Maybe this is a good reason to finally add a dedicated lint_text() (and start deprecating lint(text=...)), WDYT?

That is a good idea! Do you want me to include it in this PR?

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.29%. Comparing base (c3d1dc5) to head (dfe37ad).

@@            Coverage Diff             @@
##             main    #3017      +/-   ##
==========================================
+ Coverage   99.25%   99.29%   +0.04%     
==========================================
  Files         128      128              
  Lines        7355     7367      +12     
==========================================
+ Hits         7300     7315      +15     
+ Misses         55       52       -3     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

@MichaelChirico
For now, I just modified lint().
We can add lint_text() in a different PR.

@atusy for book-keeping, was an LLM used in whole or in part here?

I deployed Gemini to run a joint review of this and #3034 which I understand to be attempting to solve the same issue working with {lintr} in vscode.

@atusy what do you make of this caution about filename=""? I think it's not particularly relevant and can be ignored for now.

Executive Summary & Recommendation

We recommend accepting the approach of PR 3017 with improvements, and rejecting PR 3034.

• PR 3034 is incomplete. It only prevents the crash in find_package(), but fails later in exclude() when lints are found, leading to a different crash. The regression test added in PR 3034 actually fails.
• PR 3017 is a complete and robust solution that introduces a clean separation between file content and file identity. It allows full feature support (config discovery, exclusions, knitr detection) for unsaved buffers.
• Proposed Improvement to PR 3017: PR 3017 resolves empty filename "" to the directory getwd(), which causes lints to be attributed to the directory itself in the output. We propose mapping "" to "untitled" internally in lint(). This resolves to a file path (e.g., /path/to/project/untitled) which is cleaner and safer for IDE mapping.

Detailed Analysis of PR 3034

Approach

PR 3034 adds a guard at the beginning of find_package() to return NULL if the path is empty:

find_package <- function(path, allow_rproj = FALSE, max_depth = 2L) {
+  if (!nzchar(path)) {
+    return(NULL)
+  }
   path <- normalize_path(path, mustWork = !allow_rproj)

Assessment

• Correctness (Low): It is incomplete. If the linting process actually finds any lints, lint() calls exclude() at the end. In exclude(), the empty filename "" is passed to normalize_exclusions(), which treats it as an unnamed exclusion and fails with:

  Error in `normalize_exclusions()`:
  ℹ Full file exclusions must be <character> vectors of length 1.
  ✖ Items at the following indices are not: 1.
  

• Robustness (Low): The regression test added in PR 3034:

  test_that("object_length_linter does not crash with empty filename", {
    expect_length(
      lint("", text = "x = 1", linters = assignment_linter(), parse_settings = FALSE),
      1L
    )
  })

  fails on this branch due to the normalize_exclusions error.
• Maintenance Burden (Low): Very small change, but doesn't solve the problem.

Detailed Analysis of PR 3017

Approach

PR 3017 refactors lint() to support "Mode 3": lint(filename, text = content) where filename is the identity of the file (which may not exist on disk) and text is the content to lint.
Key changes:

1. find_package() tolerance: Changed to use mustWork = FALSE in normalize_path so it returns NULL instead of erroring for non-existent paths.
2. normalize_identity_path(): Introduced in lint() to resolve relative paths of non-existent files by prepending getwd() before normalizing.
3. Exclusions Refactoring: Splits normalization of source exclusions (from # nolint comments in the text) and config exclusions (from .lintr). Source exclusions are normalized with normalize_path = FALSE to prevent them from being discarded because the identity file doesn't exist on disk.
4. Caching Fix: Fixes a bug where inline data caching used inconsistent keys between saving and retrieving.

Assessment

• Correctness (High): Fully addresses the crash and ensures all features (caching, exclusions, config discovery) work correctly for non-existent files.
• Suitability (High): This is the correct design for IDE integration, as it allows linting unsaved buffers while respecting the package context they belong to.
• Robustness (High): Handles relative/absolute paths and non-existent files correctly. All tests pass, including comprehensive new tests for the new mode.
• Maintenance Burden (Medium): The changes in exclude.R and lint.R increase complexity, but it is necessary for correctness.
• Efficiency (Good): Avoids creating temporary files when text is provided with a filename.

Edge Case: Empty Filename ""

In PR 3017, if filename is "", it is resolved to getwd() (absolute path of current directory).

• Result: Lints are attributed to the directory itself in the output:

  /path/to/project:1:3: style: [assignment_linter] Use <- for assignment, not =.
  x = 1
    ^
  

• Impact: This is confusing for command-line users and might break IDE mapping (IDE might not be able to map diagnostics to the unsaved buffer if the filename is a directory).

Proposed Path Forward

We should adopt PR 3017 but apply the following modifications:

1. Map "" to "untitled" in lint():
   At the beginning of lint(), map empty filename to a dummy filename like "untitled":

   if (!missing(filename) && !nzchar(filename)) {
     filename <- "untitled"
   }

   This ensures that:

   • The filename is resolved to /path/to/project/untitled (a file, not a directory).
   • The output shows untitled (or absolute path to it), which is cleaner.
   • All other PR 3017 logic (config discovery starting from getwd(), exclusions, etc.) still works.
   • We tested this modification, and it successfully resolves the directory output issue while passing all tests.

2. Keep the find_package guard (Optional but recommended):
   Add the !nzchar guard from PR 3034 to find_package as defensive programming, in case it is called with "" from other entry points.

Code Diff for Proposed Improvement (relative to PR 3017)

diff --git a/R/lint.R b/R/lint.R
index b4ef2251..df719036 100644
--- a/R/lint.R
+++ b/R/lint.R
@@ -47,6 +47,9 @@ lint <- function(filename, linters = NULL, ..., cache = FALSE,
   dot_names <- if (getRversion() %in% c("4.1.1", "4.1.2")) names(list(...)) else ...names()
   check_dots(dot_names, c("exclude", "parse_exclusions"))
 
+  if (!missing(filename) && !nzchar(filename)) {
+    filename <- "untitled"
+  }
   needs_tempfile <- missing(filename) || re_matches(filename, rex(newline))
   inline_data <- !is.null(text) || needs_tempfile
 
diff --git b/R/settings_utils.R b/R/settings_utils.R
index d9327e79..262071f1 100644
--- a/R/settings_utils.R
+++ b/R/settings_utils.R
@@ -8,6 +8,9 @@ has_rproj <- function(path) {
 }
 
 find_package <- function(path, allow_rproj = FALSE, max_depth = 2L) {
+  if (!nzchar(path)) {
+    return(NULL)
+  }
   path <- normalize_path(path, mustWork = FALSE)
   if (allow_rproj) {
     found <- function(path) has_description(path) || has_rproj(path)