[Video] add the read-only H3 performance history API and list / 新增只读的 H3 性能历史 API 与列表

[edwingao28]

Summary

Reads already published H3 runs from Vercel Blob into a per-cell history projection (GET /api/video-runs?format=history) and lists them under /video?view=history with bilingual filters and result links, keeping execution identity and failure/qualification status. No new measurement. First of two stacked PRs; the /video dashboard follows.

AI model disclosure

• Model/version: claude-fable-5-1 (Claude Code)
• Role: implementation, tests and docs; the author directed and reviewed.

Validation

• typecheck, lint, fmt, vitest (video-runs route and history, 27 pass) and the navigation e2e spec on the E2E_FIXTURES=1 server, run on the stacked dashboard branch that includes these commits; not re-run in isolation here.
• I have completed the AI model disclosure and kept it current

中文说明

把已发布到 Vercel Blob 的 H3 运行读成按 cell 的性能历史投影（GET /api/video-runs?format=history），并在 /video?view=history 列出：中英文筛选、直达结果链接，保留执行身份与失败、资格状态。不产生新的测量。两个堆叠 PR 的第一个，/video 仪表板随后提交。

AI 模型：claude-fable-5-1（Claude Code），负责实现、测试与文档；作者把关方向并审阅。

验证：typecheck、lint、fmt、vitest（video-runs 路由与 history，27 项通过）及 navigation e2e（E2E_FIXTURES=1 服务器）在包含本分支提交的仪表板分支上运行；未在本分支单独重跑。

---

Note

Medium Risk
Changes default /video load behavior and adds blob index scanning plus new client navigation races; read-only and guarded by repo visibility, but UX and API surface for the hidden viewer expanded materially.

Overview
Adds read-only H3 performance history end to end: /video defaults to a Performance history tab instead of auto-scanning GitHub Actions, while Browse CI runs keeps the legacy Actions discovery path.

API & storage: GET /api/video-runs?format=history&page=N returns a versioned projection from existing Vercel Blob index metadata via publishedVideoHistory (paginated, newest publication first, no media download or publish). videoHistoryEntry maps stored artifacts into per-cell observations using existing serving/tradeoff logic, with explicit errors and null metrics where data is missing. E2E_FIXTURES=1 replays video-history.json with X-VideoGenX-Replay.

UI: New VideoHistory lists published runs with bilingual copy, URL-backed filters (history-hardware, history-concurrency, history-query), pagination, and links into the video viewer. VideoCIRuns adds ?view=history|results, published-first loading for Videos & result, cancels in-flight selection when returning to history, and preserves share URLs including history filters.

Docs & tests: docs/video-history.md documents read paths and evidence boundaries; Cypress/Vitest cover navigation, retries, stale-response guards, and catalog ordering.

^{Reviewed by Cursor Bugbot for commit 1d5084a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Sep 20, 2026 6:19pm UTC

[github-actions]

Claude finished @edwingao28's task in 3m 12s —— View job

---

Review complete

• Read changed files (route, history projection, storage, components)
• Check CI status
• Runtime verify /video?view=history and API
• Review Chinese copy
• Post findings

🟢 LGTM — no blocking issues found.

Runtime check (localhost:3000): /video?view=history renders, the Performance history tab toggles and shows pressed state, filters (Hardware / Client concurrency / workload search) render, and the error path degrades cleanly to "Could not load published history." + Retry. The only console error is a 502 from /api/video-runs?format=history, which is expected in this local env — no BLOB_READ_WRITE_TOKEN, so the dev branch proxies to SITE_URL and that fetch fails. Not a PR defect; the component handles it gracefully.

CI: Lint & Format ✅, Tests (Unit) ✅, Code Quality ✅ (E2E still running at review time).

What I checked closely and found sound

• videoHistoryEntry / history.ts — null-safe throughout (at/number/text coerce cleanly; failed/uncalibrated cells preserved without inventing zeros), and per-source try/catch keeps one bad artifact from poisoning the page. Well covered by history.test.ts.
• publishedVideoHistory — full-scan is intentional and documented (the ponytail: comment: materialize a catalog if inventory grows); index published last so partial uploads aren't observed; sort + pagination correct. video-storage.test.ts verifies ordering, 10-index read cap, and no media writes.
• Route validation for format=history (rejects run/artifact/unsafe page), the dev-proxy identity checks, and the fixtures replay branch all look correct; route.test.ts covers the no-scan/no-publish contract.
• VideoCIRuns view/deep-link wiring — initialView resolution, changeView aborting in-flight downloads, openResults/openPublished fallbacks, and the added setSourceId(source ?? '') reset (fixes opening a published source) all trace correctly. Overlay path N/A (this is the H3 video viewer, not the inference/eval chart path).
• Test coverage present for every new module — no missing-test gap.

Non-blocking (Chinese copy, advisory)

• no: '未经验证' renders under qualification: '发布验收' ("Release qualification"). 未经验证 reads as "not verified" rather than "not qualified"; 未通过验收 / 未达标 would track the source more closely. Low-confidence wording nit only — not pinging the maintainer. Overall the zh dictionary is natural and the fail→未通过 / failed→失败 status split is a nice distinction.
  · branch feat/videogenx-performance-history

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 1d5084a. Configure here.}