Skip to content

[PR] fix(relay): the egress fuse 413 says "declared model window", not "effective" — one word, two numbers - #7123

Open
ranxianglei wants to merge 2 commits into
masterfrom
2026-11-18_fuse-window-caliber
Open

ranxianglei wants to merge 2 commits into
masterfrom
2026-11-18_fuse-window-caliber

Conversation

@ranxianglei

@ranxianglei ranxianglei commented Nov 18, 2026 •

Copy link
Copy Markdown
Owner

What

Field data from #5172 (DSH 0.2.2-rc.3 + bili 0.1.215): within the same request, nudge judges against 792,576 (the reserved window) while the egress fuse's 413 message judges against 1,048,576 (the declared/unreserved wall) — a gap of exactly 256,000 (= the max_tokens reservation). The 413 calls 1,048,576 "the effective model window" — but "effective window" in [headroom]/nudge/preflight means the lower RESERVED judging window. One word, two numbers.

Layer judgment

The fuse using the looser unreserved wall is the INTENDED #4205 split (gap bodies forward → upstream 400 → refold rescue), not a bug. The real defect is the 413 mislabeling the unreserved wall as "the effective model window", conflating it with the [headroom] reserved caliber.

Fix

Log/observability only — no behavior change (the fuse still judges declaredWindow, nudge/preflight still judge reservedContextLimit).

Regression

  • tests/issue4079-egress-window-fuse.test.ts: pins "declared model window" + "unreserved upstream wall" and asserts the ambiguous "effective model window" is gone.
  • tests/issue5191-arm-judge-line.test.ts: pins the new [headroom] fuse clause.

This PR fixes the FUSE-path instance of the "effective window" word-referencing-two-numbers family (the floor-3 field data in #5172). The original preflight-deadlock root cause in #5172 remains open and is tracked separately.

Related to #5172

…fective" — one word, two numbers (#5172)

The 413 egress-window-fuse message called the UNRESERVED wall
(metadata.effectiveContextLimit) "the effective model window", but
"effective window" elsewhere ([headroom]/nudge/preflight) means the lower
RESERVED judging window. One word, two numbers: the nudge line (judges
792,576) and the 413 (judges 1,048,576) look like a contradiction but are
the intended #4205 split (the fuse rides the unreserved wall; gap bodies
forward → 400 → refold rescue).

The looser fuse bar is intended, not a bug — the defect is the mislabel.
Relabel both 413 messages ("declared model window (the unreserved upstream
wall — nudge/preflight judge against the lower reserved window, see
[headroom])") and add the fuse's caliber to the [headroom] line.

Log/observability only; no behavior change (the fuse still judges
declaredWindow, nudge/preflight still judge reservedContextLimit).

Fixes #5172
@ranxianglei ranxianglei changed the title fix(relay): the egress fuse 413 says "declared model window", not "effective" — one word, two numbers [PR] fix(relay): the egress fuse 413 says "declared model window", not "effective" — one word, two numbers Nov 18, 2026
@github-actions

github-actions Bot commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-11-18_fuse-window-caliber (3fdafc1)

Latest build — install this exact version

npm install -g billion-context@0.0.10821-pr.7123.44c6f7c

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Recent builds (commit → version)

commit version build
3fdafc1 billion-context@0.0.10821-pr.7123.44c6f7c build
e84d300 billion-context@0.0.10748-pr.7123.c1cb9da build

Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr7123.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

Verified PR #7123 against the live branch (2026-11-18_fuse-window-caliber, e84d300). The fix is correct and log/observability-only; its own regression tests pass; the single red lane is an unrelated Windows flake that I've re-run.

What I checked

Targeted tests (local, on the PR head)

  • tests/issue4079-egress-window-fuse.test.ts + tests/issue5191-arm-judge-line.test.ts: 5/5 pass.

CI (head e84d300)

  • 51 lanes: 37 success, 14 skipped, 0 real failures — except one: test (windows-latest, 22, 5) failed on tests/server-process-listeners.test.ts (file-level testCodeFailure, exit 1; its subtests pass, the process exits non-zero).
  • That test is the recently-landed [PR] fix: release process handlers when an embedded server closes #7026 signal-handler / graceful-shutdown test (sends real SIGINT/SIGTERM, stubs process.exit) — a different subsystem this PR never touches. It's a flake: master's run of the same base commit (4b77eab) is fully green, including that Windows lane.
  • Re-ran the failed job (run attempt 2, queued) to unblock the merge.

Conclusion: PR #7123 is correct and ready to merge once the re-run goes green. It fixes the #5172 reader trap (one word, two numbers) with wording only — no behavior change.


结构化八问

  1. 是什么问题:preflight and [headroom] disagree on the effective window, and preflight folding silently creates no block #5172 里同一个请求,nudge/preflight 按预留窗口(792,576)判,而 egress fuse 的 413 文案却把未预留墙(1,048,576)叫成 "the effective model window"——一个词两个数,读者会误以为两者矛盾。
  2. 什么原因导致的:413 文案措辞不精确。fuse 用未预留墙判是 "Output reservation extended to all wires" (#4184) landed without test coverage: widespread red on master is blocking th | 「输出预留扩到所有 wire」(#4184) 落地没带测试面: master 大面积红卡住 v0.1.204 发版 (已 revert, 重做要求见内) #4205 的故意设计(gap body 放行→上游 400→refold 救援),不是 bug;真正的缺陷是文案把"未预留墙"标成了 "effective window",与 [headroom] 的预留口径混用。
  3. 影响范围与严重程度:纯日志/可观测性——只影响 413 报错文案和 [headroom] 日志行,不影响任何判定逻辑、转发行为、线格式。体验瑕疵级(读者困惑),无功能影响。
  4. 怎么修复的:只改文案。两条 413 fuse 消息 "the effective model window" → "the declared model window (the unreserved upstream wall — nudge/preflight judge against the lower reserved window, see [headroom])";[headroom] 行补一句 "the egress fuse sizes against N (the unreserved wall, "Output reservation extended to all wires" (#4184) landed without test coverage: widespread red on master is blocking th | 「输出预留扩到所有 wire」(#4184) 落地没带测试面: master 大面积红卡住 v0.1.204 发版 (已 revert, 重做要求见内) #4205)",让两窗口拆分一行可见。
  5. 是否有回归问题:无。判定逻辑零改动(fuse 仍判 declaredWindow,nudge/preflight 仍判 reservedContextLimit);本地定向测试 5/5 过;CI 除一个无关 Windows flake 外全绿。
  6. 是否破坏缓存:不涉及(纯文案)。
  7. 是否改了设计方向:不涉及(纯文案,不动架构/边界/数据流)。
  8. 是否引入了新的配置:不涉及。

判定:PR 正确,可合(等 flake 重跑转绿)。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

Correcting my earlier report — the Windows lane is NOT a one-off flake. The re-run (attempt 2) failed again, and the CI's own diagnostic says "deterministic break, not a flake". I dug in; here's the accurate picture.

The red is not this PR's

It's a master-level born-red on Windows, hitting many PRs

  • The exact same test fails with the identical signature in other unrelated PRs: runs #44245921438, #44246105190, #44249183921 — all on windows-latest, 22, 5.
  • ci-fixes has no fix for it (and is already an ancestor of this PR's head, so merging it is a no-op).
  • The automated tracker for this exact failure is issue Scheduled full test failed: 0e475ceb | 定时全量测试失败: 0e475ceb #7108 ("Scheduled full test failed: 0e475ce", windows-22-5 + windows-24-5). I've posted my findings there.

Mechanism: the #7026 test's 4th case sends real SIGINT then SIGTERM to the process and stubs process.exit; on Windows these signals are emulated and the file exits non-zero deterministically. That's a master/ci-fixes concern (Windows-compat in the test, or a genuine shutdown-code bug), not this PR's.

Conclusion: PR #7123 is correct and its own tests pass (5/5). The CI red is an inherited master-level Windows failure in the just-merged #7026 test — not caused by this PR. It's mergeable on its merits; the Windows lane stays red until the #7026 test is fixed on master/ci-fixes (tracked in #7108).


结构化八问

  1. 是什么问题:[PR] fix(relay): the egress fuse 413 says "declared model window", not "effective" — one word, two numbers #7123 本身正确(纯日志文案,自有测试 5/5 过)。CI 红是 windows-latest, 22/24, 5 分片在 tests/server-process-listeners.test.ts([PR] fix: release process handlers when an embedded server closes #7026 优雅关闭/信号处理器测试)上确定性失败——文件级 exitCode 1,隔离复现,全量重试仍红。
  2. 什么原因导致的:[PR] fix: release process handlers when an embedded server closes #7026(commit 894835a)今天 01:46 UTC 才合入 master,其测试第 4 例向进程发真实 SIGINT/SIGTERM 并 stub process.exit,Windows 上信号是模拟的,文件确定性非零退出。属 master 自身 born-red(Windows),非 [PR] fix(relay): the egress fuse 413 says "declared model window", not "effective" — one word, two numbers #7123 引入。
  3. 影响范围与严重程度:所有在 windows-latest 跑 test 分片 5 的 PR 都继承此红(已确认 #44245921438/#44246105190/#44249183921 同一测试同一签名);自动跟踪在 Scheduled full test failed: 0e475ceb | 定时全量测试失败: 0e475ceb #7108。阻塞 Windows 分片合入,但 Linux 分片与 PR 自身逻辑不受影响。
  4. 怎么修复的:[PR] fix(relay): the egress fuse 413 says "declared model window", not "effective" — one word, two numbers #7123 无需修(没碰该测试/信号代码)。真正修复属 master/ci-fixes 范畴(测试 Windows 兼容化,或 shutdown 代码真 bug),已在 Scheduled full test failed: 0e475ceb | 定时全量测试失败: 0e475ceb #7108 留结论;按政策不在本 PR 分支修(master 级修复收敛到 ci-fixes/哨兵)。
  5. 是否有回归问题:[PR] fix(relay): the egress fuse 413 says "declared model window", not "effective" — one word, two numbers #7123 无回归(纯文案,判定逻辑零改动)。CI 红是继承的 master 级问题,与本 PR 无关。
  6. 是否破坏缓存:不涉及。
  7. 是否改了设计方向:不涉及。
  8. 是否引入了新的配置:不涉及。

判定:#7123 代码正确、自有测试全绿,可合;CI 红为继承的 master 级 Windows 问题(#7026 测试,今日合入),跟踪在 #7108,非本 PR 应修。

…ndows red

The only red (test windows-latest/22/5) was inherited: this branch's base 4b77eab predates master's 3186e54 (fix(tests): make the shutdown-signal test portable to Windows). Same signature as #4599 floors 1553/1554/1560. No PR-file changes; targeted local: server-process-listeners 4/4 + PR test files 5/5, typecheck clean.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

Inherited red cleared — no code change needed. The only failing check (test (windows-latest, 22, 5)) was not caused by this PR's diff; I merged current master into the branch and pushed the unmarked final head.

What I did this run:

  1. Re-verified live state first (mirror cache + PR API): at head e84d3006b0 exactly one check was red — test (windows-latest, 22, 5); the other 28 were green/skipped.
  2. Pulled the job log. The failure is tests/server-process-listeners.test.ts: file-level exitCode 1 / ERR_TEST_FAILURE with all subtests passing — the TAP stream dies after subtest 2, and the runner's own isolated serial re-run reproduced it deterministically ("REPRODUCED in isolation (deterministic break)").
  3. Diagnosed the root cause. This branch's base 4b77eab409 is exactly one commit behind master's tip 548e396022, and the missing commit is 3186e54ad2 — "fix(tests): make the shutdown-signal test portable to Windows (it killed its own child on win32)". The test it fixes (added by 894835ac94, the [PR] fix: release process handlers when an embedded server closes #7026 review commit, which IS in this branch) assumed POSIX signal semantics: on Windows process.kill(self, SIGINT) terminates the child via TerminateProcess without running the handlers, so the CI child dies mid-file. This is the exact signature already documented in CI Control Center: Centralized CI Failure Handling & Status Sync (Comment all CI issues here; do not open new issues) | CI 总控: CI 失败集中处理与状态同步(所有 CI 问题在此评论, 勿开新 issue) #4599 (multiple PRs red in the same window for the same reason).
  4. Applied the standard inherited-red convergence (same as the CI Control Center: Centralized CI Failure Handling & Status Sync (Comment all CI issues here; do not open new issues) | CI 总控: CI 失败集中处理与状态同步(所有 CI 问题在此评论, 勿开新 issue) #4599 precedent for this signature): git merge origin/master (548e396022) into the PR branch — clean, no conflicts (the PR's 4 files don't overlap the master window's hunks) → new head 3fdafc1e3b.
  5. Targeted local verification before pushing: tests/server-process-listeners.test.ts 4/4 + this PR's two test files (issue4079-egress-window-fuse, issue5191-arm-judge-line) 5/5 = 9/9 green; npm run typecheck clean.
  6. Pushed 3fdafc1e3b as the unmarked final head (this branch's one full CI pass) and left a one-line status in CI Control Center: Centralized CI Failure Handling & Status Sync (Comment all CI issues here; do not open new issues) | CI 总控: CI 失败集中处理与状态同步(所有 CI 问题在此评论, 勿开新 issue) #4599.

中文结构化八问

  1. 是什么问题:本 PR 唯一的红是 windows-latest/22/shard5 上 tests/server-process-listeners.test.ts 的确定性失败(文件级 exit 1、子测试全过、TAP 流在第 2 个子测试后中断)——继承红,不是本 PR 代码的问题。
  2. 什么原因导致的:分支 base 4b77eab409 恰好早于 master 的 Windows 修复 3186e54ad2 一个提交;被修的测试(由 894835ac94 新增,已在分支内)假设 POSIX 信号语义,而 Windows 上 process.kill(self, SIGINT) 直接终止子进程、不执行 handler,CI 子进程因此中途死亡。
  3. 影响范围与严重程度:仅影响 base 早于 3186e54ad2 的 PR 的 CI 红绿判定(同窗口多个 PR 同签名红);产品运行行为零影响。
  4. 怎么修复的:不改代码——把 origin/master(548e396022,含 3186e54ad2)merge 进 PR 分支,新头 3fdafc1e3b,即 CI Control Center: Centralized CI Failure Handling & Status Sync (Comment all CI issues here; do not open new issues) | CI 总控: CI 失败集中处理与状态同步(所有 CI 问题在此评论, 勿开新 issue) #4599 同签名既定的继承红收敛路径。
  5. 是否有回归问题:无。merge 只带入 master 窗口两个提交(687941b83d side-lane 降级 + 3186e54ad2 测试修复),与本 PR 的 4 个文件无文本冲突;本地定向 9/9 绿 + typecheck 干净。
  6. 是否破坏缓存:不涉及。
  7. 是否改了设计方向:无。
  8. 是否引入了新的配置:无。

判定:红已清除,3fdafc1e3b 全量 CI 重跑中;转绿后本 PR 即可进入 review/合并流程。

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant