Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
Desktop App — headful journey tests
Status snapshot of the optional Playwright QA recorder journeys for the
Evenfire Desktop App. These drive the real Electron window headfully (visible UI,
recorded WebM + PNG) against a live example-dev port-forward. They are separate
from the CI Playwright suite and never run in CI by default.
Recorder runs are local and opt-in. Configure the exact QA identity in
.env.qa-recorder, use only approved development endpoints, and enable a
QA_RECORDER_CONFIRM_* flag only for the side effect you intend to exercise.
Recordings stay under the git-ignored .local-notes/qa-recorder/ directory and
must be reviewed before sharing.
Coverage
Every Desktop App user journey has at least one headful spec. Each uses the exact
QA identity from .env.qa-recorder (no first-available fallback), shares launch /
login / proof plumbing in qa-recorder-helpers.ts, and records a WebM + PNG per
test under .local-notes/qa-recorder/runs/desktop-app/ (git-ignored).
| Journey | Spec | Tests | Confirm flag | Status |
|---|---|---|---|---|
| Auth & session | qa-recorder-auth.spec.ts |
3 — sign-in screen, authenticated shell, logout | — | ✅ |
| Navigation | qa-recorder-navigation.spec.ts |
3 — primary sidebar, footer settings menu, resources pages | — | ✅ |
| Agents | qa-recorder-agents.spec.ts |
3 — fleet, chat workspace, workspace routes | — | ✅ |
| Connectors | qa-recorder-connectors.spec.ts |
1 — MCP/connector health inventory | — | ✅ |
| Shared files | qa-recorder-shared-files.spec.ts |
1 — filesystem + directory browser | — | ✅ |
| Plugins & workflows | qa-recorder-plugins.spec.ts |
1 — inventory, detail, runs/artifacts (+optional trigger) | QA_RECORDER_CONFIRM_MUTATIONS (trigger only) |
✅ |
| Apps | qa-recorder-apps.spec.ts |
1 — catalog + embedded session | — | ✅ |
| Settings | qa-recorder-settings.spec.ts |
2 — appearance/notifications, information/configuration (+optional endpoint switch) | QA_RECORDER_CONFIRM_MUTATIONS (switch only) |
✅ |
| Inbox & search | qa-recorder-inbox-search.spec.ts |
2 — inbox, global search | — | ✅ |
| Chat | qa-recorder-chat.spec.ts |
1 — composer, thread, task progress | QA_RECORDER_CONFIRM_CHAT (required) |
✅ |
| Settings + chat smoke | qa-recorder-settings-chat.spec.ts |
1 — Settings tabs + one chat message | QA_RECORDER_CONFIRM_CHAT (required) |
✅ |
| Image capabilities | qa-recorder-image-capabilities.spec.ts |
2 — capability gate on a text-only model, fixture-backed image answer | QA_RECORDER_CONFIRM_CHAT (required) |
✅ |
| Model selector | qa-recorder-model-selector.spec.ts |
1 — a multi-model catalog lists rows without capability tags | — | ✅ |
| Document upload | qa-recorder-document-upload.spec.ts |
1 — attached text file read through the read tool (fixture lane only); a small file is answered with its whole-file digest, and a 6 MiB file with the digest of the first page the Host delivered (byteRange and truncated: true from the fixture ledger) |
IMAGE_CAPABILITIES_RUN_ID (runner-set) |
✅ |
Totals: 14 specs, 23 tests. Read-only journeys need no confirmation flag; every
journey still calls the loopback health guard on both API URLs — except against a
non-loopback target, where QA_RECORDER_ALLOW_REMOTE=1 skips the precheck, so a
wrong URL surfaces as an Electron timeout instead of a fast refusal. Mutating / paid
steps (chat message, workflow trigger, endpoint switch) run only under their
QA_RECORDER_CONFIRM_* flag and are otherwise skipped — the read-only part of the
journey always runs.
Run
cd desktop-app
# dev Evenfire must be closed (process-wide single-instance lock)
# keep control-api / external-rest-api / rpc-proxy port-forwarded to loopback
# (standard single minikube profile: make -C .. minikube-pf-desktop,
# fixed ports 8090 / 8091 / 8094)
npm run qa:recorder:navigation # one journey
npm run qa:recorder:all # all 14 specs / 23 tests
On a branch-owned profile the forwards use profile-owned ports instead of 8091 / 8094. Hold them on the host and point the recorder at them:
MINIKUBE_PROFILE=<owned-profile> \
make -f scripts/minikube-profiles/branch.mk branch-profile-pf # from the repo root
MINIKUBE_PROFILE=<owned-profile> \
make -f scripts/minikube-profiles/branch.mk branch-profile-info # prints the ports
cd desktop-app
EXTERNAL_REST_API_BASE_URL=http://127.0.0.1:<external-rest-api-port> \
RPC_PROXY_BASE_URL=http://127.0.0.1:<rpc-proxy-port> \
npm run qa:recorder:navigation
qa:recorder:all includes qa-recorder-model-selector.spec.ts, which fails loudly
against a catalog of fewer than two models. Run it against an environment whose
catalog has several, or run the other journeys individually.
Proof artifacts (local only, git-ignored)
find .local-notes/qa-recorder/runs/desktop-app -name '*.webm' # WebM recordings
find .local-notes/qa-recorder/runs/desktop-app -name '*.png' # full-page screenshots
Videos may contain emails, agent names, endpoint URLs, prompts, or model responses and must stay local unless reviewed and shared intentionally.
Last verified
2026-07-17 against example-dev (GKE) port-forwards — the then-current suite
passed with WebM + PNG proof for every test. This is a point-in-time snapshot; re-run
npm run qa:recorder:all to refresh.