{
  "batch": "read2-013",
  "plugins": [
    {
      "pr": 0,
      "name": "done-bell",
      "does": "Adds turn-completion and approval notifications, a chime, and a Desktop status chip.",
      "useful": 4,
      "reliable": 3,
      "polish": 2,
      "creative": 2,
      "beyond": 4,
      "total": 19,
      "best": "Uses real lifecycle hooks and Desktop busy state without sending chat content or needing keys.",
      "worst": "README installation uses an author-specific Windows path and manual Desktop copying; tests cover Python state but not Desktop notifications.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 0,
      "name": "homeassistant",
      "does": "Connects Hermes to Home Assistant device tools and filtered, event-driven gateway conversations.",
      "useful": 4,
      "reliable": 5,
      "polish": 4,
      "creative": 2,
      "beyond": 5,
      "total": 25,
      "best": "A genuine platform adapter with scoped secrets, reconnect handling, service restrictions and focused security regression tests.",
      "worst": "Requires a configured Home Assistant installation, and entity restrictions are opt-in rather than a default least-privilege policy.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 151450,
      "name": "rehearse",
      "does": "Runs spoken practice interviews and presentations with recorded replays, local delivery measurements and model-generated coaching.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 4,
      "total": 25,
      "best": "Combines camera and microphone capture, local speech and face analysis, live presence cues and scenario-specific coaching in a finished Desktop workflow.",
      "worst": "Eye-contact scoring remains a head/pupil heuristic and filler detection is English-tuned; first-use speech and face models add setup weight.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 152692,
      "name": "hermes-parallel",
      "does": "Creates three isolated file-change candidates, compares their diffs and HTML previews, and applies a human-selected candidate with rollback.",
      "useful": 4,
      "reliable": 3,
      "polish": 3,
      "creative": 3,
      "beyond": 4,
      "total": 21,
      "best": "Checks source and candidate hashes, serializes mutations, and tests fail-closed recovery while reserving apply and rollback for slash commands.",
      "worst": "The README manual install omits proof.py and worktree_review.py used by registered features; planning is sequential and constrained to move/write operations, not parallel agents.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 149244,
      "name": "lineage",
      "does": "Records agent-produced files and their sources in a provenance graph with integrity receipts and staleness checks.",
      "useful": 5,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "total": 28,
      "best": "Ten observer hooks capture model, tool and subagent provenance automatically, with redaction and explicit observed/evidenced/inferred distinctions.",
      "worst": "Terminal attribution uses a bounded mtime scan that can miss large trees or include concurrent external edits; install-local signatures are not independent attestations.",
      "standout": true,
      "blocked": null
    },
    {
      "pr": 153257,
      "name": "conductor",
      "does": "Launches and supervises local or SSH coding agents, delivering completion, failure, stall and approval notifications back to Hermes.",
      "useful": 4,
      "reliable": 2,
      "polish": 4,
      "creative": 3,
      "beyond": 5,
      "total": 23,
      "best": "Real asynchronous supervision and resumable run records let Hermes coordinate existing coding-agent apps without spending model tokens while waiting.",
      "worst": "At the field-pinned 0.4.2 code, arbitrary executable names are accepted and children inherit the process environment, with no explicit Hermes approval request before launch.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 149365,
      "name": "say-do-gate",
      "does": "Checks final agent claims against recent tool receipts and either appends discrepancy notes or nudges the next turn.",
      "useful": 5,
      "reliable": 4,
      "polish": 4,
      "creative": 4,
      "beyond": 5,
      "total": 27,
      "best": "Deterministic hooks address a common agent trust failure without model calls, with adversarial and real-Hermes finalizer integration tests.",
      "worst": "Regex-based claim and command matching cannot establish semantic truth, misses unsupported shell forms and external work, and annotation mode sacrifices streaming.",
      "standout": true,
      "blocked": null
    },
    {
      "pr": 149915,
      "name": "obol",
      "does": "Drafts, approves and tracks Stripe invoices or local bank-transfer PDFs, with invoice-risk checks and audit logs.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 4,
      "total": 25,
      "best": "Combines scoped secrets, per-write approval requests, cron refusal, idempotency and tested invoice-specific scam checks into a practical cross-chat product.",
      "worst": "Human approval still depends on Hermes policy: YOLO or approvals.mode off auto-approves its prompts; pattern-based scam checks and local hash chains have documented limits.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 149978,
      "name": "clearvoice",
      "does": "Adds a local push-to-talk speech bridge with editable transcriptions and user-approved spoken output in seventeen languages.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "total": 27,
      "best": "A substantial accessibility interface with in-memory audio, explicit cloud opt-in and single-use speech tokens that keep the user in control of spoken words.",
      "worst": "General Whisper is not adapted to atypical speech, only seven languages have sentence-rebuilding packs, and local voice/model installation adds friction.",
      "standout": true,
      "blocked": null
    },
    {
      "pr": 153970,
      "name": "defuzz",
      "does": "Audits Hermes settings and data, scaffolds or adopts projects, routes new file writes and offers manual project lint.",
      "useful": 3,
      "reliable": 4,
      "polish": 3,
      "creative": 3,
      "beyond": 5,
      "total": 23,
      "best": "Changes real agent file placement through trust-aware hooks and includes extensive tests for write boundaries, redaction, audit uncertainty and creation workflows.",
      "worst": "Linux/Python 3.13 requirements and a large lint toolchain narrow adoption; audits can read other profiles and checks execute project code without a sandbox.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 149668,
      "name": "skillcheck",
      "does": "Snapshots skill edits and compares on-demand regression-case results against a manually promoted baseline.",
      "useful": 4,
      "reliable": 3,
      "polish": 3,
      "creative": 4,
      "beyond": 5,
      "total": 24,
      "best": "Directly improves Hermes self-improvement with revision tracking, zero-tool child sessions, timeout cancellation and reversible user-invoked restores.",
      "worst": "Its optional judge uses bool(payload[\"passed\"]), so a JSON string \"false\" becomes a pass; tool-less, truncated skill replays also cannot verify real tool behavior.",
      "standout": false,
      "blocked": null
    },
    {
      "pr": 153900,
      "name": "annotary",
      "does": "Adds persistent document review threads and versions across Desktop and Webapp, waking the appropriate agent on review actions or mentions.",
      "useful": 5,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "total": 28,
      "best": "Deep profile-aware integration turns comments into actionable agent collaboration, with opt-in wake-ups, loop prevention and real-server integration tests.",
      "worst": "Setup requires a downloaded local server and per-profile enablement; the model-callable review_over=true flag bypasses the open-review push refusal.",
      "standout": true,
      "blocked": null
    }
  ]
}