{
  "batch": "final-03",
  "ranking": [
    {
      "rank": 1,
      "pr": 150644,
      "name": "hermes-canvas",
      "total": 29,
      "why": "U5/R4/P5/C5/B5: Runtime-derived plan status, mid-turn user steering, subagent attribution and a finished Desktop canvas form the strongest broadly useful coding companion; the shell-write detector remains heuristic and live Desktop interaction was not independently exercised (pinned source: https://raw.githubusercontent.com/0xbyt4/hermes-canvas/d6eeff0d6f7ec653262537bdd7bb001cc6d3c7f8/README.md)."
    },
    {
      "rank": 2,
      "pr": 152458,
      "name": "housemate",
      "total": 29,
      "why": "U5/R4/P5/C5/B5: A real gateway platform, paired browser satellites, room arbitration, persistent timers and actor-aware approvals most directly realize Hermes on users' own hardware, narrowly behind Canvas because audio hardware and model downloads add setup risk and live room audio remains unverified (pinned source: https://raw.githubusercontent.com/RaYYeR220/housemate/09c1a1d0a88bfa8762cee6b3ccf94271b0f172c5/README.md)."
    },
    {
      "rank": 3,
      "pr": 153020,
      "name": "breadcrumbs",
      "total": 29,
      "why": "U5/R4/P5/C5/B5: The memory provider and context engine share exact-addressed crumbs, deterministic zoom and recall receipts, making a distinctive daily improvement to the agent itself; append-only retention and additional profile/sharing machinery make it less effortless than the leaders, and published benchmark advantages were not independently reproduced (pinned source: https://raw.githubusercontent.com/breadcrumbs-agency/hermes-breadcrumb-plugin/b8975645e5174edbcc00d5dc86a78e536bce938b/README.md)."
    },
    {
      "rank": 4,
      "pr": 150596,
      "name": "ariadne",
      "total": 29,
      "why": "U4/R5/P5/C5/B5: Recording and replay middleware, request diffs, contract checks and a Desktop debugger substantially improve Hermes itself, with an independently successful offline roundtrip, but debugging and upgrade verification serve fewer everyday users than the three above (pinned source: https://raw.githubusercontent.com/miracle352/hermes-ariadne/128c95967a0761ad24d3d224441d1502c43bbb69/README.md)."
    },
    {
      "rank": 5,
      "pr": 149640,
      "name": "hermes-ark",
      "total": 28,
      "why": "U5/R4/P5/C4/B5: Automatic pre-write snapshots, authenticated encryption, credential exclusions and previewed item-level recovery protect a user's accumulated agent state more broadly than most specialist tools; local recovery code was checked but live S3 recovery remains unverified (pinned source: https://raw.githubusercontent.com/medomohamed22/hermes-ark/a23e529c1f636aa49d974576cdc5cb15fc8242d3/README.md)."
    },
    {
      "rank": 6,
      "pr": 150556,
      "name": "hermes-ui-cards",
      "total": 28,
      "why": "U5/R4/P5/C4/B5: Schema-checked interactive cards, persisted drafts, answer receipts, undo and text fallbacks make ordinary Hermes conversations more usable across many tasks, with the documented eventual synchronization and advisory confirmations preventing a perfect reliability score (pinned source: https://raw.githubusercontent.com/robbyczgw-cla/hermes-ui-cards/c40df4d41cd5439cf4902995fd4a9e5505483f25/README.md)."
    },
    {
      "rank": 7,
      "pr": 153851,
      "name": "foreman",
      "total": 28,
      "why": "U5/R4/P5/C4/B5: Diff-based test-integrity checks, bounded second-model review and person-only exceptions improve coding agents at the tool and finish boundaries, but its openly documented command and merge bypasses keep it behind the more general interaction and recovery plugins (pinned source: https://raw.githubusercontent.com/PedValme/foreman/c15453cb2fc07fa674d6f7b548a553b3ce9e6f5d/plugin/README.md)."
    },
    {
      "rank": 8,
      "pr": 153714,
      "name": "tenure",
      "total": 28,
      "why": "U4/R4/P5/C5/B5: Per-action probation, evidence-based trust transitions, authenticated undo snapshots and an autonomy passport are an unusually complete change to agent oversight, although the large heuristic classifier and Hermes approval-mode semantics limit the strength of its always-ask language (pinned source: https://raw.githubusercontent.com/otsoa-dubief/tenure/f4a62855e592b9e170ab0c54d9278aabef05f370/tenure/README.md)."
    },
    {
      "rank": 9,
      "pr": 150840,
      "name": "kata",
      "total": 28,
      "why": "U4/R4/P5/C5/B5: Mining repeated tool sequences into owner-approved deterministic tools, rechecking rendered argument effects and detecting schema drift create a real learning capability beyond skills, while needing repeated sessions and review before the first useful compiled routine narrows immediate adoption (pinned source: https://raw.githubusercontent.com/opeblow/kata/cc59b08f1bbeeb02524c060faac04edb34de0564/README.md)."
    },
    {
      "rank": 10,
      "pr": 153844,
      "name": "fine-print",
      "total": 28,
      "why": "U4/R5/P4/C5/B5: Its hooks expose otherwise silent context and tool limits without extra model calls, including a directly verified full-search-page annotation, and conservative verified-versus-fixture detector handling earns strong reliability, while the disclosed Desktop notice limitation reduces polish (pinned source: https://raw.githubusercontent.com/Joko-F/hermes-fine-print/7b574f1b5c8dc8eafe3ea9f87ae6b0111853f0ba/README.md)."
    },
    {
      "rank": 11,
      "pr": 151887,
      "name": "cmi",
      "total": 27,
      "why": "U5/R4/P4/C4/B5: Deterministic context selection, cache-aware folding and byte-exact recall address a recurring cost of using Hermes itself and win this score tie on daily retention, but private compressor/pricing dependencies and less cohesive plugin-specific onboarding reduce confidence and polish; advertised savings were not independently reproduced (pinned source: https://raw.githubusercontent.com/jeremiebergel/context-manager-intelligence/399468f0687c051476961e4bb045eb84263be6b1/README.md)."
    },
    {
      "rank": 12,
      "pr": 152691,
      "name": "grimoire",
      "total": 27,
      "why": "U4/R4/P5/C4/B5: The locked read-only DuckDB engine, numerical provenance checks and Desktop data exploration make a polished analytical workspace, with CSV aggregation and write rejection verified inside Hermes, but data analysis is a narrower installation case than general context efficiency (pinned source: https://raw.githubusercontent.com/bettersandeep/grimoire/4824773040b131f3179e5896d71d1435908758a7/README.md)."
    },
    {
      "rank": 13,
      "pr": 150251,
      "name": "gemini-live",
      "total": 27,
      "why": "U4/R4/P5/C4/B5: A full-duplex voice engine that delegates into Hermes while preserving scoped secrets and human approvals is useful and polished, but overlaps familiar voice interfaces and depends on Google's service; backend error paths were checked while live Gemini audio remains unverified (pinned source: https://raw.githubusercontent.com/cineglobe/hermes-gemini-live/b6c2f24354aeed1a8b5651b62944d14ed96a472d/README.md)."
    },
    {
      "rank": 14,
      "pr": 150800,
      "name": "hypnos",
      "total": 27,
      "why": "U3/R4/P5/C5/B5: Actual SFT/DPO orchestration, held-out promotion gates, LoRA routing and lineage-based forgetting are among the most original extensions to Hermes, but a separate trainer and demanding setup make this a specialist install, and synthetic demonstrations do not establish everyday gains or independently verified training (pinned source: https://raw.githubusercontent.com/0xNoramiya/hypnos/0a809b9085fdeaed63fd909b927afcc41de512de/README.md)."
    },
    {
      "rank": 15,
      "pr": 150353,
      "name": "fluent",
      "total": 27,
      "why": "U4/R4/P5/C4/B5: Installed-version command cards and a pre-tool guard can prevent invented CLI flags deterministically, a practical improvement to Hermes's execution loop, but explicit adoption, conservative long-option coverage and local-backend restrictions reduce breadth relative to the entries above (pinned source: https://raw.githubusercontent.com/Bruno-coder21/fluent/182c4301f2437e784e9022f1f38576e1d68c3f5f/README.md)."
    },
    {
      "rank": 16,
      "pr": 152908,
      "name": "contract-necromancer",
      "total": 27,
      "why": "U4/R4/P5/C4/B5: Person-frozen test contracts, a real pre-verify test runner, write guards and verifiable receipts strengthen coding completion beyond a prompt, but setup, test-runtime budgets and heuristic shell protection make it less broadly retainable than the higher-scoring coding and workflow tools (pinned source: https://raw.githubusercontent.com/vijaygopalbalasa/hermes-necromancer/0ff7bec40aba10091a84f94025d4471984fd7386/README.md)."
    },
    {
      "rank": 17,
      "pr": 153809,
      "name": "living-spaces",
      "total": 27,
      "why": "U5/R3/P5/C4/B5: Persistent sandboxed apps, transactional records, revisions and native upkeep add a compelling Desktop capability, with creation and traversal rejection verified, but the pinned README explicitly discloses private cron integrations and a raw Desktop bridge that weaken catalog compatibility and reliability (pinned source: https://raw.githubusercontent.com/akinduroifedayo/living-spaces/e6145ccbd028edf55417822020426f4830343eb6/src/living-spaces/README.md)."
    },
    {
      "rank": 18,
      "pr": 153955,
      "name": "shadow",
      "total": 26,
      "why": "U4/R3/P5/C4/B5: Blind model trials in both answer orders, isolation guards and reversible switching form a polished agent-selection product, but replaying current files is only an approximation of past work and provider-dependent trial behavior remains unverified, so it falls below simpler daily improvements (pinned source: https://raw.githubusercontent.com/akinduroifedayo/shadow/38173bc9c696df3b1f87b5b9b60ccecd36a260ab/README.md)."
    },
    {
      "rank": 19,
      "pr": 152651,
      "name": "workroom",
      "total": 26,
      "why": "U4/R2/P5/C5/B5: Its integrated document workspace is ambitious and polished, but at the field's pinned commit adapters/snapshotter.py runs an unbounded _CD.search before target handling, and a direct probe with 300000 newline characters exceeded a three-second cutoff, materially undermining hook reliability at the judged revision (pinned source: https://raw.githubusercontent.com/Pratiikpy/workroom-for-hermes/c5f6bf334c1163951e203dc7c5777faadfe49795/adapters/snapshotter.py)."
    },
    {
      "rank": 20,
      "pr": 149738,
      "name": "trust-ladder",
      "total": 25,
      "why": "U5/R2/P4/C4/B5: Turning corrections into tool-boundary guards is broadly useful, but the pinned compiler accepts and example-verifies (a|aa)+$ with its full search window, then evaluate_match on 40 a characters followed by ! exceeds a three-second cutoff, so the claimed regex vetting can still stall the agent (pinned source: https://raw.githubusercontent.com/merdoks/trust-ladder/fae4e9df58767e5e3832c8f65936b97868bc98fc/compiler.py)."
    }
  ]
}