fix(api): support agentic-only runs in unofficial-run previews / 修复 agentic 专属 run 的非官方预览

[cquil11]

Summary

?unofficialrun=<run-id> chart overlay previews returned 404 for agentic-only sweeps. Root cause: the benchmark repo's run-sweep.yml only triggers the collect-results merge job for fixed-seq (1k1k / 8k1k) sweeps, so agentic-only runs upload per-config bmk_agentic_<config> artifacts and never produce the merged results_bmk artifact that /api/unofficial-run exclusively looked for.

• When results_bmk is absent, the route now falls back to downloading every per-config bmk_* artifact (same-name re-uploads deduped, newest artifact id wins) and normalizes their rows through the existing mapBenchmarkRow path — which already handles agentic v2-flat/v3-nested agg schemas.
• The fallback stays off when results_bmk exists: the merge job downloads the bmk_* pattern, so the merged artifact already contains every per-config row and fetching both would double-count.
• Per-config downloads run in bounded batches of 8 to stay clear of GitHub secondary rate limits (big agentic sweeps upload dozens of artifacts).
• 404 error message now mentions the per-config bmk_* fallback.

No frontend changes were needed — the overlay pipeline (buildChartData → transformBenchmarkRows → applyAgenticMetricAliases, processOverlayChartData with isAgentic, legend/points-dialog/dismiss paths) already supports agentic rows; it just never received any.

Overlay support

This IS the overlay path — verified manually against a real agentic run from InferenceX PR #2109: loaded /inference?unofficialrun=28861274175 (DSv4 FP4 MI355X vLLM agentic sweep, kvnone + LMCache offload variants) on a local dev server and confirmed:

• NON-OFFICIAL banner shows the run's branch (amd/agentx_dsv4_vllm) with working dismiss
• Model auto-switches to DeepSeek V4 Pro, scenario Agentic Traces, latency-percentile selector active
• All 5 overlay points (conc 32/40/40/56/72, offload ON/OFF split) render as ✕ markers on both the Interactivity and E2E Latency charts alongside official series
• Legend entry with hide toggle + "show all data points" dialog lists the overlay rows with correct conc / parallelism / offload / tput / p50–p90 values matching the raw artifacts

Test plan

• 3 new route tests: per-config bmk_* fallback (agentic rows, correct conc mapping, sibling agentic_*/server_logs_* blobs NOT downloaded), fallback disabled when results_bmk exists, same-name dedupe keeps newest artifact
• pnpm test:unit — 2539 app + 25 mcp tests pass
• pnpm typecheck, pnpm lint, pnpm fmt clean
• Manual verification against live run 28861274175 (screenshots above description; API returns 5 agentic benchmark rows)

中文说明

?unofficialrun=<run-id> 图表叠加预览对 agentic 专属 sweep 一直返回 404。根本原因：基准测试仓库的 run-sweep.yml 仅在固定序列（1k1k / 8k1k）sweep 时触发 collect-results 合并任务，因此 agentic 专属 run 只上传按配置拆分的 bmk_agentic_<config> 产物，从不生成 /api/unofficial-run 唯一查找的合并产物 results_bmk。

• 当 results_bmk 缺失时，路由现在回退为下载所有按配置拆分的 bmk_* 产物（同名重复上传按最新 id 去重），并通过现有的 mapBenchmarkRow 路径归一化（该路径已支持 agentic v2 扁平 / v3 嵌套 agg schema）。
• results_bmk 存在时不启用回退：合并任务本身就下载 bmk_* 模式，合并产物已包含全部按配置行，两者都取会重复计数。
• 按配置产物以每批 8 个的有界批次下载，避免触发 GitHub 二级速率限制。
• 前端无需改动——叠加渲染管线已支持 agentic 行，此前只是拿不到数据。

已用 InferenceX PR #2109 的真实 agentic run（28861274175，DSv4 FP4 MI355X vLLM）在本地手动验证：NON-OFFICIAL 横幅、模型自动切换、Agentic Traces 场景、5 个叠加点在 Interactivity 与 E2E Latency 图表上均正确渲染，图例开关与数据点弹窗数值与原始产物一致。单元测试（2539 + 25）、typecheck、lint、fmt 全部通过。

🤖 Generated with Claude Code

---

Note

Medium Risk
Adds many parallel GitHub artifact downloads and changes benchmark resolution for a live API path, but behavior is gated on missing results_bmk and covered by new tests.

Overview
Fixes 404 on ?unofficialrun= for agentic-only CI sweeps that never upload merged results_bmk.

When results_bmk is missing, /api/unofficial-run now collects bmk_* per-config artifacts (deduped by name, highest artifact id wins), downloads them in batches of 8, and normalizes rows through the existing mapper. The fallback is skipped if results_bmk is present to avoid double-counting. Non-bmk_* siblings (e.g. agentic_* trace blobs) are not fetched.

404 copy now mentions the per-config fallback. Three route tests cover fallback behavior, merged-artifact precedence, and dedupe.

^{Reviewed by Cursor Bugbot for commit 274fc59. 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 (UTC)
inferencemax-app	Ready	Preview, Comment	Jul 7, 2026 10:32pm