Skip to content

release v0.1.164 - #1541

Merged
ranxianglei merged 1 commit into
masterfrom
2026-09-28_release-v0.1.164
Sep 28, 2026
Merged

ranxianglei merged 1 commit into
masterfrom
2026-09-28_release-v0.1.164

Conversation

@github-actions

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 v0.1.163

@github-actions

Copy link
Copy Markdown
Contributor Author

📦 Built Package Artifact

Branch: 2026-09-28_release-v0.1.164 (1a9d5e4)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.1481-pr.1541.84fcb22-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-pr1541.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 收到,开始审查这个 release PR(不重做工作,只做核对):

  • 分支结构 ✅:2026-09-28_release-v0.1.164 恰好领先 master 一个提交 1a9d5e49 release v0.1.164,diff 仅 package.json + package-lock.json 的版本行(0.1.163 → 0.1.164),符合「one version, one commit, one branch」。
  • CI(head sha 1a9d5e49):version-guard / build-artifact / e2e / hermetic registry / codex+opencode+pi+dsh 各真实客户端 lane 均已 ✅;4 个 test 矩阵 job(ubuntu/windows × Node 22/24)仍在跑。
  • 正在做版本区间回归审计(v0.1.163 → master,共 23 个提交),完成后给出完整结论。稍等。

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 release v0.1.164 — version-window regression audit complete (v0.1.163 → HEAD)

Branch structure ✅ — Release branch is exactly one commit ahead of master: 1a9d5e49 release v0.1.164, diff touches ONLY the package.json/package-lock.json version lines (0.1.163 → 0.1.164). One-version-one-commit-one-branch satisfied.

CI ✅ — All 13 check-runs on head sha 1a9d5e49 are completed/success: full test matrix (ubuntu/windows × Node 22/24), hermetic registry e2e, real codex / dsh / opencode / pi lanes vs deterministic fake upstreams, version-guard, build-artifact. PR mergeable_state is now clean.

Window audit — every one of the 17 content commits in v0.1.163..master read individually (diff-level, against the pre-change code's existing-behavior inventory):

Major behavior changes (all intentional, each verified):

  1. Resume inheritance ON by default (Claude 上 resume 导致消息错乱 #1486) — identified clients resuming under a new session id (cc --resume style): OLD = compression state restarted at m00001, stale model citations mis-hit renumbered messages → NEW = byte-exact prefix match (≥8 msgs, strictly longer than tracked chain) inherits the parent's ref assignments + derivedFrom lineage; blocks only when forkAdoption is also on. Opt-out: BILI_RESUME_INHERITANCE=0. Anonymous sessions' world stays fully separated (identified chains excluded from anonymous resolution, # [Bug/Regression] 0.1.59 关闭内容指纹会话兜底 (#286):无会话头第三方 harness 请求被直接 400,不转发上游 #309 intact).
  2. clientError disposition ([0.1.158] MITM 链路上客户端侧疑似闲置连接拆除 RST+上游短周期停滞无看门狗——附一次客户端进程崩溃的完整取证,请评估是否需要加固 #1452/clientError drain path lacks terminal backstop: silent peer holds half-open socket indefinitely (#1528 item-1 follow-up) #1529) — malformed request: OLD = Node default (bare 400 + destroy → RST/ECONNRESET or stranded pooled conn depending on buffer state) → NEW = drain-then-clean-FIN, plus a 30s terminal backstop destroying peers that never FIN after our end() (BILI_CLIENT_ERROR_BACKSTOP_MS, 0 = hold-until-peer-death). The 400→FIN wire delta is explicitly disclosed in the code comments (PR fix: client-facing socket lifecycle hardening + stall guard (#1452) #1528 disclosure). Destroy-after-drain carries no unread residual bytes, so the [0.1.158] MITM 链路上客户端侧疑似闲置连接拆除 RST+上游短周期停滞无看门狗——附一次客户端进程崩溃的完整取证,请评估是否需要加固 #1452 RST signature cannot recur.
  3. OMP native reported windows now honored ([Bug][OMP native][0.1.163] contextWindow 上报后仍使用 registry 窗口,runtime-info 存在跨会话覆盖 #1531) — OLD: header-less omp native always missed the agent-keyed runtime table → reported contextWindow/maxOutput silently ignored (registry-peek fallback); sub-agents clobbered the main session's entry → NEW: conversation-scoped runtime table (model-gated; divergence degrades to miss, never wrong hit), header-carrying clients resolve exclusively via the agent table as before; client dedup widened from model-id to full-field fingerprint so corrective re-reports fire.
  4. Admin origin gate (压缩缓存命中率大幅度降低,看是什么原因 #1537) — Host matching relaxed from exact host:listen-port to bare loopback hostname (port ignored), fixing SSH-forwarded Web UI 403s. Rebinding defense preserved by construction: an attacker's rebound domain can never equal a loopback NAME; regression test pins forwarded-port accept, same-port backward-compat, rebinding reject (with+without Origin), and cross-site-Origin reject.
  5. Graceful shutdown retries dirty CCR content stores — OLD: a transient write failure left the content store dirty-and-lost at SIGTERM flush → NEW: resident dirty stores re-saved before the kernel drain. Production wiring verified live (flushAllSessions() passes sessions.values()); signature backward-compatible, persistence format unchanged.
  6. New knobs, zero-change defaults — BILI_STREAM_STALL_MS (opt-in, off by default; abort flows through the existing Plugin mode: upstream stream truncated before terminal event → client gets bare cut-off SSE (no completion/error event) #721 truncation chain — same wire outcome as an upstream mid-stream FIN, no new wire shape), BILI_KEEP_ALIVE_TIMEOUT_MS (default 5000ms == Node implicit default), BILI_EXPOSURE_LOG_INTERVAL_MS (default 1h — note: adds one periodic info-level [exposure] telemetry line per hour by default), plus debug-level [conn] lifecycle-ledger close lines.
  7. Negative BILI_CLIENT_ERROR_BACKSTOP_MS now falls back to the 30s default instead of clamping to 0 (= hold forever) — matches documented semantics (review fix of feat: terminal backstop for the clientError drain path (#1529) #1530).

High-risk pattern checks: ref-space invariant HOLDS — resume seeding only ADDS entries to the new session's messageRefs (no map ever shrinks), cursor holds at the parent max so pruned-message numbers are never re-issued (unit-pinned by the new tests/fork-adoption-plan.test.ts, which crafts the exact numbering-hole hazard the e2e scenario cannot cover); no fail-fast converted to silent swallowing (adoption/report failures warn-log and fall back to the exact old behavior); thresholds stable; src/update.ts, acp-kernel pin, and release workflow untouched; en/zh CONFIGURATION rows verified in sync with code defaults. Test quality: the #1452 suite asserts real socket/stream behavior (peer error state, timer-leak counts, ledger reason classification, knob-parse matrix incl. negative/garbage values), not tautologies.

本版本引入回归:无。

Verdict: mergeable — clean single-bump branch, full window audited, CI green. Merging is human-only: #1541

一句话中文摘要:逐项审完 v0.1.163 以来全部 17 个内容提交,所有行为变更均为有意且带开关或测试钉住(resume 继承默认开启、clientError 改为排空后 FIN、OMP 窗口生效、admin 门限放宽至回环主机名),未发现回归,CI 全绿,可以合并。

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