Developing Linters

Developing Linters in r-lib/lintr

When implementing, extending, or refactoring linters in lintr, adhere to the following architectural, design, and production-readiness guidelines established by project maintainers:

1. Minimalist API Surface & Defaults ("On by Default")

• Avoid unnecessary configuration knobs: When extending a linter with a new check (e.g., adding namespace import checking to namespace_linter), do not reflexively add new boolean parameters (like check_imports = TRUE) to the public function signature unless there is a clear, compelling user need to toggle only that specific sub-check independently.
• Bleeding-edge upstream R features (TODO(R>=x.y.z)): When recommending new syntax or parameters introduced in recent or upcoming upstream R releases (fixed = TRUE added to list.files() in R 4.6.0), package authors may not wish to immediately incur that bleeding-edge dependency. In such cases, providing a toggle (check_file_listing = TRUE) allows authors supporting older R versions to opt out (FALSE). Always mark such transitionary options or workarounds with an explicit TODO(R>=x.y.z) comment right above the function signature (# TODO(R>=4.6.0): Deprecate check_file_listing once R 4.6.0 is the minimum supported version.) indicating exactly when the option should be deprecated and enforced universally.
• Natural zero-lint fallback: If a check naturally produces zero lints when not applicable (e.g., inspecting namespace_imports() on a file outside an R package or in a directory without a NAMESPACE file returns empty_namespace_data()), run the check unconditionally as part of the standard linter flow.
• Keep signatures clean: Preserving existing function signatures (namespace_linter(check_exports = TRUE, check_nonexports = TRUE)) reduces API complexity and documentation churn.

2. Helper Functions, Readability, & Safe Fallbacks

• Encapsulating non-trivial logic in helper functions is encouraged: While trivial single-use wrappers purely around basic line lookups or single vapply(...) steps should generally be avoided if they merely fragment linear control flow, helper functions that encapsulate non-trivial logic (such as get_seq_lint_message() or indent_lint_metadata()) are highly valued, even if invoked only once inside a linter callback. The immense readability improvement and clear domain encapsulation far outweigh their single-use nature.
• Minimalist return types for linter helpers: Helper functions that format linter messages (e.g. get_seq_lint_message(seq_expr)) should return a crisp, simple character vector of messages (character()), rather than leaking internal multi-column data structures back to the main linter callback. Keeping helper return types minimal ensures the linter body stays clean (lint_message <- get_seq_lint_message(seq_expr); xml_nodes_to_lints(seq_expr, source_expression, lint_message, type = "warning")) and keeps PR diffs compact.
• Encapsulate empty-state checks inside helpers: Place domain 0-length guards directly at the top of helper functions (if (length(seq_expr) == 0L) return(character())) rather than polluting the main linter body with if (length(...) == 0L) wrappers.
• Namespace-aware string formatting in helpers: When stripping parentheses or formatting extracted function calls (e.g. converting length(x) -> length(...) but preserving 0-argument calls like n()), always ensure exclusions are namespace-aware (!grepl("(^|::)n\\(\\)", funcalls) or !funcalls %in% c("n()", "dplyr::n()")). Checking funcalls != "n()" naively fails when passed namespace-prefixed calls like dplyr::n().
• Vectorized argument counting via xml_find_num_(): Delegate node and argument counting directly to libxml2 using as.integer(xml_find_num_(nodes, "count(...)")). Never use as.integer(xml_find_chr_(nodes, "string(count(...))")).
• Early returns on domain empty states: Check for domain empty states early and exit (if (nrow(ns_imports) == 0L) return(lints) or if (nrow(lint_line_df) == 0L) return(list())) rather than executing checks over empty data structures.
• Avoid distracting checks for 0-line files: Do not recommend or implement defensive guards for truly empty zero-line files (if (length(source_expression$file_lines) == 0L)). Encountering a 0-line file in practical active extraction and parsing is virtually impossible; suggesting such checks during code reviews creates unnecessary distraction and code clutter.
• Rely on existing safe fallbacks: Do not write defensive wrappers around functions that already handle NULL cleanly. For example, namespace_imports(NULL) safely returns empty_namespace_data(), so if (!is.null(pkg_path)) namespace_imports(pkg_path) is redundant.
• Avoid redundant trimws() on XML node text: Text extracted via xml_find_chr_(..., "string(...)") or xml_text() is already clean and trimmed by xml2. Avoid calling trimws() redundantly on extracted XML string vectors.
• Use numbered format specifiers (%1$s, %2$s) for repeated variables in sprintf(): When formatting dynamic lint messages where placeholders are repeated (e.g., displaying both the recommended replacement and the observed violation using the exact same function name and parameter string), use numbered format specifiers (%1$s, %2$s, %3$s). This cleanly avoids redundant argument repetitions across the sprintf() call (sprintf("expect_shape(x, %1$s = %2$s) is better than %3$s(%1$s(x), %2$s)", shape_function, shape_arg_var, matched_function) over repeating shape_function, shape_arg_var twice).
• Lint message design & discernment: The ultimate purpose of lint_message is to provide a helpful, actionable cue when a user receives a lint result in their IDE or CI. When considering whether to create separate message variants for related syntactic cases (e.g. sprintf("hello") vs sprintf(foo())), exercise discernment with user-friendliness as the north star. Avoid introducing distinct message classes if the information conveyed to the user does not meaningfully differ, as extra message complexity adds code to maintain without delivering practical user value. However, use distinct messages whenever doing so provides significantly clearer guidance to the user.

3. Minimal Diffs & Logical Execution Order

• Structure execution to avoid intermediate mutations: Structure the execution order of sub-checks within a linter to avoid mutating, subsetting, or filtering shared XML node lists and symbol vectors midway through the function.
• Inline condition-dependent vector additions over intermediate state: When conditionally appending items to a list or character vector, inline the check right inside c() (c("strsplit", if (check_file_listing) c("dir", "list.files"))) instead of assigning an intermediate temporary variable. Because if (FALSE) evaluates to NULL (which automatically drops out of c()), inlining cleanly eliminates unnecessary temporary variable assignments.
• Append new checks cleanly: When adding a check to an existing linter (check_exports, check_nonexports), place the new check cleanly after existing checks so that existing code blocks and variables (packages, symbols, ns_nodes) remain untouched. This keeps diffs small, readable, and easy to review.

4. Self-Linting, Repository Health, & NEWS.md

• Verify zero new violations and dogfood new linters across lintr itself: Whenever a linter is made more strict or a brand-new linter is created, immediately run it across lintr's own repository (R/ codebase and tests/testthat/ suites). Dogfood the new checks and clean up any newly triggered violations across existing package code and tests before proposing the PR. For example, for expect_shape_linter(), some existing usages of expect_identical(nrow(x), n) were updated to expect_shape(x, nrow = n) in the same PR.
• Keep lintr 100% lint-free: Immediately clean up any newly triggered violations across the repository (e.g., changing redundant glue::glue() or cli::cli_warn() calls to glue() and cli_warn()) before proposing the PR.
• Descriptive VCS branch / bookmark naming: Avoid non-descriptive, issue-number-only branch names like fix-issue-NNNN. Always choose short, descriptive branch or bookmark names that clearly convey the functional work or feature being developed (e.g. any-duplicated-highlight-range, unreachable-switch, vector-logic-body).
• Informative, commit-specific descriptions in stack/chain VCS workflows (e.g. jj): When working in stack or chain VCS workflows where commits build sequentially on top of each other, never reuse a single generic feature description across all commits in the stack. Each commit description must concisely and accurately summarize the specific change introduced by that commit (e.g., Gemini: Unify single-argument sprintf/gettextf lints under single argument message vs Gemini: Optimize argument counting using xml_find_num_() count XPath).
• Use author's public GitHub handle in NEWS.md: Always use the author's public GitHub handle (e.g. @MichaelChirico, @AshesITR).
• NEWS.md entry scope: Internal range/span highlighting fixes for existing linters (which only adjust where a lint lands without altering public functions, parameters, or lint messages) do not require a NEWS.md entry. Reserve NEWS.md entries strictly for new features, new linters, deprecations, breaking changes, or user-facing parameter/message additions.
• Categorize NEWS.md entries under hierarchical section headings: Whenever implementing a new user-facing linter enhancement, parameter option, brand-new linter, or significant bug fix, always append a concise, well-formatted entry right under # lintr (in development) in NEWS.md. Ensure entries are nested cleanly under level 2 (##) and level 3 (###) section headings matching standard hierarchy (## New and improved features -> ### New linters or ### Linter improvements; or ## Bug fixes). Never insert uncategorized top-level bullet points straight beneath # lintr (in development).

5. Respecting Contract Boundaries & Avoiding Unexported Internals (:::)

• Strictly respect contract boundaries: Take seriously that ::: means private and avoid violating contract boundaries across packages. While lintr provides %:::% (p %:::% f) to bypass self-lint checks when calling private functions, reaching across package boundaries into unexported internals (knitr:::parse_params, knitr:::file_ext) should be avoided except in rare, justified cases.
• Look under the hood for exported alternatives: When tempted to invoke a private internal function from another package, look under the hood first—very often those internals are simple wrappers around base R functions or exported utilities from imported dependencies (such as xfun::csv_options(), xfun::divide_chunk(), or xfun::file_ext()).
• Pragmatic ~95% correctness over fragile 100% correctness: Even if private upstream helpers include extra edge-case handling or nuance, choose to ignore that nuance for lintr's purposes. lintr checks do not have to be demandingly and globally 100% correct in every obscure scenario; achieving ~95% practical correctness using clean, stable, exported APIs is far superior. If a user ever complains about a missing edge case in the future, the implementation can be refined at that time.
• Pragmatic namespace scoping & AST robustness: Exercise taste and judgment regarding defensive namespace handling. For standard base functions (such as is.numeric, is.integer, c, length), do not add complex, verbose AST predicates (e.g. expr[1][not(expr/SYMBOL_PACKAGE) or expr/SYMBOL_PACKAGE[text() = 'base']]) merely to guard against hypothetical third-party packages masking base functions (e.g. pkg::is.numeric). Such defensive guards add visual noise, XPath complexity, and maintenance overhead without delivering practical user value. Reserve explicit namespace guards for cases where package masking is a known, realistic scenario in the R ecosystem.

6. Simple & Idiomatic AST and Condition Checks

• Avoid over-engineered evaluation constructs: When checking parsed parameter values or AST expressions (such as eval options from chunk headers), do not write complex, over-defensive constructs like tryCatch(eval(..., envir = baseenv())) to handle theoretical runtime expressions.
• Check exact parser representations: Inspect and match the exact R objects produced by the parser (xfun::csv_options() produces logical FALSE for eval=FALSE and symbol quote(F) for eval=F). A direct, concise check such as if (identical(eval_value, quote(F))) return(TRUE) followed by isFALSE(eval_value) is simpler, safer, and much easier to maintain.
• Annotate symbol checks for self-linting: When comparing against quote(F) or quote(T), add # nolint next: T_and_F_symbol_linter. immediately above the line to prevent lintr's self-linting checks from flagging the symbol name.

7. Data Engineering, Centralized State, & Readability First

• Use cohesive data frames (line_metadata / internal working state): When computing many parallel properties (such as line-by-line formatting attributes or multi-argument expression parsing), structure intermediate state as a single data.frame. This enforces parallel alignment and makes code far cleaner to inspect, query, or update via parallel matrix subsetting (df[is_1arg, c("a", "b")] <- list("val", df$expr[is_1arg])).
• Internal data.frame vs minimalist return types: Use data.frame strictly as an internal working data structure within helper functions when reading and modifying multiple parallel vectors. Avoid returning heavy multi-column data frames to main callbacks when only a single output vector (like a character vector of lint messages) is needed.
• Omit redundant stringsAsFactors = FALSE: In modern R (>= 4.0.0), data.frame() defaults to stringsAsFactors = FALSE. Omit stringsAsFactors = FALSE in internal helper data frames.
• Descriptive domain variable names for message assembly: Use clear, explicit domain-specific variable names when building linter messages: preferred_usage (the recommended replacement), observed_usage (the code pattern flagged), is_decreasing, not_seq_along.
• Prioritize with() to eliminate repetitive visual noise: When writing compound logical filtering conditions over multi-column metadata frames (find_bad_lines()), embrace with(line_metadata, !is.na(line) & indent_level != expected_level & !in_str_const) over repetitive $ accesses (line_metadata$line, line_metadata$indent_level). The immense readability benefit of eliminating repetitive visual noise far outweighs negligible environment evaluation overhead.

8. Idiomatic R Vectorization over Loops

• Prefer vectorized operations: Avoid using for loops to iterate over lines or XML nodes if a vectorized alternative exists. Use Map(), vapply(), or logical subsetting.
• Vectorized sequence generation: For example, to generate a sequence of indices to mark as inside a string constant:

  line1 <- as.integer(xml_attr_(multiline_strings, "line1"))
  line2 <- as.integer(xml_attr_(multiline_strings, "line2"))
  is_in_str <- unlist(Map(`:`, line1, line2))
  in_str_const[is_in_str] <- TRUE
  

  This is much cleaner and faster than a for (string in multiline_strings) loop.

9. Handling Sequences, Ranges, Directionality, & Boundary Conditions

• Directionality & boundary rules for sequence linters: Linters inspecting sequence generation (such as seq_linter inspecting 1:n, seq(), nrow(x):1) must explicitly distinguish increasing, decreasing, and boundary structures:
  • Increasing sequences: 1:n, seq(1, n), or seq(from = 1, to = n) -> recommend seq_len(n) or seq_along(x).
  • Decreasing sequences: n:1, seq(n, 1), or seq(to = 1, from = n) -> recommend rev(seq_len(n)) or rev(seq_along(x)).
  • Single-element boundaries: seq(1, 1) or seq(from = 1, to = 1) -> recommend seq_len(1) (or seq_len(1L)), NEVER rev(seq_len(1)). Ensure is_decreasing checks exclude dot_expr1 %in% c("1", "1L") or are scoped exclusively to reverse colon calls (!is_seq & dot_expr2 %in% c("1", "1L")).
  • Non-positive lower bounds: seq(0, 1), seq(0, 10), seq(-1, 1), seq(-1, 10), 0:1, -1:10 are non-standard sequences; they must NOT be flagged as countdowns or converted to seq_len(). Guard in XPath against zero constants and unary minus: and not(expr[NUM_CONST[text() = '0' or text() = '0L'] or OP-MINUS]).

10. Robust Handling of Literate R Formats & Readable Transitions

• Expect NA_character_ in line content: When linting files, remember that literate programming formats (like .Rmd, .qmd) extract R code by masking non-R lines with NA_character_ to preserve line numbers.
• Avoid NA propagation in logical expressions: Ensure logical expressions checking conditions over lines immediately guard against NA (!is.na(line) & ...) to keep resulting booleans clean (FALSE & NA -> FALSE).
• Prioritize clear consecutive logic (diff()) over micro-optimizations: Because is_bad evaluates to FALSE (not NA) across non-code gaps, computing consecutive state differences using simple arithmetic (diff(is_bad) == 0L) cleanly isolates block transitions across gaps. Never replace clear, readable difference computations with obscure logical shifting constructs solely to avoid implicit conversions or save negligible compute. Readability is paramount.

11. Direct Documentation Workflow & roxygen2 Example Style

• Crisp parameter and linter descriptions: Keep @description and @param documentation concise and conceptual. If a linter checks a family of spiritually equivalent expressions (e.g. 1:length(...), seq(1, n), seq(10, 1), nrow(x):1), summarize the intent at a high level (This linter checks for expressions like \code{1:length(...)} (and many other spiritually equivalent variations) which are a common source of bugs...) rather than enumerating every syntax permutation in the @description.
• Roxygen example structure (@examples) for new or extended linters: When creating or extending a linter, adhere strictly to lintr's standard roxygen examples layout:
  1. Section Division (# will produce lints and # okay): Always partition @examples into distinct sections beginning with # will produce lints at top, directly followed by # okay.
  2. Explicit lint() calls: Format each demonstration cleanly across lines using lint(text = ..., linters = ...):

     #' lint(
     #'   text = "list.files(pattern = 'RDS')",
     #'   linters = fixed_regex_linter()
     #' )
     

  3. Multi-line or escaped code (writeLines): For simple single-line invocations without complex escapes, pass text = "..." directly inside lint(). For multi-line snippets (containing \n) or regex patterns containing heavy backslash escaping ("\\\\."), assign the target code to code_lines <- "...", display it first using writeLines(code_lines), and then pass text = code_lines to lint(). This allows users executing help examples to view the rendered target code exactly as seen by the parser:

     #' code_lines <- 'gsub("\\\\.", "", x)'
     #' writeLines(code_lines)
     #' lint(
     #'   text = code_lines,
     #'   linters = fixed_regex_linter()
     #' )
     

  4. Mirrored 1:1 problem-to-solution pairs: Maintain strict 1:1 structural correspondence between cases under # will produce lints and their corresponding resolutions right under # okay in the same order. Do NOT introduce orphan or duplicate examples under # okay (e.g. avoid repeating seq_along(x) multiple times).
  5. Standard footer tags: Always conclude linter roxygen blocks right after @examples with @evalRd rd_tags("<linter_name>"), @seealso [linters] for a complete list of linters available in lintr., and @export.