{
 "batch": "final-04",
 "ranking": [
  {
   "rank": 1,
   "pr": 152458,
   "name": "housemate",
   "total": 28,
   "why": "A real Hermes platform adapter that turns old phones and browsers into whole-home voice speakers, with spoken approvals tied to Hermes's own approval card, SHA-pinned model downloads, ~30 test files and screenshots; it puts Hermes on people's own hardware, limited by no speaker identification and non-commercial wake models."
  },
  {
   "rank": 2,
   "pr": 149640,
   "name": "hermes-ark",
   "total": 27,
   "why": "Encrypted, deduplicated, versioned backups of the whole Hermes home via tool_execution middleware, with keys from agent.secret_scope, restores that need an explicit yes and agent admin off by default; many users would install it day one, though each memory or skill write adds a synchronous checkpoint."
  },
  {
   "rank": 3,
   "pr": 153955,
   "name": "shadow",
   "total": 27,
   "why": "Replays real past turns on a candidate model, judges the pairs blind in both orders and switches Hermes's default model with undo, using a spend cap, write-blocking during replay and ~70 test files; it changes the model Hermes runs, but it spends real money and has many moving parts."
  },
  {
   "rank": 4,
   "pr": 153714,
   "name": "tenure",
   "total": 26,
   "why": "Trust per kind of action, earned and lost by statistical rules, routed through Hermes's own approval gate; it fails safe to an ask, always asks for money and off-machine actions, and has ~950 tests, but the default 'trusted' start lets rm/mv run behind only a 24h undo, and commands are classified from their text."
  },
  {
   "rank": 5,
   "pr": 150840,
   "name": "kata",
   "total": 26,
   "why": "Mines routines that recur across conversations into validated tools that a person must approve (/kata approve), with per-call approval for side effects, a fake-Hermes harness and real-run logs; a fresh way for Hermes to grow its own tools, though the dynamic kata_* tools are missing from the manifest and it wants tool_search off."
  },
  {
   "rank": 6,
   "pr": 150644,
   "name": "hermes-canvas",
   "total": 26,
   "why": "A live Desktop plan canvas whose box status comes from real writes and exit codes, with mid-turn user edits and a pre_verify gate on 'done'; genuinely new, but the README's 'the model cannot fake an exit code' overclaims, since the model writes each box's check command."
  },
  {
   "rank": 7,
   "pr": 150596,
   "name": "ariadne",
   "total": 26,
   "why": "Offline record and replay of Hermes sessions (cassettes, compare, bisect across Hermes commits, fork) that runs without a model or tools, and even found a real Hermes bug; novel and only possible as a plugin, but its audience is mostly plugin developers and in-chat play blocks synchronously."
  },
  {
   "rank": 8,
   "pr": 150556,
   "name": "hermes-ui-cards",
   "total": 25,
   "why": "17 schema-checked interactive card types drawn by Desktop/Webapp, with a text fallback elsewhere, an undo countdown, a screenshot gallery and Python, jsdom and e2e tests; broadly useful day one, but it builds on an existing card format, the demo video slot is empty, and model-chosen image URLs leak the user's IP."
  },
  {
   "rank": 9,
   "pr": 152908,
   "name": "contract-necromancer",
   "total": 25,
   "why": "Freezes tests by hash and uses pre_verify and pre_tool_call to stop Hermes calling coding work done until they pass, failing closed with person-only freeze/approve, process-group timeouts and an e2e run in real Hermes; well built, but no Desktop UI, macOS-only testing, and overlap with other test-gate plugins."
  },
  {
   "rank": 10,
   "pr": 152691,
   "name": "grimoire",
   "total": 25,
   "why": "DuckDB analytics over local data files with a locked-down engine, a 30 s query cap and every number in the reply checked against query results, plus a Desktop Data page and benchmarks; polished, but there is no tests folder (only a selftest) and the manifest lacks license and manifest_version."
  },
  {
   "rank": 11,
   "pr": 153851,
   "name": "foreman",
   "total": 25,
   "why": "A deterministic diff detector on pre_tool_call and pre_verify that blocks skipped tests, edited checks, secrets and agent-run merges, with human-confirmed merges pinned to the head SHA and ~6,000 lines of tests; strong, but the text-match guards can be bypassed and it uses a private Hermes import."
  },
  {
   "rank": 12,
   "pr": 153844,
   "name": "fine-print",
   "total": 25,
   "why": "Detects Hermes's own silent limits (truncated web_extract, dropped images, capped search results) through 11 hooks and tells the user and model, with fixtures from real Hermes v0.23.0; original and careful, but 19 of 35 detectors are fixture-only, it is pinned to one Hermes version, and its cards fail on Desktop by the author's own check."
  },
  {
   "rank": 13,
   "pr": 150800,
   "name": "hypnos",
   "total": 25,
   "why": "Trains a local LoRA 'brain' overnight from captured approvals and corrections, in a network-less Docker sandbox with an exam gate, then routes mastered tasks to it via llm_execution middleware; the most original idea here, but all results are synthetic, the screenshots are mock-ups, and a 0.6B model answering live turns is risky."
  },
  {
   "rank": 14,
   "pr": 150251,
   "name": "gemini-live",
   "total": 24,
   "why": "Full-duplex Gemini Live voice calls in Desktop and the dashboard, with the key read only via agent.secret_scope, single-use ephemeral tokens and an off-by-default read-only shell; polished and securely keyed, but less novel, and its composer integration needs a 375-line patch to Hermes Desktop."
  },
  {
   "rank": 15,
   "pr": 149738,
   "name": "trust-ladder",
   "total": 24,
   "why": "Compiles user corrections into proven guards (block, check, rule or hand-off to approval), with TLA+ models and ~85 tests; clever, but the install line is a '<owner>' placeholder, auto-learned guards install from chat with no confirmation, and there is no Desktop UI."
  },
  {
   "rank": 16,
   "pr": 153809,
   "name": "living-spaces",
   "total": 24,
   "why": "Lets the agent build persistent sandboxed mini-apps in a Desktop tab, with SSRF-blocked networking and large test suites; the most novel Desktop capability, but it relies on private Hermes cron modules and the raw desktop bridge (the README admits it is not catalog-approved), Space internet is on by default, and there are no screenshots."
  },
  {
   "rank": 17,
   "pr": 152651,
   "name": "workroom",
   "total": 24,
   "why": "An Office-document workspace that ties each figure to its source cell, with owner-only approval, a tripwire and ~90 test files; inventive but niche, and its safety net reads .env and key-file bytes and reverts other plugins' changes, which is surprising."
  },
  {
   "rank": 18,
   "pr": 151887,
   "name": "cmi",
   "total": 23,
   "why": "A context engine that folds old tool results into recallable digests and changes what Hermes sends to the model; but its bundled console reads the auth.json credential pool, passes API keys on the command line, and offers account switching to stretch quota, against the catalog's credential rule, and the manifest is bare."
  },
  {
   "rank": 19,
   "pr": 150353,
   "name": "fluent",
   "total": 23,
   "why": "Checks shell commands against the real --help and man pages of installed programs and blocks long options that do not exist, carefully and conservatively; useful but narrow, its 'tokens saved' figure is a fixed estimate, and it does nothing on remote terminal backends."
  },
  {
   "rank": 20,
   "pr": 153020,
   "name": "breadcrumbs",
   "total": 23,
   "why": "A memory provider and context engine with BM25 recall and cache-stable folding, plus many extras; it has real plugin depth, but its LoCoMo comparison against mem0 and others has no benchmark harness in the repo, its scope is sprawling, the hive server is unauthenticated, and the README is hard to follow."
  }
 ]
}