Skip to content

[PR] chore(batch): land 29 open fragmented fixes in one history-preserving batch - #6482

Merged
ranxianglei merged 79 commits into
masterfrom
batch4-dry
Nov 14, 2026
Merged

ranxianglei merged 79 commits into
masterfrom
batch4-dry

Conversation

@ranxianglei

@ranxianglei ranxianglei commented Nov 13, 2026 •

Copy link
Copy Markdown
Owner

chore(batch): land 29 open fragmented fixes in one history-preserving batch

29 small, mostly single-file fix PRs from 21 external authors, line-by-line
reviewed together, cherry-picked with -x (original authorship and messages
preserved), verified as one unit. No features; every change is a bug fix, a
diagnostic improvement, or a regression pin.

Contents

PR One-liner
#4263 (Zhishulo) DashScope "Range of input length" 400 classified as context-overflow so the probe path can arm (#4262)
#4386 (omovomo) Fleet inspector: CRLF tool labels + visible-width background bar
#4683 (Maziar123) Side-window inherit: child conversation sized from parent runtime info, not the parent wall (#4682)
#4756 (z0d0dOr) Release input-only request data (parsed body, affinity cache) before the upstream stream starts; retry bodies stay intact (#4753)
#4861 (yanhuojiyusheng) Handoff: persist the completed assistant reply when the client closes on the terminal frame — canSave no longer coupled to res.destroyed (#4860)
#4875 (j0nsh1n) Edited user instructions re-enter the wire unchanged: user-role anchors exempt from order/simhash reconcile tiers (#4874)
#4935 (fathonix) Bare announcement-tail terminal turn retried once with a continuation nudge; second degeneration emits an in-band error (#4944)
#5027 (adlternative) pi direct-send fast path gains the window guard — proxy outage can no longer egress an oversized history (#5025)
#5074 (elight) Tool-less structured-output requests forwarded as side requests, not main-lane turns (#5073)
#5117 (elight) hermes plugin: plain-http provider base_url routed through the proxy (#5116)
#5166 (mzbgf) kernel: sync-mode block liveness decided by coverage shape, not existence
#5177 (TimBriseby) Nested transport timeout: 504 cause chain no longer masquerades as the client idle timeout (#5176)
#5384 (LkxPro) Proxy tool failures classified by the status line only — a summary quoting an earlier FAILED result no longer flips anyCompressFailed
#5482 (Maziar123) 413 diagnostics distinguish window-short from hard provider size limit, bounded cause string
#5615 (xulongwu4) Responses-WebSocket cache diagnostics (sha256 fingerprints only, no payloads)
#5709 (Louie0806) Single-shot dsh sidecar lane requires a bounded one-shot body — a tool-less MAIN turn-1 residual is no longer mistaken for a sidecar (#5687)
#5712 (yanhuojiyusheng) pi manifest entry → dist/agent/billion-context.js re-export shim + identity pin
#5743 (mariopasquali) opencode: repeated-content message IDs continued across delta turns from both seams (#5741)
#5954 (mariopasquali) Fold reap keeps position fingerprints (foldPositions ids/canon) in lockstep (#5953)
#6103 (atrokhym) Google: stripped proxy-call thoughtSignature promoted onto the surviving sibling functionCall (#6079, #5252)
#6110 (gongfpp) dsh 0.2.2 auto-review sidecar shape recognized (#6104)
#6113 (SajjadPSavoji) OMP ≥18.18: default PI_ANTHROPIC_MESSAGE_THREADS=0 with a warn-once beta notice (#6088)
#6190 (aftabkh4n) Windows: double trailing backslashes when quoting for cmd.exe (MSVCRT eats 2n→n); drop the odd-run refusal
#6300 (vladk1m0) OMP owns truncation recovery: raw truncation error back to the host for a budget-increment retry ≤10, never for context-overflow (#5951)
#6340 (myNameIsZhangYu) Five-part context PR: OpenAI omitted cache fields ⇒ 0 cached tokens (#6329); atomic directory swap for in-place updates (#6309); auto_vacuum before WAL (#6323); upsert skips unchanged large columns + regenerated goldens (#6325); archive exports serialized + newest dump wins restore (#6317)
#6362 (sorasoras) Generation-boundary detection also arms the native-compaction ref-ledger rebase (reuses the #4931 mechanism)
#6428 (raydocs) kernel: heal unpaired server_tool_use / *_tool_result blocks in coreToAnthropic before forwarding (#4954)
#6461 (xiaozhu1337) responses functions.exec namespaces: bili-injected tools get a "call directly" note; host tools byte-exact (related to #5592 — not auto-closing it)

Excluded: #5028 (same-PID container-generation gate for served-markers) —
master 7a923f8 (2026-11-07) already landed the identical #3923 fix
(SAME_PID_GENERATION_MARGIN_MS), so the PR's net delta is zero; it will be
closed as superseded.

Excluded (reverted in this branch): #6345 (guowenjie1992, kernel: span-safe
gap bridging at the mega-turn merge gate) — it breaks two pinned invariants
under the CI matrix (node 22/24): #1592 fold round-2 renders a double
fold-summary ("mid-history rewrite") instead of the consolidated shape, and the
#2265 e2e loses its folded block across the model-switch re-stamp. Isolated by
kernel bisection (reverting just this commit turns both green on node 22 with
dist-verified builds). Reverted here as 7d7631f so the batch can land; #6345
stays open for follow-up (a comment was left on the PR).

Merge notes (deviations from the raw PR diffs)

All 29 PRs cherry-picked -x oldest-first. Substantive unions/adaptations:

Follow-up commit (CI fix, 2026-11-13)

Verification

  • npm run typecheck: clean.
  • Touched tests: all 51 touched/added test files green (1239 tests,
    1206 pass; the 3 reds are host-detection tests that fail on master in this
    environment too); kernel suite 39/39; pi-subagents 19/19;
    golden-storage-render 64/65 (the 1 red fails on master too).
  • Per-PR added-line audit: every source line added by each PR is present on
    this branch (documented exceptions: superseded hunks listed above, three
    reworded comment lines, and the skipped version-bump lines).
  • Full GitHub CI (all platforms, full suite) runs on this branch.

Closes #4262, #4682, #4753, #4860, #4874, #4944, #5025, #5073, #5116, #5176, #5252, #5386, #5687, #5711, #5741, #5953, #6079, #6088, #6104, #6309, #6325, #6329

(#5531, #5951, #6317 are already closed by earlier batches; #5592 is deliberately
not auto-closed — #6461 only addresses its proxy-prompt gap.)

kernel-change: land two reviewed kernel fixes from external PRs — #6428 (heal unpaired server_tool_use/*_tool_result in coreToAnthropic, #4954), #5166 (sync-mode block liveness by coverage shape); line-by-line reviewed, kernel suite green locally and in CI. (#6345's kernel change was excluded — see Excluded.)

Zhishulo and others added 30 commits November 13, 2026 10:28
…probe path can arm (#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
#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 #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 #2313 e on the probe path
(fails before this change with the 400 passed through, passes after).

(cherry picked from commit 931a100)
…#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 (#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.

(cherry picked from commit 1fafccf)
The handoff snapshot was gated on the client socket still being writable
(relay passed canSave=() => !res.destroyed, and both pipes required
!res.destroyed before reply.save()). Clients that close the connection the
instant they hold the terminal frame — pi does exactly this — therefore lost
every assistant reply: the pipe threw 'This operation was aborted' after
sawTerminal=true, took the client-abort branch and returned without saving.

Persistence now answers 'did the upstream turn finish?', not 'is the socket
still open?': the pipes save on sawTerminal/replyCompleted even from the
client-abort path, and the relay no longer couples canSave to res.destroyed.
Partial turns (no terminal) are still never captured.

Refs #4247

(cherry picked from commit 63e13be)
Restore formatting-only mass re-anchoring by requiring the unwrapped candidate to match the stored normalized identity. Keep user messages out of approximate matching and preserve changed punctuation and instruction order. Pin the non-user order-bracket path and quiet folded resends.

Implementation drafted by Grok 4.6; design, review and verification by Codex.

(cherry picked from commit c5279e6)
Align-refactor (review of #4935): this PR is scoped to the bare-announcement
terminal-turn retry (#3905 symptom), but it also carried the LoopCtx
streamMarkers field and the emitMarker consumption gate that belong to the
streamMarkers setting feature. That feature has its own PR (#4934, which
wires the config surface but lacks these two hunks) — keeping them here makes
between two independent features. The hunks move to #4934 so each PR is a
closed loop.

(cherry picked from commit fa60f8f)
…he PR body describes

Align-refactor (review of #4935): the PR body promises detection is
'accompanied by common action announcement verbs' and requires 'no markdown
code fence open', but the shipped detector was tail-char + length only — any
turn under 400 chars ending in ':' or ellipsis qualified, including legitimate
short answers ('The answer is 42:') that would now cost an extra billed
completion call each. Implement both promised gates: an ANNOUNCE_CUE word list
(progressing forms + intent phrases only, mirroring the body's examples) and a
reject on odd backtick-fence count (an unclosed fence is the #1755/#2190
truncation class, not a stall). English-first by design: non-English
announcements keep the pre-#3905 no-retry behavior until traffic evidence
justifies extending the cue list.

(cherry picked from commit aec1e1b)
…ling; pin the 400-char cap

Align-refactor (review of #4935): the suite only covered unit + compress-loop
lanes while the change also rewires pipePluginChatWithStrip and
pipePluginResponsesWithStrip terminal classification — the #2303 sibling
suite covers all three lanes, so match that bar: plugin chat positive
(#3905 log line, one [DONE], single finish_reason stop), colon-answer and
ellipsis-answer negatives, tool-call exclusion, retry-degenerates in-band
error (#870); responses-pipe mirror set. Also fix at-l3, whose 'long answer'
fixture did not actually end on the tail char and passed for the wrong reason
— the rewritten fixture (441 chars, cue-bearing, colon tail) pins the length
cap itself at loop level. Drop dead imports.

(cherry picked from commit 206ddc9)
…roseAcc (#3905)

proseAcc accumulates every streamed text field including the thinking/
reasoning channel. endsWithAnnouncement(proseAcc) therefore saw the tail of
the model's internal monologue: a thought ending mid-sentence on ':' or an
ellipsis followed by a whitespace-only answer was reclassified as a bare
announcement (the new #3905 branch) instead of the #2756 degenerate shape —
the retry still fired but under the wrong log line, and a whitespace-only
turn on an unclean finish lost its 'not retried ... not a clean stop'
refusal log entirely (stalledTail suppressed it). Two subtests in
tests/issue2756-degenerate-gate.test.ts went red on every shard-3 CI leg.

Operand swap to answerProse in both plugin pipes (chat + responses). The
compress-loop twin already tests text-only assistantText and is unaffected.

(cherry picked from commit 76c98b8)
…not egress an oversized history (#5025)

Without a live proxy, directSendOrRefuse measured the raw body against a
window it read from the stamped x-bili-plugin-context-window header or from
state.declaredWindowForDirect. On the pi lane BOTH were empty on exactly that
path: before_provider_headers returns at `proxyBase === undefined` before it
stamps the header, and the lane never armed the slot (#4894 armed dsh only).
The check then failed open by construction, so a long session's unfurled
history (multi-million tokens) went straight to the provider, whose
"input exceeds the context window" 400 repeated on every retry and made the
session unrecoverable for as long as the proxy stayed down.

Record the per-request ctx.model.contextWindow at the TOP of the headers hook
(before the readiness return) and arm it as declaredWindowForDirect in
pi-native.ts. An unknown window records undefined — the guard keeps failing
open instead of guessing — and a healthy proxy is unchanged, since the
stamped header still outranks the reader.

Verification: tests/issue5025-pi-direct-window.test.ts (5 cases: recording
before the early return, unknown-window reset, the armed reader, local
oversize refusal with no upstream egress, and no false refusal when the body
fits); typecheck; issue4894/pi-native/issue3433/issue2072/fix-1897 suites;
unused-export and test-import-graph gates.

(cherry picked from commit cfcb3f9)
A request with no tools that asks for a JSON reply (response_format
json_schema/json_object, or text.format on the Responses wire) is a
utility call, not an agent turn. Child processes of a host inherit the
proxy through HTTP(S)_PROXY, and adding the ACP tools to their
extraction calls makes backends that refuse tools plus response_format
answer 400. Tool-carrying requests and an explicit main persona are
still main turns.

Fixes #5073

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit ef508e9)
The helper is only called by isSideRequest() inside the same module; no
other file imports it, so the unused-exports gate rejects the leading
export. Behavior is unchanged.

(cherry picked from commit 7a29d54)
Review supplement on 7a29d54 (#5074): every classification rule in this file carries its issue number as a comment anchor (#2500/#2503/#3149/#4438); the two comments added by ef508e9 did not, so the rule's provenance was not greppable. Comments only - no behavior change.

(cherry picked from commit 17e82b7)
Ship the classification as the exported isStructuredOutputCall predicate with
the #546 no-tools veto built in, so handle() can consult it independently of
isSideRequest and skip restoreOutputBudget for the call — its 64000 budget
sizes a utility response, not a main turn, and must not seed
outputBudgetHighWater (#1897/#2500 class).

Give the #388 side passthrough an honest log label, and extend the e2e to
assert kernel state, stats, snapshot, request counter and high-water are all
untouched, mirroring the #2500 suite. The new high-water assertion fails
without the restore skip and passes with it.

(cherry picked from commit a828e59)
…5116)

The native plugin set only HTTPS_PROXY, so a plain-http base_url (a LAN
llama.cpp, vLLM or ollama server) bypassed the proxy and its sessions were
never compressed. apply_env now sets the HTTP pair too and defaults NO_PROXY
to loopback, matching the launcher's own plain-http routes. All of them sit
in _ENV_ROUTED_KEYS, so the #4644 watchdog unroutes and restores them along
with the HTTPS pair.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MMUmTHLFoNYr4iAsprTFfk
(cherry picked from commit 461504c)
A compressed block's summary may only be injected while it still describes
content on the wire. The current predicate asks whether ANY covered message is
still present, so a host rewind that truncates part of a span keeps the whole
summary alive and the model reads a description of history that no longer
exists. Reap (#2695) does not catch it either: it needs a strict majority of
covered ids absent for three consecutive passes, and a partially truncated span
keeps its majority.

The gap's shape, in the block's own coverage order, is the discriminator:

- survivors a prefix, missing the newest end   -> the rewind shape; deactivate
- survivors a suffix, missing the oldest end   -> the host dropped old history;
  the summary is the only record of it, so keep it active
- interior/ragged holes                        -> the id-churn shape (ids are
  content hashes); keep it and let fold-reconcile re-key the churned ids
- nothing covered visible                      -> defer to the positional
  witness (#4316, unchanged): a surviving ledger position is the churn /
  merged-pair shape, not a deletion

Everything except the tail gap keeps today's keep-alive behaviour, so the only
behaviour that changes is the shape this change names. The positional witness
is not touched: narrowing it to exact coverage would false-kill merged content
(a pair collapse covers fewer ids than positions).

Tests: sync-config gains the four shapes plus the positional-witness
composition and the no-witness deletion case.

(cherry picked from commit 7c4d4b9)
… rejection cause

Review follow-up on #5482. bound() collapsed only C0 + DEL before clamping,
so C1 control characters survived into the appended log suffix — in particular
U+0085 (NEXT LINE) and U+009C (file separator), which render as line/field
breaks. A relay that echoed such a byte into attempts[].details.error.message
could forge a second log field, contradicting the module's own "no newline may
forge a second log field" invariant (pinned by the \r\n test). Widen the strip
to C1 (U+007f-U+009f) and extend the control-shape regression test to pin NEL
and file-separator. No behavior change beyond the log-line sanitization.

(cherry picked from commit c749eb8)
…ile existence (#4599)

The receiver fixture creates its .port file before the write lands; on a loaded runner the existsSync poll can win that race and read back an empty file, so Number("") === 0 tripped the assertion (CI: ubuntu-24 shard 2, run 42202673910 — second occurrence of this signature after PR #5327's one-off, recorded in #4599 as record-only). Poll until the content parses as a positive integer instead; the 15s deadline and the diagnostic message are unchanged.

(cherry picked from commit 8d283b9)
Correlate exact egress fingerprints, socket generations and send attempts with terminal usage-field presence. Preserve reported numeric zero versus absent cache usage without logging raw payloads or headers. This adds diagnostics, not a provider cache-miss fix.

(cherry picked from commit 1d59e82)
request_sha256 serialized and hashed the whole effective request a second
time on every send, including delta sends. A synthetic benchmark on
Node 24 put that at about 2.4 ms per MB of body, run synchronously.

The field added little. On full sends it hashed the same text as
frame_sha256. On delta sends, ResponsesWsHistory.continuation() has
already matched the settings and every prior input item, so mode=delta
attests the same thing. The hash also changes every turn, so it cannot
show prefix stability.

The remaining send fields are unchanged. The test now pins the exact
key set of the send record.

(cherry picked from commit 30ee96b)
The generic single-shot gate added in #5687 matches "no tools + one user
message + any system text", and both of its call sites opened on
`agent !== "main"`. dsh's native client never stamps `x-bili-plugin-agent`,
so that header reads `undefined`, and `undefined !== "main"` is always
true: the FIRST unstyled main turn of a conversation matched the sidecar
shape and rode the #388 verbatim side lane. The kernel never saw it, so the
durable-state carriers (skills, workspace instructions, MCP catalog,
runtime context) never received their BLOCKED refs (#2419/#2446/#2447/#2481),
and the wire-byte contracts of #2817/#5418 reset.

A genuine utility sidecar is distinguishable from that turn in two ways, and
this patch requires both:

- it pins a TINY fixed output budget (prompt-enhance sends 4096), whereas a
  main main turn either omits the field (openai wire) or sizes it from the
  conversation (#5656's own residual turn sends 64000) -- hasSingleShotBudget;
- it never carries a dsh durable-state marker -- carriesDurableState, the
  same marker table fold-reconcile.ts already screens with.

The #5656 e2e residual that pinned "tool-less turn 1 rides the side lane" is
revoked and re-pinned to the corrected contract; the side-lane boundary test
allowlists the new predicate like every other pure shape predicate.

Tests: issue2419/2446/2447/2481 (4/0 each), issue2817 (1/0), issue5418
watermark (1/0), issue5656 (5/0), side-lane-invariants (15/0), issue5658
(2/0). Negative control: reverting only the two source files re-reds 2446
(3/1), 2817 (0/1), 5418 (0/1).

(cherry picked from commit 427778b)
This was referenced Nov 13, 2026
…native import cannot leak INLINE_ORIGIN egress

The #5025 test statically imports pi-native.ts, which arms at module load.
Since the #5428 inline default flip, plan.mode is 'inline' with no explicit
env/file choice, so armNativePi registers the origin waiter with INLINE_ORIGIN.
The waiter persists into the shared factory's awaitNativeProxyOrigin(), which
makes the before_provider_headers handler resolve proxyBase=INLINE_ORIGIN and
call registerTools -> fetchManifest, egressing to the inline origin inside the
hermetic suite. Add a _resetNativeOriginWaiterForTest hook (native-bootstrap)
and call it in the suite's beforeEach so the waiter is cleared per test.

Fixes the 6 red test lanes (issue5025-pi-direct-window.test.ts 1630/1631) on
head e484c03.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Fixed the 6 red test lanes on head e484c0322f — pushed d04a8adcb9 to batch4-dry.

Root cause: All 6 lanes fail the same 2 tests in tests/issue5025-pi-direct-window.test.ts (1630, 1631). The tests expect no egress to the inline origin, but an egress to http://billion-context.inline/__bili/plugin/manifest happens. I traced it: the test statically imports pi-native.ts, which arms at module load. Since the #5428 inline default flip, plan.mode is "inline" with no explicit env/file choice, so armNativePi registers the origin waiter with INLINE_ORIGIN. The waiter persists into the shared factory's awaitNativeProxyOrigin(), which makes the before_provider_headers handler resolve proxyBase=INLINE_ORIGIN and call registerTools → fetchManifest, egressing inside the hermetic suite.

Fix: Added a _resetNativeOriginWaiterForTest hook to native-bootstrap.ts (clears the module-level origin waiter) and called it in the test's beforeEach. Verified locally: all 5 tests pass, related suites (issue4894-direct-window-guard, pi-native) pass, typecheck clean.

Still owed from the earlier wake: the kernel-guard authorization checkpoint. The PR body already carries the kernel-change: declaration (commit 91db32defe) and the kernel version is bumped to 0.0.197 (1b15692af4), so the CI job should pass on the new head — but the kernel changes are human-gated and need your explicit authorization. I have NOT edited the PR body for this.

CI status: runs on d04a8adcb9 are still queued (CI backlog). I will confirm green in a follow-up once they complete.


八问

  1. 是什么问题:6 条 test 车道全挂在同一文件 tests/issue5025-pi-direct-window.test.ts 的 1630/1631 两个用例——断言「无外发」,实际多了一条到 http://billion-context.inline/__bili/plugin/manifest 的外发。
  2. 什么原因导致的:测试静态 import pi-native.ts,模块加载即 armNativePi。kernel authorization request: release v0.1.211 bundles kernel 0.0.176 (#5415 abort-absorb nudge debt) #5428 把无显式配置时的默认模式翻成 inline,于是 waiter 被注册成 INLINE_ORIGIN 并泄漏进共享工厂的 awaitNativeProxyOrigin(),handler 拿到 proxyBase=INLINE_ORIGIN 后走 registerTools → fetchManifest 外发。
  3. 影响范围与严重程度:仅测试套件(非生产路径)。6 条车道全红,阻塞本 PR 合并;生产行为不受影响。
  4. 怎么修复的:在 native-bootstrap.ts 加 _resetNativeOriginWaiterForTest(清空模块级 origin waiter),测试 beforeEach 里调用,使 handler 的 proxyBase 回到 undefined、不再触发 registerTools。
  5. 是否有回归问题:无。本地跑了 issue4894-direct-window-guard(5/5 过)、pi-native(20/20 过)、typecheck 干净。改动只加了一个测试钩子并在测试里调用,不动生产行为。
  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:修复已推送,待 CI 跑完确认全绿。kernel-guard 仍需 owner 显式授权(PR body 声明与版本 bump 已就位,CI 侧应过)。

@ranxianglei
ranxianglei merged commit 2153152 into master Nov 14, 2026
50 of 56 checks passed
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

Re-verified the live state before touching anything: the watermark regression this event reported has already been fixed and merged — PR #6640 (commit 3728f968fd, merged to master as e5e1647ba0) landed minutes ago with exactly this fix. No follow-up PR is needed, and I did not open one.

What I did this run:

  1. Re-verified live state. The mirror cache for [PR] chore(batch): land 29 open fragmented fixes in one history-preserving batch #6482 is stale (frozen at the pre-merge head d04a8adcb9, mergeable_state: blocked); the PR itself is already merged (2153152f08). The current master is e5e1647ba0.
  2. Root-caused all 6 red lanes (3 distinct failure signatures):
  3. Implemented and verified the fix locally (delayed the bodyText release to the upstream response boundary + realigned the A/B turn-0 pin to the [PR] fix: unknown dsh one-shot sidecars steal the raw-key anchor and kill compression (context-explosion root cause) #5687 budget-gate residual). When I went to push, master had moved and the identical fix was already merged (PR [PR] fix: watermark fastpath (plugin full-resend watermark fastpath) silently disabled by input release mutation | fix: 水印快路径(plugin full-resend watermark fastpath)被输入释放改动静默禁用 #6640, authored by the owner). I aborted the rebase, confirmed the merged fix is green locally, deleted my redundant branch, and stopped — no duplicate PR.

Bottom line: the one real batch regression (watermark fastpath) is fixed and merged (PR #6640). The three Windows lanes are documented environmental flakes, not batch regressions — they clear on re-run. No further code action from this lane.


结构化八问

  1. 是什么问题:批次 PR [PR] chore(batch): land 29 open fragmented fixes in one history-preserving batch #6482 合入后,3 条 shard-3 测试车道(ubuntu 22/3、ubuntu 24/3、windows 22/3)因 watermark 快路径回归而红;另有 3 条 Windows 车道因环境性 flake 而红。
  2. 什么原因导致的:批次 cherry-pick 的 [PR] fix: release input-only request data before streaming #4756 在解析期(handle.ts:750)提前释放了 bodyText,导致 prepareOpenai 的 [Bug][dsh] 20k-message / 81 MiB-inbound session: host aborted 6× in 2 h — the #3050 resident-AST OOM is not bounded by sessions.hardMaxBytes (Windows, 0.1.210) #5403/[Bug][dsh] Plugin lane performs a full re-parse and resends the entire body every round (81 MiB → hundreds of MB AST), w | [Bug][dsh] plugin 车道每轮全量解析重发 body(81 MiB→数百 MB AST)是 #5403 OOM 的生产者:用水位线断言替换逐条复核,服务端只解析新增尾段 #5418 watermark 扫描门(bodyText !== undefined)永远不触发,watermark 快路径整体失效。Windows 3 条是已记录的 AV/索引器锁竞争(test(windows): driveWire golden-harness 清理期 index.db EBUSY 复发(#2799 族,#2808 后残留) #2908)与 abort 拆除预算超时,非批次回归。
  3. 影响范围与严重程度:watermark 回归影响所有 plugin-mode openai 请求——快路径失效、每轮全量重解析重哈希(性能退化,功能仍可用);Windows flake 仅影响 windows-latest CI 的偶发重跑,不影响产品。
  4. 怎么修复的:已合入的 PR [PR] fix: watermark fastpath (plugin full-resend watermark fastpath) silently disabled by input release mutation | fix: 水印快路径(plugin full-resend watermark fastpath)被输入释放改动静默禁用 #6640 把 bodyText 的释放收窄到 fastpath 不可能触发的切片(非 openai wire 或 flag 关),strict gate(pluginMode + !titleGen)仍由 prepare-openai 权威裁决;A/B pin 的 turn-0 改为真实 sidecar 形状(4K 预算)以继续钉住「side lane 不盖章」。
  5. 是否有回归问题:无。修法是条件释放(只在 fastpath 不可能触发时释放),不改变 strict gate 的裁决;A/B 字节相等与 residency band 断言保持。高危模式逐项过:无语义漂移、无 fail-fast 变吞错、无阈值/时序/输出格式变化。
  6. 是否破坏缓存:不涉及。watermark 快路径是前缀缓存的优化车道,修复只是让它重新可用,不改变缓存失效策略。
  7. 是否改了设计方向:不涉及。修复恢复了 [Bug][dsh] 20k-message / 81 MiB-inbound session: host aborted 6× in 2 h — the #3050 resident-AST OOM is not bounded by sessions.hardMaxBytes (Windows, 0.1.210) #5403/[Bug][dsh] Plugin lane performs a full re-parse and resends the entire body every round (81 MiB → hundreds of MB AST), w | [Bug][dsh] plugin 车道每轮全量解析重发 body(81 MiB→数百 MB AST)是 #5403 OOM 的生产者:用水位线断言替换逐条复核,服务端只解析新增尾段 #5418 的既有设计(KDD docs: env vars primary config method + cleanup #15),没有转向。
  8. 是否引入了新的配置:不涉及。没有新增 env/config/默认值。

判定:无需进一步代码动作。唯一真实的批次回归(watermark 快路径)已修复并合入(PR #6640);3 条 Windows 车道是环境性 flake,重跑即清。本 lane 收工。

ranxianglei pushed a commit that referenced this pull request Nov 14, 2026
…merge commit

These files (issue-number lists recording the superseded closings of the
previous four batches #6439/#6455/#6464/#6482) are present in neither
parent of the #6574 merge commit yet appear in its result, and survive to
the branch tip. They are not part of any source PR, are referenced
nowhere, are not mentioned in the PR description, and the previous four
batch merges never committed such files to master. Process scratch data
does not belong at the repo root; the closing record already lives in
GitHub history.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet