Repository navigation
[PR] fix(dsh): preserve verified seeded context inheritance - #5962
2772780896 wants to merge 11 commits into
Conversation
(ai: vllm-qwen/qwen3.8-27b) 标签: 查重:不重复,但同家族,需与在途 PR #5602 协同。
完整审核结论(八问 + 回归清点)稍后在本线程给出——CI test lanes 仍在跑,我正在本地做定向验证与代码审查。 |
(ai: vllm-qwen/qwen3.8-27b) PR #5962 审核结论(八问) 1. 是什么问题 2. 什么原因导致的 3. 影响范围与严重程度 4. 怎么修复的
5. 是否有回归问题
6. 是否破坏缓存 7. 是否改了设计方向 8. 是否引入了新的配置 行为变更披露(old → new,均有意)
质量门(⑤b)结论:测试是真测试——正负双向(gate 纪律 #3617)、HTTP 级 E2E、临时端口( 合并协调:与在途 PR #5602 都动 判定:代码审查通过,方向正确(修的是守卫本身而非症状),质量达标。§7.4 自动合并不适用(触碰插件 wire 协议新字段 + 会话绑定逻辑 = human-only tier),请人工合并;建议等 CI test shards 全绿后合入。 |
9508aeb to
3702a3e
Compare
(ai: vllm-qwen/qwen3.8-27b) PR #5962 冲突已解决:4 个 PR 提交 rebase 到当前 master 并推送,mergeable 已恢复。 1. 是什么问题 2. 什么原因导致的 3. 影响范围与严重程度 4. 怎么修复的 5. 是否有回归问题 6. 是否破坏缓存 7. 是否改了设计方向 当前状态:PR mergeable = True,merge state = blocked(CI 正在新 head 3702a3e 上重跑,checks 排队/运行中)。若 CI 出现红项,我会在后续轮次(自动修复 lane,本轮 1/3)继续处理。 |
(ai: vllm-qwen/qwen3.8-27b) 本轮为重复唤醒,冲突已在上一轮解决——本轮只做活体核验,未重做 rebase。 唤醒消息里的 "merge state: dirty" 是排队时的旧快照。本轮对 GitHub 活体核验(非记忆、非旧评论):
七问(本轮)
当前状态:等待新 head |
…unused-exports gate (ranxianglei#5962)
(ai: vllm-qwen/qwen3.8-27b) CI 红项已修:unused-exports 门禁失败,一行修复已推上 PR 分支(a11b67f)。 本轮先活体验证:镜像缓存(新 head 3702a3e)显示 30 项检查 0 pending、恰好 1 红 = unused-exports,其余全绿——与唤醒报告一致,未重复处理。 七问
当前状态:修复 commit |
(ai: vllm-qwen/qwen3.8-27b) 重复唤醒:报告的红项已在上一轮修复,本轮活体验证后不再动手。 唤醒消息指向旧 head
|
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) 3 个 test shard 红(shard 3 × ubuntu-24/22 + windows-22)已修,新 head 七问
schema 变更声明(门禁要求 golden 重生成附理由):唯一 schema 变更是 fork capability 块新增 CI 状态:新 head 已推送(git 确认 |
[bot] 🏷 APPROVE — 已独立审完 host 侧全部改动(23 文件,+1742/-93,无 kernel/ 或 pi-subagents/ 改动,不触发 kernel 人工审核档;head 9413219 CI 绿)。未照抄 issue 侧结论,diff 逐段读过(fork 协议 / durable 身份哈希 / 块选择 / 漂移重验四个高危域)。 方向(①): 修的是根因——缺一条能按稳定 rawId + 不透明内容身份做比对的继承路径,不是治标。DSH forkSource 子会话在父会话含多模态/不透明内容时,旧文本快照比对必然 409 SNAPSHOT_UNAVAILABLE → 放弃继承 → 全量重放 → 不可重试 413。新路径( 回归(②) 逐项过:
代码(③): 测试(④): 客户端/服务端 durable 身份哈希有 parity 测试钉死( diff 整洁度(⑤/⑤b): 23 文件全部与本 PR 目的相关,无跑题文件/生成物/锁文件连带变更;无 8问:
判定: 可合。无小修项(代码/注释/测试已整洁,未发现 typo/坏谓词/小边缘)。 |
ranxianglei
left a comment
There was a problem hiding this comment.
APPROVE — host-side change (no kernel/ or pi-subagents/). Reviewed the full diff (23 files, +1742/-93). The durable-fork protocol is correct and fail-closed: raw mode bypasses only the opaque text-comparability guard, not consistency checks; durableFork capability is gated by replicasAvailable(); client/server durable identity hash is parity-pinned (dsh-fork-adopt.test.ts). Regression surface checked: legacy text path unchanged, non-DSH callers unaffected, behavior changes B1/B3 disclosed. Tests cover parity, opaque detection, tool-pair fail-closed, raw fallback, capability gating. No small findings. Full 8-question checklist in the issue thread.
[bot] 🏷 复核完成,已 APPROVE(非卡住,工作已交付)。
协调 issue:任务未给编号,我按 "coordination review lane" / "5964" 等关键词搜不到对应线程。如本 review lane 有专门协调 issue,请给编号,我补一行完成报告。 |
9413219 to
7000d49
Compare
…unused-exports gate (ranxianglei#5962)
(ai: vllm-qwen/qwen3.8-27b) PR #5962 冲突已解决:6 个 PR 提交 rebase 到当前 master 并推送,mergeable 已恢复(True)。 七问
当前状态:PR mergeable = True(冲突已消除),state = behind——master 在我 rebase 后又前进了 14 个提交到 e1f31ad(纯 SQLite 默认等),无冲突,合并时自然收敛。CI 在新 head |
7000d49 to
eb0557d
Compare
…unused-exports gate (ranxianglei#5962)
…ns {3,5,11,14}
The probe anchor PROXY_TOOL_AT=6 was calibrated on 2026-10-31 for ranxianglei#4521
(schedule {4,6,8,9}). The nudge scheduling has shifted again — nudge now
lands at {3,5,11,14}. Re-calibrate the anchor to the next nudge-carrying
request pair (5 → tool round, 6 → fold round). The fold/proxy-tool-round
mechanics under test are unchanged.
Closes: master-red on test (windows-latest, 22, 3) and (ubuntu-latest, 22, 1).
…ic wire shape The probe's hardcoded request-index anchors (PROXY_TOOL_AT/FOLD_AT) broke every time the nudge gate's schedule shifted. The last recalibration (245ecb1) was measured on a direct local run (timestamps ON by default since ranxianglei#4313), but CI runs the suite preload pin BILI_MESSAGE_TIMESTAMPS="0" (timestamps OFF) — a different wire shape, a different schedule (no nudge at index 5) — so it passed locally and failed on every CI shard. Root-cause fix instead of another anchor whack-a-mole: - Arm the fold on the first request carrying the ephemeral nudge with >=11 tagged refs; fold on the re-request after the acp_status round. The probe now materializes the ranxianglei#2929 scenario wherever the schedule places it. - Pin compress.messageTimestamps:false so the probe's wire shape is hermetic regardless of how the suite is launched (matches the CI preload pin). Supersedes the 245ecb1 recalibration. Verified: passes with and without the preload pin, same deterministic schedule (nudge at inputs[7], fold at inputs[8]).
Verify bounded child identities and complete tool pairs against parent history, preserve covered reasoning omissions, and revalidate inherited state before every send. Keep legacy storage and non-DSH prefix semantics intact. Refs ranxianglei#5595.
…unused-exports gate (ranxianglei#5962)
eb0557d to
22be19d
Compare
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) PR #5962 冲突已解决:6 个 PR 提交 rebase 到当前 master(301e1a54)并推送,mergeable 已恢复(True / blocked,不再是 dirty)。 七问
当前状态:PR mergeable = True,mergeable_state = blocked(不再是 dirty——冲突已消除;blocked 是分支保护要求 CI 通过 + review,属正常待办)。推送时 master 恰好停在 301e1a5(未再移动),所以 head 基于当前 master,不是 behind。CI 在新 head 说明:本会话内 master 移动了三次、工作环境重置了两次(需重新建立 remote/分支),均已完成处理。本轮按 CI-hold 策略以「最终 head」推送(未带 [ework:ci-held]),因为本地已验证且 master 停在 301e1a5,这是应触发一次完整 CI 的最终提交。若 CI 运行期间 master 再移动导致重新 dirty,自动修复 lane 下轮(2/6)继续 rebase。 |
…be fix + arch-docs module map
(ai: vllm-qwen/qwen3.8-27b) CI fix — PR #5962 (mirror #5964). Re-verified live first: the 3 reported shards were all This is an inherited master failure, not a PR regression. The PR diff does not touch that test file. #4599 (CI control) confirms master Action taken: 八问:
判定:本 PR 的 3 红全部为继承性 master 红,已按政策 merge ci-fixes 收敛,本地验证绿,等 CI 重跑即转绿。 |
Motivation
Refs #5595. This extends the existing opaque/raw adoption fix rather than replacing its commits.
DSH ordinary forks, sidechats and context-inheriting subagents can replay the same inherited history in different provider projections. Matching only a contiguous provider prefix can stop after a few messages; model switches can also omit covered assistant reasoning. A successful fork receipt can therefore inherit no useful folds, replay several MiB of raw history and hit the local non-retryable 413 guard. Merely copying the parent's latest blocks is unsafe for historical forks and changed payloads.
Changes
Interface and behavior
isSeededplus a validated source parent (forkSource, or legacyparentSession). Fresh subagents with only an ownership parent do not inherit. Ownership, lifecycle, catalog/workspace relationships and seed construction remain unchanged.mode=rawfallback. Add the optional protocol-2durableFork.messagesidentity evidence only when the parent snapshot advertises it. An inserted first child message may produce an empty prefix while durable evidence still proves inherited content.Scope and review
This is host-side adoption and persistence plumbing, not general server-side copy-on-write/reference storage. No changes under
kernel/, no storage-schema/version migration, no config surface change, and no change to non-DSH callers' default prefix semantics. Wire/identity/persistence changes require human review; do not auto-merge.The additive repair was applied and validated on the existing PR head
77a5a3861. The separate local Responses EOF meter-recovery commitf0b9a91b4is deliberately NOT included. This branch has not been rebased onto current master; local checks are not a remote full-CI verdict or a merge-readiness claim.Verification
On the exact clean PR-head baseline plus this additive patch:
npm run typecheck— passed.npm run build— passed, including dist import-annotation/external checks; existing ignored-bare-import warning only.node tools/check-architecture-docs.mjsandgit diff --check— passed.The matrix covers the common seeded gate, fresh-child negatives, historical cuts, zero-prefix insertion, later reasoning omission, nested originals/cold reload, file and signature drift before/after publication, tool-result ownership swaps, archived whole-pair ID reuse, broken ref proofs, unresolved retired media, legacy storage fallback, and non-DSH/proxy/plugin compatibility. Expected-red reproductions were run before fixing the unsafe cases.
A targeted read-only audit also ran the actual selector/hash helpers against the reported large-history shape: the large T2 blocks and complete nested closures become eligible when all hard-content witnesses match and only covered reasoning is absent. This is offline selector evidence, not a deployed fork/model test. No transcript contents or production store files are included.
Limitations / deployment boundary
127.0.0.1:8199/v1) unreachable. The gated core test was skipped, not passed.