audit-coverage
Read coverage artifacts a build already wrote (lcov, Cobertura, coverage.py JSON, a Go cover profile) and report line coverage per file and function, plus CRAP per function, for a change, a path, or the tree. Runs no tests, never gates. Use when: 'coverage of this change', 'coverage per function', 'CRAP score', 'which functions are complex and untested', 'read the lcov report', 'coverage.xml', 'how covered is this file'. Complexity alone: /code-metrics:audit complexity.
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 30
- Forks
- 7
01 前言 Front Matter
- shell
- bash
- metadata
- { "summary": "Coverage and CRAP read from build artifacts, no verdict", "workflow-stage": "anytime" }
- description
- Read coverage artifacts a build already wrote (lcov, Cobertura, coverage.py JSON, a Go cover profile) and report line coverage per file and function, plus CRAP per function, for a change, a path, or the tree. Runs no tests, never gates. Use when: 'coverage of this change', 'coverage per function', 'CRAP score', 'which functions are complex and untested', 'read the lcov report', 'coverage.xml', 'how covered is this file'. Complexity alone: /code-metrics:audit complexity.
- allowed-tools
- [ "Bash(${CLAUDE_SKILL_DIR}/scripts/audit-coverage.sh:*)", "Bash(git branch --show-current:*)" ]
- argument-hint
- [--json] [--all] [--base <ref>] [--artifacts <path>]... [<path>...]
- user-invocable
- true
- disable-model-invocation
- false
02 资源文件 7
- 参考文档evals/evals.json2.7 KB
- 脚本scripts/audit-coverage.sh8.0 KB
- 脚本scripts/audit-coverage.test.sh16.3 KB
- 脚本scripts/crap.py2.4 KB
- 脚本scripts/join.py47.5 KB
- 脚本scripts/test_crap.py3.4 KB
- 脚本scripts/test_join.py63.7 KB
03 SKILL.md 正文
Pre-computed context
Current branch: !git branch --show-current 2>/dev/null || echo "unknown"
Purpose
Coverage is a dynamic measure. This skill reads it statically: it parses an artifact a test run already wrote and joins it to the functions in scope. Running the tests is the project's own job and stays there, so nothing here executes a test command, starts a runner, or installs a tool.
Two numbers come out. Line coverage per file and per function is read from the artifact. CRAP per
function is derived: comp^2 * (1 - cov/100)^3 + comp, where comp is the function's cyclomatic
complexity from the audit skill's audit-complexity.sh script, run over the same scope.
| Format | Written by | What the join gets from it |
|---|---|---|
lcov .info |
most JavaScript and C/C++ toolchains, coverage lcov |
line hits; function records in either lcov record shape, which is where a function end line can come from at all |
| Cobertura XML | gcovr, coverlet, kcov (which is how a Bash lane gets an artifact) | line hits, <sources> prefixes, and <method> regions with their hit flag |
| coverage.py JSON | coverage json |
executed and missing lines, and per-function regions |
| Go cover profile | go test -coverprofile |
statement counts over line ranges, which is not a line table: the profile never says which lines carry the statements, so a Go file reports the statement ratio go tool cover -func prints, labels it cov_source: statement-ratio, and reports lines_executable and lines_hit as null. The format names no functions either, so a Go function reports coverage_pct and crap as null, and its line counts as null unless a line-measuring artifact also covered the file |
The record shape and version behind each format are its row under "Coverage artifacts" in
${CLAUDE_PLUGIN_ROOT}/reference/collectors.md, which carries each claim with its basis, date and
recheck trigger.
The SQLite data file coverage.py writes while measuring is never read; why is the coverage.py
SQLite row of that file. Run coverage json or coverage lcov and point this skill at the result.
Run it
"${CLAUDE_SKILL_DIR}/scripts/audit-coverage.sh" # the change, with the artifact auto-discovered
"${CLAUDE_SKILL_DIR}/scripts/audit-coverage.sh" --artifacts build/lcov.info src/ # an explicit artifact and an explicit scope
"${CLAUDE_SKILL_DIR}/scripts/audit-coverage.sh" --all --artifacts coverage.xml --artifacts cover.out
"${CLAUDE_SKILL_DIR}/scripts/audit-coverage.sh" --json --all # the code-metrics/v2 document instead of markdown
With no --artifacts and no coverage.artifacts in the config, the well-known names are looked
for under the repository root, at most two directory levels deep and never inside node_modules,
.git or vendor: coverage/lcov.info, lcov.info, coverage.xml, cobertura.xml,
coverage.json, coverage.out, cover.out. Each artifact's format is detected from its content,
not its extension.
Present the markdown report as printed: the scope, the "Coverage of this run" table, the
references with their provenance, then one row per file and per function. Keep the --json
document when the numbers feed a comparison: pass it to /verification:measure metrics when the
verification plugin is installed, treating status: empty on either side as INCONCLUSIVE;
otherwise keep the JSON beside your notes and compare by hand.
Getting a first artifact
A run that finds no artifact reports every lane unavailable, lists the paths it searched, and
ends its markdown with a line pointing here, because the remedy is the project's to apply: this
skill runs no test and installs nothing. One command per lane produces an artifact the next run
can read. Each row restates a producer's documented default, verified against the page in its
Basis column on 2026-09-12; the producer's page wins over the row. Recheck trigger: a release note
or changelog entry of that producer naming the flag or the output path in its row, or a read-time
fetch of the Basis page that no longer states what the row does; either re-derives the row from the
page and refreshes the date.
| Lane | Producer | Command shape | Writes | Basis |
|---|---|---|---|---|
| TypeScript/JavaScript | vitest | npx vitest run --coverage --coverage.reporter=lcov |
lcov .info at coverage/lcov.info, auto-discovered |
Vitest coverage guide, coverage.reporter |
| TypeScript/JavaScript | jest | npx jest --coverage |
lcov .info at coverage/lcov.info, auto-discovered; the default coverageReporters list carries lcov |
Jest CLI, --coverage; Jest configuration, coverageReporters default ["clover", "json", "lcov", "text"] |
| TypeScript/JavaScript | c8 or nyc, over any test runner | npx c8 --reporter=lcov <test command> |
lcov .info at coverage/lcov.info, auto-discovered |
c8 README, --reporter, which takes any Istanbul reporter, lcov among them |
| Python | coverage.py | python -m coverage run -m pytest && python -m coverage json |
coverage.py JSON at coverage.json, auto-discovered; coverage xml writes coverage.xml, also auto-discovered; coverage lcov writes coverage.lcov, which needs --artifacts |
coverage.py command pages run, json, xml, lcov |
| Python | pytest-cov | pytest --cov=<package> --cov-report=json |
coverage.py JSON at coverage.json, coverage.py's default name for the json report, auto-discovered; --cov-report=json:<path> moves it |
pytest-cov reporting, --cov-report; the default file name is coverage.py's, per its json page |
| Bash | kcov | kcov <outdir> bash <test script> |
Cobertura-compatible XML under <outdir>; pass the cobertura.xml it writes with --artifacts |
kcov README, "Kcov will also write cobertura-compatible XML output" |
| Go | go test |
go test ./... -coverprofile=coverage.out |
Go cover profile at the path -coverprofile names; coverage.out and cover.out are auto-discovered; file rows carry the statement ratio, function rows need a line artifact too |
go command, testing flags, -coverprofile |
| C# | coverlet, as the dotnet test collector |
dotnet test --collect:"XPlat Code Coverage" |
Cobertura XML at TestResults/<run id>/coverage.cobertura.xml; pass it with --artifacts. The C# complexity lane is deferred, so the lane reports file rows and no CRAP |
coverlet VSTest integration, --collect:"XPlat Code Coverage" |
Auto-discovered means the output lands on one of the well-known names the run looks for with no
--artifacts; anything else is named explicitly or listed under coverage.artifacts in the
configuration.
Reading the numbers
- A file row is the artifact's own line table:
coverage_pct,lines_executable,lines_hit. A file a Go cover profile covers is the exception: the profile weighs statements, so the percentage is its statement ratio,cov_sourcereadsstatement-ratio, and both line counts are null. - A function row exists for every function whose complexity collector reported a real end line.
It carries
coverage_pct,cyclomatic,crap, and, in the JSON,cov_sourceandhit. - The markdown table is ordered so the rows worth reading survive its 200-row cap: rows over a
reference first, then function rows by CRAP descending with a null CRAP last, then file rows by
coverage ascending with a null percentage last, and file and line only as the tie-break. The
first function row is the highest-CRAP function in scope, however its file name sorts; the rows
the cap drops are counted on the table's last row and are all in the
--jsondocument, which is not capped. cov_sourcesays where the coverage came from:artifact-regionwhen the artifact carried the function's own region (coverage.pyfunctions, a Cobertura<method>, an lcov 2.2FNLend line),line-rangewhen the range came from the complexity collector, in which case nested function ranges are subtracted from the parent first,statement-ratioon a file row whose percentage came from a Go profile's statements rather than from a line table, andambiguouson a function the join refused. The markdown table renders the values; the per-rowcov_sourceandhitfields are in the--jsondocument.- A function the artifact names by a short name binds to the function whose line range holds the
lines the artifact recorded, not to whichever function the name happens to end at. Two
runmethods in one file and a record namedrunare one function's coverage, and the range says whose. Where nothing separates the candidates, because the artifact placed none of them at a line or because two of them fall inside the same range, the join is refused rather than guessed: the row readscov_source: ambiguous; coverage, line counts, hit, and CRAP are null because those numbers would be a guess, while cyclomatic from the complexity row is kept because that collector already measured it; it carries acoverage-ambiguouslabel, and the lane's run row turnspartialand names the functions it left unjoined. A missing coverage number an operator can see beats a wrong number they cannot. hitis the artifact's function-hit flag. When it says the function was never entered,coverage_pctis 0 rather than the 1/N a declaration line executed at import would produce.- A function with no executable lines in the artifact reports
coverage_pct: nullandcrap: null, never 0: a fabricated 0 percent would produce a fabricated maximal CRAP. Itslines_executableandlines_hitare 0 when an artifact that measures lines covers the file and carries none in that range, and null when no artifact covering the file measures lines at all, which is every function of a file only a Go profile covers. - CRAP is Savoia and Evans' formula (Agitar Labs, 2007). It is a description of two numbers you
already have, not a validated predictor of change risk, and no standard sets a threshold for it.
/code-metrics:principlescarries the provenance, the rename history, and the caveats; read it before you quote a CRAP number at anyone. - A lane whose resolved complexity collector reports no function end lines gets a
<lane>/craprow withstatus: not-applicableand that reason. Bash is that lane in this version: no maintained Bash collector reports an end line, so Bash CRAP is a stated gap, not a null. - A lane whose cyclomatic collector did not resolve, or ran and produced nothing parseable, gets a
<lane>/craprow withstatus: unavailablenaming that collector and its reason, whatever the lane's coverage row says. CRAP needs both numbers, and the missing one is the cause worth reading; when the coverage row is notokeither, its reason follows the collector's on the same row, so a partial count or a no-artifact search never hides a collector failure and the collector never hides them.Functions: 0in the summary reads with that row. - A lane matched by fewer than all of its scope files carries
status: partialand the reasonpartial, N of M scope files present in the artifacts, so a total miss never reads as "no executable lines" and the document cannot settle ascompletewhile a row saysN of M. A lane no artifact covers isunavailablewith that count; when no artifact was found at all, the reason lists every path searched. A lane that left a function unjoined ispartialfor that reason too. - A row carrying that
N of Mcount also says which files: its reason names the first five scope files no artifact mentions (; missing: a, b, c, d, e, +N more in the JSON) and the JSON row'smissingkey lists all of them, root-relative and sorted. The count alone cannot tell a test file an artifact never records from a source file the suites never reach; the paths can. - Whether the covered code is actually checked by its tests is a different question, and coverage
alone cannot answer it: a line can execute under a test that asserts nothing.
/mutation-testing:auditowns that question when themutation-testingplugin is installed; without it, say that the coverage number is an execution count and stop there. - Exit 0 whenever a report was produced, including an empty one; exit 2 for a usage error, which
includes a named artifact or scope path that does not exist; exit 3 when a complexity collector
ran and produced nothing parseable. On exit 3 the collector's stderr is on the
<lane>/craprow, and the summary carries anExit 3:line naming the lane and collector, so the markdown alone says why the run has no function rows.
Configuration
Everything tunable resolves through the code-metrics layers (user-global, team
docs/conventions/code-metrics.yaml, personal override; per-key override; keys in ${CLAUDE_PLUGIN_ROOT}/reference/config.md):
| Key | Default | Effect |
|---|---|---|
coverage.artifacts |
[] |
Explicit artifact paths; empty means auto-discover. A listed path that does not exist is a usage error, so a stale entry is loud |
coverage.path_prefix_strip |
[] |
Prefixes removed from artifact paths before the join, for compiled-output layouts (dist/, build/) |
coverage.reference |
null |
No bar. A number counts rows below it as over_reference and nothing else |
coverage.crap.reference |
null |
No bar; no standard sets one for CRAP |
/code-metrics:setup writes the team file and probes the collectors.
What this skill does not do
- It never runs tests, starts a test runner, or installs a coverage tool. It reads what a build already wrote; if there is no artifact, the report says so and stops.
- It never reads the SQLite data file coverage.py writes while measuring.
- It never judges. A reference here is a value to count against, and no finding, severity, or exit-code gate follows from one.
- It does not measure branch, condition, or MC/DC coverage. Those records are parsed past, not reported, because the join is a line-level one.
Next
- The numbers feed a before-and-after comparison:
/verification:measure metrics. - Coverage looks healthy but the tests may assert nothing:
/mutation-testing:audit. - A number is about to be quoted at someone:
/code-metrics:principles.
Gotchas
- The complexity collector decides which functions can carry CRAP at all. With no collector
resolving for a lane, that lane has no function rows and its
<lane>/craprow isunavailablenaming the collector's own reason. File-level coverage is unaffected: the scope the join keys on is the one the audits measure, not the files a complexity collector emitted rows for, so a file with no functions still reports its lines. - Path shapes differ between artifact and source tree. The join normalizes both sides, strips the
repository root and the configured prefixes, then falls back to a component-wise suffix match
and finally to a unique basename (which is how a Go profile's module path resolves). Both
fallbacks require a single winner: an artifact path that fits two scoped files equally well
(two services vendoring the same
pkg/a.py) stays unmatched and shows up in the partial count, because attributing it to either would credit one service with the other's coverage. The basename fallback needs a single winner on the artifact side too, counted across every artifact read rather than within one: when two services each ship ahandler.goand only one is in scope, the out-of-scope path stays unmatched instead of merging its coverage into the scoped file, even though the scope names exactly one candidate. - An artifact older than the source it describes joins by line number and will be wrong without saying so. Re-run the tests before reading the numbers when the diff has moved lines.
- Two artifacts covering one file are merged by keeping the larger hit count per line, so a line
is never counted twice. Two artifacts covering one function are folded the same way: the hit
flag is the larger of the two, so a suite that never entered the function cannot report it at 0
percent beside a file row the other suite already showed as covered. Two Go profiles covering
one file merge block by block on the profile's own
start.col,end.colidentity and are totaled after that, so two test shards that entered different blocks report what they covered between them rather than what the luckier shard covered alone. - A Go profile and a line artifact naming the same
.gofile do not average or overwrite each other: the profile's statement ratio is that file's exact native measure and wins the percentage, the row saysstatement-ratio, and the line counts stay null rather than reading as the counts behind a number they did not produce. The lane'scollectorstill names both formats read. Function rows are line-based, so a Go function does take its region from a line artifact when one covers it. - A Cobertura report with several
<source>roots is read against all of them: a relative class filename takes the first root under which that path exists in the scanned tree, and falls back to the first root when no candidate exists. Two roots that both hold the file is an ambiguity the report cannot settle, so the first one wins and the class lands under it. The probe settles anything only when the roots are relative, or absolute and present on the machine running the audit. A report built elsewhere declares absolute roots (the usual coverlet, kcov and gcovr shape for a CI run) naming directories this machine does not have, so every candidate misses and every class lands under the first root. Rewriting a root from another machine onto the local tree needs a mapping the report does not carry. - Two functions in one file whose qualified names share a tail (
A.runandB.run) can bind to each other's artifact region when the artifact carries only the short name; the file-level numbers are unaffected.
04 同仓库更多 Skills
track
Backlog reads through the bound tracker: search open and closed items, view one item with its comments and state history. Use when another skill searches or reads tracker items.
reviewer-media
Reviewer media for one pull request: decides whether it needs a narrated walkthrough, builds the storyboard from /review:explain-change's digest, has /explainer-video:produce render it and hands the manifest to /source-control:pull-request media. Use when asked for 'reviewer media', 'a walkthrough video for this PR', 'a produced video of this change', or when pull-request ready or monitor hands a PR over.
route
Model and effort per agent role (orchestrator, worker, verifier, retrieval) for a multi-agent run, resolved through the bundled defaults and the user, team and personal override config layers into fenced JSON with single-agent and fan-out variants and the layer behind each value. Use when asked 'which model for the subagents', 'route the roles' or 'model routing', or before a skill passes a role's model on a spawn.
route
Model and effort per agent role (orchestrator, worker, verifier, retrieval) for a multi-agent run, resolved through the bundled defaults and the user, team and personal override config layers into fenced JSON with single-agent and fan-out variants and the layer behind each value. Use when asked 'which model for the subagents', 'route the roles' or 'model routing', or before a skill passes a role's model on a spawn.
principles
Module-design vocabulary for architecture findings, and the checks built on it. Use when: 'what is a deep module', 'what is a seam', 'deep vs shallow module', 'what is the deletion test', 'leverage and locality', 'module design vocabulary', 'which word do we use for X in architecture findings'. Scanning a codebase for friction is /architecture:improve.
write
Structured five-field bug report (title, steps to reproduce, expected vs actual, justified severity, suggested fix location) from an informal description, read-only. Use when the user names a defect they observed, or asks for one written up ('report a bug', 'file a bug', 'write this up as a bug').