diff --git a/CHANGELOG.md b/CHANGELOG.md index 978e98698a..ce0599984c 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -16,6 +16,7 @@ Versions follow the merge of a `*_release-v*` branch; CI publishes to npm on tag - **系统架构文档体系落地(owner 指令 2026-11-08):`docs/architecture/` 成为实时维护的架构枢纽**:新增 `docs/architecture.md` / `docs/architecture.zh-CN.md` 总枢纽(整体架构图中英双版转正 + 8 模块总表),首个样板模块页 `docs/architecture/storage.{md,zh-CN.md}`(统一存储:职责/原理/工作机制/硬不变量,配模块图中英双版)。文件级地图 `reference/architecture.md` 补齐 125 个缺失条目(新存储层 / marker / fastpath / lanes 等全部新模块)并升级为 CI 强制全量覆盖。新增 CI 车道 `architecture-docs`(`tools/check-architecture-docs.mjs`):地图覆盖全 `src/` 文件、总表↔页面成对、图 svg 优先 + png 备选齐备、SVG 基础良构(未转义 `&` 曾让英文图线上破图)。AGENTS.md「Documentation Layout」新增实时维护规则:改结构 ⇒ 同 PR 更新页 + 图 + 总表。其余 7 个模块页(pipeline / lanes / kernel / identity / fastpaths / webui / update)同一 PR 全部补齐——总表 8 行全 ✅(每页 = 中英双版文档 + 中英双版图)。 ### Fixed +- **原地更新后,长驻宿主重拉的 lane 代理不再被打上旧构建的代码指纹(#5354)**:`entryScriptFingerprint()` 按路径把入口脚本的 sha256 缓存到进程退出,之后再也不看盘上的文件。omp 插件就跑在长驻的 agent 宿主进程里——`npm i -g` 把新 `dist/index.js` 原地换上之后,宿主重拉 lane 代理时仍把**旧**摘要经 `BILI_SPAWN_FINGERPRINT`(#4125)交给子进程:子进程跑的是新代码、登记的却是旧指纹,其他宿主一律判「代码指纹不一致」接不上,只好各自另起代理,每重拉一次端口就往后跳一格、`port-zone.json` 的 `lanes.omp` 随之漂移(实测一天之内 18792 → 18797,三个同构建代理各占一口)。现在缓存以文件身份(dev/ino/size/mtime/ctime)为键:文件被替换或改写即重算,读不到时返回 `undefined` 走无指纹路径。tests:`tests/issue5354-fingerprint-follows-the-file.test.ts`(改名替换/原地改写/删除三种退化)。 - **纯 SQLite 姿态下证实损坏不再静默降级 JSON,改为响亮拒启(owner 策略 2026-11-21,#7156 生产事故)**:quick_check 证实 index.db 损坏(含「首个 post-open 写入」撞损坏的大库/缓存判决分支,#7156 FIX4)且进程有效姿态是**纯 SQLite**(有效 mode full + sqlite 后端 + 无加密 + jsonWriter off = 2026-11-13 缺省翻转后的缺省,或显式 off)时,旧契约(响亮告警 + 不挂索引照常启动、JSON 信封服务)在这里会服务一个近乎为空的非持久视图,其后每一轮都只落在非持久副本里——静默丢持久历史,比损坏本身更糟(实测现场:0.1.216 的 3.2GB index.db 损坏后自动「降级 JSON」,面板显示 requests:2/上下文 0,而真实 1563 请求/124k 上下文/$12.41 成本全在索引里)。现在该姿态 REFUSES TO START:文件原样不动(不搬不拷不重建),日志与诊断面只给可操作的手工修复(备份 index.db/-wal/-shm → 手工移开 → 再启动,从双写期信封重建);`BILI_STORAGE_ALLOW_LEGACY_DEGRADE=1` 可有意识地恢复旧降级路径(永不缺省、不被任何配置值隐含)。其余姿态(显式双写 full、shadow、legacy、网络后端、加密)信封仍是真恢复来源,旧契约逐字节保持。姿态判定是纯函数(与 persist 的 jsonWriterOff 同一公式,两开关永不分歧);拒启被记录进诊断面(`/__bili/status` 与 diagnostics 可见「REFUSING TO START (pure-sqlite stance)」前缀)且每次启动都重新生效(latch 只短路进程内重扫)。回归:tests/issue7156-pure-sqlite-refusal.test.ts(姿态真值表 + 两种纯 sqlite 拼法与 file 层的损坏拒启/逃生阀/重复拒启 + 字节不变断言 + 双写姿态保留降级契约)、tests/issue4816-integrity-throw-quarantine.test.ts、tests/acp-status-storage-degraded.test.ts、tests/web-storage-banner.test.ts、tests/storage-durability.test.ts 与 storage 全量套件绿。 - **信封导入不再顶掉 index 已有的压缩块谱系与生命周期统计(#7156 事故第二现场,2026-11-21)**:#7156 的生产现场里,降级窗口(不挂索引、纯信封服务)内的一次会话保存把信封里的 32 个粗块结构写进了 `importRecord`——它把信封当成权威谱系:按 incoming 集合先删后插,index 里 103 个更细的压缩块(及其 summary/content 引用链)被整批删除,生命周期统计也从 requests:1563/124k 上下文/$12.41 成本回退成 requests:2/0。之后每一轮 preflight 都在 32 粗块的地基上折叠,上下文装不进模型窗口、会话不可用。现在 `ingestSession` 新增 `envelopeImport` 标记(仅 `UnifiedShadow.importRecord`/`importRecordAsync` 两个导入入口传):存储侧已有块谱系时**只采纳信封里 seq 更高的新消息**,块谱系与 stale 清理一律跳过(index 无块的真正首开导入不变,仍整体采纳信封结构);stats_json 同带单调守卫——envelopeImport 下已存 requests ≥ 传入 requests 时保留已存值(仅导入路径生效,organic 保存仍是权威且零额外开销)。回归:tests/envelope-import-block-lineage.test.ts(细块保留 + stats 只升不降 + 首开导入采纳 + organic 权威语义不变 + importRecord wiring 五用例)、tests/storage-ingest-watermark.test.ts。 - **dsh 运行时注册的 preset 继承压缩完全禁止(#7190,关 #7178 / #7062 / #7169)**:dsh 的 preset 注册表 `AgentPresetRegistry.register(definition)` 是**启动后才出生**的 preset 行唯一的出生点,它把嵌套行直接挂进内存 PresetTree(其 `commit()` 为空实现),从不经过 loader 的 `internal/config` 变换——bili 的 #4147 结构性 disable 重写(`bili-compaction-gate`,#3354/#3662)永远碰不到这些行。于是运行时注册 preset 的压缩引擎是全副武装的(每步压力触发 + 模型切换前置压缩 + 溢出恢复三条触发器都在 dsh 的 surface 计量上点火),而 bili 侧 wire guard 拒绝每一次压缩调用:#7062 的逐步 422 风暴、#7178 的模型切换死局(`MODEL_SWITCH_COMPACTION_FAILED` 永不满足前提)、#7169 的 landed-checkpoint 触发 rebase 抹掉 bili 自己的折叠块再 DIGEST DRIFT 冻结信封,全是这台引擎在开火。修法与 boot-time 行同一判决、同一机制:gate 入口(always-on,bili-native 关着也活着)现在包一层 `agentPresets.register`——定义里的 compaction-basic / command-compact 行在挂入前就地改写为 `disabled: true`。判决按每次注册调用检查(bili-native 没宣告持线 = 定义原样通过,dsh 原生压缩照常);无 compaction 行 / 畸形定义按身份原样通过(注册表自己的校验看到字节级相同的输入);Symbol-for 标记保证重 apply 不双重包裹;宿主没有该注册表 = 无运行时 preset 可注册,静默不包裹。遥测新增 `runtime` / `runtimeWrapped` 维度,#6770 覆盖一行据此报告「register() 已覆盖——尚无运行时注册」或「N 个运行时注册被改写」,wrap 未知时不报假缺口。**非模型可见文本改动**(只动 dsh 侧 web banner 的覆盖一行与日志)。dsh 源码依据(deepseek-ai/deepseek-harness @ 2e39e70,与现场 0.2.2-rc.3 逐点核验一致):`packages/preset/agent-preset-registry/src/index.ts:96`(register 立即激活、重复 id 抛错)、`mount.ts:22-23`(PresetTree 空 commit)、per-fiber `internal/config` waterfall(`vendor/cordis/src/fiber.ts:691-700`)——preset 作用域 fiber 在自己的上下文里解析配置,根上注册的 global hook 不触发。回归:tests/issue7190-dsh-runtime-preset-coverage.test.ts(新增 10 条行驱动正反用例:改写/身份保持/判决关/重包裹/`this` 转发/无注册表降级/get 回退/遥测去重 + 覆盖文案四形态)、tests/issue3354-dsh-compaction-disable.test.ts、tests/issue6770-compaction-coverage-text.test.ts、tests/issue4147-dsh-overlay-structural.test.ts 全绿。 diff --git a/src/instance.ts b/src/instance.ts index 7fa0dd5749..7b872a29cb 100644 --- a/src/instance.ts +++ b/src/instance.ts @@ -321,20 +321,31 @@ export async function findPresentUnverifiedSameLaneTwin( * compares it against the hash of the script it would spawn — so "same * version" is never enough: two installs of 0.1.x with different dist * contents (local rebuild, npm link, unpublished branch) must not be - * confused for one codebase. Unreadable file → undefined → never attach. */ -const scriptFingerprints = new Map(); + * confused for one codebase. Unreadable file → undefined → never attach. + * #5354: the cache follows the FILE, not the process — it is keyed on the + * script's identity (dev/ino/size/mtime/ctime) and re-hashes when that + * changes. The omp plugin lives inside the long-running agent host; a + * path-only cache kept the first build's digest across an in-place update + * (`npm i -g` renames the new dist over the old path), and #4125 then + * stamped every respawned child with it — same code, different registered + * fingerprints, no attach, a laddered lane port per respawn. */ +const scriptFingerprints = new Map(); export function entryScriptFingerprint(scriptPath?: string): string | undefined { if (!scriptPath) return undefined; const resolved = path.resolve(scriptPath); - const cached = scriptFingerprints.get(resolved); - if (cached !== undefined) return cached; + let identity: string; let digest: string; try { + const stat = fs.statSync(resolved); + identity = `${stat.dev}:${stat.ino}:${stat.size}:${stat.mtimeMs}:${stat.ctimeMs}`; + const cached = scriptFingerprints.get(resolved); + if (cached?.identity === identity) return cached.digest; digest = createHash("sha256").update(fs.readFileSync(resolved)).digest("hex"); } catch { + scriptFingerprints.delete(resolved); return undefined; } - scriptFingerprints.set(resolved, digest); + scriptFingerprints.set(resolved, { identity, digest }); return digest; } diff --git a/tests/issue5354-fingerprint-follows-the-file.test.ts b/tests/issue5354-fingerprint-follows-the-file.test.ts new file mode 100644 index 0000000000..567e495c02 --- /dev/null +++ b/tests/issue5354-fingerprint-follows-the-file.test.ts @@ -0,0 +1,48 @@ +// #5354: entryScriptFingerprint() cached a script's digest by path for the +// life of the process. The omp plugin lives inside the long-running agent +// host, so after an in-place update (`npm i -g` renames the new dist over the +// old path) the host kept the FIRST build's digest — and #4125 then stamped +// every child it respawned with it (BILI_SPAWN_FINGERPRINT). Children running +// the same new code registered different fingerprints, could not attach to +// one another (code-fingerprint-mismatch), and each respawn laddered onto the +// next lane port. The digest must follow the file on disk, not the process. +import { test, after } from "node:test"; +import assert from "node:assert/strict"; +import { createHash } from "node:crypto"; +import fs from "node:fs"; +import os from "node:os"; +import path from "node:path"; +import { rmrf } from "./tmp-rm.ts"; +import { entryScriptFingerprint } from "../src/instance.ts"; + +const DIR = fs.mkdtempSync(path.join(os.tmpdir(), "bili-5354-fp-")); +after(() => rmrf(DIR)); + +const sha256 = (text: string): string => createHash("sha256").update(text).digest("hex"); + +function script(name: string, text: string): string { + const file = path.join(DIR, name); + fs.writeFileSync(file, text); + assert.equal(entryScriptFingerprint(file), sha256(text), "first read hashes the bytes on disk"); + return file; +} + +test("a script replaced by rename — how npm installs an update — yields the new build's digest", () => { + const file = script("rename.js", "// build A\n"); + const staged = `${file}.new`; + fs.writeFileSync(staged, "// build B\n"); + fs.renameSync(staged, file); + assert.equal(entryScriptFingerprint(file), sha256("// build B\n")); +}); + +test("a script rewritten in place yields the new build's digest", () => { + const file = script("in-place.js", "// build A\n"); + fs.writeFileSync(file, "// build B, a longer one\n"); + assert.equal(entryScriptFingerprint(file), sha256("// build B, a longer one\n")); +}); + +test("a script removed after it was fingerprinted is unreadable → undefined, never the old digest", () => { + const file = script("removed.js", "// build A\n"); + fs.rmSync(file); + assert.equal(entryScriptFingerprint(file), undefined); +}); diff --git a/tests/resume-miss-timestamp-seed.test.ts b/tests/resume-miss-timestamp-seed.test.ts index c73731bd74..12a0dba6a2 100644 --- a/tests/resume-miss-timestamp-seed.test.ts +++ b/tests/resume-miss-timestamp-seed.test.ts @@ -275,6 +275,7 @@ test("resume-inheritance MISS seeds the timestamp ledger from the closest tracke store2.closeQuietly(); // shared singleton hygiene: leave no chain behind for later tests in this file prefixAffinity.forget("new"); + prefixAffinity.forget("old"); await closeServers([proxy2, upstream], "miss-seed phase 2"); if (prevStampEnv === undefined) delete process.env.BILI_MESSAGE_TIMESTAMPS; else process.env.BILI_MESSAGE_TIMESTAMPS = prevStampEnv; @@ -334,6 +335,10 @@ test("weak-adoption gate: equal-depth replays and truncated subviews start FRESH const prevStampEnv = process.env.BILI_MESSAGE_TIMESTAMPS; delete process.env.BILI_MESSAGE_TIMESTAMPS; _resetSessionsForTest(); + // Defense in depth: drop any stale "old" chain a previous test left in the + // shared singleton tracker so this test's phase-1 chain is the only "old" + // the phase-2 hydration can see (the #5608 Windows-CI stale-tracker race). + prefixAffinity.forget("old"); // Ten distinct messages: the miss diagnostics only judge chains at or above // MIN_RESUME_PREFIX (8) — the truncated subview below stays just above that // floor so it still produces a closest candidate.