ci(workflow): use internal cache on self-hosted runners

[quanruzhuoxiu]

Summary

• route eligible Linux, macOS, and Windows jobs through repository runner variables, while retaining GitHub-hosted runners where KVM or interactive desktop constraints require them
• route cache operations to GitHub storage on hosted runners and the internal TOS cache service on self-hosted runners, following the web-infra-dev/rspack pattern
• cache the Node toolcache, pnpm store, Playwright dependencies, Puppeteer Chrome for Testing, and Electron artifacts
• replace repeated full browser/system setup with reusable setup actions and install only missing runtime dependencies
• use focused Nx builds, shallow checkouts where history is unnecessary, docs-only path filters, and cheaper trigger conditions
• remove duplicated report and extension AI scenarios while preserving dedicated AI coverage
• limit model-heavy platform jobs to one mobile lane and one desktop lane, and split Headless extension jobs into dependency chains so queued work does not occupy runners or exceed the shared model TPM quota
• make report validation select a successful Playwright retry instead of a failed first attempt when the suite ultimately passed

Timing results

All runner durations exclude queue time.

Job	Before	Cold cache	Hot cache
Chrome Extension	36:54	7:05	5:03
Bridge	17:42	9:57	8:01
Playground	19:18	14:13	12:56
Recorder	11:38	9:09	10:17
Settings	19:03	14:23	12:10
AI Todo	15:03	8:14	11:22
AI e2e-web	26:31	13:25	10:47
AI e2e-report	20:36	14:32	16:04
AI unit	16:04	13:49	11:18

Final validation on 7474a3c8d after model-quota scheduling:

Job	Actual runner time	End-to-end	Result
e2e-web	13:43	14:33	passed
e2e-report	10:46	11:36	passed
AI unit	15:50	16:39	passed
iOS	6:32	7:02	passed
Android	10:36	17:48	passed after iOS lane
macOS	14:15	14:20	passed
Windows	8:14	22:39	passed after macOS lane
Chrome Extension	5:32	7:29	passed
Recorder	11:35	12:36	passed
Settings	10:57	26:40	passed after Recorder
Bridge	11:08	22:10	passed after Chrome Extension
Playground	31:05	53:40	passed after Bridge

The serialized lanes intentionally trade some end-to-end latency for lower concurrent runner usage and model stability. In the unthrottled run, iOS, macOS, and Recorder all failed after three 429 ModelAccountTpmRateLimitExceeded responses. The throttled run completed 16/16 checks without a TPM failure. Playground remains the wall-clock bottleneck and its 31:05 runner time is model variance rather than queue time.

The e2e-web browser setup itself dropped from 4:06 to 0:23. A separate run had a 4:59 checkout outlier, so checkout, setup, and AI-test variance are reported separately rather than attributed to runner performance.

Validation

• pnpm run lint
• YAML parsing for all changed workflows and report scenarios
• pnpm exec nx run-many --target=build --projects="@midscene/report,@midscene/core,@midscene/web,@midscene/cli" --skip-nx-cache
• pnpm exec nx run-many --target=build --projects="@midscene/report,chrome-extension" --skip-nx-cache
• pnpm exec nx run-many --target=build --projects="@midscene/report,@midscene/computer" --skip-nx-cache
• pnpm exec nx test @midscene/report (15 files, 82 tests)
• NODE_OPTIONS=--max-old-space-size=8192 pnpm --dir packages/computer exec vitest run tests/unit-test (13 files, 88 tests)
• pnpm --dir packages/core exec vitest run tests/unit-test/yaml-player-output.test.ts
• validated report retry selection against the failed run's real merged HTML artifact
• git diff --check
• final GitHub Actions run: 16/16 checks passed