CI measures a different set of files, or runs a different set of tests, than your laptop. The gap is not missing tests, it is a measurement mismatch: a subset run, a shard, or source files that exist in CI but not locally. Diff the measured file sets and the executed test sets before writing a single new test.

```text
pytest coverage "Required test coverage of 90% not reached" failing only in CI: how to find the gap
```

## Steps

1. Generate the coverage report in CI as an artifact (or print the missing-lines table) and compare the per-file numbers against your local report. Find which files lose coverage in CI.

   Expected: You have the list of files whose CI coverage is lower than local.

2. Check whether CI runs fewer tests: look for -k filters, deselections, sharding, or a fail-fast that aborts the run before the later test files execute.

   Expected: You know exactly which tests CI skips that your laptop runs.

3. Check whether CI measures more files: generated code, platform-specific modules, or vendored sources that exist in the CI checkout but not in your local tree will drag the percentage down.

   Expected: You find the files CI measures that your local run never sees.

4. Align the measurement: include or omit the same source paths in both configs, and make sure CI runs the full test set that the threshold was calibrated against.

   Expected: Local and CI now measure the same files over the same tests, and the numbers agree.

5. If a real gap remains after alignment, write tests for the uncovered lines the report names. Rerun CI to confirm the threshold passes with the aligned config.

   Expected: The CI gate passes and the report shows which lines the new tests cover.

## Use this when

- coverage threshold passes locally but fails in CI
- CI and local coverage percentages disagree
- the coverage gate fails after sharding or subsetting tests

## Not for this skill when

- coverage is below the threshold in both environments (write more tests)
- the coverage tool itself crashes or errors
- the threshold was just raised and nothing changed (calibrate it first)

## Variant phrasings

- pytest coverage threshold failing only in CI
- required test coverage not reached CI but passes locally
- coverage mismatch CI vs local how to find gap

## Why it happens

Coverage percentage is measured files divided by executed lines over total lines. CI and local can disagree on both halves: CI may run a subset of tests (shards, filters, fail-fast), which lowers the numerator, or it may measure extra files (generated code, platform modules), which raises the denominator. Either way the percentage drops without a single line of product code changing. Writing new tests before checking the measurement just adds coverage to files CI may not even be counting.

## Edge cases

- Coverage config files can differ by environment: check for separate CI config, environment-variable-gated settings, and rc files the CI template injects.
- Parallel runs need their coverage data combined before the report; an uncombined shard reports its own slice as the whole.
- Omit generated and vendored code from measurement explicitly rather than covering it; thresholds should measure code humans maintain.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_fOJb6CwAFcdEBIHbT4rDXQ