[github-actions]

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 v0.1.161

• fix: gate the clamp baseline on usage-grade source; pin the gate in plugin mode: no hard backstop when the agent ignores the compress nudge — context grows past the window and upstream 400s #453 backstop
• test: pin the FAILED-branch dropped-entry note (review follow-up)
• docs(license): MIT plus one attribution term for user-facing products
• fix: gate stale estimate baselines out of nudge/preflight sizing (emergency shrink / preflight 的本地估算用错了视图：超长会话（客户端历史 127 万，压缩后 outbound 仅 ~16 万）在上游失败无 usage report 时，bili 按原始历史估算 #1492)
• fix: surface parse-dropped compress entries; normalize single-object content (compress failed may be caused by params #1494)

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-27_release-v0.1.162 (c27f000)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.1443-pr.1510.c57e7a7-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-pr1510.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 收到。这是 one-click release 的自动 fallback PR(直推 master 被 branch protection 拦下)。初步核对:

• PR head c27f0006bd 已拉取核对:相对 master 仅 改 package.json + package-lock.json 版本行(0.1.161 → 0.1.162),release commit 干净、符合 one-version-one-commit 规范。
• CI 全绿(13/13):test ×4 矩阵(ubuntu/windows × node 22/24)、build-artifact、version-guard、e2e(real codex)、hermetic registry e2e(local verdaccio)、各 client lane 零 token 回归全部通过。
• 正在按发版审计规范对 v0.1.161 以来的全部变更(5 个内容提交)做区间回归审查,结论稍后跟帖。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete. I pulled the branch (2026-09-27_release-v0.1.162, head c27f0006bd), cross-checked the diff against CI, and ran the version-range regression audit over v0.1.161..PR head (all 5 commits) rather than just the two version lines.

PR state

• Open, not draft, mergeable, base = current master (a1a9dba0), linear history (merge-base == base).
• Diff vs master is exactly the release commit: package.json + package-lock.json version lines 0.1.161 → 0.1.162 only — one-version-one-commit discipline holds.
• CI 13/13 green on head: test ubuntu/windows × node 22/24, build-artifact, version-guard, real-codex e2e, hermetic registry e2e (local verdaccio self-update chain), dsh/opencode/pi lane zero-token checks, codex CODEX_HOME overlay.
• npm latest = 0.1.161 = master's version → no drift; 0.1.162 is the correct next patch. acp-kernel pin unchanged in the window (0.0.95) — no cross-repo dependency change.

Window audit — existing behaviors walked, item by item

Change	Existing behavior touched	Verdict
ea531307 fix #1494	compress tool call parsing/receipts	Intentional, disclosed
c741120f fix #1492	nudge sizing, preflight trigger, failure-shrink arming	Intentional, disclosed
ee125241 clamp gate (#1503 follow-up)	#453 output clamp baseline, count_tokens preview, forged-compact handoff, absorb view	Intentional, disclosed
11c3a161 test pin	none (test-only)	No behavior change
42202e0e license docs	none (docs/legal only)	Policy change, no code

Behavior changes (old → new, all intentional):

1. compress failed may be caused by params #1494 — previously hard-failing compress param shapes (single-object content, bare top-level array, string-encoded JSON args) now normalize and fold (old: whole call failed with kind=content-not-array/not-object; new: normalized to canonical form). Canonical inputs pass through byte-identical (verified in the diff). On partial parse-drops the receipt now carries an explicit [N of the submitted entr(ies) were REJECTED…] note in both the success receipt and the FAILED message, plus a server-side compress PARTIAL log (old: silent under-reporting — the bug fixed).
2. emergency shrink / preflight 的本地估算用错了视图：超长会话（客户端历史 127 万，压缩后 outbound 仅 ~16 万）在上游失败无 usage report 时，bili 按原始历史估算 #1492 — estimate-grade lastInputTokens baselines (failure-shrink arming) no longer floor nudge sizing (effectiveTokenCount), the preflight trigger, or the preflight loop floor while the payload is measurable; usage-grade baselines unchanged (old: any positive baseline floored all three consumers; new: only usage-sourced floors when measurable). Arming itself now measures with defaultCountTokens instead of estimateTokensFast (CJK-aware lower bound; old chars/4 understated CJK up to ~4×). I verified against the pinned acp-kernel 0.0.95 that the two counters are numerically identical on ASCII input, so the "ASCII byte-identical" claim holds — the delta is CJK-only and confined to the failure-arm path.
3. ee12524 — the plugin mode: no hard backstop when the agent ignores the compress nudge — context grows past the window and upstream 400s #453 clamp baseline (estimateInputTokens) now floors only on usage-grade source, closing the fail-open hole where an estimate-grade poison masked the payload estimate through max() (old: any positive lastInputTokens floored the clamp baseline; new: usage-grade only). count_tokens preview / forged-compact handoff / absorb view feed tokenCount=0 ("unknown") for non-usage baselines so the kernel counts messages itself — steady-state (usage reported) behavior is byte-identical.
4. License — MIT plus one additional attribution term for user-facing products (commercial or OSS: attribute billion-context/acp-kernel with repo links in homepage/docs/About; server-side/embedded use satisfied by shipped docs). Docs-only; worth carrying into the release notes as a policy change downstream users should see.

Regression checks performed: all 10 lastInputTokens write sites at the PR head are consistently source-tagged (usage vs estimate vs preserved-on-decrement); all six gate sites consume the tag uniformly; persist.ts round-trips the source across restarts and legacy sessions without a source safely fall back to per-turn estimates; zero remaining estimateTokensFast usages in src/; both compression modes reasoned (normalize/applyRanges changes sit on the proxy-side compress path; plugin mode suppresses proxy injection and is unaffected); preflight gates alter only poison scenarios, steady-state identical in both modes.

Conclusion

本版本引入回归：无。 All five window changes are intentional and individually disclosed in their commits; no silent behavior drift found. Release commit is clean, CI fully green including the real-client e2e, base is current master.

可合并 (mergeable) — ready for a human merge; merging publishes v0.1.162 via release.yml.

中文摘要：对 v0.1.161→v0.1.162 全部 5 个变更做了版本区间回归审计（#1494 参数容错+显式拒收回执、#1492/钳位门控把过期估算基线挡在 nudge/preflight/钳位之外、许可条款、测试钉桩），逐项确认均为有意且已披露的变更，未发现无意回归，CI 全绿，可以合并。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Closing status — no further work needed on my side:

• Review complete (details in my previous comment): pulled 2026-09-27_release-v0.1.162 (c27f0006bd), ran the full version-range regression audit over v0.1.161..head (all 5 commits: compress failed may be caused by params #1494, emergency shrink / preflight 的本地估算用错了视图：超长会话（客户端历史 127 万，压缩后 outbound 仅 ~16 万）在上游失败无 usage report 时，bili 按原始历史估算 #1492, plugin mode: no hard backstop when the agent ignores the compress nudge — context grows past the window and upstream 400s #453 clamp gate follow-up, test pin, license docs). 本版本引入回归：无 — every behavior change is intentional and disclosed old→new; CI is 13/13 green on head including real-codex e2e and the hermetic registry self-update suite; base is current master; release commit is clean (version lines only); npm latest 0.1.161 = master version (no drift); acp-kernel pin unchanged (0.0.95).
• Verdict: 可合并 (mergeable). The freshly posted build-artifact note is routine CI plumbing for human smoke-testing of the exact build — nothing actionable for me, and it confirms the artifact pipeline succeeded.

Next step is yours: a human merges the PR and release.yml publishes v0.1.162 to npm automatically. No agent action remains.