---name: nushell-prodescription: |---

Nushell Pro

Write secure, idiomatic, portable, and testable Nushell. Keep the main workflow in this file; load detailed references only when the task needs them.

Operating Workflow

1. Identify the task: new script, debugging, review, refactor, module design, migration, performance work, security audit, or Bash conversion.

2. Read the target .nu files, nearby tests, AGENTS.md, and existing project conventions before changing code.

3. Establish both the project's supported Nu range (from CI, documentation, or project configuration) and the active local version. Do not assume the local binary defines the compatibility target. Prefer the official English docs; translations can lag. When local docs/source checkouts are available, inspect them and record their revision separately from the installed Nu version. From any shell, prefer:

   nu --version
   

   When already inside Nushell or when structured version data is needed, use:

   version | get version
   

4. Load the smallest relevant reference set:

   Task	Reference
   Nu 0.116 migration, completions, flag spreading, cleanup	Nu 0.116 Migration
   Nu 0.115 migration, YAML, CLI args, command changes	Nu 0.115 Migration
   Nu 0.114 migration and version compatibility	Nu 0.114 Migration
   Strings, regex/globs, generated JS/JSON	String Formats
   Security, paths, credentials, destructive operations	Security
   Script/code review	Script Review and Anti-Patterns
   Bash/POSIX conversion	Bash to Nushell
   Modules, exports, scripts, tests	Modules & Scripts
   Daemons, background jobs, E2E smoke tests	Daemon & E2E Smoke Tests
   Types, records, lists, conversions	Data & Type System
   Streaming, closures, performance, diagnostics	Advanced Patterns
   Large columnar data	Dataframes
   Common mistakes	Anti-Patterns

   Only the three newest migration guides are listed here. For earlier versions, consult the migration archive.

5. Apply the cross-cutting guardrails below before style or performance cleanup.

6. Before a generated script's first real run, use nu-check --debug in the intended scope. Assert exact output for tricky strings/serialized payloads, then run the relevant tests; parsing alone cannot detect unintended but valid interpolation. Use --ide-check when structured diagnostics are needed.

7. Report security/correctness findings before style and performance notes.

If a referenced file is unavailable, say so and continue with this file rather than inventing its contents.

Cross-Cutting Guardrails

These rules apply across task types. Load the routed reference before relying on syntax or behavior that changed between Nushell releases.

Types and pipeline contracts

• Declare pipeline input in the I/O signature rather than as a positional parameter. Capture $in once when it must be reused because streams can be single-pass.
• Type exported command parameters and input/output signatures.
• Treat external/config records as untrusted; use optional access such as $record.field? and validate the resulting type/value.
• Remember that default evaluates its fallback argument eagerly. Make the fallback null-safe or branch explicitly when it can fail or is expensive.
• Keep return types consistent. Avoid any unless the function is genuinely polymorphic.
• Use external_arg on main parameters only when the CLI must preserve the caller's token spelling instead of applying Nushell literal coercion. It produces glob/list<glob> values in Nu 0.115, so validate or convert them before treating them as structured application data.
• Prefer match for several branches on one value; use if for one-off boolean predicates.
• On Nu 0.116, record spreads pass named flags to internal/custom commands. Null omits flags unless their type accepts nothing; switch false/null also omits the flag. Validate forwarded keys and preserve default/null intent.

External commands and errors

• Pass external arguments as separate values, never as an interpolated command string. Use complete and check exit_code when status matters.
• Prefer try/catch and $err.details for structured in-process errors.
• On Nu 0.115, a spanned error make label has the shape {text: ..., span: {start: ..., end: ...}}. Do not use the obsolete flat {text, start, end} form.
• On Nu 0.115.0, a try/finally nested directly inside an outer try whose handler is catch silently skips the inner finally. Give the outer block a finally, or put the inner one behind a do/command boundary, and assert that the owned state is gone. This defect is fixed in 0.116; keep the workaround only when the supported versions need it.
• Treat rendered nested-Nu diagnostics as presentation text, not a stable protocol. For CLI tests, normalize ANSI styling, gutters, and PTY wrapping before matching a long, domain-specific phrase.

Security stop checkpoint

Before approving code that executes commands, deletes files, reads credentials, or accepts paths/patterns, confirm these boundaries:

• Never pass untrusted strings to nu -c, source, run, ^sh -c, ^bash -c, or ^cmd.exe /C.
• Pass external command arguments as separate values, not an interpolated shell command string.
• For existing paths restricted to a base directory, expand both paths and use path relative-to to prove containment. A string starts-with check is unsafe because /safe/base2 starts with /safe/base.
• Prefer Nu's built-in mktemp/mktemp --directory; it is portable and returns a path directly. Do not use predictable temp names.
• Scope secrets with with-env; do not log them or pass them in argv when a stdin/config-file mechanism exists.
• Guard destructive paths against root, $nu.home-dir, unexpected types, and untrusted globs. Consider TOCTOU and partial-success behavior.

For output paths that do not exist yet, validate the existing parent directory with the same containment rule, then join only a validated leaf name.

Strings and formatting

• Prefer simple literals and raw regex strings; use double quotes only when actual escapes are required.
• Remember that $'...' interpolates but does not process escape sequences.
• A literal ( directly inside interpolation requires $"...". Since $'...' does no escape processing, every ( inside it opens an interpolation expression and \( cannot prevent that. Write $"\(abc)($var)", never $'\(abc)($var)'. The failure is often silent rather than an error: $'(1 + 1) items' evaluates to 2 items. See String Formats. For generated code with many delimiters, join plain literal fragments and serialized data instead.
• Keep code and data separate across interpreters: pass argv/stdin/JSON to a fixed program. If an API only accepts source, use a fixed, trusted template with target-language serialization; a single argv string is still code when the receiving tool evaluates it. Do not concatenate raw values into source.
• Use kebab-case for commands/flags, snake_case for variables/parameters, and SCREAMING_SNAKE_CASE for environment variables.

Data formats and grouping

• from xml loses XML namespace identity (reproduced on Nu 0.115.1): it stores only local names for elements and attributes, with no namespace URI or declaration in the result. For example, <x xmlns:r="urn:r" id="plain" r:id="rel"/> becomes an attributes record containing only id: rel; reversing the two attributes leaves id: plain. Later attributes overwrite earlier ones when their local names collide. Do not use this record for namespace-sensitive XML lookups, edits, or round-trips. Use a namespace-aware XML parser or, when the query xml plugin is available, XPath with --namespaces (and --output-names when names matter). Match namespace URI plus local name, not prefix spelling.
• Nu 0.115 makes from yaml default to YAML 1.2 with strict non-string keys and tags. Pin --spec, --multiple, tag handling, and key resolution when the input contract is controlled by another system instead of inheriting defaults accidentally.
• Let to yaml reject non-round-trippable values by default. Opt out only when that data loss is part of the documented contract, and quote the value: --non-roundtrip 'null' works across 0.115/0.116; bare null works from 0.116. --non-roundtrip 'lossy' alone is still rejected on 0.116.0, so use --serialize when a lossy encoding is genuinely wanted. On 0.116 replace --compact-list-indent with --list-indent compact|indented.
• group-by record output cannot represent a null key and omits that group in Nu 0.115. Use group-by --to-table when null groups must be retained or distinguished from empty strings.

Data flow and performance

• Prefer pipelines and immutable let bindings.
• Use where, select, update, insert, items, transpose, reduce, and enumerate instead of manual parsing and mutable accumulation.
• Do not capture mut variables in closures.
• for is appropriate for sequential side effects but is not a transforming expression; use each when a list result is required.
• Use par-each only when concurrency is safe and beneficial; preserve each when order or sequential side effects matter.
• Add lines before parse when line-by-line stream parsing is intended.
• On Nu 0.116, streamed errors reach length, columns, is-empty, and each while; consume streams in tests and fix upstream failures. Put flags before predicates in row-condition commands such as take until.
• Prefer direct row conditions for simple any/all predicates on Nu 0.115; retain closures when the predicate needs setup, destructuring, or reuse.
• Use native tables for small interactive data and Polars for large columnar group-by/join/aggregation workloads.

Modules and scripts

• Export only the intended API; keep helpers private.
• Use export def main when the command should match the module name.
• Use def --env/export def --env for caller-visible environment changes.
• Do not name commands, aliases, modules, or exports after parser keywords. Also treat $ans as reserved in Nu 0.115; rename older ans bindings.
• source, use, and run targets must be trusted and available at parse time.
• Test at the correct seam: direct functions for stable structured errors, CLI subprocesses for argument parsing/process boundaries, and both when needed.
• For Nu 0.116 completers, use named token/place/buffer inputs and place.command for shell words. Test via commandline complete; load the 0.116 migration guide for output/fallback rules and @interactive ownership.

Review Order

When reviewing code, report findings in this order:

1. Security: injection, traversal, credentials, destructive operations, temp files, environment poisoning.
2. Correctness: types, null handling, parse-time constraints, exit codes, cleanup/rollback, data-format contracts, platform/version behavior, stable tests.
3. Maintainability: naming, module boundaries, duplication, documentation.
4. Performance: streaming, unnecessary collection, safe parallelism, Polars.

Skip issues already enforced by the project's formatter/linter unless the tool output shows they are currently failing.

Validation

For ordinary script/module validation, prefer nu-check --debug. It parses without running the target's top-level commands and reports parse-time type errors as well as syntax errors. Use the matching parse mode:

nu --no-config-file -c 'nu-check --debug path/to/script.nu'
nu --no-config-file -c 'nu-check --debug --as-module path/to/module.nu'
nu --no-config-file path/to/test-script.nu

• Without --debug, invalid content returns false and the surrounding nu process can still exit 0; consume the boolean explicitly. With --debug, a parse error throws, so an uncaught failure exits nonzero. Missing files are errors. Neither mode proves runtime correctness or provides a sandbox.
• Pass the file path directly when checking a saved script with relative imports. Piping its text to nu-check checks an anonymous source; the optional path does not restore the file's import context when pipeline input is present.
• For editor/CI integrations needing spans and diagnostic records, use nu --no-config-file --ide-check 100 path/to/script.nu.
• --ide-check emits JSON Lines on stdout and may still exit with code 0 when a record has type: "diagnostic" and severity: "Error". It also exits 0 with empty output for missing files on older versions; 0.116 fixes that case with a nonzero exit. Verify the path for cross-version checks. Parse every non-empty line with from json; do not use the process exit code alone. Treat severity: "Error" diagnostics as blockers, surface other severities such as Warning without blocking, and ignore type: "hint" records. Handle a non-zero exit code or stderr separately as a CLI startup or I/O failure.
• Prefer --no-config-file for reproducible standalone checks. Omit it when the script intentionally depends on commands or environment from user configuration, and document that dependency.
• source executes top-level code; it is not a syntax check. Use it only when intentionally loading audited definitions/setup. Validate side-effecting scripts with nu-check/--ide-check, then execute in an owned temp fixture. When saving structured values, use a recognized data extension or serialize explicitly (to json, to yaml, to nuon, and so on) before save.
• Do not use nu --testbin in Nu 0.115+. Recreate the required behavior with Nushell itself or a purpose-built test fixture.
• Reproduce terminal-sensitive tests under a narrow PTY when diagnostics or tables are involved, for example stty cols 24 && nu tests/example.nu.
• Check diffs for debug markers and accidental changes before finishing.
• If validation fails, fix the smallest reproducible issue and rerun the exact failing command before broadening the test suite.