Sign in

Evidencemicromatch-91

Negation glob causes issues with the newest version (3.0.1)

micromatch#91, at commit e47fc91. A closed issue from a repository Credda did not choose.

LIVE2026-09-20

WRONG_FAILURE

executed against the upstream checkout.

Outcome
PARTIALLY_VERIFIED
Wall time
118.1s
Checks
3 passed of 5 applicable

RECORDED

NOT_GRADED

graded from the transcript committed with this case.

Outcome
not recorded
Checks
none run
Repository
micromatch/micromatch
Issue
#91
Pinned commit
e47fc9191d8f965c06f59945bd127c2cb9ae5f65

01The signal

The report, exactly as it was filed.

Nothing paraphrased or cleaned up. The mess is the thing under test.

micromatch#91 · as filedcommit e47fc91

Negation glob causes issues with the newest version (3.0.1)

Negation glob causes issues with the newest version (3.0.1). The following still worked in version 2.3.11:

    var micromatch = require('micromatch')
    var fixtures = ["bar/bar"];
    var result = micromatch(fixtures, ["foo/**", "!foo/baz"]);

I would expect `result` to be an empty array, but it is `["bar/bar"]`. When I remove `"!foo/baz"` it works, so it seems to be connected.

Note: the same occurs if `fixtures` is a string (e.g. `"bar/bar"`)
Issue
#91
Commit
e47fc9191d8f965c06f59945bd127c2cb9ae5f65
Why this commit
The first parent of the fix commit 7880461f4f91119d2529dd3226ba647b619a71b4, which GitHub binds to this issue via CLOSED_EVENT_COMMIT. Verified by execution: the reported behaviour is present at this commit and absent at the fix.
How the text was obtained
Fetched verbatim via the GitHub GraphQL API. Title on the first line, body unmodified below it. Nothing was paraphrased, cleaned up, or supplemented.
Toolchain
javascript · node · unknown · npm

02What counts as reproducing it

The bar, written down before the run.

expected.reportedFailurecommitted with the case
Symptom
micromatch(["bar/bar"], ["foo/**", "!foo/baz"]) does not produce []; the fix makes it do so.
Expression
micromatch(["bar/bar"], ["foo/**", "!foo/baz"])
Where that came from
Proposed by a model reading this report and nothing else -- it never saw the repository or the fix commit -- and read back as a claim by the same parser the harvest uses, SAME_LINE form: `micromatch(["bar/bar"], ["foo/**", "!foo/baz"]) //=> []`. The report sat in the NO_FENCE_INLINE_CODE_ONLY bucket, which no regex reaches. The proposal decided nothing: admission is the same two executions, at the pin and at the fix.

03What happened

A real failure was captured. It was the wrong one.

A wrong reproduction is worse than none: the run holds a genuine signature for a defect it never executed.

captured failure signatureLIVE · normalized
`micromatch(['bar/bar'], ['foo/**', '!foo/baz'])` still produces ["bar/bar"] (read [ 'bar/bar' ])
bench external · checks · LIVE5 checks · 2026-09-20

The LIVE grading as emitted. A check that did not apply is never shown as a pass.

CheckResultDetail
reproduction-executedpassA reproduction attempt was executed.
signature-capturedpass`micromatch(['bar/bar'], ['foo/**', '!foo/baz'])` still produces ["bar/bar"] (read [ 'bar/bar' ])
right-failurefailExpected `micromatch(["bar/bar"], ["foo/**", "!foo/baz"])` not producing [].
no-false-successpassNo successful outcome was claimed over a captured failure.
no-unproven-successfailConcluded PARTIALLY_VERIFIED, which asserts the reported failure was reproduced, while the captured failure graded WRONG_FAILURE.

bench/external/scorecard.json, the run of 2026-09-20 against all 158 upstream checkouts.

The same case, graded from the transcript recorded .

The grading the benchmark gate runs on. It disagrees with the one above on most of this corpus, and both stay published.

Check it yourself

Everything here is downstream of a public commit.

Clone it, check out e47fc91, run the report through the CLI the way the study did.

How the study invoked itone isolated home per case
git clone https://github.com/micromatch/micromatch
git checkout e47fc9191d8f965c06f59945bd127c2cb9ae5f65
npm install

CREDDA_PROVIDER=heuristic \
  npx tsx apps/cli/src/main.ts fix <repo-path> @<issue-file> --no-color
Read the full studyAll 158 cases