Run and triage

Running (only on request)

Do not start E2E runs on your own. When the user asks:

1. Pick the narrowest scope: one page folder, or one case with -g "<ID>".
2. Prefer running through qa-kit: it picks the package script or Playwright project from the configured scopes (qa-kit config get e2e.scopes), captures a JSON report, and updates history:

qa-kit run --dry-run -- e2e/billing/invoices            # show the resolved command first
qa-kit run -- e2e/billing/invoices -g "INV-04"          # script/project chosen from scopes
qa-kit run --script test:e2e:admin -- users             # or name the package script
qa-kit run -- --project=chromium tests/checkout         # or pass Playwright flags directly

Any other Playwright command also works; then record results with qa-kit ingest (the Claude Code plugin does this automatically after the command). 3. Never set the env vars listed in e2e.protectedEnv (data-mutation or production switches), never change the target environment, never print or copy credentials or session files. If a test needs a switch the user has not set, report that it was skipped and why. 4. A new or changed spec runs twice; both results must match.

Read what happened

qa-kit history --filter failing      # current failures with last error
qa-kit history --filter flaky        # status flips over the last runs (p F ~ s)
qa-kit memory search "<test ID or page>" --kind bug
qa-kit cases --id <ID>               # expectation, known issue, catalog finding

For each failure open the error and the screenshot/trace in the project's output dir (qa-kit config get e2e.outputDir; error-context.md holds the error and an accessibility snapshot).

Classify (decide before changing anything)

Never "fix" an app bug by weakening an assertion, adding skip/fixme, or copying the current wrong behaviour into the expectation.

Report

• Group confirmed app bugs per feature, not per test: test IDs, run command, environment and role (no passwords), expected vs actual, evidence path. Use a tracker MCP if the user wants it filed, otherwise give a ready-to-paste report.
• Update the page catalog: Automation column (e.g. failing: bug, in the catalog's language) and Findings with run date and environment.

Learn (mandatory)

qa-kit memory add --kind bug   --scope <page dir> "<IDs>: <bug>, confirmed <date> on <env>, issue <ID>"
qa-kit memory add --kind flaky --scope <page dir> "<ID> flips because <cause>; fixed by <change>"
qa-kit memory add --kind gotcha "<environment lesson for this app>"
qa-kit memory add --global --kind gotcha "<lesson true for every app, e.g. a Playwright behaviour>"

If an earlier memory entry was wrong (e.g. a "bug" that was a permission problem), retire it: qa-kit memory forget <ID> --reason "<why>" or add the corrected entry with --supersedes <ID>.