Plugin catalog: nvidia-app and nvidia-broadcast (official, Windows only)

[alt-glitch]

Two NVIDIA plugins join the catalog: nvidia-app and nvidia-broadcast, Windows only, official tier.

What they are

Portable Agent Plugins (no Python of their own) from NousResearch/hermes-nvidia, one per NVIDIA application,
each in its own subdirectory of the repo, both pinned to 5023f3a4ea. Each carries one MCP server entry
(the vendor's own local server), one skill, and an application declaration under
plugin.json → extensions["com.nousresearch.hermes"].servers.

entry	app it needs	how the server is reached	tools seen live
nvidia-app	NVIDIA App ≥ 11.0	endpoint + token re-read from the app's own McpServer/server.json on every connection	12
nvidia-broadcast	NVIDIA Broadcast ≥ 2.3	loopback gateway, 18100-18109; needs the app running in the user's desktop session	10

What gates them

flowchart LR
  A["hermes plugins install nvidia-app<br/>(or the catalog card / deep link)"] --> B{"platforms: [windows]<br/>catalog gate"}
  B -- other OS --> R1["refused: unsupported_os"]
  B -- Windows --> C{"declaration:<br/>app present ≥ min version?"}
  C -- no --> R2["refused before the swap:<br/>missing_app / version_too_old"]
  C -- yes --> D["installed + enabled"]
  D --> E{"liveness at connect time"}
  E -- app running --> F["tools registered"]
  E -- not running --> G["card pill + catalog line:<br/>'start NVIDIA Broadcast'"]

Loading

The catalog platforms field is the coarse gate (plugins_cmd_catalog.py::_refuse_unsupported_catalog_platform);
the per-server declaration is the fine one (NS-942, #119049). Neither launches anything.

Evidence

Two real Windows machines, real sessions, upstream main:

machine	result
ARM64 laptop (vendor drop builds)	MCP: registered 22 tool(s) from 2 server(s)
x64 desktop (RTX 5090)	same 22/2 from the temp-home run; then installed from the desktop deep link into the real profile, model called NVIDIA App tools; Broadcast correctly reported not running (it cannot start inside an RDP session)
this Mac	hermes plugins validate → server availability: … unsupported_os, install refused

Rows with timestamps are in the plugin repo's EVIDENCE.md.

Expected CI state

• validate job (schema + static checks): green, 240 entries.
• pinned-sha job: red for these two entries until the repo is public. It does a plain git clone on a
  public runner; the repo is private until release. That is the only expected failure. Flip the repo public, re-run,
  and the job clones and validates at the pin.

Notes for the reviewer

• Tool names: on current main these register as mcp__agent_plugin_nvidia_app_<sha8>__nvidia_app__…, over 64
  chars and hash-clamped. Portable plugin MCP tools keep their names instead of a hash (64-char provider cap) #119263 fixes the loader namespace so they come out as
  mcp__nvidia_app__nvapp_client_get_driver_status. Independent of this PR; better together.
• security scan — caution on nvidia-broadcast: one line in the vendor skill (dump_all_env, a documented
  troubleshooting step). Read at admission per catalog rule 7; a catalog install accepts it without a prompt.

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 47b1077 — feat(plugin-catalog): nvidia-app and nvidia-broadcast (offic}

debug info

CI timings

CI timings · View report · View job

Wall time 5m51s vs 6m19s (-7.4%). 7 job(s) slower, 3 faster, 2 unchanged.

• Python tests / Run tests: -60.0s
• OS-specific tests / macOS-only tests: -16.0s
• Python tests / e2e: +13.0s
• Python lints / Windows footguns (blocking): +10.0s
• Detect affected areas: +10.0s