 5,
      "reliable": 4,
      "polish": 4,
      "creative": 4,
      "beyond": 5,
      "best": "Fail-closed execution middleware, chat-scoped in-memory secrets and explicit time-limited release implement a useful boundary that prompting cannot provide.",
      "worst": "Auxiliary model calls bypass its middleware, CLI history retains raw input, and encoded or split secrets escape its heuristic detectors.",
      "standout": false,
      "blocked": null,
      "total": 27
    },
    {
      "pr": 153938,
      "name": "syrinx",
      "does": "Parks agent questions during focus blocks, batches a return briefing, retains standing answers and queues late-night requests.",
      "useful": 4,
      "reliable": 3,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "best": "Clarify interception, counted assumption debt and a finished Desktop pane turn attention management into an integrated agent feature.",
      "worst": "Its outward-action protection is a fixed matcher that fails open, and after a focus block ends approval of held actions is requested in text rather than enforced by the plugin.",
      "standout": false,
      "blocked": null,
      "total": 26
    },
    {
      "pr": 152207,
      "name": "codecoach",
      "does": "Teaches C and Python through incrementally generated projects with sandbox-executed examples and tests, plus grounded study of imported notes.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 4,
      "best": "Reference-solution, progression and hidden-output checks make the tutor materially more dependable than a teaching prompt.",
      "worst": "Projects require a working Linux bubblewrap sandbox and can take minutes per step, while passing generated tests cannot establish teaching quality.",
      "standout": false,
      "blocked": null,
      "total": 25
    },
    {
      "pr": 153796,
      "name": "thumbs-fix",
      "does": "Turns Telegram thumbs-down feedback into a reviewed skill-file repair with a diff, explicit approval, backup and undo.",
      "useful": 4,
      "reliable": 4,
      "polish": 4,
      "creative": 4,
      "beyond": 5,
      "best": "It connects feedback to the skill actually loaded and applies approved edits through Hermes tooling with stale-draft checks.",
      "worst": "Message-range attribution can mistake unrelated bot messages for replies, and undo overwrites subsequent skill edits while staged writes are not tracked to final approval.",
      "standout": false,
      "blocked": null,
      "total": 26
    },
    {
      "pr": 152258,
      "name": "kumi",
      "does": "Gives Hermes project-aware control of Ableton Live, REAPER and Renoise, local audio measurement, sample mapping and sound creation.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 5,
      "beyond": 5,
      "best": "A substantial readable runtime, DAW bridges, sound-map Desktop workspace and code-bound approvals turn the agent into a practical music-production interface.",
      "worst": "Setup downloads a platform runtime and installs DAW components, and undo depends on the live runtime with some changes requiring the DAW undo or remaining irreversible.",
      "standout": true,
      "blocked": null,
      "total": 28
    },
    {
      "pr": 151376,
      "name": "joust",
      "does": "Races coding subagents in Git worktrees, evaluates their patches and rival tests, and offers winner merging with rollback.",
      "useful": 4,
      "reliable": 4,
      "polish": 4,
      "creative": 4,
      "beyond": 5,
      "best": "Lifecycle-bound tool routing, repeated checks and Crossfire provide a real orchestration layer with readable results and Desktop review.",
      "worst": "Worktree scoping rewrites paths and terminal working directories but does not confine shell command text, so the isolated-contender claim has a significant boundary.",
      "standout": false,
      "blocked": null,
      "total": 26
    },
    {
      "pr": 153524,
      "name": "shadow-army",
      "does": "Requires a code-bound review from a different observed model family before agent-driven Kanban completion and requests approval for unreviewed publishing.",
      "useful": 5,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "best": "It binds verdicts to observed model identity, successful check evidence and the current tree while surfacing removed assertions and stale reviews.",
      "worst": "Only agent tool completion is gated, fingerprinting and completion are not atomic, and Git fingerprint subprocesses inherit the full process environment.",
      "standout": true,
      "blocked": null,
      "total": 28
    },
    {
      "pr": 150678,
      "name": "precedent-log",
      "does": "Stores human rulings per profile and Kanban board, injects them into worker prompts and checks repeat escalation against those rulings.",
      "useful": 4,
      "reliable": 3,
      "polish": 4,
      "creative": 3,
      "beyond": 5,
      "best": "Durable ruling records, bounded veto history and a Desktop retirement interface make human decisions reusable across workers.",
      "worst": "Default enforcement can leave a worker card running indefinitely after a veto, and matching and compliance rely on word overlap and text markers.",
      "standout": false,
      "blocked": null,
      "total": 24
    },
    {
      "pr": 151098,
      "name": "coffee-vram",
      "does": "Queues Hermes calls to an external Ollama instance using memory estimates, idle-model eviction, keep-warm behavior and Desktop telemetry.",
      "useful": 4,
      "reliable": 3,
      "polish": 3,
      "creative": 3,
      "beyond": 5,
      "best": "Execution middleware and a FIFO scheduler manage local-model contention that a prompt cannot control, with tests for eviction and host-mode handoff.",
      "worst": "Timeout or missing telemetry deliberately admits calls anyway, so memory safety is not guaranteed, and the install command still contains a placeholder repository.",
      "standout": false,
      "blocked": null,
      "total": 23
    },
    {
      "pr": 151604,
      "name": "plur1bus",
      "does": "Replaces Hermes persistent memory with profile-scoped semantic capture and recall, reranking, memory dynamics and a Desktop management interface.",
      "useful": 5,
      "reliable": 4,
      "polish": 4,
      "creative": 3,
      "beyond": 5,
      "best": "A genuine memory provider combines bounded recall, scope-isolation tests and confirmed storage/model operations with searchable persistent memory.",
      "worst": "The large native dependency and model stack makes setup heavy, and Windows ARM64 needs a separate installer because catalog dependencies cannot supply its native wheels.",
      "standout": false,
      "blocked": null,
      "total": 26
    }
  ]
}