lint_cmd::run always returns Ok(()) and discards the linter process exit status. This breaks CI usage because rtk lint will exit 0 even when ESLint (or another linter) reports errors. Capture output.status and propagate a non-zero exit code when not using --from-file (while still allowing fixture mode to avoid exiting during benchmarks).

This workflow runs on pull_request but grants contents: write and pull-requests: write at the workflow level. That exposes write-scoped GITHUB_TOKEN to untrusted PR code during the benchmark step. Consider defaulting to read-only permissions for PRs and elevating permissions only for the master/main push job (e.g., via job-level permissions and/or splitting the PR and push workflows).

update-readme-metrics.sh currently only checks for the README markers and never updates README.md from benchmark-report.md, but the workflow later opens a PR claiming metrics were updated. Either implement the marker replacement (extract table from the report and rewrite the README section) or adjust the workflow/script messaging so it matches the actual behavior.

get_db_path() ignores the RTK_TEST_DB env var that the test helper sets, so any code path that calls Tracker::new() (e.g., track()) will still write to the real user DB during tests. Consider honoring an env override in get_db_path() (and/or disabling tracking in tests) to keep tests hermetic.

truncate() slices strings by byte index (&s[..max_len - 3]), which can panic for non-ASCII/UTF-8 text when max_len - 3 lands in the middle of a multi-byte character. Consider truncating by char_indices() (or using unicode-segmentation) to avoid runtime panics on Unicode output.

Doc comments in utils.rs are in French while most other Rust module docs/comments in this repo are English (e.g., tracking.rs, config.rs). For consistency and maintainability, consider translating these docs to English or aligning with the project’s prevailing documentation language.

This test no longer asserts any of the expected transformed content (it only checks ANSI stripping). That makes it much less likely to catch regressions in filter_vitest_output. Consider re-adding at least one stable assertion about the filtered summary output in addition to the ANSI check.

In Bash, echo "$ cat $fixture_file" expands $cat as a variable (usually empty), so the generated markdown command line will be wrong. Escape the literal $ (e.g., \$) or use single quotes for these prompt lines.

Same issue as above: echo "$ $rtk_cmd --from-file $fixture_file" treats $ as variable interpolation, producing an incorrect command line in the markdown output. Escape the $ (e.g., \$) or use single quotes.