Test · Skills

The opt-in Test family ships three skills, extracted alongside the Code-family skills for every detected agent target on heal init. They only act on findings produced by the test observers.

For installation and the drift-aware update model, see Code › Skills — the mechanism is shared.

/heal-test-reporter-setup — wire up lcov

One-shot setup skill. Detects your project’s language stack and proposes the lcov reporter configuration plus CI integration so lcov.info lands at one of heal’s default lcov_paths.

Stack	Proposes
Rust	cargo llvm-cov --lcov --output-path lcov.info
Python	pytest --cov=src --cov-report=lcov (pytest-cov)
JS / TS	nyc --reporter=lcov mocha / vitest --coverage --coverage.reporter=lcov
Go	go test -coverprofile=coverage.out + gcov2lcov
Scala	scoverage plugin + lcov reporter
Mixed	per-stack proposals + a CI step producing a single lcov.info

Read-only on the codebase — proposes commands and config edits without running them. You run the commands; heal reads the resulting lcov.info on the next heal status.

Trigger phrases: “set up coverage reporting”, “configure lcov for heal”, “wire up coverage”, “/heal-test-reporter-setup”.

/heal-test-review — the audit skill

Read-only. Reads heal status --json, filters to the [features.test] slice, and returns:

1. An architectural reading of the test suite — is the dominant axis “no unit tests”, “tests aren’t keeping up with their source”, “important paths uncovered”, or “skipped flakes that solidified into permanent skips”?
2. A prioritized test-fix TODO list — coverage gaps on hotspot files first, then drifting tests, then skip-ratio outliers.

Never edits source. After reading the review you can act on any item right away — ask the agent in the same session (“write the missing tests for the top three”, “re-enable the skipped tests under auth/”). Mechanical fixes flow through /heal-test-patch; judgment calls — should this skip stay because the underlying flake is real? is this uncovered file something we test, or something we delete? — wait for your direction.

Why review and patch are split

Patch handles the mechanical: write a unit test for a clearly uncovered branch, sync a drifted test back to its source, unskip a test whose recorded reason (“waiting on issue #123”) is now resolved. Review also surfaces the judgment calls — suspiciously weak assertions, skips that probably hide a real flake, files that maybe shouldn’t be tested at this layer at all. Mixing the two would either paper over a real flake or refuse to write the easy tests.

Trigger phrases: “review the test health”, “where should we add tests”, “which tests should we unskip”, “/heal-test-review”.

/heal-test-patch — the write skill

Drains the test slice of .heal/findings/latest.json one finding at a time, in Severity order. One commit per fix.

Pre-flight (refuses to start otherwise):

• Clean worktree.
• Cache exists (runs heal status --json to populate if missing).
• [features.test] enabled in .heal/config.toml.
• lcov.info present at one of lcov_paths when coverage_pct findings are in scope (otherwise the skill points the user at /heal-test-reporter-setup).

Per-metric moves:

Metric	Default move
coverage_pct	Write or extend a unit test for an uncovered hot path; re-run the reporter so lcov.info updates.
skip_ratio	Re-enable a skipped test whose reason no longer holds; remove the skip marker, run the test, fix any failure inline.
change_coupling.drift	Align the drifted test with its source — the skill surfaces both files together.

Refusals (encoded in the skill body, won’t budge with prompting):

• Assertion-weakening — never converts assert.equal(x, 5) to assert.ok(x) to pass a stale test. If the assertion is wrong, the test gets removed with a commit message naming the reason.
• Skip-the-flake — never adds a skip marker to make a flaky test go away.
• Scaffold-without-running — every commit runs the test suite. Tests that “look right” but haven’t been executed never land.

Constraints: one finding = one commit, Conventional Commit subject + Refs: F#<finding_id> trailer, never push / amend / --no-verify. Findings whose metric belongs to the Code or Docs families are skipped.

Trigger phrases: “fix the test findings”, “drain the test cache”, “add tests heal flagged”, “/heal-test-patch”.