[ranxianglei]

What this fixes (D2 of #5095, tracked as #5308)

A collapsed block committed under a shared conversation id during the ≤v0.1.208 era reappears forever in the parent session after upgrading to ≥v0.1.209:

• The #2695 zombie-reap exemption shelters it. A block whose covered substrate was never verified present on this session's wire keeps e==0, so it never accrues a death streak and is never reaped (src/fold-reconcile.ts). That exemption exists exactly so healthy proxy-mode folds (whose covered ids are permanently replaced by the anchor and never reappear) are not reaped. A foreign block from the shared-id era has the same signature — "covered ids never on this session's wire" — so the same exemption gives it permanent shelter.
• Meanwhile the render side (kernel/src/sync.ts stillPresent) keeps the block active as long as any covered-id hash collides with the current request or its anchor is present — which happens readily when the parent and the dead child read the same files. So the foreign summary keeps being re-rendered into the parent transcript, polluting context indefinitely.
• Once rendered into the client's persisted history the bytes become the client's own; the resent copy carries an h_ hash (not acp_summary_bN), so no proxy-side stale-copy drop can remove it afterward. There is no automatic cleanup path for provably-foreign legacy blocks.

Only upgraded sessions that lived through ≤v0.1.208 are affected (the lane split in #4774 / #5184 / #5163 means new foreign blocks can no longer be produced).

The fix — host-side sweep, option A (owner-pre-approved in #5095 / #5308)

New command bili sessions orphan-sweep [id] [--fix] [--json]:

• Provably-foreign criterion, not a heuristic: a block is flagged when its set of covered message ids has zero overlap with its own session's unified-store transcript — reading active AND retired rows (a folded message's row is retired/tombstoned, never deleted, so it still proves the substrate belonged to this session). Zero overlap ⇒ the block references ids that never existed in this session's durable record ⇒ provably external/orphan.
• Legit folds are structurally safe: any genuine fold of this session leaves its covered rows in the store (active or retired), so its overlap is non-empty and it is never swept. This is the negative gate that protects normal compression.
• Dry-run by default — reports findings only, changes nothing. --fix archives.
• Reversible archive via a metadata exclusion list: --fix records each swept block in session.metadata.orphanExcludedBlocks (blockId → {at, reason}) and persists it to the session's index.db row (advancing updated_at so the write wins the full-mode load staleness guard). buildSession drops listed blocks on load. No kstate rewrite, no schema change — deleting the entry restores the block. Idempotent (already-excluded blocks are filtered out before detection).
• Requires full unified storage (BILI_STORAGE_UNIFIED=full); in shadow/legacy mode it degrades to read-only reporting and refuses to perform ("cannot prove", never guess).

This clears the b33-class stock. Option B (block carries creation-time conversation-key lineage — precise, no heuristic) is a kernel-boundary design item tracked separately: it needs a kernel-change declaration + owner approval and is deliberately out of scope here.

Config surface

New CLI subcommand (no new env var, no new config-file field):

Surface	Added	Why the existing surface can't express it	Compat
CLI	bili sessions orphan-sweep [id] [--fix] [--json]	No existing maintenance command covers block-level provenance cleanup; this is a per-session operation, so it sits under bili sessions <sub> next to rm / archive / restore, reusing the existing --fix perform-token convention from bili doctor --fix.	Additive. Dry-run by default (report-only). No-op unless invoked with --fix. Gated on full unified storage. Reversible.

Placement note: the issue proposed this as "bili doctor --fix 孤儿块清扫". I placed it under bili sessions because (a) it operates on individual sessions' blocks, matching sessions rm/archive/restore; (b) doctor --fix currently means dsh-ledger repair specifically, and overloading it would blur semantics. Same capability, same --fix token, better home — trivial to move under doctor if you prefer.

Behavior change

None at runtime. The sweep runs only when explicitly invoked. The single load-path change — buildSession now filters out blocks listed in metadata.orphanExcludedBlocks — is a no-op fast path when that key is absent (the default for every existing session), so loading behavior for all current sessions is unchanged until an operator runs orphan-sweep --fix. Old→new: state: mergeState(parsed.state) → state: filterExcludedOrphanBlocks(mergeState(parsed.state), metadata?.orphanExcludedBlocks); identical output whenever the exclusion list is empty.

Verification

• tsc --noEmit clean on current master (rebased onto f6e69aee1; the intervening commits touched kernel/, so a full rebuild + typecheck ran).
• New suite tests/orphan-sweep.test.ts green against a real full-mode session: a zero-overlap block is flagged while an overlapping fold is not (both directions, Test Blind Spot Retrospective: Reinforcing Negative Assertion Discipline, Tightening Fake Upstream Strictness, Thin Inst | 测试盲区复盘:补反向断言纪律、假上游严格化、薄安装冒烟、形态矩阵与生命周期覆盖 #3617 gate discipline); dry-run is a no-op; --fix archives only the foreign block; reload confirms the foreign block is gone and the legit one stays; idempotent; reversible.
• Targeted regression tests/compress-from-store.test.ts (the buildSession load path) still 4/4.
• Full-suite verdict is left to CI on the pushed head.

Closes #5308

Related: #5095 (source, D1 closed) · #5184 / #5163 (D1 two-layer fix) · #2695 (e-gate origin) · #3640 (anchor parking) · #4774 (lane split)

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-31_orphan-block-sweep (29fde22)

Latest build — install this exact version

npm install -g billion-context@0.0.7472-pr.5334.2c0aa80

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

Recent builds (commit → version)

commit	version	build
29fde22	billion-context@0.0.7472-pr.5334.2c0aa80	build

Download artifact

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

tar xzf billion-context-pr5334.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

查重: 不重复 —— 本 PR 是 #5308-D2(#5095「孤儿块清理」第 2 项)的唯一实现;按失败签名与关键词检索 open+closed issue/PR,无并行工作。

标签: bug + scope/storage + p2(从原 p1 下调:这是 ≤v0.1.208 时代的有限存量污染,不是缓存泄露类只增不减的失控增长;受影响会话仍可用、有绕行,符合 p2 口径)+ 保留 pr。

中间状态(代码审核已完成,CI 红分账已完成):

1. 代码侧:7 个文件的实现已逐项对照既有行为核验(归档会话不会误伤、retired 行不删故零重叠判定成立、staleness guard 保证 --fix 不被旧 JSON 回灌等),未发现缺陷;新测试带双向断言(Test Blind Spot Retrospective: Reinforcing Negative Assertion Discipline, Tightening Fake Upstream Strictness, Thin Inst | 测试盲区复盘:补反向断言纪律、假上游严格化、薄安装冒烟、形态矩阵与生命周期覆盖 #3617 纪律)。
2. CI 红(head 29fde22)共 3 条静态门,已在纯净 master f6e69ae 与 PR head 两棵树上本地复跑确认分账:
   • lifecycle-ledger-guard = 继承 master 红(缺账 BILI_SESSION_CONTENT_STORE_PIN / BILI_SPAWN_FINGERPRINT,两树同红)→ 已按哨兵指令 merge origin/ci-fixes 进分支;
   • docs-sync = 继承 master 红(New built-in chat prompt pack: default compression rules for long-term companionship / multi-session memory scenarios (e | 新增内置 chat 提示压缩包：长期陪伴 / 多会话记忆场景的默认压缩规则（compress.promptPack="chat" 启用） #5319 手改 CONFIGURATION.md 未同步种子 website/config-reference/compress.yaml,EN/ZH 各 ±"chat" 一行,两树同红;ci-fixes@0371089 尚不含此修复)→ 待哨兵收敛,已在 CI Control Center: Centralized CI Failure Handling & Status Sync (Comment all CI issues here; do not open new issues) | CI 总控: CI 失败集中处理与状态同步(所有 CI 问题在此评论, 勿开新 issue) #4599 留言;
   • unused-exports = 混合:1 条继承(CodexWindowSource)+ 2 条本 PR 自带(OrphanBlockFinding / OrphanSweepPlan——exported 但全仓无模块边界引用)→ 已按门禁自身处方在分支上直接去掉 export(先例 [PR] fix: surface unsigned functionCall/thought parts on the Google lane before they 400 upstream #5277、[PR] fix(observability): surface suspected DNS-resolver wedge as logged + panel alarm (#5221 phase 2) #5322)。

修复合入后等 CI 重跑,再给最终八问结论。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

最终结论(八问)

1. 是什么问题:≤v0.1.208 共享会话 id 时代,折叠块可能被提交到「不属于它」的会话名下;升级后这些块在父会话里永久复活渲染——每次请求都带着幻影内容,污染上下文、白烧 token。此前没有任何清理路径(Claude Code sub-agents of the same type share a single compaction lane: unrelated sessions bleed content into each other | Claude Code 同类型子代理共用一条压缩车道:无关会话互相串内容、折叠暂停;旧版共享父 id 的折叠块升级后仍残留在父会话 #5095 的 D2)。本 PR 新增 bili sessions orphan-sweep(默认 dry-run,--fix 归档)并在加载时过滤被排除的块。
2. 什么原因导致的:因果链两层——(a) 产源:共享 id 时代把 A 会话的折叠状态写进了 B 会话的记录;(b) 为何杀不掉:两个机制互相庇护——kernel e-gate 对「substrate 从未出现过」免计死亡连击 + render 侧 stillPresent 只要任一 covered id 撞上父会话读过的消息就判 active;而唯一能收尸的 orphan-gc 只看当前 wire,这类碰撞恰好落在 wire 上。本 PR 的判别器改用存储存在性(与 wire 存在性独立):covered ids 与会话 unified-store transcript 行零重叠 ⇒ 可证外来;合法折叠必然保留其 covered 行(retire 不 delete),故永不被误伤。
3. 影响范围与严重程度:仅 ≤v0.1.208 升级来的存量会话(新版本不再生成新孤儿);受影响会话仍可用,只是上下文带幻影内容 + token 成本;有限存量、非只增不减 → p2(已从 p1 下调,理由见上条回复)。--fix 仅在 full 统一存储模式生效,shadow/legacy 只报 performable=false 不动数据。
4. 怎么修复的:检测 = 逐会话比对 block.effectiveMessageIds 与 store transcript msg_ids(active+retired,两侧经 baseIdOf 归一);transcript 为空的会话直接跳过(判据空真,防误报)。修复 = 写 metadata.orphanExcludedBlocks 排除清单(可逆、不改写 kstate),buildSession 加载时按清单过滤 state.blocks。测试是真行为断言且双向齐全(Test Blind Spot Retrospective: Reinforcing Negative Assertion Discipline, Tightening Fake Upstream Strictness, Thin Inst | 测试盲区复盘:补反向断言纪律、假上游严格化、薄安装冒烟、形态矩阵与生命周期覆盖 #3617):外来块被抓/重叠的合法块不被抓、dry-run 无副作用、--fix 只归档外来块、重载后外来块消失且合法块保留、幂等、清掉清单即还原——全部跑在真实 full-mode store 上,不是恒真断言。
5. 是否有回归问题:逐项审过的既有行为——① 归档会话:loadAll() 枚举含 tombstone 的全量行,但归档会话无 transcript 行 → 命中跳过分支,不会误伤;② 「retire 永不 delete」前提:src/storage/store.ts 全部 4 处 DELETE FROM messages 逐一核验——ingest-repair 只删 retired 字节重复且 live twin 同 msg_id 存活、另两处是整会话删除/整会话归档,均不产生「行没了但会话还在」→ 零重叠判定成立;③ staleness guard:full 模式 buildForLoadAll 仅当 JSON savedAt > dbAt 才回灌旧 JSON,sweep 写入 updated_at=now 使 DB 行胜出,排除清单不会被旧文件回灌;④ 排除清单缺失/畸形/空 = no-op fast path,对无此键的正常会话零影响;⑤ 未触碰 kernel/、wire 格式、配置面、消息 ref/id 逻辑。已知局限(与既有 sessions rm/archive、storage gc 同类,非本 PR 引入):持有 --fix 之前内存态的宿主进程若在 --fix 之后落盘会覆盖排除清单,直到该会话下次 fresh load——PR 正文已声明「next load/restart 生效」,属离进程维护命令的共同模型。
6. 是否破坏缓存:不影响前缀缓存/重试/去重键——过滤发生在加载期(state.blocks 收缩),出站渲染照常走服务端正史;排除清单存 metadata_json,不参与 wire 身份与 checkpoint。
7. 是否改了设计方向:没有。Option B(块携带创建时 lineage)明确留作 kernel 侧独立项(Legacy foreign fold blocks linger forever in parent session after upgrade | 旧版共享父 id 的外来折叠块升级后在父会话永生残留 #5308),本 PR 只做 host 侧维护命令 + 加载期过滤,与现有 storage-maintenance 命令族一致。
8. 是否引入新配置:无 env/config 字段。新 CLI 子命令 bili sessions orphan-sweep [id] [--fix] [--json];复用既有 --fix flag 名(与 doctor 的 --fix 互斥,因两条命令本身互斥)。

CI 现状与处置(head 29fde22 → 新 head 2d8f680 已推送):3 条静态门红 = 双树本地复跑分账——unused-exports 里 2 条本 PR 自带(OrphanBlockFinding/OrphanSweepPlan 无模块边界引用)已按门禁处方在分支修掉(commit 978ab77);lifecycle-ledger-guard 继承红已 merge ci-fixes@0371089db 转绿(本地 OK);docs-sync(#5319 ±"chat" 种子漂移)与 CodexWindowSource 两条继承红待哨兵收敛进 ci-fixes,已在 #4599 留言。test 矩阵此前因门红被 skip,尚无全量套件判定——等收敛后重跑即为终审证据(本地未跑全量,遵本地测试纪律)。

判定:代码本身审核通过、可合;当前被 2 条继承红阻塞,哨兵把 docs-sync 种子修复与 CodexWindowSource 收敛进 ci-fixes 后,分支再 merge 一次即可全绿合入。