workflow-e2e

name	workflow-e2e
description	Run Smelter real-interface E2E checks and capture concrete evidence for user-visible behavior.
version	0.55
type	workflow
consumes	src/** built
produces	artifacts/ (video, screenshots, logs, trace, io-samples)
default_pattern	A
default_agent	qa-tester
supports_patterns	["A"]
can_delegate_to	["e2e-testing-patterns"]
gate	{"postcondition":[{"artifacts_exist":true},{"build_clean":true},{"scoped_tests_pass":true},{"real_interface_invoked":true},{"no_interface_mocks":true},{"per_surface_artifact_present":true},{"effect_observed":true},{"e2e_infra_present":true}]}
surface_mapping	{"UI":"Playwright real browser","CLI":"subprocess stdin/argv/exit_code","HTTP_API":"real server + curl/fetch","Database":"real or in-process test DB","Hook":"stdin JSON -> stdout JSON"}
exempt_if_surface	["style","typography","typo","dialogue"]

workflow-e2e

Drive the real interface and save artifacts proving the target effect happened. Test-runner stdout alone is not E2E.

Announce: I'm using workflow-e2e to drive the real interface and capture artifacts proving the target effect materialized.

Surface Contract

Surface	Real interface	Required artifact
UI	Playwright real browser	video + screenshot + console log
CLI/script	real subprocess	argv/stdin/stdout/stderr/exit transcript
HTTP API	real server on a port	request/response log + follow-up observation
Database	real compatible DB engine	SQL log + SELECT/readback evidence
Hook	actual stdin/stdout pipe	input JSON + output JSON + exit code

No mocks for the interface under test. No dry-run. No handler-only calls when the real surface is browser, shell, HTTP, DB, or hook I/O.

Execution Checklist

1. Identify affected surfaces from $ARTIFACTS_DIR artifacts, diff, and plan.
2. Verify the required harness exists. If missing, record cause: e2e_infra_missing and route to workflow-coding to install/configure it; do not skip E2E.
3. Build/run prerequisite scoped tests.
4. Run the real scenario through the real interface.
5. Save artifacts under $ARTIFACTS_DIR/e2e/.
6. Assert both acknowledgement and effect. Ack-only signals (toast, 2xx, exit 0, banner) are insufficient unless the effect is a local UI toggle.
7. For UI layout/position requirements, assert the target element's bounding box against the viewport, then save the screenshot/video proving it.
8. Record scenario evidence with non-empty references.

Effect Evidence

• UI: DOM state or screenshot diff proving the target state changed; for placement, include bounding box + viewport assertion.
• API: follow-up GET/list query or DB SELECT confirming state.
• CLI: follow-up command or filesystem diff showing the result.
• DB: SELECT reads back the inserted/updated value.
• Hook: stdout JSON parses and matches expected shape.
• Local UI toggle: absence/presence is the effect; declare it explicitly.

UI Fidelity Rules

• Exercise the real affected route with real backend data. Do not replace it with debug routes, synthetic SVG/images, or mock fixtures unless the task is explicitly about that harness.
• Prefer the project's documented E2E command and .env.e2e setup when they exist; do not skip them for an ad-hoc Playwright script.
• Match the user's visual conditions for artifacts: viewport, browser, theme, zoom, and DPR/device scale factor. A 2x screenshot does not prove a 1x rendering issue is fixed.
• Validate custom measurement code against known positive and negative cases before using it as evidence. If the measurement transform erases the target defect, the result is invalid.
• Open and inspect screenshots/videos directly. Artifact existence, non-zero file size, and passing test output are not visual confirmation.
• Do not claim a subjective visual issue is fixed until the user confirms in their environment; report local evidence as "changed and locally observed" instead.

Credentials

If auth/API secrets are required, ask the user to populate .env.e2e with only needed E2E_* keys. Never read .env, echo secrets, fabricate credentials, or stub auth to pass.

Forbidden

• Claiming pass with zero artifacts.
• Treating unit/integration runner output as E2E.
• Using mocks/stubs for the exercised interface.
• Skipping because infra is absent.
• Using ack-only evidence for durable effects.

Terminal State

On pass, invoke workflow-e2e-review. On gate fail, route to the producer (workflow-coding) with cause and artifact evidence.