{
 "batch": "final-02",
 "ranking": [
  {
   "rank": 1,
   "pr": 152458,
   "name": "housemate",
   "total": 27,
   "why": "Adds a whole new Hermes platform (a gateway adapter, a TTS provider, 7 tools, a Desktop Rooms page) that turns old phones and laptops into wake-word room speakers, matching Nous's push onto people's own hardware; it passed the catalog check, enable and doctor here, failed cleanly with 'hub not running' when called without the gateway, and 537 of its tests passed (28 failed, 4 errored on this machine), but setup is heavy and the wake phrase is 'Hey Jarvis'."
  },
  {
   "rank": 2,
   "pr": 150556,
   "name": "hermes-ui-cards",
   "total": 26,
   "why": "Gives Hermes Desktop and the Webapp 17 interactive answer cards with an answer round-trip a prompt cannot draw, finished like a product (hero, gif, 23 card images, two languages); catalog check, enable and doctor passed, bad calls returned a clear 'missing card' error, and 463 tests passed with 4 parity failures here; generative UI is a known idea."
  },
  {
   "rank": 3,
   "pr": 153955,
   "name": "shadow",
   "total": 26,
   "why": "Replays your own recent turns on a candidate model, blind-judges both, gives a statistical verdict and then switches the model gradually with undo, which only middleware can do and is new for agents; it passed the catalog check, enable and doctor (6 tools, 22 hooks), but the check warns of many private core imports that can break on any release, and I could not get its test suite to collect on this machine."
  },
  {
   "rank": 4,
   "pr": 149738,
   "name": "trust-ladder",
   "total": 25,
   "why": "Turns a user's correction into an enforced guard, which is the plugin most Hermes users would install on day one: run inside Hermes it learned 'never push straight to main' and then blocked 'git push origin main' while letting 'echo hello' through; 95 of 96 tests passed here, there are no secrets or network code, but the 'minimal call' its own tool description promises (a bare command string) is refused by its schema, and it has no Desktop UI."
  },
  {
   "rank": 5,
   "pr": 151887,
   "name": "cmi",
   "total": 25,
   "why": "Replaces Hermes's context engine with a deterministic one that stubs old tool results and lets the model recall them, so it works on every turn once installed; 498 tests passed here and catalog check, enable and doctor passed, but the check warns of private core imports (it wraps Hermes's own compressor), the savings figures are the author's own replays, and the README leads with a proxy for other agents."
  },
  {
   "rank": 6,
   "pr": 150353,
   "name": "fluent",
   "total": 25,
   "why": "Checks the long options of the agent's terminal commands against the installed program's own help before they run, a pre-execution veto a skill cannot give; the cleanest engineering in the batch (402 passed, 140 skipped here, no catalog warnings, bounded help calls, every hook fails open), though only five programs have tuned parsing and the saving it reports is a flat estimate."
  },
  {
   "rank": 7,
   "pr": 150644,
   "name": "hermes-canvas",
   "total": 25,
   "why": "The agent draws its plan as boxes bound to file globs and check commands and hooks set each box's status from real writes and exit codes, with a Desktop pane for steering; 306 tests passed here, the catalog check had no warnings and bad calls returned clear errors inside Hermes, but shell-write parsing is best effort and the value is mostly for Desktop coding sessions."
  },
  {
   "rank": 8,
   "pr": 150251,
   "name": "gemini-live",
   "total": 25,
   "why": "Adds full-duplex voice calls to Desktop and the dashboard where a Gemini Live voice model delegates work to the real agent, with the key read through Hermes's secret store and a single-use browser token; 126 of 127 tests passed here (the one failure was a file-read rule on this machine), but it needs a Gemini key, sends audio to Google, and its place in the built-in voice picker needs the author's patch to Hermes Desktop."
  },
  {
   "rank": 9,
   "pr": 153714,
   "name": "tenure",
   "total": 25,
   "why": "Per-action-type trust that asks first after a mistake and earns autonomy back, with file undo, a signed passport and the most polished Desktop page in the batch (five locales); 2,401 tests passed and 6 failed here, but it rests on a 5,200-line heuristic shell classifier, the catalog check raised a security-scan caution and private-import warnings, and its own README says the study behind its rule is pending."
  },
  {
   "rank": 10,
   "pr": 153809,
   "name": "living-spaces",
   "total": 25,
   "why": "Lets the agent build persistent sandboxed mini-apps with records, revisions and cron upkeep in a Desktop tab, something new for agents (15 tools registered, 394 tests passed here with one failing on a missing module), but the catalog check warns of private cron imports and a raw Desktop bridge, it is qualified on macOS only, its READMEs show no screenshot, and Spaces can copy home-folder files and post to any public host."
  },
  {
   "rank": 11,
   "pr": 149640,
   "name": "hermes-ark",
   "total": 24,
   "why": "Encrypted, deduplicated, versioned backups of the Hermes home with item-level restore and a guard that snapshots before memory writes, which nearly every long-running user wants; 174 tests passed here and secrets go through Hermes's secret store, but backup is a familiar idea, the passphrase (and so encryption) is optional for a local repository, and its ark_checkpoint tool was not offered in my run."
  },
  {
   "rank": 12,
   "pr": 152691,
   "name": "grimoire",
   "total": 24,
   "why": "DuckDB question answering over the user's data files with a post-reply check that every number traces to a query; inside Hermes it attached a CSV, answered a SELECT, refused DROP TABLE and refused /etc/passwd with clear errors, and its self-test passed, but SQL over files is a familiar tool and its tests are one self-test file."
  },
  {
   "rank": 13,
   "pr": 153851,
   "name": "foreman",
   "total": 24,
   "why": "A deterministic gate that stops a coding agent rigging its own green check plus a second-model diff review, enforced by blocking hooks; 384 tests passed here (the one failure is its deliberately broken example cart) and it answered NO_REPOSITORY cleanly, but it only helps people who let Hermes commit to git and its README lists bypasses it cannot see."
  },
  {
   "rank": 14,
   "pr": 150840,
   "name": "kata",
   "total": 24,
   "why": "Mines tool-call sequences that recur across conversations and, after the user approves, registers each as one deterministic tool, a new way for an agent to grow its toolset; its tests passed here and kata_list answered cleanly, but the payoff needs strictly repeated workflows, so fewer users would keep it."
  },
  {
   "rank": 15,
   "pr": 153020,
   "name": "breadcrumbs",
   "total": 24,
   "why": "A content-addressed append-only memory provider and context engine with zoomable recall that replaces Hermes's memory layer; 669 tests passed here, but it is sprawling and jargon-heavy, the catalog check warns of private core imports, it stores every turn without redaction of its own, and its comparison benchmark could not be checked from the code."
  },
  {
   "rank": 16,
   "pr": 152651,
   "name": "workroom",
   "total": 24,
   "why": "Keeps Word, PowerPoint and Excel figures linked to workbook cells with owner-approved releases and a tripwire that reverts direct agent edits; the deepest engineering here (31 tools, 14 hooks, 730 tests passed) and an excellent README, but the audience is narrow and Windows-centred and the reviewer sent it back several times over slow guard patterns."
  },
  {
   "rank": 17,
   "pr": 150800,
   "name": "hypnos",
   "total": 24,
   "why": "Captures approvals and corrections, trains a small LoRA overnight and routes mastered tool-call turns to that local model, the most original idea in the batch and one only a plugin can do; but it needs a hand-built torch environment and a GPU or patience, its results are from a synthetic persona, the real-training test is skipped by default, and few users would keep it running."
  },
  {
   "rank": 18,
   "pr": 152908,
   "name": "contract-necromancer",
   "total": 23,
   "why": "A person locks the project's tests by hash and the plugin runs them itself when Hermes claims done, blocking edits to them; it answered cleanly with no contracts frozen and its suite was passing when my run timed out, but its own README names a similar test referee and it adds no UI."
  },
  {
   "rank": 19,
   "pr": 150596,
   "name": "ariadne",
   "total": 23,
   "why": "Records a Hermes session to a cassette and replays it offline to show what another Hermes build would do differently, a novel middleware tool for plugin authors and upgrade checks; only 21 unit tests (all passed here), a narrow audience, and its upgrade check downloads and installs upstream Hermes."
  },
  {
   "rank": 20,
   "pr": 153844,
   "name": "fine-print",
   "total": 23,
   "why": "Detects where Hermes silently truncates, downscales or swaps what the model sees and tells the user or the model; careful and candid (its tests passed here, no network or subprocess), but it is pinned to Hermes v0.23.0 output strings, its own README marks many detectors fixture-only, and it serves power users debugging answer quality."
  }
 ]
}