Dashboard QA bug hunt

Read the shared QA contract first. Read dashboard/CLAUDE.md. Use Playwright MCP to test real browser behavior.

Arguments: optional server URL, journey names, wN, SHIP=1, DRY_RUN=1. Default to production and report mode. A target URL changes the client under test; inspect its API origin independently. Record whether the running client matches local HEAD before declaring a source fix missing from production.

Run in the background

• Default to a headless browser with an isolated QA profile. Keep the user's windows, tabs, mouse, keyboard, and system clipboard undisturbed. A fresh context inside a headed browser can still open a visible window; it does not satisfy this requirement.
• Launch Playwright MCP with --headless --isolated. For Claude Code, check mcpServers.playwright.args in the repo's .mcp.json; for Codex, check the active [mcp_servers.playwright] configuration, normally in ~/.codex/config.toml. Inspect effective launch settings, including overrides and any attached-browser endpoint, before the first browser action. A config edit takes effect on a new MCP server/browser session, not an already-running headed session. Do not attach to the user's regular browser.
• If the current MCP session is headed, establish a separate headless session before continuing UI checks. An already available Playwright runtime in a QA-owned process is an acceptable fallback when MCP cannot be restarted safely. Preserve the same login and evidence workflow. If no background session is available, report the UI checks as blocked and continue independent source/API investigation; do not silently launch a visible browser or stop another task's browser server.
• Use Playwright's page/context APIs for clicks, typing, keyboard navigation, focus assertions, viewport changes, screenshots, and downloads. Do not call bringToFront, activate desktop apps, drive the OS mouse/keyboard, or open downloaded artifacts in desktop viewers. Browser-local focus checks remain part of accessibility QA.
• For uploads, use file inputs/file-chooser APIs with authorized fixtures. For print checks, inspect print media or a headless PDF where supported; this does not verify the native print dialog. Run visible-browser, native-dialog, or actual screen-reader checks only when explicitly requested. Otherwise record that coverage as unverified and continue the background journeys.
• Record the browser mode and session ownership with the run evidence. Reuse the owned headless browser with fresh contexts as needed, and close only this run's contexts, browsers, and helper processes during cleanup.

Prepare the browser

• Record the initial browser session. Authenticate using QA_EMAIL and QA_PASSWORD from dashboard/.env, unless the user supplies another path. From the repo root, run node .agents/skills/qa-shared/scripts/qa-login.mjs --env-file dashboard/.env. The helper hides the password and transfers session cookies in memory. Inspect features/auth/hooks/use-login-form.ts, graphql/query/auth.graphql, and the backend auth resolver/cookie helper for the current contract. Verify the resulting account and role, not merely a 200.
• Use an ordinary desktop viewport (for example 1440×1000), then repeat the relevant controls at a narrow viewport. Record locale, theme, and active filters. Prefer accessible roles and names from snapshots to brittle selectors.
• If the account is unavailable, continue public-ledger and signed-out journeys and explicitly leave authenticated behavior unverified.
• For local reproduction, install existing dependencies with yarn install --immutable inside dashboard/; start yarn dev. Never assume the local server uses a sandbox backend. Existing scripts/perf-fixture-api.mjs can isolate loading behavior, but results against its synthetic owner/fixtures must be labeled local simulations.

Sweep whole journeys

Use selected journey names to narrow this table; otherwise survey the available surfaces and deepen the journeys that show failures. Report each skipped group.

Journey	Observable promise
auth	Sign-in reaches the intended safe next route; cancellation/back and expired-session states remain usable. Do not send reset or invitation email without authorization.
discovery	Gallery search narrows, pagination changes results, a public ledger opens, and returning restores a coherent list. Test ledger switching without leaking prior data.
overview	Date/account filters affect cards and charts; recent activity opens the corresponding journal item; empty and loading states are truthful.
journal	Type, flag, date, account, payee/tag filters compose; pagination/count agree; postings expand; back/reload preserves the stated filters. Authorized synthetic edits persist exactly once.
accounts	Search and hierarchy expansion work; a linked account shows matching entries and balances; closed or absent accounts have honest states.
reports	Balance Sheet, Income Statement, Trial Balance, Cash Flow, holdings/statistics and budget honor the selected period and units; drill-down retains intended context.
export	CSV/Markdown and Print / Save as PDF match the filtered supported statement, dates, entity, commodity units, and sign conventions. Read features/reports/CLAUDE.md before interpreting financial signs.
files	Tree navigation, text preview, branches, commit/diff links, and downloads work. Test create/edit/upload/delete only with authorized disposable files and check the resulting commit.
query	BQL executes, errors are actionable, result paging/export matches the result, and editing the query does not show stale output as new.
import	Receipt/file/bank entrypoints and validation are usable. Actual imports and bank linking need the applicable authorization; inspect previews without booking real transactions.
settings	Available profile/ledger/settings screens render; owner/member/public controls match permissions. Inspect billing, credentials and collaborators without changing them.
accessibility	Keyboard access, dialogs/focus return, accessible primary controls, narrow screens, theme and locale changes. Confirm names with the accessibility tree.

After a surprising state, capture the screenshot, console errors, request status, and operation response (sanitized). Use page.evaluate with fetch for an authenticated API control when appropriate, returning only relevant non-secret fields. Do not export an unfiltered HAR containing login bodies or cookies. Reproduce once from a fresh page load before retaining a finding.

Research map

• Thin route wrappers: dashboard/src/routes/; implementation: dashboard/src/features/<feature>/.
• Cross-feature infrastructure: dashboard/src/common/, particularly Apollo links, auth, filters, navigation and ledger layout.
• Report behavior and exports: features/reports/CLAUDE.md; importer and ledger-data have their own nested guides. ADRs are in docs/adrs/.
• Shared server causes: read backend-cluster/backend-v2/CLAUDE.md before tracing src/features/<domain>/api, services, schemas, and ledger adapters. Read the ledger package guide if the response originates there. Keep fixes assigned to the package that owns the defect; do not create cross-package imports.

Apply the contract's reproduce → research → dedupe → report/optional filing steps. Finding a bug does not require changing product code or adding a test that merely repeats the implementation. If the user also requests a fix, reproduce it with a meaningful regression check and run the owning package's normal gates.