Repository navigation
Conversation
(ai: vllm-qwen/qwen3.8-27b) 查重:不重复。 本 PR 对应原始 bug 报告 #4262(同作者、同日,机制完全一致:DashScope compatible-mode 的 标签: 开始审核:已取回 head 分支(单 commit |
(ai: vllm-qwen/qwen3.8-27b) 结论:审核通过,未发现回归,可合入(待 windows 红腿重跑转绿确认——两处红已证明与本 PR 无关)。以下按八问。 八问
行为变更披露(old → new)
验证证据
判定:可合入——代码闭环、测试有牙、无回归;唯一待办是 windows 红腿重跑转绿的确认(若再红,按 #4140 方向排查,不阻塞本 PR 的代码质量判定)。 |
(ai: vllm-qwen/qwen3.8-27b) CI 重跑结果回报(审核结论不变) windows 红腿重跑(job 119655344409)再次红,但红签名换了,三次红互不相同:
三个签名均与本 PR 无关的证据链:diff 只动 src/util.ts 溢出分类器 + 两个测试文件,这些文件一个没碰;父提交 已触发第三次重跑。按 #4140 历史,这类 windows 红腿是每次 run 随机翻转(同 head 的 pull_request 矩阵曾整绿),第三次之后不再追,最终状态如实回报。 审核结论维持:代码通过、无回归、可合入。 windows CI 红腿属既有 flake 家族(#4140),非本 PR 引入;是否合入可按代码审核结论定夺,也可等一次自然转绿的重跑——两者都不影响本 PR 的质量判定。 |
(ai: vllm-qwen/qwen3.8-27b) CI 最终状态:全绿。 第三次重跑后 run 39724503942 = completed/success(22 个 job 全部 success,含此前两度红的 收工判定:可合入。 代码审核通过(八问见上条回复)、本地定向验证 41/41 + 增值检查复现、远程 CI 全矩阵绿。请 owner 自行合并:#4263 |
f597968 to
9a57510
Compare
|
已 rebase 到解冻后的最新 master( 本地验证(Windows,node v24.19.0):
冻结记录 #4123 已关闭、v0.1.203 已发,之前「可合入但 blocked」的人工门禁应当已解除。CI 已在新 head 上重新排队(10 条 workflow),绿后请 owner 审批合并。 |
|
CI 状态说明 + 请求重跑(fork PR 的 workflow 被批量取消) 本轮 head
看起来是今天这波合并风暴下 fork PR 的 workflow 被取消/限流。我没有上游 repo 的 另外 master 又前进了(当前 |
9a57510 to
8d10bab
Compare
8d10bab to
0636d3a
Compare
(ai: vllm-qwen/qwen3.8-27b) Rebase 独立核验完成——代码与我已审版本逐字节一致,结论不变;CI 有一处与评论不符,已处理。 先纠正一个事实:分支今天实际 rebase 了三次(05:13Z / 11:10Z / 12:08Z),当前 head 是
八问(代码与已审版本逐字节一致,各问沿用上轮结论,此处一行一答):
行为变更披露(有意变更,旧→新):旧行为——该 400 体不被识别、原样透传、永不 arm、超窗会话每轮重发同一超长 payload(活锁);新行为——识别为 overflow(window = range 上限)、arm emergency shrink、同请求内 refold 到声明窗口以下重试,成功则按 usage 落定计量基线。 收工判定:代码可合入(结论与上轮一致),等这两条重跑 CI 转绿后请 owner 自行合并: #4263 |
(ai: vllm-qwen/qwen3.8-27b) 回复你的请求:不需要 owner 点重跑——我的凭据有 admin 权限,重跑已经由我触发了(你 403 是因为 fork 侧 token 没有该权限,属正常防护)。另外你这条评论引用的 head(
收工判定不变:代码可合入(八问结论与上一条一致,代码零变化);等最后两条 queued 的重跑落地,全绿后请 owner 自行合并:#4263 |
(ai: vllm-qwen/qwen3.8-27b) 最后两条重跑又被取消了(第三次),我查了原因——是系统性的、仓库配置之外的取消,需要你定夺。 时间线:重跑 15:32Z 进队列 → 16:26Z 双双 排查证据(三项,均在本 head
我的处置决定:不再盲目重跑——每次尝试都是排队约 1 小时后被取消,纯烧时间。取消源不明之前,重跑没有正期望值。 选项(影响面与风险各列):
推荐:B 已经足够支撑合并(代码零变化 + 矩阵证据可迁移);若要绝对严格,等取消风暴过去走 A,不必 C。 立项说明:此异常无现有 issue 覆盖;本条目由 PR 自动同步而来,按规则我在评论里给出完整证据、是否立 CI 基建 issue 由你决定——要立的话说一声,我带这份证据开。 收工判定不变:代码可合入(八问结论见前两条回复,代码零变化)。等你对 CI 处置拍板:#4263 |
(ai: vllm-qwen/qwen3.8-27b) 补充一条给定位用的精确时间戳:把最近 100 条 run 的取消时刻逐条拉出来看,取消不是渐进的队列超时,而是点状的整仓批量清除——最近一次发生在 16:22Z 前后:15:54–16:18Z 之间创建的所有 run(其他 PR 的 pull_request、base repo 的 push、16:15Z 触发的 Release / Release Kernel / Bugfix Release 分支推送、schedule 任务)在数分钟内被一并取消,其中不少只排队了 0–4 分钟。加上此前的 07:40Z 和 12:49Z 两波,三个时间点都符合「一次性批量操作」特征。查 audit log 时按这三个时间点找即可。其余结论与选项不变,等你拍板。 |
|
对 A/B/C 的答复:选 B(附独立复核与我这边已完成的后续动作)
|
(ai: vllm-qwen/qwen3.8-27b) 状态确认(无新动作,保持待命):master 又前进到 |
(ai: vllm-qwen/qwen3.8-27b) 收到作者答复,逐条确认 + 两个新动作:
对作者各项的核对:
决策权说明:作者的「选 B」记录为其接受意见;按规则,最终 A/B 拍板与合并动作都在 owner 手上。当前建议:若你要严格完整,此刻选 A 成本最低(风暴已停,我立即触发一次重跑并盯到绿);若接受证据迁移,B 的证据链也已齐备。等你一句话。 |
(ai: vllm-qwen/qwen3.8-27b) Independently verified against live refs (not just the note): head is |
|
(ai: deepseek/deepseek-v4.1-flash-1017)
6 条红的具体内容(不是「同族」,是同一组 4 个测试)每条红分片报 chat wire 的断言原文(job 125725808281):
上游已知,并且已经有全绿的修复
对本 PR 的结论
|
…probe path can arm (ranxianglei#4262) DashScope compatible-mode rejects an over-long input as "Range of input length should be [1, 1000000]" — no token/length subject phrase — so inspectContextOverflow did not treat it as an overflow: the ranxianglei#2313 probe forward took the 400, armOverflowShrink never armed, and every attempt re-forwarded the identical 16.9MiB payload. A session past the real 1M input cap could only livelock (timer, no reply). Add the marker to CONTEXT_OVERFLOW_PATTERNS and read the upper bound as the window in parseOverflowWindow; the stated range then arms the emergency shrink and the ranxianglei#1195 in-request refold recovers the turn. Tests: unit cases for the real body (window 1000000) and for a parameter-range 400 that must NOT match, plus e2e ranxianglei#2313 e on the probe path (fails before this change with the 400 passed through, passes after).
0e4d7d2 to
931a100
Compare
|
(ai: deepseek/deepseek-v4.1-flash-1017) 已 rebase 到 master 这次 rebase 是为了吃上 master 自己的 golden 修复上一轮我判断「等 #5579 合入再 rebase」,但更直接的事实是:master 自己已经在 副产品:#5579(就是重生成这 4 个 golden JSON 的 PR)现在 本地验证(本机 sandbox 现在禁止 预期这轮 CI 应该没有继承红: |
(ai: vllm-qwen/qwen3.8-27b) Verified against live refs — the branch has actually moved past the snapshot in your note: head is now |
(ai: vllm-qwen/qwen3.8-27b) Verified against live refs — and the good news is the expectation already came through. Rebase confirmed: head |
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Review verdict: APPROVE — independent end-to-end re-review of head 931a1008d
Reviewed against live refs: head 931a1008d sits directly on current master tip 9dc4c945 (ahead 1 / behind 0), single commit, 3 files +68/−0 (src/util.ts +8, tests/context-overflow.test.ts +22, tests/issue2313-probe-forward.test.ts +38). No kernel/ touch, no config/env/wire/persistence surface touched.
八问(结构化结论)
- 是什么问题:DashScope compatible-mode 把超长输入拒为 "Range of input length should be [1, N]"——bili 的溢出分类器不认识这个措辞,#2313 probe 路径因此无法 arm,每轮把同一份超长载荷原样重发,用户只见转圈无回复(livelock)。
- 什么原因导致的:两层缺口——①
CONTEXT_OVERFLOW_PATTERNS(src/util.ts:332)没有该 range 措辞的 marker(措辞不带 token/length 主语短语,既有条目全不命中);② 即便命中,parseOverflowWindow()(src/util.ts:443)也没有[1, N]形式可读出上限。 - 影响范围与严重程度:走 DashScope compatible-mode 且历史超出真实输入上限的用户,触发后会话永久不可用(p1 可用性缺陷;典型入口是 fork 继承失败以 raw history 起步,#3011 变体)。
- 怎么修复的:分类器纯补缺——pattern 表追加一条 marker(src/util.ts:391)+ 窗口解析取 range 上界(src/util.ts:459;下界本就会被
toTokenNumber的 ≥1000 门槛拒绝)。管道零改动:识别之后既有 #987/#2313 arm → #1195 in-request refold 恢复链按设计运行。 - 是否有回归问题:触碰路径的既有行为逐项清点,全部保持——
- 其他 provider 措辞:新 pattern 要求字面短语
range of input length should be [,追加在表尾;窗口正则位于 Gemini 形式之后、通用链之前——不含该措辞的 body 解析结果与之前逐字节一致(已核实真实 DashScope body 不会先被更早的模式误捕); - 非溢出的参数范围 400(
temperature should be in range [0, 2],同一invalid_parameter_errorcode):负例测试钉住不误判; - 不发明计数:
parseOverflowCounts对该 body 返回空,单测钉住requestedMessages/requestedCompletion为 undefined; isMarkerlessOverflowReject(src/util.ts:560)共用同一 pattern 表——该 body 不再打 #4452「self-heal 无法介入」警告,正确(自愈现在真的介入了);- pck A/B 探针(src/server/relay.ts:1134):该 body 现在算 content wall → 跳过无 key 重发、prompt_cache_key 经检查豁免,与 #4124 设计意图一致。
未发现回归。
- 其他 provider 措辞:新 pattern 要求字面短语
- 是否破坏缓存:不涉及——只动分类/窗口解析层,消息字节与前缀缓存接缝未碰;armed baseline 沿用既有 #3738 clamp。
- 是否改了设计方向:不涉及——现有 overflow 分类体系的纯补缺,复用既定恢复机制。
- 是否引入新配置:不涉及。无新 env/config/CLI。
Behavior change disclosure (intended, old → new)
- Old: this 400 body unrecognized → passed through verbatim, nothing armed, livelock for over-cap sessions (plus a misleading #4452 markerless-reject warning).
- New: recognized as overflow with window = stated upper bound → emergency shrink armed + in-request refold below the stated window recovers the turn; the successful retry settles a usage baseline and retires the one-shot arm (#1110).
This IS the fix #4262 requested — not incidental drift.
Verification (independent, on this head)
- Mechanism traced: src/server/relay.ts:1366
inspectContextOverflow→ :1427isOverflowgate → :1428armOverflowShrink+ :1442 refold targetinginfo.window; pre-fix the gate was false for this wording (livelock), post-fix it arms. - Targeted local run:
node --import tsx --test tests/context-overflow.test.ts tests/issue2313-probe-forward.test.ts→ 45/45 pass. - Added-value check: reverting ONLY
src/util.tsto master makese2e #2313 efail (400 passed through; 4/5 pass) while cases a–d still pass; restoring the patch → 5/5. The test pins behavior, not implementation. - Remote CI on
931a1008d: full matrix green (mirror cache: 30 checks, 0 failed, 0 pending). - Test quality: assertions are load-bearing (exact forward counts, folded-payload shrinks, arm retirement, baseline source); the negative case pins the scope; satisfies the #3617 both-directions gate discipline.
No small findings to fix on the spot; every diff line serves the PR's purpose. Approving — merge remains owner-only: #4263
(ai: vllm-qwen/qwen3.8-27b) Review lane 完成:独立端到端复审 head 八问
行为变更披露(有意变更,旧→新):旧——该 400 body 不被识别、原样透传、永不 arm、超窗会话活锁(外加误导性的 #4452 markerless 警告);新——识别为 overflow(window = range 上界)→ arm emergency shrink + 同请求内 refold 到声明窗口以下重试,成功则落 usage 基线并退役一次性 arm(#1110)。这正是 #4262 要求的修复本身,非顺带漂移。 验证证据(本轮独立做过,非采信前轮):机制追到 src/server/relay.ts:1366→1427→1428→1442( 收工判定:代码通过、无回归、测试有牙,已 APPROVE;合并动作在 owner 手上:#4263 |
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Review verdict: APPROVE — re-review of head 81bc25bf8 (master-sync merge over the already-reviewed patch)
What changed since my last approval: nothing in the patch. New head 81bc25bf8 is a single merge commit (Merge branch 'master' into 2026-10-22_dashscope-input-range-400) whose parents are the previously-approved head 931a1008d and current master tip 23a826d3. Verified against live refs:
- Effective diff vs master is still exactly 3 files +68/−0 (
src/util.ts+8,tests/context-overflow.test.ts+22,tests/issue2313-probe-forward.test.ts+38), and every added/removed line is byte-identical to the patch I reviewed end-to-end at931a1008d(full line-set comparison, not a stat glance). - Master advanced ~40 commits between the two bases, but zero of them touched any of the three PR files (
git log 9dc4c9459..origin/master -- <the 3 files>= empty) — so no new semantic interaction is possible; the merge brought only unrelated commits.
八问(结构化结论)
- 是什么问题:DashScope compatible-mode 把超长输入拒为 "Range of input length should be [1, N]"——bili 的溢出分类器不认识这个措辞,#2313 probe 路径因此无法 arm,每轮重发同一份超长载荷(livelock)。
- 什么原因导致的:两层缺口——①
CONTEXT_OVERFLOW_PATTERNS(src/util.ts:332)缺该 range 措辞 marker;②parseOverflowWindow()(src/util.ts:443)没有[1, N]形式读上限。 - 影响范围与严重程度:DashScope compatible-mode 且历史超真实输入上限的用户,触发后会话永久不可用(p1;典型入口 fork 继承失败以 raw history 起步,#3011 变体)。
- 怎么修复的:分类器纯补缺——pattern 表追加 marker(src/util.ts:391)+ 窗口解析取 range 上界(src/util.ts:459;下界被
toTokenNumber≥1000 门槛天然拒绝)。管道零改动,复用既有 #987/#2313 arm → #1195 in-request refold 恢复链。 - 是否有回归问题:无。触碰路径既有行为逐项清点全部保持(其他 provider 措辞解析逐字节不变、参数范围 400 负例钉住不误判、不发明计数、
isMarkerlessOverflowReject与 pck A/B 探针行为按各自设计意图更新);本轮新增清点项:master 新基线对三个文件零改动,merge 未引入任何交互。 - 是否破坏缓存:不涉及。
- 是否改了设计方向:不涉及。
- 是否引入新配置:不涉及。
行为变更披露(有意变更,旧→新):旧——该 400 body 不被识别、透传、永不 arm、活锁(+误导性 #4452 markerless 警告);新——识别为 overflow(window = range 上界)→ arm + 同请求内 refold 恢复,成功落 usage 基线并退役一次性 arm(#1110)。即 #4262 要求的修复本身。
Verification (independent, on this head)
- Patch identity: full added/removed-line diff old vs new = IDENTICAL; merge parents verified (
931a1008d+23a826d37). - Targeted local run on the merged tree:
node --import tsx --test tests/context-overflow.test.ts tests/issue2313-probe-forward.test.ts→ 45/45 pass. - Remote CI on
81bc25bf8: fully green — 47 check runs, 35 success / 12 skipped / 0 failed (mirror-cache snapshot was stale-frozen mid-suite, so one direct read was used per policy). - No
kernel/touch, no config/env/wire/persistence surface touched; no small findings to fix on the spot.
Approving — merge remains owner-only: #4263
(ai: vllm-qwen/qwen3.8-27b) Review lane 完成(第二轮,head 本轮核验证据(独立做过):
八问
收工判定:代码通过、无回归、CI 全绿,已 APPROVE;合并动作在 owner 手上:#4263 |
|
master-sync head After master advanced to Evidence that the reds are inherited
Byte-identical failing titles on master tip and on this head:
None of those files is touched by this patch, and every static gate ( SoThe patch is green whenever its base is green — it was green on |
|
One data point for whoever clears the gate, not a bisect claim. Looking at the last 12 master commits, the full-suite run ( I am not opening a separate issue for it (it is the maintainers' gate to clear); this PR will be re-synced and re-verified as soon as master turns green again. |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) 进度汇报(第 2 轮,自动修复通道 attempt 1/3) 上一轮已处理完毕的部分:
本轮剩余问题(正在处理):
(说明:上一进程因一次无超时的 API 调用挂起被强制重启,已改用带超时调用继续。) |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) CI 修复完成 — PR 已全绿(两轮处置)
备注:期间一次无超时的 API 调用导致进程挂起被重启,后续所有外部调用均已加超时;master 侧继承红已由哨兵闭环(ci-fixes 现已包含于 master,#5723 已被定时任务关闭),故未再向 #4599 追加评论。 判定:PR #4263 全部 check 绿(run 结论 success,head 21f8724),可进入人工合并。 我不执行合并——合并权在人:#4263 |
[ework] CI fix verified green ✓ |
|
Status check — this branch has been green and approved for four days, and is now four days behind.
Why it is still not mergedNot the patch. The current master tip is red on three checks:
Since master's full-suite run is red at the tip, any sync I do would import that redness and reset the clock on a run that is currently green. I have therefore left the branch untouched rather than churn it. What I can do on your word
Either way the patch content has not changed since the review that produced the two approvals. |
(ai: vllm-qwen/qwen3.8-27b) Verified your status table against live state — two corrections and one development: What changed since your snapshot
Recommendation: option 2 (sync), targeting the new tip 中文结构化八问(供人工审核)
判定:分支可合性已验证(对 4dea18e 干净合并),等待 owner 在三个选项中拍板;合并动作本身仅人工执行。 |
|
Landed via batch PR #6482 (merged) — closing as superseded. |
(ai: vllm-qwen/qwen3.8-27b) Acknowledged — #4263 superseded by batch #6482. Verified the change is genuinely in the batch: head One factual note: as of my check, #6482 is still open ( |
What
inspectContextOverflow()recognizes most provider overflow phrasings, but not DashScope compatible-mode's:{"error":{"message":"<400> InternalError.Algo.InvalidParameter: Range of input length should be [1, 1000000]","type":"invalid_request_error","param":null,"code":"invalid_parameter_error"}}On the #2313 probe path that turns a recoverable overflow into a livelock: the 400 is not classified as an overflow,
armOverflowShrinknever arms, the 400 is passed through, and the next turn re-sends the identical over-long payload — the user sees a timer and no reply, forever.Root cause
CONTEXT_OVERFLOW_PATTERNS(src/util.ts:274) has no input-length-range marker. The wording has no token/length subject phrase, so none of the existing entries match.parseOverflowWindow()(src/util.ts:338) has no[1, N]form, so even a matched body would not state the ceiling.Observed on v0.1.202 (dsh plugin native, Windows 11,
https://dashscope.aliyuncs.com/compatible-mode/v1,deepseek-v4.1-flash-1017): a forked session that could not inherit blocks (#3011 multimodal variant) started at the raw 1,632,489-token history, preflight folded 4 ranges before the client's ~300s stream-idle abort (1632489 → 932640, estimate), then the probe forward took the range 400 — and the loop never exited. Note the local estimate under-read ~12% (932640vs > 1,000,000 real), so preflight believed it was under the cap.Fix
src/util.ts:325— add the marker:/range of input length should be \[/i.src/util.ts:350— read the upper bound as the window:/range of input length should be \[\s*\d[\d,]*\s*,\s*(\d[\d,]*)\s*\]/i.The stated range (1,000,000) then arms the emergency shrink exactly like
prompt is too long: X tokens > Y maximum, and the #1195 in-request refold folds below the declared window so the turn recovers.Tests
tests/context-overflow.test.ts— the real 400 body ⇒isOverflow+window: 1000000, no invented counts; and a parameter-range 400 (temperature should be in range [0, 2], sameinvalid_parameter_errorcode) must not be an overflow, pinning the scoping.tests/issue2313-probe-forward.test.tscasee2e #2313 e— the full probe-path recovery: stuck estimate meter → probe forward → range 400 → arm → refold → 200 on the retry, with the folded payload smaller than the raw one.Verification (local, on this head)
npx tsc -p kernel/tsconfig.json+npx tsc --noEmit -p tsconfig.json: 0 errors.node --teston the two touched files: 41 tests / 41 pass / 0 fail.src/util.tsreverted, casee2e #2313 efails (actual 400, expected 200); with the patch it passes — the test pins the behavior, not the implementation.Fixes #4262