nvsh

You already have an agent CLI — claude, codex, qwen, kiro, agy or a local Nemotron. nvsh puts it behind your bash prompt on NVIDIA hardware: commands run exactly as they do today, and when one fails, the agent you already trust diagnoses it and proposes a fix.

Works like a shell. Helps when things break. Fixes when you let it.

Recorded on a DGX Spark with the scripted demo adapter — the diagnosis text is a committed fixture, the hook, daemon, panel and approval are real; the same session on a Jetson AGX Thor and an AGX Orin.

Install

uv tool install nvsh        # or: pipx install nvsh
nvsh setup

nvsh setup inserts one marked block into your .bashrc (with a timestamped backup) that appends a function to bash's PROMPT_COMMAND. It never wraps bash and never becomes your login shell.

Set up

Work with it

A failing command opens an inline diagnosis panel. Nothing runs without your approval — nvsh proposes, you accept, edit or reject.

Exit codes that aren't errors — Ctrl-C (130), SIGPIPE (141), grep finding nothing — never trigger it, and automatic calls are rate-limited.

Safety first

Nothing leaves the machine except to the agent you trust. On a successful command nvsh makes no network call at all. On a qualifying failure a redacted, bounded context slice — the command line, its exit status, at most 64 KB of its output, and the detected platform block — goes only to the agent you chose, and nowhere else.

With claude or codex, that agent is a third-party hosted service: your failure context leaves your network. With pi/Nemotron or openai-compat pointed at localhost or your LAN, it does not.

Redaction runs first, before anything leaves the process: HF_TOKEN=, --api-key, Authorization: headers and .env-style assignments are scrubbed from the context. nvsh context --show prints the post-redaction bytes so you can check before you trust it. nvsh never edits a harness's own settings or trust files — it passes launch flags and protocol-level policy, and reports what it finds.

What nvsh never does

• Run an agent-suggested command without your confirmation.
• Add latency or a network call to a successful command.
• Wrap or replace bash — it hooks into the bash you already run.
• Replace jetson-cli / dgx-spark-cli; it calls them when installed.

Not yet: macOS and zsh are untested — #11.

What lands where

More:

• architecture — hook over wrapper, and the parked login-shell mode.
• platforms — where each detected device value comes from.
• shell integration and daemon — the @target grammar and the warm session.
• CLAUDE.md — contributor conventions.
• demos — how the recordings above were made, and the re-record loop.

License

Apache 2.0 — see LICENSE.