| name | e2e |
| description | Use for real UI E2E smoke/regression proof across web, Flutter, browser, Chrome, or device flows, with artifacts. |
E2E
Run real UI only. Unit tests, typechecks, static scans, and curl can support the run, but never count as E2E proof.
Default to auto-full-safe: infer the target, run the fullest non-destructive coverage, capture proof, and ask only for missing auth, target, data mode, or risky side effects.
Load
Read the focused reference that matches the task:
references/defaults.md for underspecified runs, onboarding questions, scope, and regression gates
references/project-pack.md before the first run in a repo, and before reusing saved auth/flows/log commands
references/browser-first.md before choosing Browser, Chrome, device tooling, or standalone Playwright
references/capture-artifacts.md before running/delegating flows or judging evidence
references/runbook.md for plans, runner prompts, triage, fixes, and final reports
references/dogfood.md for dogfood fixtures and artifact checks
Rules
- Use Codex Browser first for local/public web when callable; use Chrome extension for signed-in browser state when available; use Flutter/device tooling for native flows; standalone Playwright is last resort or CI artifact work
- Stop only the failed UI driver after one denied/non-profile-lock probe; continue through
references/browser-first.md fallbacks before accepting missing visual evidence
- Before first-run setup or saved auth/flow reuse, check the project pack when the helper script is available; saved auth reuse still needs an automated E2E command
- Never mark resolved from screenshots alone; reproduce, patch only after cause/ripple checks, and rerun impacted E2E plus regression checks
- In auto mode, fix actionable click-time violations within risk limits, rerun that step/flow, and continue; stop only for unsafe side effects or unclear ownership
- Every checked action needs UI evidence: action event, settled assertion, screenshot or video frame, and artifact path
- Every checked flow must use or leave a runnable automated E2E command; manual-only runs are incomplete
- Confirm data mode before running flows when it is unknown: mock/seeded test data is default; prod data requires explicit approval and must be read-only unless exact writes are separately approved
- No prod writes/deletes, email/SMS, payments, or sharing unless explicitly approved
- UI E2E requires desktop and mobile 2x video artifacts so exact appearance is reviewable; native phone counts as mobile
- Capture by default:
events.jsonl, desktop+mobile cursor/click videos, step screenshots, supported logs/traces, and desktop+mobile 2x cursor recaps. If video is blocked after driver fallbacks, report it as an artifact limit instead of a clean pass
- Artifacts go under
docs/e2e/<RUN_ID>/; never overwrite prior runs and never count zero UI calls as a pass
- Keep plans short; split flows across agents only when independent
Runbook
Read references/runbook.md after the relevant policy references above.
Exit
Final output: report path, driver used/fallbacks, flow/step/action totals, artifact links, fixes, unresolved issues, regression commands, skipped checks and why.