e2e

Live verification

Per-repo setup (read, never guess)

Read the repo's verification config: AGENTS.md, CLAUDE.md, TESTING-NOTES.md, docs/operations/, .agents/verify.md, or .claude/verify.md. It should define: boot command (compose file), health checks, ports, seed data, test credentials per role, newman collection paths, Playwright config, DB connection. If none exists, offer to create docs/verification.md; keep real credentials in a gitignored local file or secret store.

Layers - run what the change touches

• Boot: use the repository's boot command (often docker compose up -d), wait on its health checks, confirm migrations applied, and capture logs on failure.
• API: newman collections when they exist; otherwise curl the changed endpoints - auth flow, happy path, validation errors (RFC 9457 shape), and idempotent replay for money operations.
• DB: psql/mongosh state checks after writes - rows exist, balances reconcile. Ledger changes must run the zero-row invariant queries from the ledger references.
• UI: Playwright against the running app - real login with configured credentials, the changed flows, screenshot evidence. Prefer role, label, and accessible-name locators; use stable test IDs only when a semantic locator is insufficient, per ../../references/ui-selector-conventions.md.

Resource: scripts/playwright-setup.sh installs Playwright browsers. Pass --with-deps only when approved OS dependencies are needed. The repository must already declare @playwright/test; add it only when the requested scope includes test setup.

Evidence rules (default-to-reject)

• "Tests pass" without runner output = not verified. A design-tool screenshot is not evidence from the running app. Every claim needs the command plus its actual output.
• Reject missing coverage, not a low defect count: requirements marked done without a verification command fail the gate, as does a UI change without the relevant responsive/mobile check. If no defect is found, state the verified scope and remaining gaps without inventing issues.
• On failure: diagnose via debug, fix via build, re-run the failed layer, then the smallest end-to-end path that proves the layers still integrate.

Put the report (what ran, evidence, gaps) in the existing plan doc or PR description when one exists; otherwise return a concise verification report. Do not create a separate schema or ceremony file.