Test inventory — audit → write → run ⇄ fix

Why this exists

Tests grow file by file, and after a while nobody can say what a screen actually does, which of those behaviours each layer proves, and which cases nobody handled. A coverage percentage does not answer it — a line can run without its result being checked. This skill audits one screen at a time across every layer, as use cases, hunts the defects and unhandled cases, then writes only the tests you approve and runs and fixes them in a bounded loop. Every claim carries a file:line or a command, so the report can be trusted, and the token spend stays bounded by the one screen.

Contributions are welcome: https://github.com/adyusuf/claude-code-standards

One screen at a time, every layer at once (web unit · e2e · backend unit/integration · Android · iOS · Maestro). The unit of work is a use case, not a file: a report tells the reader what the screen does, which layer proves each behaviour, what is wrong and what nobody handled.

Arguments

/adyusuf:test-inventory <phase> <target>

The four phases

The 3 ⇄ 4 loop

phase 3 (full run, classify) ──red──▶ phase 4 (fix, narrow re-run) ──all fixed──▶ phase 3
        │                                   │
      green ──▶ report, STOP        unapproved product bug / ceiling ──▶ report, STOP

• One round = one phase-3 run + the phase-4 fixes that follow it. Rounds are numbered in the report (R1, R2, …).
• Ceiling: 3 rounds (global #28's three passes). If the third phase-3 run is still red, stop and report every still-open failure individually — do not start a fourth.
• The loop also stops early when a failure needs a decision that is not the agent's: a product bug whose fix was not approved, a test whose expectation contradicts a project rule, or an environment the agent cannot repair (a missing secret).
• Phases 1 → 2 → 3 are gated by the user; 3 ⇄ 4 runs on its own inside those limits.

Non-negotiables (all phases)

1. Evidence or it did not happen. Every "tested", "missing" or "wrong" carries a file:line or the command that showed it. A grep hit is a lead, not a finding — open the file and confirm before writing it down (global #15, #28).
2. Numbers are measured. Test counts, coverage and durations come from a run or a count command; otherwise the cell says "not measured".
3. A defect is not a test gap. A product bug found during the audit is reported as a defect with its own fix task. It is never quietly fixed inside a test task, and a test is never bent to pass around it.
4. A missing feature is not a missing test. If a layer does not implement the behaviour (e.g. mobile has no saved-card flow), the cell is — and the row goes to "parity", not "gaps".
5. Dead code is not tested — it is proposed for deletion.
6. The project's rules win. Read the project CLAUDE.md first: its layer layout, test commands, env vars (e.g. an integration DB string), language and doc folder. Respect the active operating mode for agent use (global #27); in mode A no agents.
7. E2E follows global #33: listed as gaps in phase 1, never written or run at the dev stage; they are written and run at the test → prod gate.
8. Reports go to files and to chat. Per-screen file + an index (see templates/screen-report.md, templates/index.md); the chat gets a short summary table (global "STATUS REPORTING"). No Artifact dashboards. Report prose is written in the user's language; file names follow the project's docs convention.
9. Every phase ends with a documentation sync (references/doc-sync.md): the findings register, every screen report, the index, and the project documents the phase made stale (CLAUDE.md counts and known gaps, SETUP.md) are updated before the chat report. A phase whose Markdown still describes the previous state is not finished.

Where the reports live

Use the project's docs folder and naming convention. Default: docs/test-inventory/ with one <screen-slug>.md per screen, a README.md index and a findings register (findings.md). If the project already has an inventory folder (for example docs/test-envanteri/), continue there — never start a second one.

Every finding is also written to the findings register the moment it is verified (templates/findings.md): a stable id (F-001…), screen, priority, class, evidence, proposal and status. The register is the one list the user approves from, phase 2 writes against and phases 3/4 close; screen reports reference its ids. A finding is never deleted — it is closed with evidence, or marked "not a finding" with the reason.

Supporting files

• references/phase-1-audit.md — screen selection, surface mapping, use-case matrix, verification
• references/unhandled-case-checklist.md — the defect and unhandled-case hunt (the most valuable step)
• references/list-tests.md — commands that list test names and counts per stack
• references/phase-2-write.md — how the approved tests are written
• references/phase-3-run.md — full run, per-failure isolation and classification, mutation proof, coverage
• references/phase-4-fix.md — fixing, narrow re-runs, the hand-back to phase 3, the ceiling
• templates/screen-report.md — the per-screen report
• templates/index.md — the index across screens
• templates/findings.md — the findings register (all findings and run failures, with status)
• references/doc-sync.md — the documentation sync that closes every phase