Repository navigation
release v0.1.187 - #2305
release v0.1.187#2305
Conversation
📦 Built Package ArtifactBranch: Option A — Install this exact build (recommended)npm install -g billion-context@0.0.2447-pr.2305.1c1b3cfThis exact version is immutable and reproducible. The rolling Option B — Download artifact
tar xzf billion-context-pr2305.tgz
npm install -g packageInstall instructions are refreshed on each push. |
…gate fails on release head (#2305 review)
[bot] 🏷 Release v0.1.187 review — 结构化结论(七问) ① 是什么问题:这是自动 fallback 的 release PR(master 直推被分支保护挡下)。审查对象 = release commit 本身 + 整个版本窗口(v0.1.186..HEAD,123 commits / 194 files,按 ②b 规则合并为一个整体审阅)。发现唯一实质问题:head 上 ② 什么原因导致的:release head(f7d3e86)上 ③ 怎么修复的:去掉 ④ 是否有回归问题:未发现回归(逐项过窗口内每个 fix/feat 的既有行为):
⑤ 是否破坏缓存:不破坏。exact-prefix 快路径未动;simhash adoption(#2265)命中后 ⑥ 是否改了设计方向:无架构转向。唯一机制级调整是 persona-anchor(#2241/#2250):id-sharing host(dsh/codex/claude-over-Responses)的 fork 触发从「system 文本变了」收紧为「system 变了 AND 历史不延续 raw key 链」——同一会话切模型不再 fork(锚迁移,压缩状态留在 raw key),辅助 blob(auto-review)仍照旧 fork;fork id 格式 ⑦ 是否引入新配置:一个—— 重大行为变更清单(有意变更,旧值→新值+原因):
流程观察(非代码阻塞):§5 要求改 判定:release commit 干净(仅 version + lockfile)、分支名合规、kernel 纪律合规、路由面/缓存/双模式语义均核验通过;唯一红灯(unused-export)已在本分支修复并本地验证。CI 在新 head 371b15b 上转绿后即可合并(kernel/identity/update.ts 内容本就属人工合并范畴,请 owner 自行 merge: #2305 )。 |
|
owner-managed lane(dsh desktop/profile、pi npm)更新检查节奏:3min(搭全局车)→ 30min+≤15min jitter,独立 throttle 文件(https://github.com/ranxianglei/billion-context/issues/2192:桌面布局曾每周期重下整包,~480 次/天/机);全局自更新节奏不变。 这个是把所有的自动更新都变成30分钟了吗 |
[bot] 🏷 不是——只有 owner-managed lane 这一类变了,bili 自身的自动更新节奏没动。 代码层面( const throttleFile = ownerLane ? OWNER_THROTTLE_FILE : THROTTLE_FILE;
const intervalMs = ownerLane ? ownerLaneIntervalMs() : CHECK_INTERVAL_MS;
改慢的原因(#2192):这三份是宿主拥有的副本,bili 只做 housekeeping;桌面布局一旦损坏,旧逻辑每 3min 重下一次完整 tarball(~480 次/天/机)。它们坏了也不影响用户正在用的 bili 本体——本体的修复路径仍是 3min 全局自更新 + advisory,没降速。 |
mNNNNN refs restart at m00001 on a rebase (resetSessionCompression clears the in-memory index/caches but old spill files survive), so a ref-only filename plus skip-if-exists silently served the PREVIOUS generation's pixels for a new image under a reused ref — wrong screenshot, correct metadata, no error. With an 8-hex sha256 of the bytes in the name, same name implies same content (skip stays a correct idempotency check) and different content lands in a different file, so a stale hit is impossible on every trigger path (rebase, crash-before-persist, BILI_PERSIST=0, manual session-file delete). Readers are pure index lookups on the stored path, so no other file changes; orphaned old-gen files age out via the existing 7d prune. Regression test: two generations, same session+ref+ media type, different bytes → two files, each round-tripping its own bytes, re-spill still idempotent per content.
…test flake on v0.1.187 lane (#2305 review)
[bot] 🏷 CI 跟进(新 head 371b15b):35 项检查里 34 绿,唯一红是 排查结论:测试自身时序 flake,非窗口回归。
已修(小问题直接修,标注 commit): 另:审查期间分支上多了一个并发提交 CI 正在新 head |
[bot] 🏷 新 head d4f2a75 的 CI 跟进(上面那条 30min 问题已在楼上回复过——只有 owner-managed lane 那三份变 30min,bili 本体自更新仍是 3min):
排查:这个文件是纯字符串单测(只 import tag-echo-filter,无端口、无 timer、无网络),但日志里它的任何子测试都没跑出来——文件级 943ms 静默失败('test failed',无断言细节),是典型的 runner 负载下子进程崩溃形态,不是断言失败。排除回归的证据链:
已触发该 job 重跑(rerun-failed-jobs),结果出来我再报一轮:若转绿则确认瞬时 flake;若同一文件再挂,那就是 Windows 下的真实问题,我会深挖而不是再等它自己好。 |
[bot] 🏷 CI 全绿确认:head 分支最终状态: 可合并(人工合并):#2305 ——合并后 CI 自动发布 v0.1.187 到 npm 并打 tag。 |
… owner cadence call) 30+15 kept broken host copies (dsh desktop / dsh profile / pi npm) stale too long relative to what a check costs (a few-KB version-doc fetch, and failure storms are already capped by the #2192 backoff, not by cadence). Constants + comment only: no download/extract/install/version-check code touched. Owner-directed default-value change per AGENTS.md §7.4 (must stay human); riding release v0.1.187's human merge.
Auto-fallback from the one-click release workflow: direct push to master was blocked by branch protection. Merging this PR publishes via the standard release.yml flow.
Changes since v0.1.186
m01233</acp>remnant through to the client reasoning stream (model echoes refs from compress confirmation markers) #2023 Model re-emits internal ACP artifacts into assistant prose (mangled render tags + absorb tool-call text) #2066 【bug】最终版issue: 模型生成的<acp …>标签会漏到用户可见输出(附根因、量化与修法建议) #2190 [bug] 出口层缺口:文本不在受管字段的帧绕过 tag-echo 状态机(E1/E5 原始转发出口 + 零日志真泄漏样本) #2248)--separator<acp …>标签会漏到用户可见输出(附根因、量化与修法建议) #2190): pin field replay seq=22409 — degenerate close strips via ref-anchored open-set rule<acp …>标签会漏到用户可见输出(附根因、量化与修法建议) #2190): pin H2-exclusion invariant — touching a well-formed pair always logs a drop<acp …>标签会漏到用户可见输出(附根因、量化与修法建议) #2190): drop acpx from attested drift names after evidence decontamination<acp …>标签会漏到用户可见输出(附根因、量化与修法建议) #2190): extend acplike name set with field-attested drift names; audit raw-forward exits<acp …>标签会漏到用户可见输出(附根因、量化与修法建议) #2190): strip model-emitted degenerate-close echo residuem01233</acp>remnant through to the client reasoning stream (model echoes refs from compress confirmation markers) #2023)m01233</acp>remnant through to the client reasoning stream (model echoes refs from compress confirmation markers) #2023)