Repository navigation
[PR] release v0.1.207 - #4585
github-actions[bot] wants to merge 12 commits into
Conversation
) Four p2 fixes from the five-dimension storage review, one PR: - identity column: cross-save JOIN key strips id/wireIndex (D4-1); content_hash stays the byte-faithful CAS key. v8 migration backfills and heals wireIndex-split generations (keep lowest seq, refund CAS refs). MySQL twin via information_schema probe + idempotent ALTER. - authoritative sweep retires absent rows (retired_at tombstone) instead of DELETE; rewind-redo reactivates in place at the original seq (remap covers completeLog-less saves too); readers filter retired rows (D4-2). - snapshot-render seam image signature hashes the bytes (24-hex sha256), not the length — same-size screenshots can no longer chain a false seam into silent truncation (D4-3). - fork lineage: parent_session_id from metadata.derivedFromSessionId (D3-2). - gc full vacuum routes through vacuum() (TRUNCATE checkpoint included, files actually shrink, D5-1); loadLegacyView counts tail pointer gaps and all three load paths warn instead of silently shrinking (D4-5). Stakes first: tests/storage-review-p2s.test.ts (7), storage family + golden/snapshot/fold-reconcile/incremental/web regressions green.
Review fix for CI reds on this branch, part 1 of 2. - loadLegacyView's new tailGaps counter (D4-5) leaked into the shadow-parity digest surface: the expect leg (live session) structurally lacks the key while the got leg (loadLegacyView output) always carries it, so canonicalize()'s key-preserving normalization made every sampled verify report a fake drift. Drop the key from all digests — same class as the existing storageMode precedent (#2671): presence-not-value, and genuine missing rows still diverge the message arrays themselves, so nothing is masked. describeViewDivergence skips it too, keeping drift diagnostics clean (issue3919 forensics pins keys deepEqual). - Two stale user_version=7 pins the v8 bump missed: tests/storage-ext.test.ts (v3→v7 title/asserts) and tests/storage-ingest-watermark.test.ts (v4→v7 title/assert) now expect v8, matching the updated pin in tests/storage-unified.test.ts.
…e degrade (#4444) On a Node.js that has neither the builtin node:sqlite (< 22.5) nor a pinned better-sqlite3 artifact (e.g. Node 21.x, ABI 120), the engine self-bootstrap is refused before any download and the unified store degrades to legacy-only. The degrade reason on the unsupported-node path carried the specific coverage detail but DROPPED the "upgrade Node.js to >= 22.5" lever that the sidecar-open path already carries — so a Node 21 user read "no engine supports this Node.js" with no actionable way out. - runtime-status degradeReason(): append nodeSqliteUpgradeHint() to the unsupported-node reason (keeps the #3894 specific-detail-first ordering). - engine-bootstrap runBootstrap(): carry the same hint on both skip warns so the lane logs match the status surface. - shadow UnifiedShadow sidecar-open catch: stop advising "delete the file to reset the index" for a no-engine failure — deleting cannot conjure an engine; that advice only fits a genuine open failure (byte-identical otherwise). - regression test: force process.versions.node=21.6.1 and assert the acp_status STORAGE line carries both the coverage detail and the >= 22.5 upgrade lever. Refs #4444
…s but bytes diverge Review fix for CI reds on this branch, part 2 of 2. The D4-1 JOIN migration made remapPlanToAbsolute reuse a seq whenever (msg_id, identity) matched, and INSERT OR IGNORE then silently kept the first generation's bytes for that seq. Any byte divergence under a matching identity became permanently unhealable by full ingest: (a) head truncation — the PR's own flagship scenario — leaves stored rows carrying the old wireIndex generation, so the parity digest never matches the live view again: every 20th save reports drift, the watermark is cleared, the next full save re-runs the same no-op remap, never converges, and after DRIFT_ESCALATE_STREAK consecutive samples #4323 freezes this process's envelope writes (legacy file goes stale = data-loss vector); (b) an externally corrupted CAS row (watermark boundary-tamper test) was healed pre-migration by the byte-faithful join minting fresh rows, and is now kept corrupted forever. Fix: when the full path reuses an existing seq via the identity join and the stored content_hash differs from the plan row's hash, UPDATE the row in place to the fresher authoritative bytes (CAS ref juggling put-new -> update-ref -> release-old, FK/refcount safe; media mappings untouched because identity includes the {$media:hash} markers). Seq stability is unchanged; the delta path stays append-only. SQLite + MySQL twins. identity.ts doc updated to match; CHANGELOG notes the behavior change.
…ning silently (#4215) C10a of the #3011 family checklist (#3025): the #4215 incident showed a seeded fork child (isSeeded + correct parentSession) with ZERO adoption trace in the whole log - eight silent early returns made "not a fork" vs "a fork we could not see" unanswerable after the fact. - forkChildOf() becomes exported forkChildVerdict(): names all eight shapes (initiator-threw/-missing, session-missing, sid-missing, header-missing, not-seeded, parent-missing, parent-is-self) - a seeded candidate blocked at an entry gate (tools-ready / body-parse / side-shaped / proxy-base) logs once per session naming the gate - malformed seeded lineage (isSeeded without a usable parentSession) logs a warn once per session - ambient-blind shapes (fork status undeterminable) log ONE info notice per process, flagging whether scoped #4082 attribution was active - the C10b identity-route split named in the log - plain non-fork traffic stays completely silent Log-only change: no judgment, ordering, or wire behavior changes. Regression suite: tests/issue4215-fork-gate-diagnostics.test.ts (verdict matrix + both directions of every gate, #3617 discipline). Identity-route widening (C10b) needs an owner decision and is NOT part of this change. (cherry picked from commit fb54f6e) (cherry picked from commit 85af00f)
#4264) When the resume-inheritance gate passed but no byte-exact chain match existed, handle.ts emitted zero logs — clone-inheritance failures (the #3011/#3025 family) were left unattributable; dsh (responses protocol) could not even reach the transcript-lineage log lines, so a miss was absolute silence. Add PrefixAffinityResolver.diagnoseResumeMiss (diagnostics only, matching untouched): it re-runs the same gates in reporting mode and classifies the closest near-miss as equal-depth (duplicate conversation by design), prefix-break (head drift desynchronizes the prefix), or truncated (incoming shorter than the stored chain). The per-item divergence scan runs once on the single best near-miss over its stored trailing window (cost-bounded); deep chains report the pre-window bound instead of an unsound index. The server logs one line: info with best candidate id + depths + first divergence cell + 8-hex digests when any eligible candidate exists, debug when nothing eligible was tracked. Reuses MIN_RESUME_PREFIX=8; no new config surface, no behavior change. Tests: the three miss shapes plus truncated, deep-chain pre-window break, scoped declared-parent states (untracked/too-shallow), scoped judging-only-declared-parent, self-exclusion, and hit-path regression (findResumeParent/findResumeParentWithin results unchanged). (cherry picked from commit 7fff50a) (cherry picked from commit c1272ae)
…4265 review) When any gate conjunct short-circuits (derivedFromSessionId, publicForkReceipt, or resumeInheritance=false) there was no [resume-inheritance] line at all, so a log without "miss:" stayed ambiguous between gate-skipped, pre-#4265 build, and miss. The gate's else-branch now logs one debug line naming the skip reason. Same failure domain as #4264's stated goal (is the fallback channel even being tried) — absorbed here rather than a separate PR. Bounded like the miss line: the else-if re-checks clientProvided && requests === 0, and stats.requests increments later in the prepare lanes, so it fires at most once per session lifetime; anonAffinity is structurally null on this path (only the anonymous lane sets it). Purely additive — the gate condition itself is untouched, matching is unchanged. New e2e pin (tests/resume-inheritance-gate-skip.test.ts): both directions — skipped gate logs exactly one bounded debug line naming the reason, and an unskipped gate keeps the existing miss line and never says skipped. (cherry picked from commit f3f7578) (cherry picked from commit ea712f8)
(ai: vllm-qwen/qwen3.8-27b) 判定:不做 rebase/merge——这是 release-bugfix.yml 里写明的「conflicted ship PR」冲突类,而它指定的解法(backport-ship 按 head tree 直接发版)已经在途。 本次(2/2)只交判定,不动分支。 查重:不适用(系统转发的 rebase 唤醒,非新问题报告);冲突机制归属 #4576 follow-up / #4596(label-publish 路径)。 八问
给 owner 的行动项:
收工判定:冲突本身无需 agent 解决,机械动作(rebase/merge/push)在此场景下都是错误动作;设计决策(何时/是否按 head tree 发布)已由 label 动作表达且执行在途,本 worker 到此为止。 |
…review) ResumeMissCandidate / ResumeMissReport are referenced only inside prefix-affinity.ts (describeResumeMiss consumes them); the leading export made them dead external surface and tripped the unused-exports gate, which fails the release-bugfix full-CI gate. Drop the export keyword only - type-level, internal references intact, no behavior change.
(ai: vllm-qwen/qwen3.8-27b) PR #4585「release v0.1.207」patch-train 审阅 — 结论:内容可发;唯一 CI 红灯已定位并修掉 这是 dual-lane 补丁列车(合并即发布 pre-merge head 树 = v0.1.206 + 本列车修复,不含 master 未发布工作)。我拉了分支逐段核对 diff、本地复跑门禁、确认了 CI。唯一红灯 查重:不重复——v0.1.207 补丁列车的发布 PR;存储修复对应已立项的 #4390/#4506/#4444,末尾三个 carry(#4215/#4264/#4265)是 master 已合并工作的回移(cherry-pick,不同 SHA:master 侧 85af00f / c1272ae / ea712f8),非新开发。无并列重复发布。 八问:
发布机制核对:package.json 0.1.206→0.1.207 ✓;package-lock 两处版本同步 ✓;release-notes/package.json 新增 0.1.207 条目(tier=recommended,2026-10-24)✓;CHANGELOG 详实且与 diff 一致 ✓。#4519/#4503(CI carry 机制本身)确认不在本 lane,按 issue 口径留给下次常规发布可接受,不影响发布物正确性。 两个待 owner 注意的非阻塞项(我不改写 bot 生成的 PR 正文,仅在此指出):
收工判定:内容可发。除已修的 unused-exports 外,无回归、无缓存破坏、无方向变更、无新配置。待 CI 在新 head |
|
不要再开这个pr 这个pr有很大问题 打赏标签后会里面发布 |
Ship PR for the patch train (dual-lane release): merging publishes the PRE-MERGE head tree — the base tag plus exactly this train’s fixes, none of master’s unreleased work — via release-bugfix.yml.
These fixes expressed backport intent but their commits are absent from
bugfix-v0.1.206— a merged source PR whose picks all conflicted (no clean carry to merge), or a PR still open against the lane at ship time (possibly still holding conflicts). Ship only if leaving them for the next REGULAR release is acceptable:Changes since v0.1.206