r-code

name	r-code
trigger	writing R functions / API design / error handling
description	Guide for writing R code. Use when writing new functions, designing APIs, or reviewing/modifying existing R code.

R code

This skill covers how to design and write R functions — including naming conventions, signatures, API conventions, input validation, error handling, and common pitfalls. For documenting functions, use the document skill. For tests, use the tdd-workflow skill.

Naming conventions

Functions

Functions use snake_case and should be verbs or verb phrases that describe what the function does:

fetch_records()
build_summary()
validate_input()

A function name should be descriptive enough to make its purpose clear without a comment. Prefer clarity over brevity — don't abbreviate unless there is a widely understood convention (e.g. df for data frame, dir for directory).

Internal helpers use a dot prefix:

.parse_response()
.validate_columns()

Parameters

Parameters use snake_case and should generally be nouns, occasionally adjectives. The same rule applies: clarity over brevity.

# Good
fetch_records(file_path, page_size, overwrite)

# Bad — unclear abbreviations
fetch_records(fp, ps, ow)

File organization

One exported function per file: R/{function_name}.R (e.g. fetch_records() → R/fetch_records.R). Internal helpers used exclusively by that function live in the same file. Shared helpers go in R/utils.R or R/utils-{topic}.R (e.g. R/utils-parsing.R).

Coding style

• Always run air format . after generating code.
• Use the base pipe operator (|>) not the magrittr pipe (%>%).
• Use \() ... for single-line anonymous functions. For all other cases, use function() {...}.

Function design

Functional core, imperative shell — pure, testable functions that accept data and return data form the core. The imperative shell orchestrates program flow, manages state, and calls the functional core.

Functions should be small and single-purpose. Each function should operate at a single level of abstraction: it either orchestrates calls to other functions, or performs a direct operation on data, but does not mix the two.

# Orchestrator — delegates to focused helpers
build_report <- function(data, output_path) {
  data <- .clean_data(data)
  summary <- .compute_summary(data)
  .write_report(summary, output_path)
}

# Worker — performs one direct operation
.clean_data <- function(data) {
  data |>
    dplyr::filter(!is.na(value)) |>
    dplyr::mutate(value = round(value, 2))
}

Name functions well enough that their purpose is obvious from the call site. When reading the orchestrator above, each step is self-documenting — no comments needed.

Simplify control flow — prefer guard clauses and returning early over complex if/else structures.

Pure conditionals — the expression inside a conditional check should not cause side effects. Extract the pure check from the impure action into separate functions if needed.

General API design patterns

Enum-like arguments — declare choices as the default vector; resolve with rlang::arg_match() at the top of the function:

summarize_data <- function(x, method = c("mean", "median")) {
  method <- rlang::arg_match(method)
  # method is now guaranteed to be "mean" or "median"
}

NULL as "not provided" — use NULL as the default for optional arguments where there is no sensible scalar fallback; check with is.null():

fetch_records <- function(x, output_column = NULL) {
  if (!is.null(output_column)) { ... }
}

S3 object construction — build as a named list, set class explicitly:

.new_summary <- function(values, method) {
  out <- list(values = values, method = method)
  class(out) <- c(paste0("summary_", method), "data_summary")
  out
}

call propagation in internal validators — helpers that validate arguments and may throw errors should accept and forward call:

.check_non_empty <- function(x, call = rlang::caller_env()) {
  if (length(x) == 0L) {
    .nectar_abort("Input {.arg x} cannot be empty.", "empty_input", call = call)
  }
}

process_data <- function(x, call = rlang::caller_env()) {
  .check_non_empty(x, call = call)
  ...
}

Return tibbles, not data frames:

summarize_data <- function(x) {
  result |> tibble::as_tibble()
}

Input validation

Use stbl::to_*() and stbl::stabilize_*() to validate parameters. These functions coerce when safe and fail with clear error messages when not.

• to_*() — simple type coercion. Use when you need to ensure a parameter is the right type but don't need additional constraints.
• stabilize_*() — coercion plus content validation (regex, ranges, etc.). Use when simple type coercion isn't enough.

Validate in the function that uses the parameter, not in a caller that passes it through. This preserves R's lazy evaluation — if a parameter is never used on a code path, it is never evaluated or validated.

# Good — validation happens where the parameter is used
build_report <- function(data, title, page_size) {
  data <- .clean_data(data)
  summary <- .compute_summary(data, page_size)
  .write_report(summary, title)
}

.compute_summary <- function(data, page_size, call = rlang::caller_env()) {
  page_size <- stbl::to_int_scalar(page_size, call = call)
  ...
}

.write_report <- function(summary, title, call = rlang::caller_env()) {
  title <- stbl::to_chr_scalar(title, call = call)
  ...
}

# Bad — validates everything eagerly, breaking lazy evaluation
build_report <- function(data, title, page_size) {
  title <- stbl::to_chr_scalar(title)
  page_size <- stbl::to_int_scalar(page_size)
  ...
}

When call is available (because the function accepts it), always pass it to stbl calls so error messages point to the user's call frame.

Internal vs. exported functions

Export a function when:

• Users will call it directly
• Other packages may want to extend it
• It is a stable, intentional part of the API

Keep a function internal when:

• It is an implementation detail that may change
• It is only used within the package
• Exporting it would clutter the user-facing API

Internal helpers use a dot prefix (e.g. .parse_response()).

Error handling

Use .nectar_abort() (defined in R/aaa-conditions.R) rather than calling cli::cli_abort() directly. This wraps stbl::pkg_abort() and ensures consistent error class formatting:

.nectar_abort(
  "Column {.field {name}} not found in {.arg data}.",
  "column_not_found",
  call = call
)

Always pass call = call (or call = rlang::caller_env()) so errors point to the user's call frame, not an internal helper.

Common package mistakes

# Never use library() inside package code
library(dplyr)          # Wrong
dplyr::filter(...)      # Right
# or `@importFrom dplyr filter` if used extensively

# Never modify global state without restoring it
options(my_option = TRUE)                    # Wrong
withr::local_options(list(my_option = TRUE)) # Right

# Use system.file() for package data, not hardcoded paths
read.csv("/home/user/data.csv")                         # Wrong
system.file("extdata", "data.csv", package = "mypkg")   # Right

Dependencies

Use existing imports first

Packages already in Imports in DESCRIPTION should be preferred over base R equivalents: purrr::map() over lapply(), rlang::is_*() predicates over is.*(), and withr::local_*() over manual on.exit() state management.

When to add a new dependency

Add a dependency when it provides significant functionality that would be complex or brittle to reimplement — date parsing, web requests, complex string manipulation. Stick with base R or existing imports when the solution is straightforward.

Adding a new dependency requires explicit discussion with the developer.