dsh-nushell-only

When driving DeepSeek Harness with Nushell, the biggest friction is not using a different shell but the model still writing commands with bash and PowerShell habits - 2>&1, $x = 1, foreach, Get-Content, ls -a - all of which land on the nu parser and produce an unfamiliar error that needs to be decoded word by word. A non-zero exit from an external command even silently interrupts the entire pipeline, leaving behind a baffling [exit code: 1].

dsh-nushell-only swaps dsh sole shell seam to Nushell and systematically eliminates the above friction across four layers: the execution layer replaces bash-sandbox and pwsh-sandbox with a nu executor, unifying all ctx.shell consumers (model tools, background tasks, hook bridges) under Nushell; the interception layer rejects bash and PowerShell habit syntax before spawn and provides the Nu equivalent directly in the error message; the teaching layer injects a Nushell grammar manual, a bash/PowerShell cross-reference table, and a real-failure case directory into the model system prompt; the correction layer appends error-code-specific fix hints after a failed execution. Rules are not invented in a vacuum - every one is derived from statistical analysis of 1503 Nushell failures across 8168 shell invocations in 109 real sessions, and the refusal test cases are taken directly from those actual failing commands.

Ideal for developers already using dsh who want Nushell as their sole shell and want the model to stop repeatedly stumbling on syntax. You do not need to manually rewrite every bash hook or manually translate every error message - the plugin governs every link in the chain from the model wanting to write bash to nu refusing to execute.