Auto_Bangumi/docs/superpowers/plans/2026-07-10-issue-driven-e2e-implementation-plan.md

Issue-Driven Hermetic E2E Implementation Plan

Status

Ready for execution. This plan implements the approved 2026-07-10-issue-driven-e2e-test-design.md.

Objective

Replace the ordered, stateful E2E workflow with deterministic runtime, real-browser, real-qBittorrent, and packaged-image checks. Fix the two current production regressions exposed by issues #1075 and #1072, make recent issue fixes executable as acceptance tests, and add four parallel pull-request checks whose p95 wall-clock time stays below 15 minutes.

Guardrails

• Preserve unrelated working-tree files. In particular, never stage backend/uv.lock, AGENTS.md, or PRODUCT.md unless their owner explicitly asks for it.
• Commit the already-verified authentication implementation separately before beginning E2E work; do not mix it into issue or test-infrastructure commits.
• Do not add a production test-only endpoint. Tests use public APIs, mock service admin APIs, or a database fixture written before app startup.
• During test execution, AutoBangumi and qBittorrent run on a Docker internal network. Playwright additionally rejects requests outside the dynamically assigned local origins.
• No fixed host port, fixed container name, fixed Compose project name, fixed sleep, test retry, numbered test name, or cross-test state dictionary is allowed.
• A required test must have a positive or negative control so a broad suppression cannot make it pass accidentally.
• Each task uses a red/green cycle: add the narrow failing assertion, confirm the expected failure, implement the minimum behavior, then run the targeted and adjacent suites.

Commit Sequence

1. Existing authentication review implementation.
2. fix: preserve tmdb language and versioned episodes.
3. fix: make release classification deterministic.
4. test: add hermetic e2e support services.
5. test: replace ordered api e2e coverage.
6. test: add browser issue regressions.
7. test: exercise real qbittorrent and release images.
8. ci: run hermetic e2e checks.

The sequence keeps production fixes, reusable infrastructure, scenarios, and CI policy independently reviewable.

Task 1: Stabilize the Existing Authentication Work

Files

• Existing modified files under backend/src/module/{api,application,database,mcp,models,ports,security}.
• Existing modified and new auth tests under backend/src/test.
• Existing modified files under webui/src/{api,components,hooks,store,utils} and webui/types.
• docs/superpowers/specs/2026-07-10-auth-review-fixes-design.md.

Steps

1. Re-run the focused backend auth, migration, and MCP suites and the focused WebUI auth/access suites against the current working tree.
2. Run backend Ruff, Black check, and mypy on the changed Python files.
3. Run WebUI type checking, ESLint, Prettier check, and the full Vitest suite.
4. Stage only the known authentication implementation, tests, and updated auth design. Explicitly inspect git diff --cached --name-only before commit.
5. Commit the authentication work without staging the three protected local files listed in Guardrails.

Verification

cd backend && uv run pytest \
  src/test/test_api_auth.py \
  src/test/test_api_passkey.py \
  src/test/test_api_users.py \
  src/test/test_auth.py \
  src/test/test_auth_persistence.py \
  src/test/test_auth_service_transactions.py \
  src/test/test_mcp_security.py \
  src/test/test_migrations_module.py \
  src/test/test_session_store.py \
  src/test/test_setup.py -q
cd backend && uv run pytest -q
cd backend && uv run ruff check src
cd backend && uv run mypy src
cd webui && pnpm test --run
cd webui && pnpm test:build
cd webui && pnpm lint
cd webui && pnpm format:check

Task 2: Fix TMDB Search Localization (#1075)

Files

• Modify backend/src/test/test_tmdb.py.
• Modify backend/src/module/parser/analyser/tmdb_parser.py.

Steps

1. Add table-driven URL tests for TV and movie searches in zh, jp, and en. Parse the URL and assert exact language, query, api_key, page=1, and include_adult=false values rather than comparing an encoded string manually.
2. Add async tests proving both the whitespace-removal retry and the TV-to-movie fallback preserve the requested language.
3. Change search_url and search_movie_url to accept language and construct query strings through urllib.parse.urlencode.
4. Thread language through _search_movie and every initial, retry, and fallback call. Keep the existing language-bearing cache key.
5. Run TMDB, RSS analyser, searcher, and config-reload tests to catch signature and cache regressions.

Verification

cd backend && uv run pytest \
  src/test/test_tmdb.py \
  src/test/test_searcher.py \
  src/test/test_config.py -q
cd backend && uv run ruff check src/module/parser/analyser/tmdb_parser.py src/test/test_tmdb.py
cd backend && uv run black --check src/module/parser/analyser/tmdb_parser.py src/test/test_tmdb.py

Task 3: Parse Versioned Episodes Before CRC Tokens (#1072)

Files

• Modify backend/src/test/test_torrent_parser.py.
• Modify backend/src/test/test_offset_scanner.py.
• Modify backend/src/module/parser/analyser/raw_parser.py.

Steps

1. Add the exact reported title as a parser regression:

   [SubsPlease] Haibara-kun no Tsuyokute Seishun New Game - 06v2 (1080p) [E931DD98].mkv
   

   Assert season 1, episode 6, group SubsPlease, and episode type episode.

2. Add controls for plain 06, bracketed 06v2, 13v2 plus a CRC token, explicit E13, and a title with no episode signal.

3. Update the regular expression so an unbracketed numeric episode may carry a one- or two-digit vN suffix. The suffix belongs to the episode token and must be matched before a later [Ehhhhhhhh] CRC-like group.

4. Give explicit E/EP episode tokens an alphanumeric boundary so a CRC such as E931DD98 cannot be accepted as an episode, while E13 and EP13 controls continue to parse.

5. Add scanner tests with a real temporary database for the exact E6/CRC title, a true-positive E13 against a 12-episode TMDB season, and an unparseable no-signal title. Assert both the boolean result and persisted review fields.

6. Run all parser, offset detector, offset scanner, and issue regression tests.

Verification

cd backend && uv run pytest \
  src/test/test_torrent_parser.py \
  src/test/test_title_parser.py \
  src/test/test_offset_scanner.py \
  src/test/test_issue_bugs.py -q
cd backend && uv run ruff check src/module/parser/analyser/raw_parser.py \
  src/test/test_torrent_parser.py src/test/test_offset_scanner.py

Task 4: Extract and Test Release Classification (#1065)

Files

• Create scripts/classify_release.py.
• Create backend/src/test/test_release_classification.py.
• Modify .github/workflows/build.yml.

Steps

1. Put classification in a pure Python function accepting event name, ref type, ref name, head ref, PR title, and whether the tagged commit is on main. The PR title is accepted only to prove it has no effect on version selection.
2. Cover normal branch pushes, a PR titled/headed Doc update, tag Doc-update, a stable tag on and off main, and X.Y.Z-beta.N.
3. Make the CLI print release, dev, build_test, and version as GitHub output lines. A stable tag off main exits non-zero with a specific error.
4. Replace the embedded shell branching in version-info with the script while retaining the current job outputs and downstream job interface.
5. Confirm a non-SemVer PR title can never become VERSION.

Verification

cd backend && uv run pytest src/test/test_release_classification.py -q
cd backend && uv run ruff check ../scripts/classify_release.py \
  src/test/test_release_classification.py
python3 scripts/classify_release.py --event-name pull_request --ref-type branch \
  --ref-name main --pr-head Doc-update --pr-title "3.3.4"

Task 5: Build Deterministic Fixture Assets

Files

• Create e2e/scripts/generate_fixtures.py.
• Create generated e2e/fixtures/files/tiny-media.mkv.
• Create generated e2e/fixtures/torrents/tiny-media.torrent.
• Create RSS fixtures under e2e/fixtures/rss.
• Create TMDB response fixtures under e2e/fixtures/tmdb.
• Create a small local player document under e2e/fixtures/player/index.html.
• Create backend/src/test/test_e2e_fixture_generation.py.

Steps

1. Generate a deterministic 64 KiB payload from a fixed byte pattern; do not store random output.
2. Implement the small bencode encoder needed for a single-file torrent using only the Python standard library. Set piece length, SHA-1 pieces, filename, length, and url-list deterministically.
3. Use the internal URL http://mock-upstream:18888/files/tiny-media.mkv as the web seed and an RSS enclosure URL on the same service for the torrent metainfo.
4. Check in SHA-256 expectations for both generated files and make the unit test regenerate into a temporary directory and compare bytes.
5. Add RSS cases for normal download, duplicate refresh, #1072, and localized TV/movie searches. TMDB fixtures use poster_path: null, which proves the parser never reaches its hard-coded public image host during E2E. Fixtures contain no public URL.

Verification

python e2e/scripts/generate_fixtures.py --check
cd backend && uv run pytest src/test/test_e2e_fixture_generation.py -q
rg -n 'https?://' e2e/fixtures

The URL scan may match only the two approved internal service URLs and the local player fixture's loopback placeholder.

Task 6: Add Mock-Upstream and Fake-qB Services

Files

• Create e2e/mock-upstream/server.py and e2e/mock-upstream/Dockerfile.
• Create e2e/fake-qb/server.py and e2e/fake-qb/Dockerfile.
• Create backend/src/test/e2e/runtime/test_support_services.py.

Steps

1. Implement both services with the Python standard library so their images do not download packages at build time. Pin each Python base image with both a patch tag and immutable digest.
2. Mock-upstream serves /health, /rss/*, /tmdb/*, /images/*, /notifications/*, /files/*, /torrents/*, and /player/*.
3. File and torrent responses implement HEAD and single-byte-range GET, returning correct Accept-Ranges, Content-Range, Content-Length, 206, and 416 semantics. This is required by qBittorrent's web-seed client.
4. Its /__admin/reset selects a named scenario and clears the request journal; /__admin/requests and /__admin/state return recorded data. Unknown routes return 501 and remain in the journal.
5. Fake-qB implements the qB Web API subset AutoBangumi uses: login/logout, application preferences/version, torrent info, and torrent delete. Admin state selects Windows-path, neighbor-path, and forced-delete-failure cases.
6. Redact authorization, cookie, password, and token values before journaling.
7. Contract-test health, range requests, scenario reset, exact query/form capture, unknown-route failure, and redaction for both services.

Verification

cd backend && uv run pytest \
  src/test/e2e/runtime/test_support_services.py -m e2e -q
cd backend && uv run ruff check ../e2e/mock-upstream/server.py \
  ../e2e/fake-qb/server.py src/test/e2e/runtime/test_support_services.py

Task 7: Add the Hermetic Stack and Test Controls

Files

• Create e2e/compose/browser.yml.
• Create e2e/compose/downloader.yml.
• Create e2e/fixtures/qbittorrent/qBittorrent.conf.
• Create e2e/scripts/build_test_image.py.
• Create e2e/scripts/stack.py.
• Create e2e/scripts/collect_artifacts.py.
• Create e2e/scripts/audit_sources.py.
• Create backend/src/test/e2e/support/{client.py,database.py,poll.py,run_offset_scan.py,stack.py}.
• Replace backend/src/test/e2e/conftest.py with isolated fixtures.

Steps

1. Build the WebUI, then assemble a temporary Docker build context containing the tracked backend plus the built dist and a generated module/__version__.py set to 3.3.999-e2e.1. Do not write either generated artifact into the working tree. Build the native app image and assert /app/IMAGE_VERSION and the Python version module use the same value.
2. browser.yml starts the app, mock-upstream, and fake-qB on one internal network. Publish app and admin ports dynamically on 127.0.0.1. stack.py creates temporary bind-mounted config, data, download, and log directories so tests can seed a database before startup and retain failure state after containers stop.
3. downloader.yml overlays real qBittorrent and shared download volumes. The image reference contains a patch tag and immutable sha256 digest; a stack validation command rejects floating references.
4. Mount a fixed, test-only qB configuration. Disable DHT, PeX, LSD, UPnP, NAT traversal, RSS, and update checks.
5. stack.py creates a unique project name, starts services with health waits, discovers published ports, emits an environment file, restarts one service, and tears down volumes in finally/signal handlers.
6. collect_artifacts.py records Compose status/logs, service journals, qB state, backend logs, and a copy of the failed SQLite database with secret values redacted where output is textual.
7. Support polling helpers accept a deadline and include the last observed state in assertion messages. They never sleep without rechecking a bounded condition.
8. The pytest fixtures consume AB_E2E_BASE_URL, mock/fake admin URLs, Compose files, project name, and artifact directory. The environment file exposes separate internal-service and dynamically published loopback mock URLs so the backend and browser never need to resolve the other's address. Fixtures create unique resource prefixes and an idempotent baseline administrator without relying on test definition order.
9. Provide two app fixture types: a dedicated fresh stack for first-run and pre-start database scenarios, and one baseline stack for independent CRUD scenarios. A bootstrap test never shares its config/data directory with a baseline test. A database-seeded scenario first completes Setup with the mock downloader, stops the app, writes fixtures into the bind-mounted current-schema SQLite database, runs any one-shot service operation, and restarts the app. It never pre-creates the database before the initial Setup boot.
10. audit_sources.py enforces the public-host allowlist, immutable image references, absence of fixed ports/container names and Playwright retries, and absence of blind waits outside the bounded polling helpers.

Verification

python e2e/scripts/build_test_image.py --version 3.3.999-e2e.1
python e2e/scripts/stack.py validate --profile browser
python e2e/scripts/stack.py validate --profile downloader
python e2e/scripts/stack.py smoke --profile browser

Task 8: Replace the Ordered Runtime E2E Suite

Files

• Create backend/src/test/e2e/runtime/test_bootstrap_auth.py.
• Create backend/src/test/e2e/runtime/test_session_access.py.
• Create backend/src/test/e2e/runtime/test_restart_persistence.py.
• Create backend/src/test/e2e/runtime/test_tmdb_language.py.
• Create backend/src/test/e2e/runtime/test_offset_review.py.
• Create backend/src/test/e2e/runtime/test_windows_path_delete.py.
• Modify backend/pyproject.toml marker declarations if lane-specific markers are introduced.

Steps

1. Bootstrap/auth runs on its dedicated fresh stack and verifies first-run setup, exact { "authenticated": true } bodies, absence of token material, HttpOnly and SameSite session cookie attributes, POST refresh, wrong-password failure, logout, and deterministic 401 responses without development bypass.
2. Session/access creates a secondary user so primary fixture credentials do not mutate. Use two cookie jars, update credentials, and prove old sessions and password fail while the replacement session works.
3. Create API and MCP tokens; prove one-time secret display, prefix-only list responses, scope isolation, Authorization precedence, account-control denial for Bearer credentials, and immediate revocation.
4. Restart persistence records user, session, tokens, RSS, and Bangumi IDs, restarts only the app service, and checks that valid data remains while revoked credentials stay revoked.
5. TMDB language selects mock scenarios and exercises real analysis for TV, movie, whitespace retry, and same-keyword language changes. Assert both the user-facing result and upstream journal.
6. Offset review completes Setup on a dedicated fresh stack, stops the app, seeds the real database, runs run_offset_scan.py against that work directory, and restarts the app. It covers E6/CRC false-positive, E13 true-positive, and no-signal control cases.
7. Windows delete points the configured downloader at fake-qB. Exercise both delete and disable through the public API, assert exact selected hash and deleteFiles, check database outcomes, and make forced qB failure visible.
8. Every test establishes its own preconditions and uses unique resource names; test order can be reversed without changing results.

Verification

python e2e/scripts/stack.py run --profile browser -- \
  uv run --directory backend pytest src/test/e2e/runtime -m e2e -q

As an isolation audit, run every collected runtime node ID once by itself and run modules in reverse order through the stack runner. Do not add a random-order dependency solely for this check.

Task 9: Install and Configure Playwright for Hermetic Test Runtime

Files

• Modify webui/package.json and webui/pnpm-lock.yaml.
• Modify webui/unocss.config.ts.
• Create webui/playwright.config.ts.
• Create webui/tsconfig.e2e.json.
• Create webui/e2e/fixtures/{api.ts,auth.ts,network.ts,test.ts}.
• Create webui/e2e/setup/{bootstrap-chromium.setup.ts,bootstrap-webkit.setup.ts}.
• Create webui/e2e/specs/production-smoke.spec.ts.
• Modify .gitignore for Playwright output directories.

Steps

1. Use pnpm add -D --save-exact for @playwright/test and @iconify-json/simple-icons, recording exact versions in both package and lock files.
2. Remove the UnoCSS https://esm.sh/ CDN fallback and load simple-icons from the installed JSON package. Prove pnpm build succeeds with outbound access blocked.
3. Define separate setup-chromium/chromium-desktop and setup-webkit/webkit-mobile dependency pairs plus firefox-nightly. WebKit uses a 390×844 iPhone-sized viewport. PR jobs select one pair; nightly selects Firefox. Set locale to en-US, one worker, zero retries, serviceWorkers: 'block', trace/video/screenshot on failure, and a base URL supplied by AB_E2E_BASE_URL. Use AB_E2E_PROFILE and spec tags so normal browser jobs exclude the @downloader scenario, the downloader job selects it explicitly, and desktop projects exclude the @mobile geometry scenario.
4. Each setup project completes the real wizard through accessible UI controls, validates the response contract and cookie flags, then writes its own browser-specific storage state. Chromium and WebKit never share one auth file. The browser profile configures fake-qB; the downloader profile uses the real internal qB endpoint. The auth path includes both browser and stack profile, so a downloader invocation never reuses browser-profile state.
5. API fixtures create secondary users and unique resources through public endpoints. Network fixtures allow only the app, mock admin, and fake-qB origins; an unexpected request fails the test with method and URL.
6. Capture browser console errors and failed responses into per-test artifact files. Redact secrets before writing logs.
7. Add a production smoke that checks /health, /#/setup, and hashed static assets before any feature spec runs. This catches a missing version module or dist staging error directly.

Verification

cd webui && pnpm install --frozen-lockfile
cd webui && pnpm build
cd webui && pnpm exec playwright test --list
cd webui && pnpm test:e2e:typecheck
cd webui && pnpm test:build
cd webui && pnpm lint
cd webui && pnpm format:check

Task 10: Add Browser Authentication and Secret-Lifetime Scenarios

Files

• Create webui/e2e/specs/bootstrap-auth.spec.ts.
• Create webui/e2e/specs/session-access.spec.ts.
• Create webui/e2e/specs/token-secret.spec.ts.
• Add accessible labels only where needed in webui/src/components/setup/wizard-container.vue, webui/src/components/setup/wizard-step-*.vue, webui/src/components/setting/config-access.vue.

Steps

1. The browser-specific setup project is the sole owner of first-run Setup. bootstrap-auth.spec.ts starts from its resulting account and covers login response secrecy, cookie persistence across reload, wrong password, logout, and browser-back protection. It never attempts Setup a second time.
2. Session access uses two independent contexts and a secondary user to prove credential updates invalidate prior sessions without destabilizing the baseline admin.
3. Token secret creates a token in the Access panel and confirms the full value appears only in the reveal modal while the list exposes only its prefix.
4. Close the reveal, navigate away, unmount the component, and return; the secret must never reappear.
5. Delay the real create-token response with Playwright routing, navigate away, then continue the request. Assert the late promise does not restore secret UI or sensitive component state.
6. Setup fields, account/token rows, and dialogs use roles, labels, and accessible names; no test ID is added for these flows.

Verification

python e2e/scripts/stack.py run --profile browser -- \
  pnpm --dir webui exec playwright test \
  bootstrap-auth.spec.ts session-access.spec.ts token-secret.spec.ts \
  --project=chromium-desktop
python e2e/scripts/stack.py run --profile browser -- \
  pnpm --dir webui exec playwright test \
  bootstrap-auth.spec.ts token-secret.spec.ts --project=webkit-mobile

Task 11: Add Browser Regressions for Recent UI and Parser Issues

Files

• Create webui/e2e/specs/tmdb-language.spec.ts.
• Create webui/e2e/specs/offset-review.spec.ts.
• Create webui/e2e/specs/mobile-rule-editor.spec.ts.
• Create webui/e2e/specs/player-local-settings.spec.ts.
• Modify webui/src/components/basic/ab-bottom-sheet.vue.
• Modify webui/src/components/bangumi-preview.vue.
• Modify webui/src/components/ab-edit-rule.vue, webui/src/components/ab-add-rss.vue, webui/src/components/basic/ab-search.vue, webui/src/components/search/ab-search-modal.vue, webui/src/components/setting/config-player.vue, webui/src/pages/index/config.vue, and webui/src/pages/index/player.vue.

Steps

1. TMDB language changes parser language from zh to jp, searches the same keyword, and proves TV, movie, and whitespace retry calls use the expected language without cross-language cache reuse.
2. Offset review creates two Bangumi through the real analysis/subscribe APIs, then PATCHes their complete public update payloads so the E13 rule is marked for review and the exact 06v2/E931 rule is not. The UI displays a warning only for the true mismatch; runtime tests remain responsible for proving how those persisted states are calculated.
3. Mobile rule editor runs at 390×844 in WebKit. Assert full-screen panel geometry, focus the title, contract viewport height, and prove the panel is not translated. Edit, save, reload, and assert persistence.
4. Player settings assert immediate localStorage writes, no config-update or restart request, same-context reload persistence, new-context defaults, and an unchanged iframe URL loading the dynamically published loopback /player/index.html fixture marker. Give the iframe a localized accessible title.
5. Name the bottom-sheet dialog/title and add the single geometry-only data-testid="bottom-sheet-panel". Add accessible labels to Bangumi and RSS fields, named regions to config sections, a named search dialog/textbox/list, and a localized iframe title. Correct any control whose explicit ARIA role masks its native button semantics. All other locators use roles, labels, and visible names.

Verification

python e2e/scripts/stack.py run --profile browser -- \
  pnpm --dir webui exec playwright test \
  tmdb-language.spec.ts offset-review.spec.ts player-local-settings.spec.ts \
  --project=chromium-desktop
python e2e/scripts/stack.py run --profile browser -- \
  pnpm --dir webui exec playwright test \
  mobile-rule-editor.spec.ts player-local-settings.spec.ts \
  --project=webkit-mobile

Task 12: Exercise Real qBittorrent and Packaged Images

Files

• Create webui/e2e/specs/rss-download-rename.spec.ts.
• Create backend/src/test/e2e/downloader/test_release_image_smoke.py.
• Create backend/src/test/e2e/downloader/test_qb_image_contract.py.
• Modify e2e/fixtures/qbittorrent/qBittorrent.conf.
• Modify e2e/fixtures/rss/tiny-download.xml.

Steps

1. Validate the pinned qB image starts with the mounted credentials and all peer-discovery/update features disabled. Resolve one supported LinuxServer image once, record the complete <patch-tag>@sha256:<digest> reference in Compose, and make the image-contract test compare that exact shape.
2. Complete Setup with the internal qB endpoint and shared download path. This occurs in the downloader invocation's dedicated setup project and fresh stack; the feature spec consumes its profile-specific storage state and does not run Setup again.
3. Add the local RSS, call the real refresh endpoint, and poll qB until the web-seed torrent is complete. No tracker, DHT, or public peer may appear in qB state.
4. Poll AutoBangumi API/database/filesystem until the media file is renamed and organized. Assert the expected season and episode in both API and path.
5. Refresh the same RSS again and prove the torrent/task count does not grow.
6. Open the WebUI and prove completed state is visible, then delete with files enabled and verify qB state, AutoBangumi state, and the shared file are gone.
7. Build the native AutoBangumi image with 3.3.999-e2e.1, start it with fresh volumes, and assert /health remains healthy, db_ok=true, and the health version equals docker exec ... cat /app/IMAGE_VERSION.
8. The image test must inspect the locally built image ID and fail if Docker attempts to pull latest or any public AutoBangumi image.

Verification

python e2e/scripts/stack.py run --profile downloader -- \
  pnpm --dir webui exec playwright test rss-download-rename.spec.ts \
  --project=chromium-desktop
python e2e/scripts/stack.py run --profile downloader -- \
  uv run --directory backend pytest src/test/e2e/downloader -m e2e -q

Task 13: Add Pull-Request and Nightly Workflows

Files

• Create .github/workflows/e2e.yml.
• Create .github/workflows/e2e-nightly.yml.
• Modify .github/workflows/build.yml.
• Create e2e/README.md.

Steps

1. Add e2e-runtime, e2e-browser (chromium), e2e-browser (webkit-mobile), and e2e-downloader-image jobs on every pull request without branch or path filters. Set timeout-minutes: 15 on each job.
2. Change the existing backend unit-test job to run pytest -m "not e2e". Keep the nested E2E skip guard as defense in depth so a normal pytest invocation cannot start Docker unexpectedly.
3. Cache uv, pnpm, Playwright browser binaries, frontend output, and Buildx layers with lockfile/source-sensitive keys. Install dependencies before starting the internal test network.
4. Give every job a unique Compose project and artifact directory derived from workflow, job, run ID, attempt, and matrix project.
5. Wrap execution so stack logs and failure artifacts upload with if: always() even when setup or tests fail. Use 14-day retention.
6. Run zero PR retries. The nightly workflow runs Firefox and repeated browser and downloader passes, uploads the same artifacts, and remains non-blocking.
7. Document exact local commands, prerequisites, expected duration, artifact locations, mock scenario authoring, fixture regeneration, and qB digest update procedure.
8. Keep branch protection unchanged during ten shadow PR runs. Record run URLs, result, and wall-clock duration; after 10/10 passes and p95 below 15 minutes, make the four named checks required in repository settings.

Verification

rg -n 'timeout-minutes: 15|retention-days: 14' \
  .github/workflows/e2e.yml .github/workflows/e2e-nightly.yml
ruby -e "require 'yaml'; YAML.load_file('.github/workflows/e2e.yml')"
ruby -e "require 'yaml'; YAML.load_file('.github/workflows/e2e-nightly.yml')"

Task 14: Remove the Legacy Workflow and Run the Full Gate

Files

• Delete backend/src/test/e2e/test_e2e_workflow.py.
• Delete backend/src/test/e2e/docker-compose.test.yml.
• Delete backend/src/test/e2e/mock_rss_server.py.
• Delete backend/src/test/e2e/Dockerfile.mock-rss.
• Delete migrated files under backend/src/test/e2e/fixtures after confirming their only replacement is under root e2e/fixtures.
• Update backend/src/test/e2e/__init__.py to describe the new lane split.
• Update e2e/README.md with the final command names and artifact paths.

Steps

1. Map every meaningful old assertion to a unit, runtime, browser, downloader, or image test. Preserve documented compatibility assertions; discard only order checks, permissive (200, 401) checks, and no-op assertions that do not describe a contract.
2. Remove the old session-scoped process, fixed ports/container names, e2e_state, and numbered methods.
3. Run audit_sources.py against the new tree for blind sleeps, public upstream hosts, floating image tags, raw secret logging, fixed ports/container names, and Playwright retries.
4. Intentionally trigger one mock 501, one browser assertion failure, and one qB timeout locally. Confirm the artifact bundle identifies the last observed state without an interactive rerun, then restore the assertions.
5. Run backend unit/runtime/downloader suites, WebUI unit/type/lint/build suites, both PR browser projects, and the native image smoke.
6. Ask a subagent to review the final diff for correctness, isolation, secret handling, flake sources, and divergence from the approved design. Fix every actionable finding and rerun the affected lanes.

Verification

python e2e/scripts/audit_sources.py
cd backend && uv run pytest -q
cd backend && uv run ruff check src ../e2e ../scripts/classify_release.py
cd backend && uv run mypy src
cd webui && pnpm test --run
cd webui && pnpm test:build
cd webui && pnpm lint
cd webui && pnpm format:check
python e2e/scripts/stack.py run --profile browser -- \
  uv run --directory backend pytest src/test/e2e/runtime -m e2e -q
python e2e/scripts/stack.py run --profile browser -- \
  pnpm --dir webui exec playwright test --project=chromium-desktop
python e2e/scripts/stack.py run --profile browser -- \
  pnpm --dir webui exec playwright test --project=webkit-mobile
python e2e/scripts/stack.py run --profile downloader -- \
  pnpm --dir webui exec playwright test rss-download-rename.spec.ts \
  --project=chromium-desktop
python e2e/scripts/stack.py run --profile downloader -- \
  uv run --directory backend pytest src/test/e2e/downloader -m e2e -q

Completion Criteria

• #1075 and #1072 are fixed with focused unit/integration regressions.
• #1069, #1068, #1067, #1065, and #1062 have deterministic acceptance tests.
• The old ordered E2E workflow and duplicate fixtures are gone.
• Runtime, Chromium, WebKit mobile, real-qB, and image tests pass locally.
• Test execution produces no public request and uses no floating container tag.
• Failure artifacts contain enough state to diagnose injected failures.
• The four PR jobs complete within 15 minutes and enter the ten-run shadow period before branch-protection enforcement.