{
 "batch": "read1-033",
 "plugins": [
  {
   "pr": 0,
   "name": "done-bell",
   "does": "Rings an OS notification and optional chime when a Hermes turn finishes or waits for approval, with a Desktop status chip/pane, a /bell command and a done_bell_status tool.",
   "useful": 3,
   "reliable": 3,
   "polish": 3,
   "creative": 2,
   "beyond": 3,
   "total": 17,
   "best": "Small, focused and safe: four observer hooks that never raise, a Desktop pane with a test button, cross-platform notifiers via argument-list subprocesses, and 9 tests.",
   "worst": "The README is a developer note (hard-coded C:\\Users\\... install path, manual copy of the Desktop pane) and the idea is a common one.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 0,
   "name": "homeassistant",
   "does": "Official Home Assistant integration: a gateway platform that forwards HA state-change events to the agent and four tools to list, read and control devices over the REST API.",
   "useful": 4,
   "reliable": 5,
   "polish": 4,
   "creative": 3,
   "beyond": 5,
   "total": 26,
   "best": "A real platform adapter plus tools, with about 176 tests, allow/deny lists that veto device control, and HASS_TOKEN read through Hermes's scoped secret store.",
   "worst": "It is a port of an integration that used to live in Hermes core, so it adds nothing new for agents.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 150186,
   "name": "reservegauge-agent-plugin",
   "does": "A skill-only package that tells the model how to review a user-supplied ReserveGauge usage CSV for coverage, reset boundaries and observed changes.",
   "useful": 1,
   "reliable": 3,
   "polish": 2,
   "creative": 1,
   "beyond": 1,
   "total": 9,
   "best": "Careful, honest instructions with no code, network or credentials, and it explicitly treats CSV cells as data.",
   "worst": "It is just one SKILL.md for one vendor's export format, which a prompt does equally well, and it has no code or tests.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 150216,
   "name": "joinsandwich-directory",
   "does": "Points Hermes at Sandwich's hosted, no-auth MCP server so it can search caregiving resources and estimate long-term-care costs by US state.",
   "useful": 2,
   "reliable": 3,
   "polish": 3,
   "creative": 2,
   "beyond": 2,
   "total": 14,
   "best": "Gives agents a public, read-only eldercare directory and cost calculator with no account needed.",
   "worst": "The package is only a remote MCP URL and a 24-line skill. All the logic is on a company server that can't be reviewed, and the README is mostly product marketing.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 151731,
   "name": "brickfact",
   "does": "Connects Hermes to Brickfact's public MCP server to find LEGO sets, compare regional retailer offers and read recorded price history, with a stdlib Python helper as a fallback.",
   "useful": 2,
   "reliable": 4,
   "polish": 4,
   "creative": 2,
   "beyond": 2,
   "total": 16,
   "best": "Well-engineered helper (fixed endpoint, no redirects or proxy, size limits, clear errors) with offline tests in CI and an unusually candid privacy and affiliate disclosure.",
   "worst": "It is a niche shopping connector to a vendor's hosted service, so the plugin adds no Hermes-specific capability.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 153078,
   "name": "cortex-zero-adaptive",
   "does": "Claims a local 1B-model loop that plans, executes and recovers browser automation steps through an external 'BAF' server.",
   "useful": 1,
   "reliable": 1,
   "polish": 2,
   "creative": 2,
   "beyond": 2,
   "total": 10,
   "best": "The observe-reason-act-adapt loop with domain/operation allowlists and failure classification is a reasonable idea for small local models.",
   "worst": "It cannot run: __init__.py has no register(ctx), and engine.py imports czengine from a hard-coded /home/professor/cortexzero-engine that the repo does not include. There are no unit tests, and the README speaks of 'proof of compromise' and web-hacking packs.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 152691,
   "name": "grimoire",
   "does": "Answers questions about CSV/JSON/Parquet/SQLite files with locked-down DuckDB SQL, saves each query as a receipt, flags any number in the reply that no query produced, and adds a Desktop Data page with charts and data-health scores.",
   "useful": 4,
   "reliable": 5,
   "polish": 5,
   "creative": 4,
   "beyond": 4,
   "total": 26,
   "best": "Checking every number in the model's reply against query results (transform_llm_output hook) on a hardened, read-only DuckDB, backed by a 22-test selftest in CI and a full Desktop page.",
   "worst": "It is large (eight tools, six hooks, about 27k lines in the repo), so it carries a lot of code and a duckdb dependency for what users may see as a data-chat feature.",
   "standout": true,
   "blocked": null
  },
  {
   "pr": 150109,
   "name": "recall",
   "does": "Keeps a private on-device SQLite FTS5 index of the user's documents and uses a pre_llm_call hook to put likely matching files in front of the model, with five tools and an @recall: reference.",
   "useful": 4,
   "reliable": 3,
   "polish": 4,
   "creative": 3,
   "beyond": 4,
   "total": 22,
   "best": "It fills a real gap (finding the file in ~/Downloads) entirely offline, with a secret denylist, symlink and decompression-bomb guards, 30 stdlib tests and an honest-limits section.",
   "worst": "recall_forget only deletes index rows, so the session-start refresh brings a forgotten folder back, and recall_open reads any non-denylisted absolute path up to 64 MB, even outside the indexed roots.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 153680,
   "name": "htmi",
   "does": "Lets people pin review comments to exact parts of HTML/Markdown documents the agent writes, notifies Hermes of new comments each turn, and lets it make checked edits and replies via 11 tools, a /htmi command and a dashboard Review tab.",
   "useful": 3,
   "reliable": 3,
   "polish": 4,
   "creative": 4,
   "beyond": 4,
   "total": 22,
   "best": "A fresh workflow: reviewing agent-written HTML on the page itself, with a localhost-only, token-protected review server, stale-edit refusal and human-only resolution.",
   "worst": "The work runs in a 70k-line bundled Node engine (htmi.mjs) whose source is elsewhere in the repo. Only 5 tests cover the plugin, and the GitHub publish/pull tools have none.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 152502,
   "name": "hermes-vault-broker",
   "does": "Fills saved website logins from a separate root-installed broker daemon (scrypt plus AES-GCM, mlock'd key) that types the password over Chrome DevTools, so the agent's tool call carries only a handle and origin.",
   "useful": 3,
   "reliable": 3,
   "polish": 3,
   "creative": 4,
   "beyond": 4,
   "total": 21,
   "best": "Real cryptography and process isolation (separate system user, peer-uid socket checks, hardened systemd unit) with 83 tests across Python versions in CI.",
   "worst": "It claims more than the code does. In default mode the password lands in a browser whose DevTools port is shared with the agent's browser tools, and the repo's own test reads it back from the field. Setup also needs root and a running daemon.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 149986,
   "name": "hermes-sanctuary",
   "does": "Shows a living, animated scene in Hermes Desktop and the web dashboard where characters wake as their tools are used and a guard reacts to errors, loops, fallbacks, budget and pending approvals.",
   "useful": 3,
   "reliable": 3,
   "polish": 5,
   "creative": 4,
   "beyond": 4,
   "total": 23,
   "best": "Exceptionally finished Desktop pane, sidebar page and dashboard strip, built only from 11 observer hooks, with bilingual docs and a demo GIF.",
   "worst": "It has no tests, it is mostly a visual layer, and the web dashboard iframe is not sandboxed although the README says it is.",
   "standout": true,
   "blocked": null
  },
  {
   "pr": 152785,
   "name": "afterimage",
   "does": "Records a computer_use click-flow once with per-step screenshots, rehearses it under a see-through HUD that compares live pixels, then replays matching steps with zero model calls.",
   "useful": 3,
   "reliable": 2,
   "polish": 4,
   "creative": 4,
   "beyond": 4,
   "total": 21,
   "best": "Turning an expensive vision-model desktop flow into a zero-token pixel-checked replay is a new and practical idea for agents.",
   "worst": "Rehearse defaults to 'repair' (the README calls it opt-in) and clicks a model-chosen coordinate without confirmation. Screenshots are left in /tmp, the overlay is never torn down, and xdotool is required but undocumented.",
   "standout": false,
   "blocked": null
  }
 ]
}