Repository navigation
Conversation
…ranxianglei#4682) Clients mint child conversation ids for work that runs on the parent's own route+model — omp opens idle-recap ephemeral turns as `<sessionId>:side:<snowflake>` (plus a `:side:conversation:<key>` variant) and subagents as `<sessionId>:sub:<id>` — and register runtime-info only under the parent uuid. The window ladder looked the id up literally, so a child missed the report and fell through to the registry peek, where a tier-gated family is capped at the built-in standard 200K without tier evidence (ranxianglei#1321). A session that reported 1M therefore had its own recap turns sized at 200K: preflight compressed ~174K of context toward a 150K headroom target, committed early SSE headers, and the client gave up while the summarizer was still running. pluginRuntimeInfoForConversation now strips one client-minted child suffix at a time and resolves the nearest reporting ancestor (bounded to 4 strips), keeping the model-equality gate at whichever level answers: a child on a different model inherits nothing, and a direct hit is never overridden by a parent. tests/side-window-inherit.test.ts drives the real ladder — a `:side:` child of a 1M-reporting parent budgets 1_000_000 (200_000 before the fix) — and tests/runtime-info.test.ts carries the unit-level cases. Docs: TECHNICAL-NOTES(+zh) records the child-id resolution and the CHANGELOG carries the incident shape.
(ai: vllm-qwen/qwen3.8-27b) 收到,正在审。 **查重: 不重复。**按「side conversation / runtime-info / window inherit」与「:side: + #1321 cap」两组关键词检索了 open+closed 的 issue 与 PR,无其他条目处理同一机制(子会话 id 查不到父级 runtime-info 而落 registry cap);在途 PR 中也没有第二个动 标签: bug / omp / scope/wiring / p2(缺陷修复:子会话被按 200K 定窗而非父会话上报的窗口;现场证据全部指向 omp 自造子 id;改动域是请求管道窗口解析;p2 = 主会话不受影响、仅大上下文下 side/subagent 回合丢轮,有绕行)。 审核进行中:独立读 diff、回归面清点、本地定向验证、CI 红分片归因(当前 head 上 7 个 test 分片红,需判定继承性还是本 PR 引入)。结论按八问体例随后补上。 |
(ai: vllm-qwen/qwen3.8-27b) 审核结论:代码可合,当前红是 master 继承红(非本 PR 引入)独立做了四件事:逐行读 diff、清点回归面、本地定向验证(含「回退修复看测试是否变红」的有牙性验证)、拉取全部 7 个红分片日志与 master 自身最新红运行的日志做失败测试名逐一比对。 一、代码审查(主关)改动只在
二、回归面清点(逐项结论)
三、测试有效性(质量关)不是假测试,有牙且能兜底:
四、CI 红归因(关键发现:100% 继承红)head
这正是 #4599 记录的 born-red 簇(storage/golden/fork 家族);其修复载体 PR #4606 已关闭未合并,所以 master 至今红着。另查证: 五、结构化八问
收工判定代码与测试:可合(无需返工,无重构项)。合并动作被 master 继承红阻塞:等 #4599 簇在 master 转绿后,作者把分支 sync 一次重触发即可(现在 sync 也只会再红一遍,不建议提前动)。PR 描述小瑕疵一条:风险节漏列 server.ts:1878 这个 maxOutput 第二调用点(同向变更,不影响判定)。 |
|
Merged master On the merged head:
Description updated per the review: the Risk section now names both |
a148ab9 to
e9fa49c
Compare
…tream teardown PR ranxianglei#4683 CI (windows-latest/node22) failed the proxy-json abort test after the 3s poll budget elapsed (closed=0). The teardown path is correct (client-abort -> fetch AbortController; exit-matrix contract 'client disconnects while awaiting the buffered JSON -> upstream request destroyed') and the three SSE siblings passed in the same shard, so this is async-propagation latency on a loaded runner, not a regression from this PR. Double the poll budget 30->60 (3s->6s) uniformly across all four abort variants as tolerance headroom; behavior unchanged.
(ai: vllm-qwen/qwen3.8-27b) 本轮状态汇总(承接我上一轮对 新增 commit 逐一对账
**flake 判定的独立意见:同意。**同一 shard 里四个同型 abort 用例(openai/anthropic/responses/json)只挂了最慢的 json 变体(中止发生在缓冲期,要取消一次未完成的 buffered read),其余三个全过——单点时序抖动特征成立。修法只放宽观测预算、不动任何产品行为/阈值/输出格式;断言仍有牙(真出现「socket 永不销毁」回归会在 ~6s 内失败,只是检出慢一拍)。本地在新 head 复跑 CI 现状:新 head 的运行已启动(镜像缓存读:25 项检查全部排队/刚起,0 failed)。注意一点:自动修复车道在本轮后已达 24h 配额上限(floor 14),若新运行再有红,将需要 owner 决策而非自动接单——届时我会先做归因再上报。 判定:分支侧完整且干净(原修复本体未动 + 一个有牙性保持的测试时序修)。等 test 矩阵出结果:全绿即 ready-to-merge(合并动作留给你);有红我再归因。 |
[bot] Merged origin/master into |
Unblocks PR #5004's test (windows-latest, 22, 2) red on head da25207: three abort cells (proxy-openai-sse / proxy-responses-sse / proxy-json) asserted closed=0 after exhausting the flat 3s (30x100ms) poll on a loaded Windows runner -- the same signature already recorded under #4599 for #4982/#4774/#4683 (rerun green / in-branch budget bump). Sibling cell abort x proxy-anthropic-sse passed in 176ms on the same shard and every other windows/ubuntu shard ran this file green on the same head, so this is runner load, not a code regression; the PR's diff touches none of these paths. Labeled CI-unblock commit; duplicates #4683's pending identical change -- drop this commit when rebasing if #4683 lands first.
…ndows runners PR #5009 CI (windows-latest/node22, leg 2/5) failed the proxy-openai-sse and proxy-anthropic-sse abort tests after the 3s poll budget elapsed (closed=0), twice (original run + rerun-failed-jobs); the two lighter siblings (proxy-responses-sse, proxy-json) passed in the same shard, and this file is green on both ubuntu legs at this head and locally (standalone 2/2). This PR does not touch the teardown path (no src changes in the head's second commit; #4998's lane-decision change is inert for plain-proxy traffic — compressReachable short-circuits on !pluginMode), so this is async-propagation latency on a slow Windows runner, not a regression from this PR. Identical 4-line diff to in-flight PR #4683 (head 9dd31d2); if #4683 merges first, drop this commit on rebase. Documented flake family: #4599 floors 350/352/358/363/365.
# Conflicts: # CHANGELOG.md
# Conflicts: # CHANGELOG.md
[bot] (ai: vllm-qwen/qwen3.8-27b) Re-synced master 58e61a3 (84 commits incl. the #4986 digest fix and fork-recovery rework) into the branch — sole conflict again additive in CHANGELOG.md (both entries kept); typecheck green locally; pushed as 980c5b9. PR now mergeable=true, state=blocked (awaiting human review only), CI running on the new head. |
…3-head # Conflicts: # CHANGELOG.md
|
The current red roll-up is a cancelled CI matrix, not an executed test failure. The GitHub job records show I requested a rerun of the cancelled workflows, but GitHub reports that rerunning them requires repository admin rights. Please rerun the workflows for head |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Cancelled matrix rerun + master sync — outcome after this round's actions:
State after push: PR open, |
|
The requested rerun has now completed with real test failures rather than cancellations. Run 41207520292, attempt 2, failed three shards:
All three fail the same two Current master contains the #5345 catalog-guard fix for exactly these failures. I also verified that the relevant catalog/test files do not differ between Please use the synced tree for the next full CI run or merge it into the PR branch before rerunning. The separate |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Synced tree now carries the next full CI run — done as requested.
State after this round: PR open, |
…erit # Conflicts: # TECHNICAL-NOTES.md # TECHNICAL-NOTES.zh-CN.md
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Independent review of head 6a291748f (fresh base: contains current master tip except 3 newer docs-only Japanese-README commits with zero file overlap; effective diff vs origin/master is exactly the 6 files, +251/−6).
What I verified myself
- Read the full diff line-by-line at head; the core change is confined to
pluginRuntimeInfoForConversationplus two constants/one regex insrc/plugin.ts. - Inventoried all four consumers of the helper: window ladder (
src/server/handle.ts:832), fallback maxOutput (src/server/handle.ts:2546),resolveKnownOutputCeiling(src/server.ts:2176), and the/__bili/statusruntimeInfodisplay (src/plugin.ts:2041) — the last is additive observability only (a child session now shows the inherited entry, consistent with how it is actually sized). - Window-ladder order,
capRegistryWindowByStandardand the #1321 tier-cap logic: untouched. - Regex edges traced by hand and pinned by tests:
sid-parentx/${PARENT}2:side:x/ empty tail / non-matching segments all stay misses; nested ids strip level-by-level (bounded 4 passes, strictly shortening, no loop risk); KDD#9 persona keys use|sub:(pipe) — no collision with:sub:. - Both-modes reasoning (§6): the conversation-table branch runs only when the
x-bili-pluginheader is absent (the agent table wins exclusively when present), so pluginMode behavior cannot change; plain proxy clients with no runtime-info entry getundefinedexactly as before. - Test teeth: I temporarily reverted the helper to its pre-fix body in a local worktree and re-ran the two new suites — the inheritance cases went red exactly as claimed; after restoring, everything green. The e2e asserts concrete
metadata.effectiveContextLimitvalues (1M vs 200K) through realstartServer+ fake upstream +x-acp-session, not tautologies. - Locally green on head: side-window-inherit 5/5, runtime-info 34/34, related window/plugin suites 133 pass,
tsc --noEmitclean. - CHANGELOG entry sits under
[Unreleased] → Fixed; TECHNICAL-NOTES en/zh updated in parity. No new config/env surface, nokernel/or lockfile touch, noas any/@ts-ignore. - CI: 0 failures observed on this head (mirror cache), lane reports the matrix green; the targeted local runs above corroborate.
Regression inventory (existing behaviors touched by this diff)
| Behavior | Verdict |
|---|---|
| Exact-hit lookup for normal ids | preserved (checked first, model-gated exactly as before) |
| Unknown / prefix-lookalike ids | preserved miss (negative cases pinned in tests) |
| Window ladder order & #1321 tier cap | unchanged |
| maxOutput fallback for child ids | intentional change, disclosed in the PR Risk section (both readers named): child inherits the parent's reported cap instead of falling to the model table — same route+model, semantics hold |
/__bili/status runtimeInfo display |
additive observability: child shows the inherited entry (matches actual sizing) |
| Memory / caches / cursors / dedup keys | untouched (read-only lookups; 32-entry table unchanged) |
No blocking findings; nothing required a spot-fix.
结构化八问
- 是什么问题:客户端(omp)为复盘回合/子代理自造子会话 id(
<parent>:side:<id>等),但 runtime-info 只按父 uuid 注册——字面 Map.get 查空后落 registry peek,#1321 的 tier cap 在无 tier 证据时把窗口压到标准 200K,合法 1M 请求被按 200K 预算,preflight 压缩途中客户端放弃,整轮丢失且摘要费用照付。 - 什么原因导致:runtime-info 会话表只做字面 id 精确匹配,不认识「子 id 是父 id 的后缀派生」这一客户端事实;#1321 的无证据 cap 放大了查空的代价。
- 影响范围与严重程度:仅会自造子 id 的 plugin 宿主(现场为 omp);主会话不受影响。触发条件 = 父会话已上报大窗口的路由 + side/subagent 回合;后果 = 该回合丢失 + 已付费摘要。p2。
- 怎么修复的:在 lookup 决策点按已知后缀锚定剥离(≤4 层)取最近上报祖先,model 门控在命中层执行;精确命中优先、祖先 model 不匹配即整体落空回旧阶梯。修的是解析规则本身,不是症状拦截。
- 是否有回归问题:无。上表逐项清点过;唯一行为变化(child id 的 maxOutput 继承父值)已在 PR 风险节披露且语义成立。
- 是否破坏缓存:否——纯只读查找,无缓存/游标/去重键接触。
- 是否改了设计方向:否,runtime-info 协议与窗口阶梯结构不变,仅补上「子 id 解析」这一缺省规则(en/zh 文档已同步记录)。
- 是否引入新配置:否。
Verdict: APPROVE. Correct and minimal, tests toothed and regression-pinning, no open findings. Ready to merge whenever you are — the merge action itself stays yours.
(ai: vllm-qwen/qwen3.8-27b) Review done: APPROVE posted (review 5823107232) on head |
|
这个方案是不是有问题?会冲突吗?另外看最新的master 有msgid+sessionid解决方案 |
(ai: vllm-qwen/qwen3.8-27b) 逐条查了最新 master(a52c300f),三个问题分开答: 1. 方案有没有问题? 2. 会冲突吗?
3. 关于「最新 master 有 msgid+sessionid 解决方案」 如果你指的是某个具体 commit/PR,给我编号我立刻核对。我的判断维持:APPROVE 有效,这是当前 proxy 侧唯一可行的修法,建议按现状合并;需要的话我可以在 PR body 补一句「身份体系获得跨会话关联后可退役本启发式」。 |
…flict sync) Resolves 2 conflicts, both in test files: - tests/wire-exit-matrix.test.ts: take master's awaitUpstreamTeardown helper (from #4683, now merged) over this branch's inline 6s poll loop — same intent (slow-runner teardown budget), helper is the converged shape. - tests/lane-decisions-matrix.test.ts: union — master's extended ROWS type (location/onlyWires from #5527) + this PR's omp row pinning observed reachability (toolless omp round goes silent, #4998). Typecheck clean; targeted suites green (lane-decisions-matrix, wire-exit-matrix, plugin-nudge-injection, issue4344-mcp-dialect: 129/129).
|
Landed via batch PR #6482 (merged) — closing as superseded. |
What & why
Client-minted child conversations look up their own id, find nothing, and get sized by the registry peek — capped at the built-in standard window for tier-gated families (#1321). The client mints these ids for work on the parent's own route+model and never re-declares the negotiated tier, so the child is a legitimate 1M request sized at 200K:
src/server/handle.ts:744resolvesruntimeWindowfrompluginRuntimeInfoForConversation(runtimeConversationId(headers, parsed, sessionHeader));Map.get+ model-equality guard (src/plugin.ts:603), so<parent>:side:<snowflake>,<parent>:side:conversation:<key>and<parent>:sub:<id>(omp's shapes) always miss;peekWindow = capRegistryWindowByStandard(model, peekRegistryContext(...), hasTierEvidence)(handle.ts:769), with no tier evidence on the wire → 200,000.Consequence in the field: ~174K of recap context compressed toward a 150K headroom target, 30 s grace, committed SSE headers, then
preflight aborted (the client disconnected during preflight compression); not forwarding— the summarizer spend is paid and the turn is lost.Change
pluginRuntimeInfoForConversation(src/plugin.ts) strips one known child suffix per pass and returns the nearest ancestor report, keeping the model-equality guard at whichever level answers::side:<id>,:side:conversation:<key>,:sub:<id>);sid-parent2:side:x,sid-parentxandsid-parent:side:stay misses.No new config key, no ladder re-ordering, no change to
capRegistryWindowByStandard.Test evidence
node --import tsx --test tests/side-window-inherit.test.ts— drives the real ladder (startServer+ fake upstream +x-acp-session, assertingmetadata.effectiveContextLimit):<parent>:side:<id><parent>:side:conversation:<k>/<parent>:sub:<id>Before the fix: 2 of 5 fail; after: 5 of 5 pass.
node --import tsx --test tests/runtime-info.test.ts— 5 new unit cases in the same direction (child inherits window +maxOutput; nested:side:→:sub:; wrong-model parent; direct hit wins; no prefix false hits). Before the fix: 2 fail / 32 pass; after: 34/34.Related suites, all green:
tests/issue1321-tier-window-cap.test.ts,tests/issue1569-window-stamp-and-anchor.test.ts,tests/issue2481-dsh-runtime-context.test.ts,tests/configured-window-outranks-registry.test.ts,tests/plugin-agent.test.ts,tests/side-request-isolation.test.ts(131 tests, 0 failures).tsc --noEmitclean.Risk
maxOutput: the same lookup feeds two output-ceiling readers: the fallback-output resolution insrc/server/handle.ts:2244-2245andresolveKnownOutputCeilinginsrc/server.ts:1886-1887(the [bug] 长会话频繁「达到输出token上限」:restore 回落到被污染的 high-water,0.1.175 未声明 output 时仍能复现 #1840 restore floor). In both, a child now inherits the parent's reported cap instead of falling through to the launcher/route/registry chain. That is the intended semantics (same route + model) and the unit case covers it.:side:ids are minted per turn, but nothing is cached per child (the lookup is pure string stripping against the 32-entry conversation table), so there is no new growth path.Fixes #4682