Verified by running both the old and new regex over the full addsInlineTest predicate (filename guard, f.patch guard, ^\+ anchor, .some(), and the alternation), 20+ inputs each way. The direction is right and the dominant #[tokio::test] form now resolves with zero regressions. Two things to fold in before merge.

Findings

• [P2] Cover the parametrized #[tokio::test(flavor = ...)] form
  .github/workflows/pr-triage.yml:78
  The ::test] alternative needs ] immediately after test, but the args form is ::test(...), so test] never appears and it's missed. This is the PR's own target class and it's live in the tree at crates/gitlawb-node/src/git/smart_http.rs:753 and :818. A PR whose only new tests use the flavor form still gets falsely labeled needs-tests. The bare form (248 hits) is fixed, so this is a partial fix rather than a wrong one.

• [P3] Anchor the match to an attribute so ::test] stops over-matching non-test code
  .github/workflows/pr-triage.yml:78
::test] is unanchored, so it also fires on ordinary added lines like let r = registry[Region::test]; and on the macro appearing in a comment or string literal. Effect is bounded (it only suppresses the advisory label, which isn't a merge gate), but it's a new false-positive class the old regex didn't have.

Both close with a single regex. I ran this against the same matrix: 0 regressions, 0 false-positives, and it catches the flavor form:

/^\+\s*#\[\s*(cfg\(\s*test|[\w-]+(::[\w-]+)*::test|test)\b/m

Anchoring to a leading #[ removes the over-match, and ::test\b (instead of ::test]) also clears #[tokio::test(...)].

Net: strict improvement as it stands, and the above makes it complete against the title. Non-gating label, so nothing here is severe, but the flavor form defeats the stated purpose on code that already exists in the repo.

I found issues that need to be addressed before this is ready.

Findings

• [P3] Keep the detector on an added diff line
  .github/workflows/pr-triage.yml:80
\s* after the diff + also consumes newlines. As a result, a Rust PR can add only a blank line immediately before an existing context-line #[tokio::test]; the expression then treats that unchanged test as newly added and suppresses needs-tests. Use horizontal whitespace (for example, [ \t]*) at this position so the attribute itself must be on an added line.

• [P3] Complete the cfg(test) predicate before accepting it as a test
  .github/workflows/pr-triage.yml:80
  The widened cfg\(\s*test\b branch also matches valid non-test predicates such as #[cfg(test = "triage_only")]. Adding only that attribute to an otherwise testless Rust PR makes addsInlineTest true and bypasses the advisory test signal; the former expression did not do so. Require the predicate to close (for example, cfg\(\s*test\s*\)) before classifying it as a test gate.

• [P3] Handle valid whitespace in the namespaced attribute syntax
  .github/workflows/pr-triage.yml:80
  Rust accepts token whitespace in attributes, including #[tokio :: test] and #[cfg (test)], but the new matcher requires both :: and cfg( to be adjacent. A PR whose only added test uses either valid form is still labeled needs-tests, so the claimed namespaced/cfg recognition remains incomplete. Allow horizontal token whitespace around these separators while keeping the match confined to the added line.

• [P3] Do not treat raw-string contents as executable test attributes
  .github/workflows/pr-triage.yml:80
  Anchoring at #[ prevents ordinary comment and string-expression lines from matching, but it does not distinguish a multiline raw string: an added fixture containing a line #[tokio::test] makes this expression true despite adding no runnable test. This is a new false positive for namespaced attributes and lets a testless Rust change suppress needs-tests; account for Rust string state (or otherwise avoid presenting the leading-attribute check as excluding strings).

Actionable comments posted: 1

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In @.github/workflows/pr-triage.yml:
- Around line 83-87: Update the patch-scanning logic around the lexical state
object and f.patch iteration so state is not carried across incomplete hunks.
Reconstruct lexical context from the complete head-version file using added-line
coordinates, or reset and initialize state from each hunk’s surrounding content
before evaluating additions, ensuring #[test] text is classified correctly when
a patch begins inside a comment or multiline literal.

I found issues that need to be addressed before this is ready.

Findings

• [P2] Preserve nesting while skipping block comments
  .github/workflows/pr-triage.yml:89
  Rust block comments nest, but this lexer represents comment state as a boolean and clears it at the first */. A source-only PR can add #[tokio::test] after an inner comment closes but before the enclosing comment closes (/* outer, /* inner */, #[tokio::test], */); Rust still treats the attribute as comment text, while this code treats it as live and suppresses needs-tests. Track block-comment depth (and cover nested comments) before accepting an attribute.

• [P2] Bound the per-file Contents API scan and do not silently classify failures as no test
  .github/workflows/pr-triage.yml:170
  Every changed Rust file with a patch now causes a serial repos.getContent call, while every failure is silently skipped. A large PR can therefore exhaust the workflow's ten-minute/API budget, and rate-limit or transient failures on the only file containing a new test make the triage job incorrectly add needs-tests (or prevent label reconciliation altogether). The previous implementation only examined the already-fetched file patches. Put a bounded strategy/fallback around this scan and surface a failure rather than treating an unavailable file as testless.

• [P3] Require test to be the terminal attribute segment
  .github/workflows/pr-triage.yml:80
  The test\b alternatives stop at a word boundary rather than a valid attribute terminator, so non-test attributes such as #[foo::test::helper], #[test::helper], and #[foo::test = "x"] match. Adding one of those to a Rust-only PR suppresses the advisory test signal despite adding no test. Require optional horizontal space followed by ] or ( after the bare/namespaced test segment.

• [P3] Recognize valid multiline test attributes
  .github/workflows/pr-triage.yml:80
  Rust permits whitespace, including newlines, between attribute tokens. A valid added test formatted as #\n[test] or #[tokio\n::test] is scanned one line at a time and neither line matches this horizontal-whitespace-only expression, so the PR still receives needs-tests. Carry enough attribute-token state across added lines (or parse the added attribute span) to cover the valid forms the workflow is intended to recognize.

I found issues that need to be addressed before this is ready.

Findings

• [P2] Do not abort triage when the bounded scan is exhausted
  .github/workflows/pr-triage.yml:199
  The cap is reached using the broad test/cfg substring prefilter rather than a real attribute match. A normal Rust PR that edits 16 files with (for example) #[cfg(feature = ...)] or even comments containing those words reaches the throw on the sixteenth file. That exits before the label reconciliation below, so every managed label (including stale labels that should be removed) and the guidance comment are left untouched. Keep the API work bounded, but use a non-fatal conservative fallback so an otherwise valid multi-file PR is still triaged.

• [P2] Handle Contents API responses that do not carry Base64 text
  .github/workflows/pr-triage.yml:206
getContent is decoded unconditionally, but GitHub returns an empty content with encoding: "none" for the 1–100 MB range unless the request uses the raw representation. An added #[tokio::test] in such a Rust file therefore gets lexed as an empty file and is incorrectly labeled needs-tests (and an absent content response would instead fail the job). Check the response encoding/size and fetch a supported raw/blob representation or take an explicit conservative fallback.

• [P3] Require an added attribute token, not merely added whitespace inside one
  .github/workflows/pr-triage.yml:226
  The expression permits newlines between # and [, then marks a match as new when any character in its span is on an added line. Thus a source-only PR can insert just a blank line into an existing multiline attribute (#, added blank line, [test]) and make addsInlineTest true without adding a test. Rust accepts that formatting, so this suppresses needs-tests for the rest of an otherwise testless change. Tie the match to an added attribute token/span rather than any whitespace it crosses.

• [P3] Preserve token boundaries while stripping literals
  .github/workflows/pr-triage.yml:136
  Removing a literal without a placeholder can join unrelated Rust tokens into a synthetic test attribute. For example, consume!(# "not an attribute" [test]); is valid input to a macro_rules! consume { ($($tt:tt)*) => {}; } matcher, but the stripper produces #[test]; the regex then treats the added invocation as a test. Preserve a separator for skipped comments/literals (or match lexed attribute tokens) so testless macro changes cannot bypass the advisory label.

• [P3] Recognize absolute namespaced test attributes
  .github/workflows/pr-triage.yml:80
  Rust accepts an attribute path beginning with ::, but the namespaced branch requires an identifier before its first ::. Consequently #[::tokio::test] is not detected even though it is a valid async-test attribute, and a PR that adds only that form still receives needs-tests. Allow the optional leading path separator when matching a namespaced test attribute.