| name | activity-coverage-gap-analysis |
| description | Measure coverage against a baseline when a provider exists, and run risk-ranked structural gap analysis when none does. Validates existing coverage artifacts before trusting them. Use when reporting test coverage and gaps. |
Activity: Coverage and Gap Analysis
Report what is tested, what is not, and what that means — with or without a coverage provider installed. Invoked by the qa-engineer agent as step 3 of its procedure.
RFC 2119 Notice: The key words MUST, MUST NOT, REQUIRED, SHALL, SHALL NOT, SHOULD, SHOULD NOT, RECOMMENDED, MAY, and OPTIONAL in this document are to be interpreted as described in RFC 2119.
Goal
Produce a gap report that is honest about its own limits. Two paths exist. The structural path always runs; the measured path runs additionally when tooling permits.
You MUST NOT return unknown, unable to determine, or an empty report. Absence of tooling is not absence of gaps.
Path A — Measured coverage
Runs only when a coverage provider is installed for the package.
- Run the project's coverage command. Do not install anything to enable it.
- Compare against the recorded baseline in
/TESTING.md.
- Report per package: measured value, baseline, delta, and whether the threshold policy is met.
- Map uncovered regions to acceptance criteria where a criterion list is available.
Gate value
| Condition | coverage_gate |
|---|
| Measured, policy satisfied | PASS |
| Measured, policy violated or baseline regressed | FAIL |
| No provider installed | SKIPPED(<reason>) |
| Provider present but the run failed | SKIPPED(<reason>) |
| User declined measurement | SKIPPED(<reason>) |
The reason string is REQUIRED and MUST be non-empty — SKIPPED() and a bare SKIPPED are both invalid. That requirement is what prevents a silent skip.
You MUST NOT emit PASS for coverage that was not measured. A crashed provider is SKIPPED, never PASS.
Path B — Structural gap analysis (always runs)
Requires no coverage provider. This path is what makes a repository with zero coverage tooling still analyzable.
- Enumerate untested surfaces. For every source file, determine whether any test file references it or its exported symbols. List files and exported symbols with no corresponding test.
- Report source-to-test size ratio per package, in lines. A package with 1,600 lines of source and 50 lines of tests is reported as such.
- Flag disproportionate surfaces. A module of 800 lines exercised by one test case has symbolic coverage, not coverage. Report it alongside wholly untested files.
- Rank by size and risk, not alphabetically and not flat. A large untested service outranks a small untested helper. Risk weighting SHOULD consider: handles authentication or authorization; touches money or persistent state; is reachable from an external entry point; has no types.
- Exclude generated and vendored code, and state the exclusions. A ranking topped by a generated artifact discredits the report.
Divide-by-zero and boundary handling
| Situation | Required behavior |
|---|
| Package with tests, no source | Report the ratio without error; do not list it as an untested gap. |
| Package with source, no tests | Rank at or near the top; this is the highest-signal gap. |
| Package with neither | Report as empty; exclude from ranking. |
| One very large generated file | Exclude and state the exclusion. |
| Very large workspace | Complete, or return a bounded partial result naming what was skipped. |
Artifact Validation
Existing coverage artifacts MUST be validated before being used as evidence. An unvalidated artifact is worse than no artifact, because it reports a number that reviewers will trust.
Check each artifact for:
| Check | Finding condition |
|---|
| Staleness | Generated before the most recent change to the code it describes, or its test count understates reality. |
| Scope narrowing | Measures a narrower scope than the package it appears to describe — for example a coverage command hard-coded to one module while presented as the package's coverage. |
| Committed state | A coverage database or report is committed to version control and not ignored. |
A stale or scope-narrowed artifact is reported as misleading and MUST NOT be used as evidence or carried forward as the package's coverage value.
Worked example
A package ships an HTML report reading 46.58%, dated three months ago, whose command is hard-coded to one of eight modules. A committed results file lists 18 tests where 112 now exist.
Correct output: the 46.58% figure is not reported as the package's coverage. Report two findings — the artifact is stale, and its scope is one eighth of what it appears to cover — plus the structural gap inventory for the seven unmeasured modules, and a note that the committed artifacts are untracked by .gitignore.
Coverage Is Not Confidence
State this in the report when a high percentage coexists with weak assertions. A suite that mocks its data layer everywhere can reach a high percentage while remaining blind to query defects. Where you observe that pattern, report it as a limitation of the measurement, not as a passing grade.
Output
coverage_gate: PASS | FAIL | SKIPPED(<reason>)- Measured coverage per package with baseline and delta, when Path A ran
- Reason for skipping measurement, when it did not
- Ranked structural gap inventory, largest and highest-risk first
- Source-to-test ratio per package
- Disproportionately tested surfaces
- Misleading artifacts, with the reason each is untrusted
- Exclusions applied
- What was not analyzed, and why
Final Instructions
- You MUST run the structural path on every invocation, including when measurement is skipped.
- You MUST NOT return
unknown or an empty report.
- You MUST NOT emit
PASS for unmeasured coverage.
- You MUST include a non-empty reason with every
SKIPPED.
- You MUST rank gaps by size and risk rather than listing them flat.
- You MUST validate coverage artifacts before trusting them, and report stale or scope-narrowed ones as misleading.
- You MUST NOT install a coverage provider — report its absence.
