release v0.1.119

[ranxianglei]

Changes since v0.1.118

Base rebased onto current master (4a1b4e7) — the original head was cut before the OpenCode/pi native stack landed.

OpenCode native stack (V2)

• native 模式总跟踪：pi / omp / dsh / opencode / qoder 等宿主的插件化发布 #820/fix(agent): restore /acp status command under OpenCode V2 (#809) #850 OpenCode 2.x native plugin mode — self-spawning proxy inside the plugin; BILLION_CONTEXT_NATIVE coexistence marker so standalone extensions back off under native mode
• opencode 2.x: no-launcher npm-plugin deployment — plugins: ["billion-context"], plugin lazily spawns the proxy #809 /acp command under OpenCode V2 + attach-to-external-proxy mode; hardened across follow-ups: no model wake (fix(agent): /acp must not wake the model under OpenCode V2 #875), idle-first registration (fix(agent): first /acp under opencode v2 shows the idle notice, not 'no status panel' #882), three-modes (fix(agent): opencode v2 preset-env attach — close the pseudo-attach 404 hole (three-mode delivery verified) #885), panel-first (fix(agent): opencode v2 acp_status renders the status panel (panel-first) #912), visible TUI status panel (fix(agent): make opencode v2 /acp status panel visible in the TUI #878)
• opencode native 插件无法走规范安装：npm 包未暴露插件入口(exports)、native 入口未随包发布，安装器只能写本机绝对路径 shim #925/feat: standard npm install path for the opencode native plugin (#925) #931 standard npm install path for the opencode native plugin; createNativeRoute 每个 model 请求前同步探测 /__bili/health（1500ms 超时）且无 TTL 缓存，与 15s respawn cooldown 的稳定性假设矛盾 #928/fix: TTL-cache the opencode-native per-request health probe (#928) #930 TTL-cache the per-request health probe

OpenCode V1 native support

• feat: OpenCode 1.x native support — V1 .server() hooks on the standard install path #919 OpenCode 1.x .server() hooks on the standard install path; x-bili-plugin-context-window stamped from config-declared limits; SDK-default provider baseURLs caught via global fetch patch
• opencode v1: keep opencode-acp legacy sessions working after switching to bili #920/feat: opencode v1 — keep opencode-acp legacy sessions working under bili (#920) #923 v1 legacy-lane routing — absorbed opencode-acp serves old sessions, bili owns new ones; host ctx passed through to absorbed acp; legacy entries stripped on bili plugin install opencode (OpenCode 1.x: native v1 support via the standard install path (near-v2 effectiveness) #918); hardening from docker E2E

pi native plugin mode

• feat: pi native 模式——fetch 拦截层 + 插件胶水（native 模式 B） #519/feat(pi): native plugin mode — self-bootstrapping proxy + fetch interception #706 self-bootstrapping proxy + fetch interception; compaction-cancel armed at event time; concurrent native respawns deduped; degrades to direct sends when respawn fails (request-construction errors not misclassified)

Cache reconciliation

• feat: per-request cache reconciliation ledger (acp_cache tool) #804 per-request cache reconciliation ledger — the acp_cache tool (held off the wire in fix: bump acp-kernel 0.0.74 → 0.0.77 (hold acp_cache off wire until adoption) #915 pending adoption) is now adopted with its execution path

Compression & context

• [Question] zocde调用子智能体的时候，子智能体有概率出错，导致子智能体停止工作 #862/feat: compress.visibilityMarkers toggle to suppress ACP visibility markers #893 compress.visibilityMarkers toggle to suppress ACP visibility markers
• fix: recover from mid-stream upstream truncations (#413 relaxation + continuation retry) #891/启动横幅硬编码 raw-HTTP-capture(on) 而真开关是 ACP_DUMP_BODY(默认关):日志永远撒谎,取证被带偏过 #413 recover from mid-stream upstream truncations with continuation retry
• [preflight] responses 协议 overhead 估算漏计 input[] 内的 developer/system 消息：codex 请求被系统性低估整个 system prompt #829/fix: count responses input[] developer/system in preflight overhead estimate (#829) #832 count Responses input[] developer/system in the preflight overhead estimate; e2e-fake overflow window recalibrated
• Codex Responses 不发送 max_output_tokens 时 output headroom 不生效，是否可 fallback 到模型 output 配置？ #924/feat: output-headroom fallback to model output when the request carries no budget (#924) #929 output-headroom falls back to the model's own output limit when the request carries no budget
• fix: strip model-echoed render tags from tool-call arguments on every wire path #934/fix: strip model-echoed render tags from tool-call arguments on every wire path (rebase of #934, Fixes #933) #935 strip model-echoed render tags from tool-call arguments on every wire path
• fix(#728): proactive nudge for upstreams that never report usage #731 feed the local outbound-payload estimate to the nudge while upstreams never report usage (fixes [Bug] 上游从不报告 usage 时 nudge 永不触发——上下文无界增长直到超窗（ChatGPT 登录态后端） #728)

Install & platform

• plugin install opencode 把 ephemeral origin 冻结进 BILI_MCP_PROXY，短路 mcp.ts 已有的验活发现；native 用 port:0 使错配必然（#403 遗留实例） #926/fix(plugin): stop freezing a stale MCP origin into opencode installs (Fixes #926) #936 stop freezing a stale MCP origin into opencode installs
• plugin install opencode 写 V1 键名 plugin 到新建的 opencode.json：2.x 下该行无效（靠目录自动发现兜住），且与用户 opencode.jsonc 配置分裂 #927 (+ CI follow-up) version-aware plugin key + jsonc-first target for opencode install; Windows-portable major-version probe; shell-mode .cmd probe + separator-preserving PATH join

Tests & docs

• pin two-arg (input, output) contract on absorbed-acp wrappers (opencode v1: keep opencode-acp legacy sessions working after switching to bili #920 review)
• README.zh-CN opencode coverage synced with English README

---

Commit release v0.1.119 touches ONLY package.json + package-lock.json (version fields, 0.1.118 → 0.1.119). Branch based on current master (4a1b4e7). Per requester instruction no extra local test rounds were run — this workflow's checks (npm ci + typecheck + test + build) gate the publish.

Requested by @ranxianglei in #772.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-18_release-v0.1.119 (a3993e4)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-922

Each push to this PR publishes a new version under the pr-922 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr922.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Verified release v0.1.119 (PR #922, head 2026-09-18_release-v0.1.119 → master):

Branch & commit ✅

• Branch based on current master 0cfde7a (zero drift); tip commit c54c3c4 carries exactly the message release v0.1.119
• Diff vs master: ONLY package.json + package-lock.json, version fields 0.1.118 → 0.1.119 (3 lines total) — nothing else bundled

Content scope ✅

• fix(#728): proactive nudge for upstreams that never report usage #731 (silent-usage nudge fallback) is master HEAD 0cfde7a
• test(launcher): isolate the host client, proxy and CA env vars #895 (hermetic launcher tests) merged as adf6d88, verified as an ancestor of master
• No extra local test rounds per requester's instruction — the release workflow's own npm ci + typecheck + test + build gates the publish

CI 🔄 (at time of writing)

• Green: version-guard, test (ubuntu-latest, 22), test (ubuntu-latest, 24), build-artifact, real codex vs deterministic fake upstream (zero tokens), codex accepts CODEX_HOME overlay + registers bili MCP
• Still running: test (windows-latest, 22), test (windows-latest, 24)

I'll follow up here once the Windows jobs finish. Merge stays human-only — please merge #922 once CI is fully green, then verify with npm view billion-context version.

中文摘要:v0.1.119 release 分支已核验(仅版本字段变更、基于最新 master、#731/#895 均已合入),CI 除两个 Windows 测试外全绿,Windows 跑完我会在此跟进;合并请人工操作。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 CI update: 7/8 green — one failure, diagnosed as a pre-existing Windows-only flake, NOT caused by this release commit.

The failure (test (windows-latest, 22) only):

not ok 101 - #152: ensureRootCA writes combined-ca.pem with MITM root + public roots
  tests/ca-combined.test.ts:36 — AssertionError: 'combined bundle must contain the MITM root CA'
# tests 1744 / pass 1743 / fail 1

Why it can't be this commit: the diff is 3 version-string lines; the identical job was green on master HEAD 0cfde7a; windows-24 and both ubuntu jobs are green on this exact commit. It's intermittent.

Root cause (verified against code + logs):

1. tests/ca-combined.test.ts isolates state by setting process.env.HOME to a temp dir — but Node's os.homedir() ignores $HOME on Windows (it uses %USERPROFILE%). So on win32, caDir() (src/paths.ts → dataDir() → homedir()) resolves to the runner's real shared dir, not the temp home. The test saves/restores XDG_DATA_HOME but never actually sets it.
2. node --test runs files concurrently, and four files touch CA paths (ca-combined, launcher, mitm, web-routing).
3. ensureRootCA() (src/ca.ts) is check→generate(2048-bit RSA, ~1s)→write, non-atomic. Two processes that both see "no root yet" each generate a different root and interleave writes of root-ca.pem / combined-ca.pem, leaving an inconsistent pair on disk (combined from generation A, root file from generation B). Only ca-combined.test.ts asserts on the file contents, so it's the sole visible victim. Linux/macOS are immune because $HOME is honored there.
4. Side note (latent, same mechanism): two bili proxies starting simultaneously on a fresh install could hit the same first-boot window in production. Low probability (keygen window, once per machine), flagging for completeness.

Action taken: I re-ran the failed job — https://github.com/ranxianglei/billion-context/actions/runs/35299920051 — will report the outcome here. If it passes, that confirms the flake and the PR is mergeable as-is.

Suggested fixes (for your call — I didn't open a separate issue per workflow rules):

• Test-side (minimal, hermetic): have ca-combined.test.ts set XDG_DATA_HOME to the temp dir (the xdg() helper checks env vars first, works identically on all platforms) instead of relying on HOME.
• Product-side (optional hardening): make ensureRootCA()'s first-boot write atomic (generate to temp + rename, or O_EXCL creation) so concurrent processes can't clobber each other's root.

中文摘要:CI 唯一失败项(windows-22 的 CA 合并测试)已定位为 Windows 专属的既有 flake——测试用 $HOME 做隔离但 Windows 上 os.homedir() 不认 $HOME,并行测试进程共享真实 CA 目录,ensureRootCA() 非原子首启写入互相覆盖;与本次 release 提交无关(仅版本字符串变更,同一作业在 master HEAD 上是绿的)。已重跑失败作业,结果稍后在此跟进;修复建议见上,由你拍板。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Rerun result: test (windows-latest, 22) → success on the identical commit c54c3c4. All jobs in the run are now green (8/8): version-guard, test ubuntu 22/24, test windows 22/24, build-artifact, plus the codex e2e / CODEX_HOME jobs from the other workflow.

That confirms the diagnosis — intermittent Windows-only CA-dir race, not a defect in this release. The PR is verified and mergeable as-is; merging stays with you per repo rules. If you want the flake killed for good, the one-line test fix (XDG_DATA_HOME instead of HOME in tests/ca-combined.test.ts) is ready to go whenever you open the issue for it.

中文摘要:重跑后 windows-22 作业通过,全绿,确认是偶发 flake 而非本 release 的问题——PR #922 可以合并(合并由你操作);根治方案(XDG_DATA_HOME 隔离)已备好,等你开 issue。