{
 "batch": "read2-024",
 "plugins": [
  {
   "pr": 0,
   "name": "done-bell",
   "does": "Rings an OS notification and optional chime and shows a busy/done/needs-you status chip when a Hermes turn finishes or waits for approval.",
   "useful": 4,
   "reliable": 3,
   "polish": 3,
   "creative": 2,
   "beyond": 3,
   "total": 18,
   "best": "Small, safe hook-based plugin (never-raise hooks, no keys, no chat access) that solves a real annoyance of waiting on long turns.",
   "worst": "Thin tests and a README whose install steps point at the author's own Windows paths and a manual copy of the Desktop pane.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 0,
   "name": "homeassistant",
   "does": "Official Home Assistant gateway platform plus four REST tools so Hermes can react to state changes and query and control smart-home devices.",
   "useful": 5,
   "reliable": 5,
   "polish": 4,
   "creative": 3,
   "beyond": 5,
   "total": 27,
   "best": "Large test suite, secret-store token, operator allow/deny lists and a blocked list of code-executing service domains make device control safe.",
   "worst": "Mostly a port of a long-existing core integration, so little is new beyond the extraction.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 152976,
   "name": "agentline",
   "does": "Gives Hermes a real US/Canadian phone number via AgentLine: place calls with a hosted voice, send/receive SMS, use SMS as a gateway chat, and answer live callers' questions mid-call.",
   "useful": 4,
   "reliable": 4,
   "polish": 4,
   "creative": 4,
   "beyond": 5,
   "total": 26,
   "best": "A full platform adapter plus a pre_tool_call hook that sends every call, text and number purchase to Hermes's approval gate and always blocks 911/premium numbers, with the key read from the secret scope.",
   "worst": "Depends on a paid third-party service (US/Canada only) and has no screenshots or demo of a real call.",
   "standout": true,
   "blocked": null
  },
  {
   "pr": 152309,
   "name": "no-manual",
   "does": "Helps people assemble furniture without instructions: finds the official manual, reverses disassembly photos into a manual, or solves flat-pack carcass furniture from part photos, then checks a photo of each step.",
   "useful": 3,
   "reliable": 4,
   "polish": 5,
   "creative": 5,
   "beyond": 4,
   "total": 25,
   "best": "Real geometry solver, step-image renderer and Desktop 3D viewer behind vision calls via ctx.llm, with child and load-bearing furniture blocked twice in code (tool handlers and a pre_tool_call hook).",
   "worst": "A home-furniture assembly helper is a niche use for most Hermes users, and its output is only as good as the vision model's reading of the photos.",
   "standout": true,
   "blocked": null
  },
  {
   "pr": 153693,
   "name": "offgrid",
   "does": "A platform adapter that lets people chat with Hermes by plain SMS, with a USSD menu fallback for basic phones, through Africa's Talking.",
   "useful": 3,
   "reliable": 4,
   "polish": 3,
   "creative": 4,
   "beyond": 5,
   "total": 24,
   "best": "A genuine new channel for people with no mobile data, with secrets read through Hermes's scoped secret store, constant-time callback token checks, an allowlist, rate limits and offline tests.",
   "worst": "Sandbox-only by its own admission (no live carrier or handset tested) and it needs a public HTTPS endpoint the user must arrange.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 153714,
   "name": "tenure",
   "does": "Replaces flat approvals with per-action-type earned autonomy: everyday work runs with 24-hour undo, a mistake puts that action type on probation, and verified evidence earns trust back, with a signed passport, Desktop page and dashboard.",
   "useful": 4,
   "reliable": 4,
   "polish": 5,
   "creative": 5,
   "beyond": 5,
   "total": 28,
   "best": "Changes how Hermes itself decides when to ask, through seven hooks, a fine-grained action classifier, an undo store and a hash-chained passport, backed by dozens of test files and five locales.",
   "worst": "Very large and complex (a 5,200-line classifier and an 8,000-line Desktop bundle), so its trust decisions rest on heuristics that are hard to audit.",
   "standout": true,
   "blocked": null
  },
  {
   "pr": 152619,
   "name": "proofly",
   "does": "Checks an agent's claimed Base mainnet ETH/USDC transfers against on-chain receipts and logs over read-only JSON-RPC, and rewrites tool results and final replies that claim a payment the chain does not support.",
   "useful": 3,
   "reliable": 4,
   "polish": 3,
   "creative": 4,
   "beyond": 4,
   "total": 22,
   "best": "Strict, read-only verifier (chain-id check, exact raw amounts, confirmations, time budget) wired into post_tool_call, transform_tool_result and transform_llm_output, tested against recorded real Base transactions.",
   "worst": "Narrow scope (Base mainnet, simple transfers, only tools the user names) and no screenshots or video, so it serves few Hermes users.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 153822,
   "name": "lights-out",
   "does": "Puts every billable AWS CLI create through Hermes's approval gate with its monthly price, records a ledger with delete commands, and prints a receipt of what is still running after each turn, with /lights out to tear it down.",
   "useful": 4,
   "reliable": 4,
   "polish": 5,
   "creative": 4,
   "beyond": 4,
   "total": 25,
   "best": "A pre_tool_call price gate that fails closed with nobody to approve, a ledger matched from real CLI output, and a benchmark plus demo video recorded against a moto AWS stand-in.",
   "worst": "Covers only the AWS CLI in one region's price table, so creates through SDKs, Terraform or other clouds go uncounted.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 150913,
   "name": "egress-guard",
   "does": "Policy-driven hooks that redact credentials in tool arguments, send data-upload shell commands to unlisted hosts to the approval prompt, and suppress small counts in tables (statistical disclosure control) in results and replies.",
   "useful": 3,
   "reliable": 4,
   "polish": 3,
   "creative": 4,
   "beyond": 4,
   "total": 22,
   "best": "Brings research-data-centre output vetting to agents in code outside the model, including secondary suppression so a hidden cell cannot be recovered from totals, with a broad test suite.",
   "worst": "Upload detection is pattern matching that a determined command can slip past, and fail_mode defaults to open on internal errors.",
   "standout": false,
   "blocked": null
  },
  {
   "pr": 149738,
   "name": "trust-ladder",
   "does": "Turns a user's correction into the strongest enforcement it can prove (a pre_tool_call block, a post-write check, an injected rule or an approval prompt), verifies each guard against the mistaken and corrected calls, gates 'done' on verifier commands and keeps a ledger of catches.",
   "useful": 4,
   "reliable": 4,
   "polish": 5,
   "creative": 5,
   "beyond": 4,
   "total": 26,
   "best": "Guards are installed only if they match the bad example and spare the good one, regexes are profiled against catastrophic backtracking, and the model can add rules but never remove them.",
   "worst": "Guards remain best-effort pattern matches on tool arguments, which its own manifest says are not a sandbox.",
   "standout": true,
   "blocked": null
  },
  {
   "pr": 152475,
   "name": "hallway-test",
   "does": "Runs five simulated testers with code-enforced limits (low vision, screen reader, colour blindness, tremor, first-time visitor) through a real Playwright browser on your page using your own model, and returns findings, step replays and a before/after comparison.",
   "useful": 4,
   "reliable": 4,
   "polish": 5,
   "creative": 5,
   "beyond": 4,
   "total": 26,
   "best": "The limits are enforced in code from computed font sizes, contrast, colours and the accessibility tree, so the model only sees each persona's filtered view rather than role-playing it.",
   "worst": "Needs Playwright browsers at 300-600 MB each and runs many model calls per test, which is heavy for small machines and local models.",
   "standout": true,
   "blocked": null
  },
  {
   "pr": 150825,
   "name": "phone-autopilot",
   "does": "Builds on phone-harness so the agent can explore an Android task once, save it as an approved routine with checkpoints, replay it without the model, heal it when an app changes, and schedule it daily.",
   "useful": 4,
   "reliable": 4,
   "polish": 4,
   "creative": 4,
   "beyond": 4,
   "total": 24,
   "best": "Every routine save and every run of unapproved code goes through Hermes's approval prompt pinned to the script's sha256, and runs happen in a killable subprocess with checkpoints and failure screenshots.",
   "worst": "Depends on a separate phone-harness plugin plus adb, and its system helpers work only on Android.",
   "standout": false,
   "blocked": null
  }
 ]
}