{
  "batch": "read2-008",
  "plugins": [
    {
      "pr": 0,
      "name": "done-bell",
      "does": "Notifies users when Hermes turns finish or need approval through lifecycle hooks, a Desktop status chip, OS notifications and chimes.",
      "useful": 4,
      "reliable": 3,
      "polish": 3,
      "creative": 2,
      "beyond": 4,
      "total": 20,
      "best": "A small, credential-free improvement to everyday agent use with working hook handlers, mute controls and a Desktop settings pane.",
      "worst": "Setup uses a developer-specific Windows path and manual Desktop copying, while the supplied tests cover backend behavior rather than notification delivery.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 0,
      "name": "homeassistant",
      "does": "Connects Hermes to Home Assistant through a filtered event-driven gateway platform, device query/control tools and standalone cron notification delivery.",
      "useful": 4,
      "reliable": 5,
      "polish": 4,
      "creative": 3,
      "beyond": 5,
      "total": 26,
      "best": "A genuine platform extension with profile-scoped secrets, reconnect logic, explicit entity policy enforcement and targeted security regression tests.",
      "worst": "Entity restrictions are opt-in, so operators must configure allow/deny lists to narrow device control; the integration mainly serves existing Home Assistant households.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 151387,
      "name": "omp-sessions",
      "does": "Delegates coding to local omp sessions with scheduling, queues, chat notifications, kanban dispatch and dashboard/Desktop supervision.",
      "useful": 4,
      "reliable": 2,
      "polish": 5,
      "creative": 3,
      "beyond": 4,
      "total": 22,
      "best": "Deep integration ties persistent external coding sessions to Hermes chats and kanban cards, with readable code, extensive tests and a finished supervision UI.",
      "worst": "At the pinned commit, Tools.start passes the model-selected mode directly to omp.new, including auto, without the agent_approvals guard used by Tools.mode; external omp setup is also required.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 153871,
      "name": "barehand",
      "does": "Adds a computer_use provider that operates QEMU or VNC consoles through pixels and keyboard input, including firmware, bootloader repair and OS installation.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 5,
      "beyond": 5,
      "total": 28,
      "best": "A distinctive extension of what Hermes can operate, with screen-text extraction, cautious action feedback, local flight recordings and fail-closed destructive-action approval tests.",
      "worst": "Reported live testing is limited to QEMU guests on one Apple Silicon Mac; Linux/Intel paths and non-QEMU VNC hardware lack equivalent validation.",
      "standout": true,
      "blocked": null
    },
    {
      "pr": 152441,
      "name": "baton",
      "does": "Hands a repository snapshot and redacted Hermes session to a remote Hermes, then retrieves a branch whose declared checks passed for the returned tree.",
      "useful": 4,
      "reliable": 3,
      "polish": 5,
      "creative": 4,
      "beyond": 4,
      "total": 24,
      "best": "Resumable git-ref handoffs, Runs API idempotency, secret scanning and tree-bound verification are backed by detailed failure-path tests and reproducible demos.",
      "worst": "Remote checks run unsandboxed when bubblewrap is unavailable, and peers._secret falls back to process environment after a scoped lookup misses or fails.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 152706,
      "name": "circuit-bench",
      "does": "Provides an ESP32 Desktop workbench and agent tools for cited pin checks, electrical calculations, schematics, firmware build/flash and serial crash debugging.",
      "useful": 3,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "total": 26,
      "best": "An unusually complete hardware workflow with deterministic engineering rules, explicit simulation, firmware-image hash checks, device identity checks and extensive regression fixtures.",
      "worst": "Its audience is ESP32 builders, and the documented session /yolo and one-shot approval gap means users should retain allow_flash=false until deliberate hardware use.",
      "standout": true,
      "blocked": null
    },
    {
      "pr": 153529,
      "name": "saywhat",
      "does": "Replaces speech transcription with local word-confidence recognition, clarification-driven personal vocabulary learning and pronunciation practice.",
      "useful": 4,
      "reliable": 2,
      "polish": 4,
      "creative": 4,
      "beyond": 5,
      "total": 24,
      "best": "A useful change to Hermes voice input itself: uncertainty survives transcription, user corrections train future recognition, and tests exercise the learning loop.",
      "worst": "confidence._readings materializes every candidate combination before keeping three, causing exponential memory/time growth on noisy or long dictation; confirmation is also model guidance rather than an enforced gate.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 151946,
      "name": "stitchslop",
      "does": "Connects Hermes to the Stitch Slop embroidery editor through a local WebSocket bridge, a streamed Chrome viewer and stitch-file export tools.",
      "useful": 3,
      "reliable": 3,
      "polish": 4,
      "creative": 4,
      "beyond": 3,
      "total": 20,
      "best": "The bridge provides real app commands and user takeover through a visible viewer, with tests for export persistence, relay authentication, origin and navigation restrictions.",
      "worst": "Chrome launches with --no-sandbox and the fixed ~/.stitchslop state and browser profile are shared across Hermes profiles; first pairing and Linux Chrome/Xvfb setup also require manual work.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 151900,
      "name": "hermes-portals",
      "does": "Creates stateful diagrams, checklists and calculators that move between a laptop and paired phone, with camera gestures and an optional source-built Mac input controller.",
      "useful": 3,
      "reliable": 4,
      "polish": 4,
      "creative": 4,
      "beyond": 4,
      "total": 23,
      "best": "Transactional state, one-use pairing, revision conflict checks and local-only camera processing make the gesture-and-device handoff more than a visual mockup.",
      "worst": "This remains a trusted-LAN HTTP prototype with one active artifact and one phone, a small fixed set of tool types and a separate multi-step companion/native setup.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 150353,
      "name": "fluent",
      "does": "Learns installed CLI help and man pages, then uses terminal hooks to reject invented options, suggest real flags and escalate selected destructive commands.",
      "useful": 5,
      "reliable": 4,
      "polish": 4,
      "creative": 4,
      "beyond": 5,
      "total": 27,
      "best": "A broadly useful improvement to the agent itself: deterministic pre-execution checks, conservative parsing, binary freshness checks and real-command corpus tests without extra model calls.",
      "worst": "The guard only covers adopted local binaries and confidently parsed long options; its own README acknowledges that a different session PATH can still be judged against the wrong binary.",
      "standout": true,
      "blocked": null
    },
    {
      "pr": 149514,
      "name": "lemma",
      "does": "Extracts scientific-paper claims, generates sandboxed numerical audits with positive-control checks, judges evidence and displays local audit trails in Desktop.",
      "useful": 3,
      "reliable": 3,
      "polish": 4,
      "creative": 4,
      "beyond": 3,
      "total": 20,
      "best": "The vendored pipeline enforces passing positive controls for decisive verdicts and refuses generated execution when no sandbox is available unless the operator explicitly opts in.",
      "worst": "Storage is resolved once at tools.py import rather than per call, and detached full audits expose status but no cancellation or hard whole-run deadline; scientific validity still depends on model-generated experiments.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 153687,
      "name": "locus-codegraph",
      "does": "Indexes workspace structure and Python call shapes to rehearse signature changes, report broken callers in edit results and offer reviewable caller patches.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "total": 27,
      "best": "A tightly integrated coding aid with AST-based binding analysis, persistent indexing, context references, Desktop/Web caller inboxes and focused tests comparing rehearsal predictions with post-edit alerts.",
      "worst": "Binding checks cannot prove dynamic dispatch, self/parameter receivers or decorated callables, and shell/external-editor changes need reindexing; inferred test links are not execution coverage.",
      "standout": true,
      "blocked": null
    }
  ]
}