Nushell support

[maxim-uvarov]

Hey guys! It would be really cool if Codex worked with Nushell.
Nushell is gaining momentum now. Just check how steadily it is growing: https://trends.google.com/trends/explore?hl=en-GB&tz=180&date=today+5-y&hl=en-GB&q=%2Fg%2F11lkrdp6pf&sni=3

I believe Nushell, in general, is a very good candidate for your reinforcement learning exercises.

[tibo-openai]

What would be needed to make it work? And out of curiosity, are you one of the maintainers or an enthusiast of it?

[maxim-uvarov]

@tibo-openai thank you for your attention!

What would be needed to make it work?

In general, OpenAI models currently don't know the Nushell language well. They often hallucinate Python commands and syntax when writing Nushell code. So, I guess something needs to be done on that level first. Later, it would be cool if Codex used Nushell for its intermediate steps and outputted Nushell's pretty data structures for user observation.

Nushell is a feature-rich shell and programming language. It has built-in support for working with JSON, TOML, YAML, XLSX, etc., as data structures. It can query SQLite databases out of the box. It supports plugins, with the most notable one being polars. Additionally, it is cross-platform, so it works on Windows in the same way it works on Unix.

And out of curiosity, are you one of the maintainers or an enthusiast of it?

I'm an enthusiast with over 3 years of Nushell experience, and I’m part of a community that communicates with the core team. Therefore, I believe it’s likely that we can get support from the maintainers if needed.

On my side, I would be happy to invest any amount of time if I could help OpenAI write better Nushell scripts.

[maxim-uvarov]

Nushell's bracket-based syntax is significantly more token-efficient for LLM development than indentation-sensitive languages.

Working with Open AI Codex on Python: constant indentation edits. Every nested change reformats multiple lines, wasting tokens on whitespace instead of logic.

Why Nushell works better:

• Explicit delimiters ({}, []) define scope, not whitespace
• Atomic edits - modify individual blocks without cascading changes
• Zero indentation overhead

Nushell is inherently more LLM-friendly by nature, as it allows writing single-word literals without quotes, while literals with spaces require quotes, and delimits list elements using spaces only, without commas.

Token efficiency ranking:

• Most efficient: Nushell, Rust, JavaScript, Go (delimiter-based)
• Least efficient: Python, YAML, Haskell (whitespace/layout-dependent)

For shell scripting with LLM assistance, Nushell's syntax is fundamentally more compatible with how these models operate.