{
  "batch": "deep-44",
  "setup": "Cloned NousResearch/hermes-agent source at fdde27e7dae174cb4ba842589e92508852b7dd9f; installed with uv sync --locked --python 3.14. Ran pinned plugins with tryit.py inside Hermes against its scripted fake model; installed test dependencies in scratch.",
  "plugins": [
    {
      "pr": 152481,
      "name": "argus",
      "validate": "passed; catalog warned about private Hermes core imports",
      "inside": "Loaded with five hooks, an auxiliary task and /argus; no direct tool. In Hermes, skills_list, a missing-file read, and tool_search returned; hermes argus status reported three tool outcomes and one completed turn.",
      "tests": "115 passed",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 5,
      "beyond": 5,
      "total": 28,
      "rank": 3,
      "best": "Intent-linked tool and skill history with human review can improve an agent’s own learning.",
      "worst": "Catalog warnings identify private Hermes imports, and deriving a busy agent’s history can require many model calls.",
      "blocked": null
    },
    {
      "pr": 151804,
      "name": "dress-rehearsal",
      "validate": "passed",
      "inside": "Loaded with two hooks and /rehearse; no direct tool. Three planned built-in calls returned, including a missing-file error. Through Hermes plugin dispatch, /rehearse on blocked and recorded a write without creating the file; show and off worked, and malformed drop returned usage.",
      "tests": "84 passed",
      "useful": 5,
      "reliable": 5,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "total": 29,
      "rank": 1,
      "best": "An armed write was blocked and recorded inside Hermes for review before a person-triggered replay.",
      "worst": "A rehearsal spans the profile, and replay cannot learn from effects of earlier recorded steps.",
      "blocked": null
    },
    {
      "pr": 149313,
      "name": "hermes-ros",
      "validate": "passed",
      "inside": "Loaded with nine deferred tools and one safety hook. Tool discovery worked; mock robot status and camera snapshot returned, a 0.1 m/s move changed pose by 0.01 m, and a 3 m/s move was blocked by over_speed_linear.",
      "tests": "70 passed, 10 live-ROS tests skipped",
      "useful": 3,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "total": 26,
      "rank": 6,
      "best": "The mock robot moved safely and the same tool path refused an over-speed command.",
      "worst": "Real ROS hardware was unavailable here, and the interlock covers ros_* tools rather than every path to a robot.",
      "blocked": null
    },
    {
      "pr": 150556,
      "name": "hermes-ui-cards",
      "validate": "passed",
      "inside": "Loaded with hui_present and three hooks; its Desktop half is supplied separately. Deferred discovery found the tool; plan and compare demos rendered useful CLI text, while an invalid demo was rejected by Hermes argument validation.",
      "tests": "Python: 463 passed, 4 failed, 12 skipped; two failures were sandbox socket permission errors and two were Python/JavaScript validator parity. Desktop JavaScript: 197 passed.",
      "useful": 5,
      "reliable": 4,
      "polish": 5,
      "creative": 4,
      "beyond": 5,
      "total": 28,
      "rank": 2,
      "best": "Plan and compare cards rendered in Hermes, with text fallback and 197 passing Desktop tests.",
      "worst": "Two invalid fixtures disagree between Python and JavaScript validators, and the Desktop half needs enabling.",
      "blocked": null
    },
    {
      "pr": 151301,
      "name": "confidential-mode",
      "validate": "passed; catalog warned about 31 private core import sites and in-process-only isolation",
      "inside": "Loaded with 11 hooks and middleware, no direct tool. Planned terminal, missing-file read, and search calls returned. The local-tier preflight refused an auto vision route, passed after local configuration, then locked; its hook blocked unattended web_search and allowed read_file.",
      "tests": "Full suite: 3080 passed, 3 failed, 5 errors, 8 skipped, 6 xfailed. All five errors were missing HERMES_HOME and passed on rerun; one failure required absent /tmp, and two expected built-in tools from an older Hermes commit.",
      "useful": 4,
      "reliable": 3,
      "polish": 5,
      "creative": 5,
      "beyond": 5,
      "total": 27,
      "rank": 5,
      "best": "Preflight detected an off-tier vision route, and the active lock blocked unattended web_search while allowing a local file read.",
      "worst": "Thirty-one private Hermes imports and in-process isolation make this sensitive lock vulnerable to host changes; shell and side-call limits remain.",
      "blocked": null
    },
    {
      "pr": 151135,
      "name": "wake-me-when",
      "validate": "passed",
      "inside": "Loaded with seven deferred stream tools, one hook and a CLI command. Deferred discovery worked; the bundled demo watch started, listed, stopped, and rejected an empty URL/condition request.",
      "tests": "Isolated standalone suite: 115 passed, 7 skipped; the Hermes-venv run had 112 passed, 3 failed, 7 skipped because Hermes plugin storage shared state between tests.",
      "useful": 4,
      "reliable": 4,
      "polish": 5,
      "creative": 5,
      "beyond": 5,
      "total": 28,
      "rank": 4,
      "best": "The bundled keynote watch started and stopped inside Hermes, and its isolated end-to-end tests passed.",
      "worst": "Real external live-stream services and message delivery were not exercised here; the plugin needs ffmpeg for media ingestion.",
      "blocked": null
    }
  ]
}