agentic-dev/docs/decisions/adr-020-coverage-architecture.md

# ADR-020 — Agent-first coverage architecture (4 layers + manifest-driven thresholds)

**Status:** Accepted
**Date:** 2026-05-13
**Builds on:** ADR-006 (vertical-feature-packages), ADR-011 (TDD foundation)
**PRD:** docs/work/prds/2026-05-13-coverage-architecture.prd.md

## Context

ADR-011 established the TDD foundation: per-package vitest configs with V8 coverage, a coverage baseline (80/75/80/80) and stricter per-layer bands (100% on `entities/`, `application/use-cases/`, `interface-adapters/controllers/`). The ESLint conformance rule `usecase-must-have-test-file` enforces _that a test file exists_. CI runs `pnpm test -- --coverage` and uploads `**/coverage/lcov.info` as an artifact.

This leaves five gaps that matter especially in an agent-first repo:

1. **No diff-coverage gate.** A slice can ship without exercising its new lines. The ESLint rule only checks file presence; the per-package thresholds catch only large drops.
2. **No aggregate visibility.** N separate lcov files; no merged view, no trend over time.
3. **Threshold declarations are duplicated** across 5+ `vitest.config.ts` files. Drift is mechanical to spot (we did it during the 2026-05-13 brainstorm: `@repo/media` had no `coverage:` block at all; `@repo/navigation` failed its declared layer thresholds in entities + controllers).
4. **100% coverage with weak assertions is invisible.** Coverage doesn't measure whether tests would catch real regressions. Mutation testing — explicitly deferred in ADR-011 — is the next signal.
5. **Coverage data isn't agent-readable.** The HTML report serves humans; the dispatch loop has no way to ask "did my slice cover its diff?".

## Decision

**1. Adopt a 4-layer coverage architecture** mirroring the 5-gate conformance philosophy (multi-latency, machine-readable, agent-first):

| Layer                              | Catches                                              | Latency            | Surface                                                      |
| ---------------------------------- | ---------------------------------------------------- | ------------------ | ------------------------------------------------------------ |
| **L0** Per-layer vitest thresholds | Drift below declared bands                           | ~5–30s per package | `pnpm test --coverage` (existing)                            |
| **L1** Diff coverage               | Changed line not exercised                           | ~5s after L0       | `pnpm coverage:diff`; CI gate; dispatch post-task            |
| **L2** Aggregate trend             | Drift across the codebase over time                  | ~10s               | `pnpm coverage:aggregate`; committed `coverage/summary.json` |
| **L3** Mutation testing            | Tests that exist + execute the code + assert nothing | Minutes            | `pnpm mutate`; on-demand, not default `pnpm test`            |

Each layer answers a distinct question; none replaces the others.

**2. Make `feature.manifest.ts` the single source of truth for coverage expectations.** A new `coverage:` section per feature:

```ts
coverage: {
  bands: {
    "entities":    { statements: 100, branches: 100, functions: 100, lines: 100 },
    "use-cases":   { statements: 100, branches:  95, functions: 100, lines: 100 },
    "controllers": { statements: 100, branches:  95, functions: 100, lines: 100 },
    baseline:      { statements:  80, branches:  75, functions:  80, lines:  80 },
  },
  mutationTargets: ["entities", "use-cases"],
}
```

Three readers consume the manifest:

- **Vitest** — `vitest.config.ts` imports the manifest's `coverage` and emits its `thresholds`. The duplicated block in 5 per-feature vitest configs goes away.
- **`assertFeatureConformance`** — reads `coverage/lcov.info` for the package at boot and asserts each band. Graceful degradation in `USE_DEV_SEED=true` (warns rather than throws when lcov is absent).
- **`pnpm coverage:diff`** — uses `baseline` for uncategorized files; stricter layer bands override per matching path glob.

This eliminates the duplication that caused the `@repo/media` drift and centralizes one decision in one place per feature.

**3. Diff coverage is cover-the-diff, not cover-the-new-code.** Every changed _executable_ line must have execution-count > 0 against the merged lcov. Modified-but-not-new lines count too — catches "agent edited code, didn't update the test." Allowlist: `*.test.ts`, `*.config.*`, `*.md`, `*.json`, `*.mjs`, plus the per-package exclude lists.

**4. Aggregate trend ships in-tree, not via SaaS.** `coverage/summary.json` is committed on merge to main. Trend readable via `git log -- coverage/summary.json`. No external service dependency; the dispatch loop can read history without a network call.

**5. Mutation testing is opt-in and narrowly scoped.** Stryker with `@stryker-mutator/vitest-runner`, runs on `entities/` + `application/use-cases/` only. Default mutation-score threshold 80% per feature (tunable per-manifest). Not part of `pnpm test`. Nightly GH Action surfaces score drift > 5%.

**6. Output format is machine-first.** `pnpm coverage:diff` emits JSON to stdout; human-readable summary to stderr. The dispatch loop reads stdout; humans read stderr or the HTML report.

## Alternatives considered

- **Codecov / Coveralls SaaS.** Polished PR comments and trend dashboards, free for OSS. Rejected as the _primary_ L2 store — adds an external dep, makes the dispatch loop dependent on a network call, and the PR-comment UX targets humans (not the primary consumer of this signal). Can be added later as gold-plating without disturbing the architecture.
- **Cover-the-new-code instead of cover-the-diff.** Lighter touch; ignores modified lines. Rejected — catches less drift. A slice that edits a use case without updating its test should fail, and cover-the-new-code wouldn't notice.
- **Keep thresholds in per-package vitest configs.** Status quo. Rejected — the 2026-05-13 audit found drift in 2 of 5 features (media had no block at all; navigation's block diverged subtly from the canonical). Manifest centralization is the only durable fix.
- **Run mutation testing in default `pnpm test`.** Rejected — Stryker on entities + use-cases takes minutes. Adding minutes to the default loop violates the constraint ("new gates must not add more than ~30s wall time"). On-demand is the right cadence; nightly catches drift.
- **Mutation testing across all layers.** Rejected for v1 — repository/controller/integration code has too many environmental dependencies to mutate cleanly. Start narrow; expand if signal is high.
- **Use ESLint or fallow for diff coverage.** Rejected — diff coverage needs runtime data (which lines actually executed), not AST or filesystem state. It belongs alongside `pnpm test`, not in `pnpm lint` or `pnpm fallow`.
- **Boot-time coverage assertion is too heavy.** Considered. Counter-argument: the assertion is `O(features × lcov-file-size)` — small numbers, ~200ms. The graceful-degradation in dev mode means contributors aren't blocked. The payoff — coverage drift caught at the same latency as TypeScript brands — justifies the machinery.

## Consequences

**Positive:**

- Every PR/task is gated on covering its own diff. Agent shipping an untested slice becomes mechanically impossible at the CI step.
- One source of truth per feature for coverage expectations. The `@repo/media`-style "no coverage block" drift can't recur.
- Trend history lives in the repo. `git log -- coverage/summary.json` answers "how has coverage moved over the last quarter?" without leaving the editor.
- Mutation testing on the highest-leverage layers (entities + use-cases — the pure-business-logic surface) raises the floor on test quality without slowing the dispatch loop.
- Machine-readable diff-coverage output integrates directly with the dispatch loop's post-task verification, completing the agent-first observability story.
- Coverage joins the 5-gate conformance philosophy as a first-class signal; ADR-020 becomes the row alongside TS brands / ESLint / boot / `pnpm conformance` / fallow / coverage.

**Negative:**

- Implementation surface is non-trivial: 6–8 stories spanning manifest schema, vitest auto-derive, two new scripts, boot-time assertion, mutation tooling, ADR + guide + glossary + generator + hook updates.
- The boot-time assertion adds a small dependency on `coverage/lcov.info` existing. Graceful degradation in dev mode handles this, but the implementation needs care.
- `coverage/summary.json` committed on merge introduces a small CI permissions surface (`contents: write`) gated to the main-branch workflow.
- Mutation testing is slow. The nightly cadence is the compromise; on-demand `pnpm mutate` is opt-in but rare in practice.

## Implementation phasing

In order (each landable independently):

1. **L0 unification** — fix `@repo/media` (missing dep + missing config block) and `@repo/navigation` (real test gaps) so every feature passes its declared bands today.
2. **Manifest schema** — extend `feature.manifest.ts` shape (Zod schema in `core-shared/conformance/`) with the `coverage:` section.
3. **Vitest auto-derive** — `vitest.config.ts` per feature imports the manifest and emits `coverage.thresholds`. Eliminates duplication.
4. **L1 diff coverage** — `scripts/coverage/diff.mjs` + `pnpm coverage:diff` script + CI gate.
5. **L2 aggregate** — `scripts/coverage/aggregate.mjs` + `pnpm coverage:aggregate` + summary.json + merge-to-main workflow.
6. **L3 mutation testing** — Stryker setup + `pnpm mutate` + nightly GH Action.
7. **Boot-time `assertFeatureConformance`** — coverage band read against lcov.
8. **Docs + generator + hook rollout** — this ADR (now landing), `docs/guides/coverage.md`, glossary entries, `pnpm turbo gen feature` template update, `.claude/hooks/prompt-context.sh` keyword group.

## Related

- ADR-006 — vertical-feature-packages
- ADR-011 — TDD foundation
- ADR-018 — audit-and-compliance (similar manifest-declared shape pattern)
- ADR-019 — sandcastle agent orchestration (the dispatch loop that reads `pnpm coverage:diff`)
- PRD `2026-05-13-coverage-architecture` — implementation seed