Best way to provide language diagnostics from a remote server resolved by a JupyterServerProvider?

[kevineger]

I've implemented a JupyterServerProvider which resolves remote kernels, running on remote machines. The servers running these kernels have some libraries pre-installed (e.g. numpy). Pylance is reporting them as missing, e.g.: Import "numpy" could not be resolvedPylance (reportMissingImports).

Is there anything available I can configure so that the user's local Python environment (where VS Code is running) is not used, and instead the remote one is? Remoting (by SSH or otherwise) into the underlying server is not an option. Will this require some kind of typing mirroring and telling Pylance to use them for the connected notebook? Sorry for the naive question, this is all new to me. I'll keep digging and update this discussion as I find things.

So far the only option sticking out to me I could explore would be to expose a language server on my remote server. Then my extension would register a new language client per server. Not sure how exactly I'd trigger/toggle it for just notebooks connected to specific servers... Need to investigate further.

Current /api/kernelspecs return:

{
  "default": "python3",
  "kernelspecs": {
    "julia": { ... },
    "python3": {
      "name": "python3",
      "spec": {
        "argv": [
          "/usr/bin/python3",
          "-m",
          "custom_launcher",
          "-f",
          "{connection_file}"
        ],
        "env": {},
        "display_name": "Python 3 (ipykernel)",
        "language": "python",
        "interrupt_mode": "signal",
        "metadata": {
          "debugger": true,
          "interpreter": { "path": "/usr/bin/python3" }
        }
      },
      "resources": { ... }
    },
    "ir":  { ... },
  }
}

The Python Language Server logs seem to be only using the local Python env:

2025-10-15 17:51:18.462 [info] (Client) Pylance client (2025.8.3) started with python extension (2025.16.0)
2025-10-15 17:51:19.605 [info] (2391463) Server root directory: file:///usr/local/google/home/keger/.vscode-server/extensions/ms-python.vscode-pylance-2025.8.3/dist
2025-10-15 17:51:19.607 [info] (2391463) Pylance language server 2025.8.3 (pyright version 1.1.405, commit d64f3fa0) starting
2025-10-15 17:51:19.613 [info] (2391463) Starting service instance "sample_nbs" for workspace "/usr/local/google/home/keger/src/sample_nbs"
2025-10-15 17:51:19.758 [info] (2391463) Starting Mcp server id:5551 ...
2025-10-15 17:51:19.799 [info] (2391463) No configuration file found.
2025-10-15 17:51:19.799 [info] (2391463) No pyproject.toml file found.
2025-10-15 17:51:19.800 [info] (2391463) Setting environmentName for service "sample_nbs": "3.12.11 (global)"
2025-10-15 17:51:19.800 [info] (2391463) Setting pythonPath for service "sample_nbs": "/usr/local/google/home/keger/.local/bin/python3.12"
2025-10-15 17:51:19.800 [info] (2391463) No include entries specified; assuming /usr/local/google/home/keger/src/sample_nbs
2025-10-15 17:51:19.800 [info] (2391463) Auto-excluding **/node_modules
2025-10-15 17:51:19.801 [info] (2391463) Auto-excluding **/__pycache__
2025-10-15 17:51:19.801 [info] (2391463) Auto-excluding **/.*
2025-10-15 17:51:19.826 [info] (2391463) Assuming Python version 3.12.11.final.0
2025-10-15 17:51:19.827 [info] (2391463) Assuming Python platform Linux
2025-10-15 17:51:19.889 [info] (2391463) Execution environment: 3.12.11 (global)
2025-10-15 17:51:19.889 [info] (2391463)   Extra paths:
2025-10-15 17:51:19.889 [info] (2391463)     /usr/local/buildtools/current/sitecustomize
2025-10-15 17:51:19.889 [info] (2391463)   Python version: 3.12.11.final.0
2025-10-15 17:51:19.889 [info] (2391463)   Python platform: Linux
2025-10-15 17:51:19.889 [info] (2391463)   Search paths:
// ...

[kevineger]

I've managed to hookup a LanguageClient to the language server I'm running on the machine hosting the remote kernel.

Now the complexity is having a connection to each remote language server (unique per JupyterServer) and mapping a given Notebook to it. In other words, only send and receive from the language server that's running on the server which the notebook has a connected kernel to. Since I don't own the notebook controller, I'm unsure the existing API (VS Code + Jupyter) will support that.