Repository navigation
[PR] fix: name the unknown nudge baseline in idle diagnostics (Part A of #5928) - #5935
ranxianglei wants to merge 2 commits into
Conversation
10f17a4 to
2652f69
Compare
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.9075-pr.5935.264ebaaThis exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr5935.tgz
npm install -g packageInstall instructions are refreshed on each push. |
(ai: vllm-qwen/qwen3.8-27b) 查重: 不重复。Searched open upstream issues/PRs for the same failure signature (blind nudge meter after resume-inheritance miss, 标签: bug + p2 + scope/kernel(诊断可见性缺陷;nudge/折叠判定域,跨车道无单一客户端归属) What this session did (verify-and-refresh, not re-implement)
Consistency notes from the audit:
CI status (live at posting)Running on fresh head Part B statusPart B (letting Verdict: mergeable once CI is green on 八问(结构化审核结论)
收工判定:代码与描述一致、定向验证全绿、diff 干净(单 commit、每行相关),CI 绿后即可合并;合并由人工执行。 |
(ai: vllm-qwen/qwen3.8-27b) CI red resolved — known flaky timing cell, re-run green, no code change. What I did this run:
七问:
判定:CI 已全绿(run 42464291482 success @ 2652f69),PR 可进入正常合并流程;本轮收工。 |
2652f69 to
0996789
Compare
(ai: vllm-qwen/qwen3.8-27b) Merge conflict resolved — PR rebased onto current master, pushed, conflict cleared. What I did this run:
No code changes beyond conflict resolution, no PR-body edits, no gates hit, no new issues filed. 七问:
|
|
(ai: vllm-qwen/qwen3.8-27b) Review verdict: APPROVE (posted as a thread comment — GitHub rejects a formal approve review from the PR's own author account) — clean, well-scoped visibility fix; one small rebase artifact fixed on the spot (below). Independent verification (not copied from the issue)Read the PR body, issue #5928, and the full diff hunk-by-hunk against base
已顺手修已顺手修:rebase 解冲突时 responses 车道的 nudge 诊断行被留在了消息变换之后(其余三车道都在变换前),导致该车道日志相对顺序漂移、且变换抛异常时丢失诊断行——已恢复回原位置,四车道一致 + commit 结构化八问
收工判定可合——等新 head |
(ai: vllm-qwen/qwen3.8-27b) Review lane complete: GitHub PR #5935 reviewed end-to-end — verdict APPROVE (posted as a thread comment; GitHub rejects a formal approve review from the PR's own author account); one small finding fixed on the spot (rebase had left the responses-lane nudge diag line after the message transform — restored to its pre-rebase position, commit |
d694e36 to
5927e65
Compare
…5928) After a resume-inheritance miss the conversation lands on a fresh key whose meter reads 0 while pendingT1 already sits far over its T1 threshold; the nudge decision is structurally unfireable for that turn but was only visible as an ordinary info line, so the oversized forward went unremarked. Part A (visibility, no behavior change to the fold decision): - diagNudge appends ', baseline=unknown' when the meter read 0 so the idle line states WHY nothing fired (byte-identical for measured turns). - that exact state (blind meter + pendingT1 >= interval + not injecting) logs at warn instead of info via nudgeDiagLogLevel(); all other states keep their level. - /__bili/stats gains per-session nudgeBaselineUnknown. The root-cause fix (letting firstSightMassReady drain a known backlog when the meter is blind) is a kernel change awaiting owner approval.
…ition The conflict rebase onto 4dea18e left the nudge diagnostic log call after repairResponsesAssistantOrdering()/applyReasoningDrop() while all three other lanes (openai/google/anthropic) keep it before the message transform, right after the recordNudgeObservation line. None of the moved code touches diagNudge's inputs, so the move only changed relative log order (reasoning-drop warnings now precede the nudge line in this lane) and dropped the line from the exception path. Restored for cross-lane consistency; no semantic change.
5927e65 to
1c13656
Compare
(ai: vllm-qwen/qwen3.8-27b) Merge conflict resolved again — branch rebased onto current master tip, conflict cleared, live-verified. What I did this run:
No code changes beyond conflict resolution, no PR-body edits, no gates hit, no new issues filed. Both commits (the fix + the position fix) are preserved as-is on top of master. 七问:
|
|
(ai: vllm-qwen/qwen3.8-27b) Review verdict: APPROVE (re-review after rebase; posted as a thread comment — GitHub rejects a formal approve review from the PR's own author account) The branch was rebased onto the current master tip ( Rebase fidelity (the actual risk of this update)
Master-advance interaction check (old base..new base on the PR's files)Master advanced ~10 merges in the window (#6150/#6166/#5991/#5809/#5609/#6132 …). On the 5 source files this PR touches, the advance consists of: the experimental fold-point feature removal (#6132 — dead-code excision of Gates on the new head
No new findings; nothing to fix on the spot this round. 结构化八问(新 head 复审)
收工判定可合——新 head |
(ai: vllm-qwen/qwen3.8-27b) Review lane complete (round 2): GitHub PR #5935 re-audited end-to-end on the rebased head |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Review lane closed out for GitHub PR #5935 — final state: Verdict: APPROVE, ready for human merge.
Nothing further pending on my side. |
Part A of #5928 — visibility for the blind nudge baseline (no change to fold-decision behavior).
Background (#5928)
After a resume-inheritance miss (
[resume-inheritance] miss: no byte-exact resume among N eligible chain(s)), the conversation lands on a fresh session key whose meter reads 0 — no usage report yet this turn, no prior-turn estimate — whilependingT1already sits far over its T1 threshold (reporter's case:usage=0% (0/356250) ... pendingT1=260393/50000). The nudge decision is structurally unfireable for that turn (every pressure band, and thefirstSightMassReadymass bypass from #194, is usage-gated), so one oversized forward goes out with ~473K prompt tokens sent uncompressed; the state self-heals on the next turn via the #728 estimate fallback. Today that turn only leaves an ordinaryinfoline, so nothing in the log explains why nothing was folded.This PR is Part A of a two-part fix: make the blind state visible. Part B (root cause: let
firstSightMassReadydrain a known backlog when the meter is blind) is a kernel change awaiting owner approval (posted as 🔶 需要决策 in the issue thread); when approved it lands on this same branch — one PR per thread.Changes
diagNudge()appends, baseline=unknownwhen the meter read 0, so the idle line states WHY nothing fired. Byte-identical for measured turns.nudgeDiagLogLevel(): the exact state (blind meter +pendingT1 >= interval+ not injecting) logs atwarninstead ofinfo; every other state keeps its level. Applied at all four prepare lanes (openai / responses / google / anthropic)./__bili/statsgains a per-session booleannudgeBaselineUnknown(lastInputTokens <= 0 && localInputEstimate <= 0)./__bili/statusstays process-level (it has no per-session list) and is untouched.Behavior changes (disclosed)
nudge idle:log line, blind meter + pendingT1 ≥ intervalinfo, no baseline tagwarn,... pendingT1=260393/50000, baseline=unknown model=.../__bili/statssessionsnudgeBaselineUnknown: boolNo config-surface changes (no new keys / env vars / flags). No wire or message-shape changes. No kernel changes in this commit.
Tests
tests/issue5928-nudge-baseline-visibility.test.ts: truth table for the warn gate with both-direction assertions (Test Blind Spot Retrospective: Reinforcing Negative Assertion Discipline, Tightening Fake Upstream Strictness, Thin Inst | 测试盲区复盘:补反向断言纪律、假上游严格化、薄安装冒烟、形态矩阵与生命周期覆盖 #3617 discipline) — warns exactly when blind + backlog ≥ threshold (boundary included); stays info for measured meters, small backlogs, injecting turns, and missing/degenerate breakdowns — plus the message-shape pin (the tag appears iff the meter read 0).tests/fork-nudge-trigger.test.ts+tests/issue4383-stale-usage-anchor.test.ts+tests/issue2155-self-heal.test.ts= 24/24;npm run typecheckclean. Full-suite verdict comes from CI on the pushed head.Related to #5928 (Part A; the root-cause Part B is pending owner approval and will extend this branch rather than open a second PR).