Skip to content

release v0.1.183 - #2058

Merged
ranxianglei merged 1 commit into
masterfrom
2026-10-04_release-v0.1.183
Oct 4, 2026
Merged

ranxianglei merged 1 commit into
masterfrom
2026-10-04_release-v0.1.183

Conversation

@github-actions

@github-actions github-actions Bot commented Oct 4, 2026

Copy link
Copy Markdown
Contributor

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 v0.1.181

@github-actions

github-actions Bot commented Oct 4, 2026

Copy link
Copy Markdown
Contributor Author

📦 Built Package Artifact

Branch: 2026-10-04_release-v0.1.183 (3d63241)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.2101-pr.2058.0cf796e-a2

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

Option B — Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr2058.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

ranxianglei commented Oct 4, 2026 •

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Release PR review — v0.1.183

Verified the branch (2026-10-04_release-v0.1.183 @ 3d63241), ran the version-interval regression audit over the full user-upgrade delta (published v0.1.182 → this build: 150 commits, 410 files, +15454/−1508, reviewed as one integrated whole), and confirmed CI. Details below.

PR shape — clean. Exactly one commit release v0.1.183 on top of current master (3a7308a); diff is only package.json + package-lock.json version line 0.1.182 → 0.1.183. Nothing bundled, branch naming and commit message conform, mergeable_state: clean.

CI — fully green. All 4 unit-test lanes (ubuntu/windows × node 22/24), version-guard, build-artifact, real-client e2e lanes (codex, opencode native, opencode V2 Responses WS, pi, dsh), hermetic registry e2e (ubuntu + windows — exercises the real self-update chain), real-image billing e2e, codex-overlay lane.

Cross-repo ordering — ok. Window contains the acp-kernel pin bump 0.0.100 → 0.0.101 (from the #1960 redacted-thinking fix). Verified 0.0.101 is live on npm before judging this PR; CI npm ci confirms resolvability.

Load-bearing areas deep-read (regression verdict: none broken):

Behavior-change ledger (old → new, all intentional, all in-window):

  1. Release-notes entry: mandatory for every stable release → opt-in (only severe releases need one) [发版可见性:分级标签的 release 元数据(模型撰写摘要)+ 用户版跨 span 更新说明 + acp_status //acp 面板送达「更新已就绪待重启」(#1577 延伸) #1870 policy revision, docs(release): release-notes entries are opt-in — only severe releases need one (#1870 policy revision) #2057].
  2. Auto-update persistent install failure: retry every 3-min cycle forever → exponential backoff after 3 consecutive failures on the same (installDir, version), 5 min base, 6 h cap; success or different key resets; forced updates bypass [auto-update installs but never activates on long-lived plugin hosts; unwritable install dir retries every 3-min cycle for 11+ days #1603].
  3. Install-dir resolution: walk-up could adopt an unrelated ancestor package root → foreign named ancestor refused, loud diagnostic naming target vs running copy [[follow-up #1603] update install-dir resolution targets the global root (/usr/lib/node_modules) instead of the running copy — root cause of the 12-day retry wall #1628/fix(update): pin install-dir resolution to the running copy; self-explaining failure diagnostic #1632].
  4. ~35 behavior knobs: env-only → env > config file > default; env parsing and defaults unchanged [彻底整治环境变了滥用,除非必要都走配置文件 #2030].
  5. Signed thinking/thought streams: tag/marker prose stripped → forwarded verbatim (cosmetic echoes inside thinking panes accepted); text channel strips unchanged [[acp-loop] In-turn compress re-request edits the latest assistant message → Anthropic 400 "thinking blocks … cannot be modified" (session stays broken) #1960].
  6. bili codex: overlay folded back into the real ~/.codex home only at the NEXT bili start → written back on clean exit (generated files excluded) [[Codex launcher] 退出时未将 overlay SQLite 数据库回写到原始 CODEX_HOME #1965].
  7. Post-response connection close: immediate destroy → linger up to 5 s for peer close signal; new BILI_POST_RESPONSE_LINGER_MS (default 5000) [建议:客户端侧连接请优雅关闭(close_notify+FIN),避免向下游客户端发出 RST #1982].
  8. Preflight summarization hard 400/413: fail-fast 502 + dead-end cooldown → bounded halving shrink-retry (cap MAX_SUMMARY_CALLS_PER_PREFLIGHT); transient and auth/routing 4xx semantics unchanged [preflight: summary call answered HTTP 400 — the response body is never logged or dumped, there is no chunk-shrink fallback, and the turn fail-fasts 502 retryable=false #1993].
  9. Plugin raw-snapshot retention: unbounded → capped at 16 MiB by default, fork refused past the cap; BILI_PUBLIC_SNAPSHOT_CAP_BYTES / plugin.snapshotCapBytes, 0 disables [feat(plugin): public isolated conversation forks and attributed context observations #2016/feat(plugin): expose isolated forks and attributed context observations #2017].
  10. MITM upstream key: non-default CONNECT port dropped from the key → preserved (one-time orphaning of stale cache entries for such upstreams; previously distinct upstreams could share a key — the fix closes a cross-talk risk, transient cost only).

Caveat (flagged, not blocking): the updater changes (#1603/#1632/#2030 touch src/update.ts) ship alongside ~100 content commits rather than after a standalone NO-OP validation release, which §5's protocol prefers. Mitigations present: the 182→183 hop itself still executes the old (proven) updater code; the new updater first acts on the next hop; hermetic registry e2e (real self-update chain) is green on both OSes. Acceptable trade-off for a routine patch — noting it for the record.


结构化七问:

  1. 是什么问题:例行发布 PR(v0.1.183,one-click 回退路径),纯版本号提交,非缺陷修复。
  2. 什么原因导致的:不涉及(无缺陷);审计对象是上一版本以来的全部变更是否引入回归。
  3. 怎么修复的:不涉及;发布流程本身——分支形态、版本纪律、CI 触发条件逐项核验通过。
  4. 是否有回归问题:无。高危模式逐项过:update.ts 三处变更语义核验(backoff 自愈、fail-fast 保留、默认值不变)、thinking 通道按 KDD#10 字节级测试钉死、preflight「fail-fast 变重试」为有意且封顶、linger/mitm-key/快照上限均为有界有意变更、持久化格式未动、acp-kernel 跨仓顺序正确(0.0.101 已上 npm)。
  5. 是否破坏缓存:否。持久化格式/版本不变;唯一缓存键变化是 MITM 上游键补入非默认端口(一次性失效,方向正确,顺带堵住跨端口串键)。
  6. 是否改了设计方向:无架构转向;knoobs 层、WS 桥、fork provenance 均为既有边界内的增量;bugfix carve-out 发布通道是 owner 决策的第二发布路径,已有生产使用记录(v0.1.182 即走该通道)。
  7. 是否引入了新的配置:有——BILI_POST_RESPONSE_LINGER_MS(默认 5000ms)、BILI_PUBLIC_SNAPSHOT_CAP_BYTES(默认 16 MiB,0 关闭)、providers[].protocol 可选字段、约 35 个既有 env knob 增加配置文件层级(默认值全部与历史一致);均有文档(CONFIGURATION.md / reference/env-vars.md / PLUGIN.md)。

判定:可合。 请人工合并合并后 CI 自动发布。(另:updater 变更未走独立 NO-OP 验证发布一事已按上文披露,供记录。)

@ranxianglei

Copy link
Copy Markdown
Owner

PR #2058 合并 Diff 回归审查报告(v0.1.181 → pr2058)

按「合并为一个大 diff 审查、重点看回归、必要时本地验证」完成。结论:未发现阻断性(blocking)回归,可以合并。 153 commits / 411 files(+15701/−1505)逐块过完,所有行为变化都能对应到有意的修复或策略决策。

一、自动更新子系统(重点)

安装管线(备份→拷贝→校验→回滚)完全未动,改动全部在外围防护与提示策略,方向都是「响亮失败优于静默错装」:

二、其他集群(按风险从高到低)

三、本地验证

验证项 结果
typecheck ✅ 干净
全量套件(443 文件) 833 ✔ / 25 ✖
25 个失败定位 全部在 tests/dsh-native.test.ts
基线对照 同样 25 ✖ 在 v0.1.181 复现(沙箱缺 dsh attach 能力,15s 超时)→ 环境问题,非本 PR 回归
opencode-native 单独跑 1 ✖,基线同样失败;全套件内通过
15 个最高风险新特性文件 129/129 通过(WS lane、fork、knobs 优先级、#1987/#1993、provider-protocol、update-backoff 等)
自动更新生态 9 文件 64/64 通过

四、非阻断跟进

对应问题已单独开 issue 跟踪:WS 背压硬失败、fork 快照每请求克隆/持久化体积、knobs+出站估计的每请求开销、#1977 可见性变化的文档说明。

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Cross-review reconciliation — two independent audits, same verdict

I audited the same PR from the user-upgrade-delta angle (v0.1.182 tag → HEAD, 150 commits); your report covers v0.1.181 → pr2058 (153 commits, the PR-body framing). The union covers both framings completely, and both audits land on no blocking regression, mergeable. Before agreeing, I re-verified the findings where our notes differ or where yours adds to mine:

1. #1977 visibility change — you're right, my ledger missed it. Verified at commit level: c4af34de + 78276c01 (merged via #1983) sit inside my audit window; getUpdateVisibility now yields visible only for critical-tier entries, recommended is record-only, and a pending restart alone no longer announces itself. It's the intentional Option-C compromise (supersedes #1979 full removal per your 「不能完全静默」 decision), docs updated in the same commits (CONFIGURATION.md en/zh, release-notes/README, reference/release.md), and the stale-probe path (fetchStaleNotice → __bili/status stale:true) is untouched — so "restart needed" still has an exit via /acp command output. Adding it to the behavior-change ledger as item 11:

  1. Update visibility on acp_status + /acp panel: any pending/recommended update surfaced → critical-tier only; routine/recommended silent by design [Release-notes visibility should be silent by default (owner decision) — advisories stay the only auto-visible channel #1977, fix(#1977): release-notes visibility surfaces only for critical-tier releases #1983].

Your classification stands: this is the most regression-looking surface in the whole window, and it's deliberate with a documented escape hatch.

2. #1985 mitm port — your characterization is the accurate one. I read the hunk (src/mitm.ts, MITM_UPSTREAM_KEY assignment): old code stamped https://${host} unconditionally, so every non-443 CONNECT was silently misrouted to :443 (connect-refused for local gateways); new code preserves :${port} when ≠443. That's a real routing bug fix, not just cache-key hygiene as I described in my ledger item 10 — correcting that entry: "MITM non-443 upstreams misrouted to :443 → port preserved (one-time orphaning of any stale cache entries keyed without port)". Same commit also carries the raw-socket marker for #1982 close-classification, consistent with your note on the monkey-patch guard conditions.

3. Follow-up issues — all four verified filed and open: #2076 (WS backpressure hard-fail at src/responses-ws.ts), #2077 (fork snapshot O(full-history) clone + per-session persistence), #2078 (knobs per-call config re-read ~5–10/request + forward() full-wireBody re-parse), #2079 (#1977 docs/UX surfacing gap). Each references PR #2058 as source. Nothing dangling.

4. Your local validation closes the one caveat I raised. My only reservation was that the src/update.ts changes (#1603/#1632/#2030) ship alongside ~100 content commits rather than after a standalone NO-OP validation release (§5 preference). Your baseline-controlled runs address exactly that: updater ecology 64/64 green, high-risk files 129/129, and the 25 dsh-native failures proven environmental by reproducing them identically on v0.1.181. Caveat stays on record, no longer load-bearing.


结构化七问(交叉核对轮):

  1. 是什么问题:两条独立审查结论的交叉核对,非新缺陷。
  2. 什么原因导致的:不涉及;差异仅来自审查窗口口径(升级增量 vs PR 正文口径)。
  3. 怎么修复的:不涉及;逐条复核分歧点(Release-notes visibility should be silent by default (owner decision) — advisories stay the only auto-visible channel #1977、fix(server): graceful post-response client close — linger for peer close signal (#1982) #1985)到提交/hunk 级并修正我方清单两处表述。
  4. 是否有回归问题:两审一致——无阻断性回归;新增确认 Release-notes visibility should be silent by default (owner decision) — advisories stay the only auto-visible channel #1977 为有意 UX 收紧(critical-only,静默有出口),fix(server): graceful post-response client close — linger for peer close signal (#1982) #1985 为真 bug 修复(非 443 CONNECT 误路由)。
  5. 是否破坏缓存:维持原判——持久化格式未动;mitm 端口入键造成的一次性缓存条目孤儿化方向正确(同时消除跨端口串键风险)。
  6. 是否改了设计方向:无转向;Release-notes visibility should be silent by default (owner decision) — advisories stay the only auto-visible channel #1977 属提示策略的产品决策(owner 拍板的 Option C),已在同一批提交中同步文档。
  7. 是否引入了新的配置:无新增配置项(前一轮清单不变:BILI_POST_RESPONSE_LINGER_MS、BILI_PUBLIC_SNAPSHOT_CAP_BYTES、providers[].protocol、~35 knob 的 file tier,默认值均与历史一致)。

判定:可合。 两份独立审计(v0.1.182 增量视角 + v0.1.181 全窗口视角)、CI 全绿、本地基线对照验证齐备;四个跟进问题已立项跟踪。等待人工合并,CI 自动发布。

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