Skip to content

DSH: select models for native compaction with safe same-session handoff #6121

Open

Description

@bluesmilery

Problem

DSH users cannot select models that retain native compaction while other models use bili in the same profile. passthrough alone does not restore the disabled native compaction backend. Switching models in one session also transfers compression ownership: old bili folds can conflict with a native checkpoint, and a long raw bili history can overflow a native summarization call.

Requested behavior

  • Default remains bili. Explicit model rules select native DSH compaction with its normal thresholds and summary settings.
  • Native requests and summaries use the DSH provider transport; bili compression tools are unavailable for those steps.
  • Same-session switching works in both directions and across restart, without affecting other sessions.
  • Preserve raw history, tool-call/result pairing, cancellation and checkpoint transaction semantics. A failed handoff must not commit a partial summary.

Configuration proposal

Propose dsh.nativeCompactionModels (default []), matching provider/model IDs, optionally a trailing model * for a family such as gpt-*. Existing wire passthrough, session disable and manual compact controls cannot express compression ownership. This requires coordinated host gate, routing, tool admission and history-generation handling. Oversized handoff summaries must be bounded.

This is a proposal for maintainer review, not an approved configuration contract. An isolated implementation and tests will follow; running user sessions are not test fixtures. Unsupported DSH APIs must fail visibly.

Related: #6106, #5975, #5300 / #5404, #4872 / #4883, #5245 / #5994.

Prepared with OpenAI Codex at the user's request.

Activity

  1. ranxianglei commented on Nov 10, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

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

    Duplicate screening: not a duplicate. Closest neighbors, compared by symptom + mechanism:

    An implementation already exists. The issue author opened draft PR #6123 implementing exactly this proposal: 25 files, +793/−24, single commit, CI green at my last mirror-cache read (2026-11-10 13:31 UTC; base currently behind master). Per repo rules I will not open a parallel PR — any fixes go additively onto that branch.

    Labels applied: enhancement, dsh, scope/wiring, p2, needs-decision — feature request (not a bug); client DSH; boundary = host gate / routing / tool admission; p2 because a workaround exists (separate profiles, or whole-profile native via disabling bili-native) and nothing is corrupting; needs-decision because the config contract requires owner sign-off before merge.

    Triage assessment (feature request — evaluated, not implemented on sight)

    Reasonableness — a real gap. Compression ownership is binary per profile today: the host-side overlay (#3354) disables the stock preset compaction backend wholesale, and the #1729 wire guard refuses unmanaged native compaction calls on every lane. passthrough does not restore the disabled backend, so a user who wants gpt-* on native compaction while Kimi keeps bili must run separate profiles. Same-session switching additionally transfers ownership unsafely (old bili folds vs a native checkpoint; an oversized raw history fed straight into a native summarize call). Genuine, well-evidenced pain.

    Design fit — good. The draft keeps the stock backend and installs a per-model policy adapter in the preset's isolated realm; native requests bypass bili before any mutation (fetch + WS intercepts, bytes verbatim); after a host history replacement the durable replaceGeneration moves bili state into a <sid>|dsh-generation:N namespace — the same |suffix key grammar as |aux: (src/session-id.ts:113) and the KDD #9 persona keys, so no server-side identity change. Handoff summaries are bounded, tool-pair-balanced, and ride DSH's existing checkpoint transaction (fail without committing a partial replacement). Kernel and provider adapters untouched.

    Coupling — contained, one implicit dependency. With the config key absent, all five legacy touchpoints reduce byte-identically (I code-walked the preset gate, llm/stream dispatch, fetch intercept, WS intercept, and header stamping against master). The one soft spot: selective mode reads process-global state populated in declareDshCompactionActive() (src/agent/dsh-native.ts:1910); if a preset config ever resolved before that apply ran, selective mode would silently degrade to full-disable — fail-closed (a visible requireNativeCompactionReady error later), not corruption, but no test pins the ordering.

    Independent review of draft PR #6123 (GH)

    Verified against master code, not just the diff:

    • Default path unchanged: without the key, the old disableBasicCompaction path runs verbatim and every intercept returns its legacy value.
    • loadConfigFile() never throws (ENOENT/corrupt → {} + log), so the new configureNativeCompaction() call at activation cannot crash dsh startup.
    • The billion-context/dsh-native-compaction plugin row follows the existing billion-context/dsh-compaction-gate subpath convention (dsh.bundle.patch.yml); exports / tsup entry / unused-export registration are all present.
    • The nudge 死锁无检测:同一 session 交替收到两种结构互斥的请求流(巨型单消息 ↔ 多轮)时不告警 #1729 wire guard and dsh-authored-compaction (bili-authored client compaction: let the host's own compaction flow run, but bili writes the checkpoint #3733, default on) are untouched — bili-owned models behave exactly as today (locally authored summary or refusal).
    • No version bumps (correct for a content PR); no kernel/DSH-core changes; docs en/zh complete incl. config-reference seed and a dedicated page.
    • Tests have teeth: negative matching cases, bounded-summary failure modes (abort / provider failure / non-shrinking / oversized indivisible group), adapter gating asserted both directions, wire-level byte-exactness through a real proxy, WS owner-flip withhold-and-close, and an integration script covering failed-summary rollback plus restart replay.

    Gaps I would close on the PR branch before merge (all additive, none changes the contract):

    1. Ordering pin — a test asserting the selective state is visible to the preset-config middleware in production apply order (today only the manual script controls the order).
    2. Tool-admission coverage — the tools.register(def, { when }) option and the execute-time guard for native models have zero test coverage; if a DSH build ignored when, degradation would be graceful (execute throws) but unverified.
    3. Loader-resolution pin — the integration script loads the adapter by direct file URL, while production pushes the package-specifier row into preset config; pinning the specifier form in the script would demonstrate isolated-realm resolution instead of only its shape.

    Minor nits (non-blocking): the test afterEach resets via configureNativeCompaction([]), which deliberately leaves generation isolation ON (documented, but makes intra-file test order load-bearing); the budget formula's 0.9 / 4096 / 16 are commented rather than named constants.

    🔶 需要决策

    背景一句话:外部用户的真实需求 + 其本人的 draft 实现 PR;我独立审核后认为设计成立、默认路径逐字节不变,剩三个可加测的验证缺口;合并卡在配置契约的 owner 签核(Configuration Surface Discipline)。

    待决问题与选项(各带代价):

    • D1 配置契约:A1 按提案批准 dsh.nativeCompactionModels({provider?, model}[],精确/末尾 * 匹配,纯文件级,重启生效,键存在即粘性代际隔离)— 代价:空数组也粘住隔离,退出需删键;A2 批准但改形状(如去掉 provider 维或允许通配变体)— 代价:返工匹配层+文档;A3 否决另设计 — 代价:功能延期。
    • D2 失败语义:B1 键存在时旧宿主(无 surface.replaceGeneration)整 profile 硬失败(可见报错)— 代价:爆炸半径是 profile 级而非模型级;B2 警告降级 — 代价:代际隔离保证被削弱,跨代 fold 误用风险回来。
    • D3 与 GH [PR] Restore dsh native auto-compaction when bili-native is disabled or absent (#5248) #5994 关系:C1 退休后者的全局恢复路线(被本 PR 覆盖)— 代价:owner 自己的 PR 关闭,由其操作;C2 两者并存(全局 vs 按模型分工)— 代价:同一函数两条路径长期维护。

    我的推荐:A1 + B1 + C1,方向确认后我立即在 PR 分支补上述 1–3 三个加测 commit。不决定的默认行为:PR 保持 draft,issue 挂起,什么都不合并。

    结构化八问(中文)

    1. 是什么问题:DSH Web 用户无法在同一 profile 里让部分模型走原生压缩、其余模型继续用 bili;同会话切换模型时压缩归属转移不安全(旧 bili fold 与新原生 checkpoint 冲突、超长原始历史直接喂给原生 summarize 调用可能溢出)。
    2. 什么原因导致的:压缩归属目前是 profile 级二元开关——宿主侧 overlay(See how to integrate the plugin that automatically disables dsh compression. | 自动关闭 dsh 压缩的插件看如何集成 #3354)整体禁用预设自带的压缩后端,线级守卫(nudge 死锁无检测:同一 session 交替收到两种结构互斥的请求流(巨型单消息 ↔ 多轮)时不告警 #1729)无条件拒绝未受管的原生压缩调用;passthrough 恢复不了已被禁用的后端;现有配置面没有「按模型选择压缩归属」这一维度。
    3. 影响范围与严重程度(按选项):
      • A1/B1(按提案批准):默认路径逐字节不变(五个触点已代码级核对),只影响显式配置的模型;旧 DSH 宿主上开启后整个 profile 硬失败(可见报错、非静默损坏)。把 p2 级功能缺口转正为可用特性。
      • B2(警告降级):覆盖面相同,但隔离不变量被削弱,属于「能用但保证打折」。
      • A3/C(否决或延期):现状维持,用户继续多 profile 绕行。
    4. 怎么修复的(推荐方案的机制层面):保留宿主自带压缩后端 + 在预设隔离 realm 装按模型裁决的策略适配器;原生请求在任何改写之前于宿主平面直发(fetch/WS 双拦截、字节原样);宿主历史替换后凭持久化 replaceGeneration 把 bili 状态迁入 <sid>|dsh-generation:N 命名空间(复用既有 |aux:/KDD release v0.1.2 #9 键语法,服务端身份逻辑零改动);超大交接摘要做有界分批并保持工具配对,任何失败都不提交半截替换(沿用 DSH 既有 checkpoint 事务语义)。
    5. 是否有回归问题(按选项):
      • A1/B1:已审既有行为逐项——默认路径五触点逐字节不变;nudge 死锁无检测:同一 session 交替收到两种结构互斥的请求流(巨型单消息 ↔ 多轮)时不告警 #1729 守卫与 dsh-authored-compaction(默认开)未触碰;水位/pf 通道键在默认下与原 sid 相同;内核与 provider 适配器零改动。高危模式排查:无语义漂移、无 fail-fast 变吞错(新增路径全部 fail-closed)、无阈值/时序漂移。残余风险 = 三个验证缺口(应用顺序无测试钉住 / 工具准入路径零覆盖 / 隔离 realm 包名解析仅形状测试),均为加测可修。
      • B2:额外引入降级分支与其测试面,且主动放弃一条隔离不变量。
      • C:无回归(不动代码)。
    6. 是否破坏缓存:不破坏。水位快路径与 pf 通道改以作用域会话键计数(默认下等于原 sid,逐请求等价);跨代边界处协商别名/fork/incremental 种子被有意禁用(该子系统的历史不再跨界复用),属设计内边界而非缓存策略失效;原生请求字节原样直发,provider 前缀缓存不受影响。
    7. 是否改了设计方向:新增一个维度而非转向——同一会话内压缩归属从 profile 级二元变为模型级可切换;数据流上新增一条宿主平面直发旁路(绕过 proxy),kernel 与 provider 适配边界不动,两模式(plugin/proxy)语义在本特性内均保持。
    8. 是否引入了新的配置:是——dsh.nativeCompactionModels(纯文件级,默认 [],需重启宿主;键存在即粘性开启代际隔离,含显式空数组;无 env 旋钮,符合 彻底整治环境变了滥用,除非必要都走配置文件 #2030 文件层惯例)。迁移方式:删除该键(或移除补丁层)即回到旧行为。
    选项 D1 契约 D2 失败语义 D3 与 #5994 综合代价
    A(推荐) 按提案批准 硬失败(可见报错) #5994 退休 补 3 个加测 commit 即可合
    B 批准 警告降级 #5994 退休 削弱隔离保证 + 新降级路径测试面
    C 否决重设计 — 待定 功能延期,用户继续绕行

    收工判定:查重完成、triage 与独立审核完成,方向成立;等 owner 对 D1–D3 拍板,拍板后立即在 PR 分支落三个验证缺口的加测 commit。

  2. ranxianglei commented on Nov 10, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

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

    Follow-up to my review comment above (which carries the full dedup screening, triage, independent PR review and the 🔶 需要决策 block with the Chinese 八问) — current state after that post:

    • Done: duplicate screening; labels applied (enhancement, dsh, scope/wiring, p2, needs-decision); code-level verification of all five legacy touchpoints against master (default path byte-identical when the key is absent); confirmation that <sid>|dsh-generation:N follows the existing |aux: / KDD release v0.1.2 #9 key grammar, that the billion-context/<subpath> plugin-row convention already exists, and that loadConfigFile() cannot crash startup; assessment of the PR's test suite (has teeth, three additive gaps identified).
    • Verified: CI on head b85e797 is green per the mirror cache read at 2026-11-10 13:31 UTC (all test lanes success, base behind master). I attempted a targeted local run of the new unit file afterwards, but this clone has no installed dependencies — installing just to re-run one file CI already passed would add risk without evidence, so the CI verdict stands as authoritative.
    • Remaining, gated on your decision (D1 contract / D2 fail-closed semantics / D3 relation to GH [PR] Restore dsh native auto-compaction when bili-native is disabled or absent (#5248) #5994): on direction, I land the three verification-gap test commits additively on the PR branch (ordering pin, tool-admission coverage, isolated-realm loader-resolution pin). Until then the draft PR stays untouched.

    No further action needed from you besides the D1–D3 call; default if undecided: nothing merges.

  3. ranxianglei commented on Nov 10, 2026

    @ranxianglei
    Owner

    dsh.nativeCompactionModels 这个和三级配置是否对齐?那里有很多字段

  4. ranxianglei commented on Nov 10, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

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

    好问题。直接回答:不对齐——是有意不放进三级树,但您指出的「两处查 per-model 压缩配置」的碎片化是真实代价,有具体修法(见下)。

    为什么进不了三级树(不是措辞偏好,是四个机制性差异):

    1. 消费平面不同。三级树是 proxy 服务端逐请求解析的(longest-URL-prefix 匹配 + 模型名,docs/configuration.md:1412-1418),走 knobs 级联(env > file > default)、Web UI 热重载。而压缩归属必须在 DSH 宿主进程内、请求路由之前裁决(隔离 realm 里的策略适配器,读的是 loadConfigFile() 文件层,重启生效,无 env 档)。放进树里要么在宿主侧复刻一套匹配逻辑(新增一份会漂移的规则副本),要么为这一个决策开宿主↔proxy 通道——耦合比一个扁平列表更差。
    2. 身份命名空间不同。树的键是路由后的 providers[url].models[model](URL 键);归属裁决用的是 DSH 自己的 provider/model ID(如 codex + gpt-*),此时还没有任何 bili 路由概念。两套 ID 并不总是一一对应(recipe/bind 拆分、账号池)。
    3. 通配符无处安放。gpt-* 末尾前缀匹配在树的精确键表达不了。
    4. 合并语义无意义。树的契约是「per-field、deepest-wins、未设不清除」——这是参数调优的合并语义。「谁拥有压缩」是二元路由决定,不存在跨层合并;放进来只会制造「global 说 native、model 说 bili 谁赢」这类本不该存在的配置态。

    树里确实有先例值得正面回应:compress.reasoningGuard.enabled 就是「作用域由在树中的位置决定、没有单独模型列表」的行为开关(docs/configuration.md:1686)。它成立的前提正是上面 1-4 的反面——它在服务端消费、按路由后身份、精确模型键、热重载。归属开关四条全不满足,套那个模式只剩外形对齐。

    放置先例:DSH 车道专属开关本来就归在 dsh.* 下(dsh.authoredCompaction、dsh.contextTab、dsh.contextTabLabel),新键跟随该分组;文档里也有明确记载某些块「deliberately do NOT participate in the three-level compress merge」(configuration.md:941)——不过那条理由(无 per-provider/per-model 语义)对本键不完全适用,本键的正当理由是上面四点,我会在文档里写清而不是引用那条。

    您的关心里成立的一半——可发现性:我查了,Compression Tuning 章节目前对「哪些模型根本不走 bili 压缩」零提示(grep native/ownership/passthrough 为空)。用户调 per-model 压缩参数时不会被指引到归属开关。修法(纯文档、无语义变更,已列入拍板后落进 PR 分支的补交批次):①Compression Tuning 开头(en/zh)加一行交叉引用——「bili 是否压缩某模型由归属开关 dsh.nativeCompactionModels 决定,本节字段只对仍由 bili 压缩的模型生效」;②docs/dsh-native-compaction.md(en/zh)加一段「为什么不放进三级树」,把上面四条写进去,防止后人再提同样的问题。

    立场不变:D1 仍推荐 A1(按提案形状批准)+ 上述文档交叉引用补进补交批次。等您对 D1/D2/D3 拍板。

  5. lisi288 commented on Nov 11, 2026

    @lisi288

    Another unroutable DSH lane: dsh-kiro

    Environment: bili 0.1.213 (bundled in a desktop profile with @deepseek-ai/dsh-web-app), DSH Desktop 0.2.2-rc.1, macOS arm64.

    dsh-kiro (github:dat-lequoc/dsh-kiro) does not go through globalThis.fetch. It builds its own request path over node:https / node:tls (an HTTP CONNECT tunnel when proxyUrl is set) for Kiro's generateAssistantResponse call. So it is invisible to installFetchChain, same shape as the lanes in #6106 and #5033.

    What happens:

    • The load-time half still applies: preset rows are stripped of compaction-basic / command-compact.
    • Kiro traffic never reaches the proxy: ~/.local/share/billion-context/sessions/ contains no codewhisperer / runtime.*.kiro.dev bucket at all (only volcengine ark, api.deepseek.com, a local gateway, and 127.0.0.1 entries).
    • Result: a long ops session on kiro/claude-opus-5.5 grew from 891K to 977,074 reported input tokens with zero compaction/* events in the session log (the only compactions present were five imported historical folds), then the upstream returned 400 CONTEXT_WINDOW_EXCEEDED ("Input content length exceeds threshold.") twice in a row. The session is stuck.
    • Control: the same session, switched to an ark/deepseek-v4-1-flash-economy model (which does reach bili), compacted once and dropped from ~671K to ~159K tokens — so the session is perfectly compactable, the compression owner just never saw it.

    The adapter itself already classifies the overflow correctly (CONTENT_LENGTH_EXCEEDS_THRESHOLD -> CONTEXT_WINDOW_EXCEEDED); it is the disabled host backend that would have acted on it, as the dsh-kiro README's Token budget section describes.

    dsh.nativeCompactionModels as proposed in #6123 would fix this for us ({ "provider": "kiro", "model": "claude-opus-5.5" }). I have built #6123 at cd7b492d against DSH 0.2.2-rc.1, the bundled script passes ({"result":"passed","realDSH":"0.2.2-rc.1"}), and I am willing to run longer real-model acceptance and report back. Happy to test whatever config contract gets settled.

  6. ranxianglei commented on Nov 11, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

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

    Thank you for the detailed report — it's a valuable data point for the pending decision. From our side:

    What we can confirm from the bili side (your account is treated as a report; I verified what this repo can prove):

    Thread status, so you know what you'd be testing against:

    • The contract decisions (D1 shape / D2 fail-closed semantics / D3 relation to GH [PR] Restore dsh native auto-compaction when bili-native is disabled or absent (#5248) #5994) await the maintainer's call; my posted recommendation is approve-as-proposed plus small doc cross-references. Your offer to test "whatever config contract gets settled" stands — once settled, your kiro acceptance run becomes the live validation data point for the unroutable-lane use case.
    • Heads-up: the PR head moved since my initial line-by-line review — b85e797 (reviewed) → cd7b492, which is exactly the sha you built. The delta is 2 commits: fix(dsh): meter summary tool schemas in host tokens (replaces UTF-8-byte subtraction in the native-summary input budget with host-tokenizer metering; fixes the negative-budget case on a 327-tool catalog) and test(dsh): cover partial summary failure and concurrent ownership (+100 test lines). I am reviewing that delta now and will report before the review conclusion is considered final.
  7. ranxianglei commented on Nov 11, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    (ai: opencode/qwen3.8-27b)

    Classifier round 1 — plan & category scheme (300 open issues)

    Approach:

    1. Pull every issue's title + body through the platform mirror API (batched), plus comments where the thread carries triage context.
    2. Classify by root cause (component + mechanism), not surface symptom. Categories below are dynamic — new ones get minted as needed and stay stable across later rounds.
    3. Where a body points at an upstream PR, record its number; CI/merge state comes from the cached upstream-checks feed (no gh polling).
    4. Source inspection in this workdir is used for the ambiguous/high-value items; clear-cut items are classified from title+body+known architecture (AGENTS.md module map, KDD ledger) and marked medium/high confidence accordingly. Deep-dive items beyond this pass's timebox ship as confidence: low rather than being dropped.

    Seed category scheme (component/mechanism families):

    Progress cadence: [progress] N/300 every 25 issues, each carrying the cumulative classifier-json block. Final reply carries the complete 300-entry array. Analysis-only lane: no PRs/issues/labels/edits from this worker.

  8. ranxianglei commented on Nov 11, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    (ai: opencode/qwen3.8-27b)

    [progress] 25/300

    [{"number":5763,"title":"notice-store pre-hydration receipt blocks legacy->kv migration","category":"storage","root_cause":"src/notice-store.ts hydrateNoticeReceipts() gates on cache!==null as its hydrated latch, but addDismissedKeyInMemory lazily creates that same cache on first dismiss: a pre-hydration receipt makes hydration bail for the whole process, permanently skipping the one-time legacy JSON->kv migration (file re-imports every boot) and hiding sibling lanes kv rows until restart","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fixed by upstream PR 5768 (matched via twin [PR] row 5770)"},{"number":5764,"title":"UNRECOGNIZED PATHS mislabels non-POST probes; counter never expires","category":"panel-ui","root_cause":"protocol derivation only runs for POST-with-body requests, so HEAD/GET probes to known endpoints (/responses, /) resolve protocol=null and trip the UNRECOGNIZED PATHS counter, which has no TTL/windowing: a one-off probe stayed resident on /__bili/stats + acp_status panel for 6 days","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},{"number":5765,"title":"acp_status full report inlined into assistant-visible marker each call","category":"panel-ui","root_cause":"buildVisibilityMarker() special-cases acp_status by inlining the entire ~40-line report into the assistant message marker (all other tools get one line), so every acp_status call leaves a permanent ~40-line panel block in the client session history (uplink already stripped; cost entirely on the visible surface)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},{"number":5766,"title":"six inline base64 images defeat tier-2 span selection -> 502 fail-fast","category":"preflight-tier","root_cause":"any candidate two-block span containing an oversized inline image exceeds TIER_DISTILL_SOURCE_CHAR_BUDGET=60_000 chars while images are current-turn protected (unfoldable), so tier-2 recovery finds 0 legal spans and 0 summary calls; the structurally-exhausted verdict was installed in src/preflight.ts without the deterministic flag its sibling futile paths carry, so the server maps it to 502 preflight_compress_failed and the host retries 5x in 8s","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},{"number":5768,"title":"feat: compression block retire + external summary rewrite","category":"kernel-compression","root_cause":"feature request: no surface exists to retire a folded block (remove from resident context while keeping decompress access) or to replace a block summary externally (acp_retire/acp_restore/acp_rewrite + export enrichment); long single sessions accumulate unrelated blocks because every block is injected every turn (113 blocks, ~320K-token view)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"feature-request, not a defect"},{"number":5769,"title":"compression conflict banner: unannounced-rewrite x2 (claude)","category":"conflict-watch","root_cause":"auto-filed diagnostic dump: claude session ledger shows 2 unannounced-rewrite events (9/19 and 5/24 incoming messages carrying pre-turn refs of known ids) with all classification buckets (sibling/suspected/displayOnly/activeConfirmed) at 0 - banner raised, rewrite source unclassified","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"low","notes":"needs session-level deep dive beyond this pass"},{"number":5770,"title":"[PR] notice-store pre-hydration receipt no longer blocks migration","category":"storage","root_cause":"fix: independent hydrated latch replaces cache!==null gate; hydrate UNIONs into existing lazy cache instead of bailing; live-cache snapshot moved immediately before publish (zero awaits between) so mid-hydration dismissals survive; failure keeps cache=null so file fallback persists and retry is possible","ci_state":"green","merge_state":"clean","pr_number":5768,"confidence":"high"},{"number":5771,"title":"[PR] mark structurally-exhausted tier-recovery deterministic","category":"preflight-tier","root_cause":"fix: when the tier-distillation candidate picker proves no legal two-block span exists (oversized inline media push every span past the 60k char budget - structural impossibility, identical payload can never recover), install the exhausted verdict with the deterministic flag so the server stops mapping it to 502 and reports the true status","ci_state":"green","merge_state":"clean","pr_number":5769,"confidence":"high"},{"number":5773,"title":"feat(design): preserve in-progress implementation state across folds","category":"kernel-compression","root_cause":"design request: folds discard just-written implementation details (incident: 101,418-token fold at step 43 -> model lost in-progress code, re-read a dozen source files, self-contradicted thresholds/protocol checks, added 10+ re-read/probe/modify loops); compress-only usage (decompress/acp_status never called) makes the loss unrecoverable in-session","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"feature-request; source: analysis of #917 user comment"},{"number":5774,"title":"[PR] gate UNRECOGNIZED PATHS counter to model-calls + 24h window","category":"panel-ui","root_cause":"fix: counter fired for any request whose derived protocol was null even though protocol is only derivable for POST-with-body, so non-POST probes/scans were mislabeled not-a-known-protocol despite hitting endpoints that ARE in the table; gate counting to model-calls and window the counter to 24h","ci_state":"green","merge_state":"clean","pr_number":5772,"confidence":"high"},{"number":5775,"title":"[PR] fastpath-v6: tunnel-origin requests without identity channel are cohort exclusions, not faults","category":"watermark-fastpath","root_cause":"fix: v6 identity hard gate (src/server/fastpath-v6.ts) is fail-loud on the premise that lane traffic always comes from a native shim stamping pf-* headers in-process, which holds only OFF-tunnel; URL-prefix/attach-mode dsh sessions (tunnel origin) lack x-bili-pf-tail-ids, so with watermarkV6MsgId rollout forced on EVERY model request was refused 400 pf_v6_identity_fault - reclassify as cohort exclusion, not fault","ci_state":"green","merge_state":"clean","pr_number":5773,"confidence":"high"},{"number":5777,"title":"[PR] claude /remote-control + compression coexist via cert-MITM lane (corrects #4693)","category":"docs","root_cause":"#4693 was closed on the premise that Claude Code undici ignores HTTPS_PROXY so /remote-control cannot coexist with compression; premise is stale (CC 2.1.301 honours HTTPS_PROXY per #3446; on 2.1.318 the cert-MITM lane carries the whole session incl. /remote-control + folds + mcp__bili__* tools) - corrects the documented limitation","ci_state":"red","merge_state":"clean","pr_number":5777,"confidence":"medium","notes":"CI red: lock-sync check failing"},{"number":5778,"title":"T2 nudge repeatedly recommends block merges rejected by turn-integrity gate","category":"kernel-compression","root_cause":"pendingByTier in kernel/src/compress.ts treats every active tier-1/2 block as a nudge target without screening against the apply-side turn-integrity/coverage rules, so T2 distillation nudges advertise ranges compress deterministically rejects (incident: 27/28 failed merges, e.g. b102..b165 could not preserve coverage+turn integrity) - repeated futile compress attempts","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},{"number":5779,"title":"[PR] screen tier merge nudges against compression eligibility","category":"kernel-compression","root_cause":"kernel-change fix: screen tier 2/3 nudge targets on the actual apply view using the same projection compression executes on; whole-range failure falls back to up to eight individual blocks, otherwise normal tier arbitration; regression tests cover rejected and valid ranges","ci_state":"green","merge_state":"clean","pr_number":5779,"confidence":"high"},{"number":5780,"title":"sessions.gc does not clean up dsh-deleted sessions","category":"persist-session","root_cause":"sessions.gc only reaps sessions it observes via incoming traffic; when DSH deletes a conversation the proxy never sees it again, so the plugin copy (~/.local/share/billion-context/openai/<hash>.json + DB rows) becomes an orphan forever (measured: 3 deleted sessions left 5.3M of copies + 2023 DB rows, plus 18M image copies)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},{"number":5782,"title":"UNCOVERED HISTORY advertises protected-boundary turn fragments compress always rejects","category":"kernel-compression","root_cause":"uncoveredHistoryRuns (src/uncovered.ts, the #4331 audit surface) never runs the shared projectUnit apply-side foldability judgment that #3236 wired into Compressible-ranges + acp_status, so the audit advertises structurally-unfoldable runs (e.g. m02080-m02088 rejected: would split 1 turn at protected-zone boundary) and the atomic batch rollback drags adjacent legal ranges down with it","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},{"number":5786,"title":"kernel authorization request: export projectUnit screening entry","category":"kernel-compression","root_cause":"decision request (owner-gated kernel boundary): export projectUnit/buildProjectionShared/collectCoverage/computeProtectedRefs + types from kernel/src/index.ts as pure re-exports (zero behavior change) so the host can tag UNCOVERED HISTORY runs with verdict-backed foldability instead of string-parsing rendered reports; host side cannot solve because buildStatusReport returns rendered strings only","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"awaiting owner A/B decision"},{"number":5787,"title":"[PR] screen UNCOVERED HISTORY runs through apply-side foldability judgment","category":"kernel-compression","root_cause":"fix: run the #4331 UNCOVERED HISTORY audit through the shared apply-side projectUnit foldability judgment (#3236 doctrine: advertise and execute must share one projection); coordinates with PR #5311 honest compose surfaces ([not foldable] tags)","ci_state":"green","merge_state":"behind","pr_number":5785,"confidence":"high"},{"number":5788,"title":"compression conflict banner: recovered id-churn-adoption lights alarm (dsh)","category":"conflict-watch","root_cause":"#4131 introduced the id-churn-adoption ConflictKind (fold-reconcile re-mints and fully re-anchors ids across builds, 0 unmatched, marks NO compaction boundary) but only wired the recording side; the two display classifiers were never updated, so a ledger whose ONLY evidence is a recovered id-churn-adoption event falls into the generic unexplained-signal bucket and lights the web alarm banner framed as outside rewrite/phantom compressor","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},{"number":5789,"title":"[PR] stop recovered id-churn-adoption from lighting the conflict banner","category":"conflict-watch","root_cause":"fix: update the display classifiers to recognize recovered id-churn-adoption ledgers (benign cross-build id re-mint, fully re-anchored, 0 unmatched) as non-alarm so they stop lighting the compression-conflict banner and phantom-second-compressor framing","ci_state":"green","merge_state":"clean","pr_number":5789,"confidence":"high"},{"number":5790,"title":"[storage] persist a schema reader floor for rolling upgrades/rollbacks","category":"storage","root_cause":"feature request: whether an OLDER build may open a store a NEWER build migrated is known only from prose in src/storage/schema.ts; multi-process shared index.db (#5082) means rolling upgrades run old+new code on one store while the version gate never refuses a newer store - propose exporting a reader floor next to STORAGE_SCHEMA_VERSION written in the same transaction as the bump","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"feature-request"},{"number":5792,"title":"[PR] Pi: opt-in structured acp_delegate notification envelope","category":"pi-lane","root_cause":"pi acp_delegate completion/failure notices reach the host as plain steering user messages whose wording third-party consumers can only identify by regex-parsing prose (fragile for pi-tidy turn-folding and UI integration); adds opt-in typed envelope (customType billion-context:delegate-notification) with versioned details carrying stable run identity + launch provenance while content stays byte-identical","ci_state":"green","merge_state":"clean","pr_number":5792,"confidence":"high"},{"number":5796,"title":"[feat][dsh] long sessions have no convergence path (compaction double-gated + incremental full-resend)","category":"dsh-lane","root_cause":"dsh compaction is gated off on BOTH sides (profile compaction/compaction-basic/command-compact/tool-result-pruner disabled:true AND bili dsh-compaction-gate rewrites those lines to disabled whenever bili is active) while bili folds only outbound (29.6k msgs -> 63K tokens upstream but host surface stays 11.2M messageTokens); incremental lane additionally falls back to full-resend on OpenAI wire shape (tool_calls/role:tool guard) - every step rebuilds+uploads the full 51.2MiB payload (host 27-41s), O(N) with no exit","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"feature/design request spanning dsh gate + wire-shape fallback"},{"number":5797,"title":"title-generation side call pollutes fold ref state -> permanent fold-reconcile drift","category":"session-resolve","root_cause":"ZCode desktop title-generation side call (fresh session start) injects refs into the fold ref state; first fold succeeds but every subsequent turn logs fold-reconcile resent-history-drifted, forcing the legacy-outbound-kept fallback - compression permanently disabled for the rest of the session (outbound~=inbound every request, e.g. 421.0KiB in / 422.8KiB out)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},{"number":5798,"title":"[perf][dsh] 25.8-32.7s of each step elapses before bili sees the body","category":"performance-memory","root_cause":"TTFT attribution on a 29,619-msg/51.2MiB dsh session: aligning DSH step/start events with bili request: logs, the observed per-step host-side gap is 25.8-32.7s (459 pairs: p50 27.0s, p90 44.5s) and ALL of it occurs before bili receives the first byte; bili local= is only 5.5-7.2s - perceived slowness lives in the host process before it sends; asks bili to expose the host<->proxy boundary before deciding optimization ownership","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"}]
    
  9. ranxianglei commented on Nov 11, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    (ai: opencode/qwen3.8-27b)

    [correction] progress 25/300 — 7 rows re-issued verbatim from the on-disk ledger

    My previous progress reply hand-transcribed the PR/CI/merge fields for the [PR] rows and drifted on 7 of them (the classification ledger itself was correct; only the posted copy was wrong). Replace those 7 objects in the prior array with these, taken directly from the ledger:

    [{"number":5770,"title":"[PR] notice-store pre-hydration receipt no longer blocks migration","category":"storage","root_cause":"fix: independent hydrated latch replaces cache!==null gate; hydrate UNIONs into existing lazy cache instead of bailing; live-cache snapshot moved immediately before publish (zero awaits between) so mid-hydration dismissals survive; failure keeps cache=null so file fallback persists and retry is possible","ci_state":"green","merge_state":"behind","pr_number":5768,"confidence":"high"},{"number":5771,"title":"[PR] mark structurally-exhausted tier-recovery deterministic","category":"preflight-tier","root_cause":"fix: when the tier-distillation candidate picker proves no legal two-block span exists (oversized inline media push every span past the 60k char budget - structural impossibility, identical payload can never recover), install the exhausted verdict with the deterministic flag so the server stops mapping it to 502 and reports the true status","ci_state":"green","merge_state":"behind","pr_number":5769,"confidence":"high"},{"number":5774,"title":"[PR] gate UNRECOGNIZED PATHS counter to model-calls + 24h window","category":"panel-ui","root_cause":"fix: counter fired for any request whose derived protocol was null even though protocol is only derivable for POST-with-body, so non-POST probes/scans were mislabeled not-a-known-protocol despite hitting endpoints that ARE in the table; gate counting to model-calls and window the counter to 24h","ci_state":"green","merge_state":"behind","pr_number":5770,"confidence":"high"},{"number":5777,"title":"[claude] /remote-control and compression can coexist via the cert-MITM lane (corrects #4693)","category":"docs","root_cause":"#4693 was closed on the premise that Claude Code undici ignores HTTPS_PROXY so /remote-control cannot coexist with compression; premise is stale (CC 2.1.301 honours HTTPS_PROXY per #3446; on 2.1.318 the cert-MITM lane carries the whole session incl. /remote-control + folds + mcp__bili__* tools) - corrects the documented limitation","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"medium"},{"number":5779,"title":"[PR] screen tier merge nudges against compression eligibility","category":"kernel-compression","root_cause":"kernel-change fix: screen tier 2/3 nudge targets on the actual apply view using the same projection compression executes on; whole-range failure falls back to up to eight individual blocks, otherwise normal tier arbitration; regression tests cover rejected and valid ranges","ci_state":"red","merge_state":"dirty","pr_number":5777,"confidence":"high","notes":"CI red: lock-sync failing; branch has merge conflicts (dirty)"},{"number":5789,"title":"[PR] stop recovered id-churn-adoption from lighting the conflict banner","category":"conflict-watch","root_cause":"fix: update the display classifiers to recognize recovered id-churn-adoption ledgers (benign cross-build id re-mint, fully re-anchored, 0 unmatched) as non-alarm so they stop lighting the compression-conflict banner and phantom-second-compressor framing","ci_state":"green","merge_state":"behind","pr_number":5787,"confidence":"high"},{"number":5792,"title":"[PR] Pi: opt-in structured acp_delegate notification envelope","category":"pi-lane","root_cause":"pi acp_delegate completion/failure notices reach the host as plain steering user messages whose wording third-party consumers can only identify by regex-parsing prose (fragile for pi-tidy turn-folding and UI integration); adds opt-in typed envelope (customType billion-context:delegate-notification) with versioned details carrying stable run identity + launch provenance while content stays byte-identical","ci_state":"green","merge_state":"behind","pr_number":5790,"confidence":"high"}]
    

    Errors fixed: 5770/5771 merge_state clean→behind; 5774 pr_number 5772→5770 (+merge behind); 5777 was NOT a [PR] row — pr/ci/merge reset to null/none/n/a (the red lock-sync CI belongs to an unrelated upstream PR); 5779 pr_number 5779→5777 with ci red/dirty (lock-sync fail + conflicts); 5789 pr_number 5789→5787 (+merge behind); 5792 pr_number 5792→5790 (+merge behind). All other rows in the 25/300 array stand as posted. From here on each progress post is emitted verbatim from the ledger file, not retyped.

  10. ranxianglei commented on Nov 11, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [bot] 🏷 (ai: opencode/qwen3.8-27b)

    Status — classification in progress, 25/300 done.

    Done so far:

    1. Plan + category scheme posted (seed categories for root-cause grouping).
    2. Batch 1 (mirror 5763–5798) classified: full per-issue rows (category, one-line root cause, CI state, merge state, upstream PR number where matched, confidence) posted as [progress] 25/300.
    3. A correction followed: my first progress post had hand-transcription drift on 7 PR/CI/merge fields; those 7 rows were re-issued verbatim from the on-disk ledger. All other rows stand.

    Method: mirror issue bodies fetched from the local mirror API; upstream PR mapping from the open-PR list (exact/contains/token match); CI state from the local upstream-checks cache (~60 s refresh); merge state from GitHub GraphQL. Every progress post will be emitted verbatim from the ledger file to prevent further transcription drift.

    Remaining: 275 issues (batches of 25), then a final reply carrying the complete 300-row array. Continuing now — next progress at 50/300.

  11. 65 remaining items

  12. ranxianglei commented on Nov 19, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [bot] 🏷 [progress] 25/73 — incremental root-cause classification round

    Category scheme: reusing the round's dynamic strings (kernel-compression, kernel-identity, fold-reconcile, throttle-governor, codex-lane, cache-invalidation, egress-fuse-pricing, wire-fidelity, persist-session, storage, panel-ui, preflight-tier, performance-memory, config-knobs, compress-gate, compress-parser, install-lane, external-provider, reasoning-drop, window-budget, dsh-lane, identity-channel, test-quality, dsh-model-routing, fork-adoption, identity-msgid, kernel-compress, window-fit, plugin-mode, session-resolve) + the reserved tinyfix (none this round). The 14 [Root Cause] rcagg boards use their board string verbatim.

    Field mapping: mergeable_state blocked → reported as dirty (schema has no blocked); unknown → clean (no known conflict, unverified). ci_state: failed>0 → red, total==0 → none, else green (with a CI incomplete (N pending) note where pending>0). Issue rows (no PR) → ci_state: none, merge_state: n/a.

    Renumbered PRs found this round (mirror → GitHub): 5683→5681, 5315→5311, 5779→5777, 5255→5250, 5336→5334, 5693→5691 (plus carried 5235→5233, 5357→5355, 5384→5382).

    Rows 1–25 (of 73, batch order):

    [
    {"number":7075,"title":"dsh ref unknown after upgrade","category":"dsh-lane","root_cause":"dsh lane: after upgrade the dsh session ref is unresolved ('dsh ref is unknown'); 15 blocks collapsed to 7 — ref resolution/inheritance broken on the dsh lane post-upgrade","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"medium"},
    {"number":7300,"title":"compress no-op loop (289 blind retries)","category":"compress-gate","root_cause":"compress gate: compressFailStreak resets on any success so the breaker never trips, and the 'too small' error text teaches the model to retry → 289 blind compress retries (kernel/src/compress.ts:625)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":7301,"title":"429 hold not auto-cleared after quota exhausted","category":"throttle-governor","root_cause":"throttle governor: bili_throttle_hold set on 429 is not auto-cleared after the upstream quota is exhausted, so the credential stays held","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":5683,"title":"print ascending ref pairs after generation reset","category":"kernel-identity","root_cause":"kernel identity: after a generation reset ref numbers are non-monotonic; compressible ranges endpoint-labeled positionally print startRef>endRef verbatim on 3 surfaces — model reads corruption; fix = render-site canonicalization","ci_state":"green","merge_state":"dirty","pr_number":5681,"confidence":"high"},
    {"number":5400,"title":"retire persistently-unmatched fold coverage","category":"fold-reconcile","root_cause":"fold-reconcile: a covered id that stays missing with no anchor match every pass lived in block coverage forever, freezing ledger accrual; fix retires after ABSENT_RETIRE_PASSES=3 evidence-less passes","ci_state":"green","merge_state":"dirty","pr_number":5400,"confidence":"high"},
    {"number":5803,"title":"kernel: relocate tool-delta systems not demote","category":"kernel-compression","root_cause":"kernel compression (outbound render): healIllegalSystemPositions demote-strips tool-delta system chains, orphaning paired tool_removal blocks → 400 tool_reference_unresolved; must relocate not demote","ci_state":"green","merge_state":"clean","pr_number":5803,"confidence":"high"},
    {"number":5652,"title":"unpriced-token ledger + health installPath","category":"panel-ui","root_cause":"panel-ui/billing: unpriced models count as free (no unpricedTokens ledger); health endpoint missing installPath; lineage column undocumented","ci_state":"green","merge_state":"behind","pr_number":5652,"confidence":"high","notes":"CI incomplete (2 pending)"},
    {"number":7295,"title":"429 Retry-After 600s freezes credential → 424","category":"throttle-governor","root_cause":"throttle governor: 429 hold takes max(jittered backoff, Retry-After) so a ~600s Retry-After dominates and freezes the whole credential across sessions, escalating to terminal 424","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":7296,"title":"Codex ordering: user separator before web_search_call","category":"codex-lane","root_cause":"codex lane: repairResponsesAssistantOrdering inserts a synthetic user separator between reasoning and web_search_call → upstream 400 'web_search_call provided without its required reasoning item'","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":7299,"title":"compress folded the user message being answered","category":"kernel-compression","root_cause":"kernel compression: compress accepted a range whose endId was past the current user message, folding it → the model replied to a fabricated user message (range boundary not guarded against the active turn)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":5478,"title":"cache-seam: attribute tools-array prefix breaks","category":"cache-invalidation","root_cause":"cache-invalidation: cache-seam monitor fingerprinted only the message list, not the tools array, so a tools-array change was misattributed as a message prefix break; fix adds toolsArrayFps","ci_state":"green","merge_state":"clean","pr_number":5478,"confidence":"high"},
    {"number":4928,"title":"preflight 502→413: doom gate prices raw bytes","category":"egress-fuse-pricing","root_cause":"egress-fuse-pricing: the 502 path's estimate writeback arms the evidence-probe regime where the doomed-probe gate prices raw inbound bytes vs max(2MB, window×7.5) → terminal 413 (dsh provider-switch death spiral)","ci_state":"green","merge_state":"dirty","pr_number":4928,"confidence":"high"},
    {"number":4594,"title":"Codex WS prewarm runs as ACP turn, folds wiped","category":"codex-lane","root_cause":"codex lane: src/responses-ws.ts only treats generate:false with empty input as prewarm; a codex context-only prewarm falls into the ACP pipeline as a zero-message turn → folds wiped + forged summaries re-injected","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":5315,"title":"honest compose surfaces (not-foldable tags)","category":"kernel-compression","root_cause":"kernel compression (compose surface): UNCOVERED HISTORY rows tagged [not foldable] + compress success receipts teach a source-of-truth split, so the model copies non-foldable spans into compress batches","ci_state":"green","merge_state":"dirty","pr_number":5311,"confidence":"medium"},
    {"number":4790,"title":"outbound tool-set flip mid-session invalidates cache","category":"cache-invalidation","root_cause":"cache-invalidation: acp_read_original_image/acp_retrieve presence rides mutable per-request session state → the outbound tool array flips mid-session → prefix cache invalidated; fix adds a noteBiliToolSurface tripwire","ci_state":"green","merge_state":"dirty","pr_number":4790,"confidence":"high"},
    {"number":7298,"title":"session GC: retain when companion stat unreadable","category":"persist-session","root_cause":"persist-session: session GC treated a failed stat(companion) as absence and deleted the session + orphaned originals; fix permits the no-companion path only on ENOENT","ci_state":"green","merge_state":"dirty","pr_number":7298,"confidence":"high"},
    {"number":7297,"title":"session GC deletes originals on companion stat fail","category":"persist-session","root_cause":"persist-session: same as #7298 — session GC deletes originals when companion stat fails (not just ENOENT); fix PR #7298 in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7298 in flight"},
    {"number":5591,"title":"snapshot-shadow lost-tail + messageRefs DIGEST DRIFT","category":"storage","root_cause":"storage: snapshot-shadow mechanical lost-tail mismatch + unified-storage shadow state.messageRefs DIGEST DRIFT on 0.1.212 (shadow write vs table digest divergence)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"medium"},
    {"number":5672,"title":"tool_search_call/output missing from tool-pair-order","category":"wire-fidelity","root_cause":"wire-fidelity: src/tool-pair-order.ts doesn't know tool_search_call/tool_search_output, so the #766 adjacency guarantee doesn't apply → tool_search_output with a missing call → upstream 400 'No tool call found for tool output'","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6215,"title":"client/cloud compaction leaves folded history unreachable","category":"persist-session","root_cause":"persist-session: client/cloud compaction (Codex cloud, DSH) closes ACP state and leaves folded blocks only as a preCompactionArchive marker → decompress refuses, search_context can't see, retrieve can't resolve","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"medium"},
    {"number":7294,"title":"diagnostics: avoid reading unused log generations","category":"performance-memory","root_cause":"performance-memory: buildDiagnostics read every retained log generation for every topic even when unused; fix selects log evidence only for the consuming topic","ci_state":"green","merge_state":"dirty","pr_number":7294,"confidence":"high"},
    {"number":6467,"title":"preflight: scope baseline + k-hat to the model","category":"preflight-tier","root_cause":"preflight-tier: the preflight size model used a cross-model usage baseline + a stale k-hat that over-corrects → spurious folds; fix scopes baseline demotion + a fresh-measurement cap per model","ci_state":"green","merge_state":"dirty","pr_number":6467,"confidence":"high"},
    {"number":6546,"title":"image usage contaminates text calibration multiplier","category":"preflight-tier","root_cause":"preflight-tier: the relay settle read a shared prepare slot that recorded text+image under the TextEstimate name, so image usage trained the text calibration multiplier; fix uses a per-request wire text estimate + image reserve and skips image-bearing samples","ci_state":"green","merge_state":"dirty","pr_number":6546,"confidence":"high"},
    {"number":7293,"title":"diagnostics reads unused log generations","category":"performance-memory","root_cause":"performance-memory: same as #7294 — diagnostics read unused log generations; fix PR #7294 in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7294 in flight"},
    {"number":5779,"title":"screen tier merge nudges vs compression eligibility","category":"kernel-compression","root_cause":"kernel compression: tier distillation nudges advertised block ranges rejected by the kernel's turn-integrity/coverage rules; fix screens tier 2/3 targets on the actual apply view","ci_state":"red","merge_state":"dirty","pr_number":5777,"confidence":"high"}
    ]
    
  13. ranxianglei commented on Nov 19, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [bot] 🏷 [progress] 50/73 — rows 26–50 (batch order)

    [
    {"number":7277,"title":"3rd large request 503, limits not configurable","category":"config-knobs","root_cause":"config-knobs: MAX_LARGE_DECODED_REQUESTS=2 and LARGE_DECODED_ADMISSION_BYTES=32MiB are hardcoded in src/request-body-budget.ts → 503 request_body_busy, not configurable; fix PR #7283 in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7283 in flight"},
    {"number":7291,"title":"web sessions keep stale dollar estimates after price change","category":"panel-ui","root_cause":"panel-ui: the web dashboard's disk session-summary cache is invalidated only by mtime+size; a price edit doesn't invalidate it → stale dollar estimates; fix PR #7292 in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7292 in flight"},
    {"number":7065,"title":"msgid storage slimming proposal","category":"identity-channel","root_cause":"identity-channel: msgid storage keeps three full id projections per session; proposal to slim to a range reference + anti-forgery fingerprint (design proposal, no fix yet)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"medium"},
    {"number":5235,"title":"kernel: preserve issued refs, quarantine collisions","category":"kernel-identity","root_cause":"kernel identity: a restart could restamp previously issued refs (allocator cursor reuse of historical numbers); fix preserves issued refs and quarantines known collisions","ci_state":"red","merge_state":"dirty","pr_number":5233,"confidence":"high"},
    {"number":5255,"title":"degrade 2nd long-lived index.db writer to legacy-only","category":"storage","root_cause":"storage: two bili installs sharing one sessions dir both RW-open index.db → corruption; fix degrades the second long-lived writer to legacy-only instead of co-residing","ci_state":"none","merge_state":"dirty","pr_number":5250,"confidence":"high"},
    {"number":7292,"title":"web: refresh cached disk summaries on price change","category":"panel-ui","root_cause":"panel-ui: fix for #7291 — refresh the cached disk session summaries when billing prices change","ci_state":"green","merge_state":"dirty","pr_number":7292,"confidence":"high","notes":"CI incomplete (6 pending)"},
    {"number":7288,"title":"carry no-op receipt regression tests from #5564","category":"test-quality","root_cause":"test-quality: the no-op compress receipt regression tests were stranded in a superseded PR (#5564); this carries them into the main branch","ci_state":"green","merge_state":"behind","pr_number":7288,"confidence":"high"},
    {"number":6785,"title":"msgid delivery × snapshot plane unified merge","category":"identity-channel","root_cause":"identity-channel: unified merge of msgid delivery (pin) + the snapshot plane (4 tables, shadow write, egress); absorbs #6564 + #5901 + #6680","ci_state":"red","merge_state":"behind","pr_number":6785,"confidence":"high"},
    {"number":7287,"title":"v1internal pre-compress envelope wrap/unwrap","category":"external-provider","root_cause":"external-provider (v1internal/Antigravity/CloudCode lane): the preflight pre-compress summary sub-request was not wrapped in the v1internal envelope and its response not unwrapped → 400/502; fix PR for #7286","ci_state":"green","merge_state":"behind","pr_number":7287,"confidence":"high"},
    {"number":5336,"title":"orphan-sweep provably-foreign collapsed blocks","category":"persist-session","root_cause":"persist-session: a collapsed block under a shared conversation id (≤v0.1.208 era) reappears forever in the parent session after upgrade; the zombie-reap exemption shelters it (same signature as healthy proxy-mode folds); fix = orphan-sweep","ci_state":"none","merge_state":"dirty","pr_number":5334,"confidence":"high"},
    {"number":7279,"title":"reuse normalized billing override rows","category":"performance-memory","root_cause":"performance-memory: resolveBillingPrice converted every override into a new registry cost row on every lookup; fix builds the normalized rows once in setBillingOverrides (fix for #7276)","ci_state":"green","merge_state":"behind","pr_number":7279,"confidence":"high"},
    {"number":5357,"title":"re-hash entry script on file change","category":"install-lane","root_cause":"install-lane: entryScriptFingerprint cached the sha256 by path for the process lifetime; after an in-place npm i -g it kept the old digest → code-fingerprint-mismatch + lane port laddering; fix re-hashes on file identity change","ci_state":"green","merge_state":"behind","pr_number":5355,"confidence":"high"},
    {"number":7290,"title":"web: select freshest disk session summaries","category":"panel-ui","root_cause":"panel-ui: the web index selected the disk summary winner by savedAt but the summary merge ignored it → a stale twin overwrote the fresh one; fix reads through the byId index (fix for #7289)","ci_state":"green","merge_state":"dirty","pr_number":7290,"confidence":"high","notes":"CI incomplete (13 pending)"},
    {"number":5384,"title":"overflow-stated counts steer local fold targets","category":"window-budget","root_cause":"window-budget: on a fresh overflow-only session the ring can't publish, so local fold judgments ran at raw-estimate calibre → re-400 loop; fix retains the upstream-stated count as a provisional estimator scale (bias stats.overflowBias)","ci_state":"red","merge_state":"dirty","pr_number":5382,"confidence":"high"},
    {"number":7286,"title":"v1internal pre-compress request not wrapped","category":"external-provider","root_cause":"external-provider (v1internal lane): same as #7287 — preflight pre-compress summary sub-request not wrapped in the v1internal envelope → 400/502; fix PR #7287 in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7287 in flight"},
    {"number":7289,"title":"dashboard summaries prefer stale copies","category":"panel-ui","root_cause":"panel-ui: same as #7290 — the file-backed dashboard summary merge can prefer a stale twin over the fresh copy; fix PR #7290 in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7290 in flight"},
    {"number":7285,"title":"registry: index relay model suffix candidates","category":"performance-memory","root_cause":"performance-memory: the relay lookup scanned all lowercase-index entries for each model suffix; fix builds the suffix buckets once (fix for #7284)","ci_state":"green","merge_state":"dirty","pr_number":7285,"confidence":"high"},
    {"number":7283,"title":"make large-decode admission limits configurable","category":"config-knobs","root_cause":"config-knobs: fix for #7277 — makes MAX_LARGE_DECODED_REQUESTS + LARGE_DECODED_ADMISSION_BYTES configurable (network.maxLargeDecodedRequests / largeDecodedAdmissionBytes) and clarifies the 503","ci_state":"green","merge_state":"behind","pr_number":7283,"confidence":"high"},
    {"number":7284,"title":"relay registry scans every row per suffix","category":"performance-memory","root_cause":"performance-memory: same as #7285 — relay registry lookups scan every lowercase-index row per model suffix; fix PR #7285 in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7285 in flight"},
    {"number":7282,"title":"registry: cap cost-row index cache at 2 tables","category":"performance-memory","root_cause":"performance-memory: the costLowerIndex single-slot cache thrashed (overrides-miss) so the registry index was rebuilt per lookup; fix caps it at an LRU Map of 2 (fix for #7278)","ci_state":"green","merge_state":"dirty","pr_number":7282,"confidence":"high","notes":"CI incomplete (10 pending)"},
    {"number":5443,"title":"kernel auth: heal FOLLOW side of system position rule","category":"kernel-compression","root_cause":"kernel compression (authorization request): healIllegalSystemPositions only checks the follower side of Anthropic's mid-conversation system position rule, not the predecessor → 400; needs kernel approval","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"kernel approval pending (Kernel Boundary)"},
    {"number":7143,"title":"GPT-6.2 prefix edit drops reasoning","category":"reasoning-drop","root_cause":"reasoning-drop: GPT-6.2+ Responses API drops reasoning items after any prefix edit (fold/reasoning-drop); only the official compaction path preserves them","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":5737,"title":"folded summary role asymmetric across wires","category":"wire-fidelity","root_cause":"wire-fidelity: the folded summary role is asymmetric across wires — OpenAI wire → user (systemToUser re-voicing), Anthropic wire → system; compressed history carries live-message authority","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":5693,"title":"dsh long-session drift comprehensive remediation","category":"dsh-lane","root_cause":"dsh-lane: four-part dsh long-session drift remediation — refuse decisive native compaction on dead-proxy direct-send paths, etc.","ci_state":"green","merge_state":"dirty","pr_number":5691,"confidence":"high"},
    {"number":3978,"title":"preflight compression too slow on ultra-large context","category":"preflight-tier","root_cause":"preflight-tier: preflight compression on a ~1.2M-token context takes >30s; the client disconnects mid-compression → abort → retry loop","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"medium"}
    ]
    
  14. ranxianglei commented on Nov 19, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [bot] 🏷 [final] 73/73 — rows 51–73 (batch order) + round summary

    [
    {"number":7278,"title":"registry cost-row index rebuilt per lookup","category":"performance-memory","root_cause":"performance-memory: same as #7282 — the registry cost-row index is rebuilt per lookup on overrides-miss (single-slot cache thrash); fix PR #7282 in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7282 in flight"},
    {"number":7280,"title":"billing: build cost table once per config load","category":"performance-memory","root_cause":"performance-memory: fix for #7276 — build the normalized billing cost table once per config load instead of per lookup","ci_state":"green","merge_state":"dirty","pr_number":7280,"confidence":"high"},
    {"number":6156,"title":"credential fingerprint: all auth headers, order-independent","category":"throttle-governor","root_cause":"throttle-governor: credentialFingerprint broke on the first recognized auth header in insertion order (order-dependent + incomplete) → 429 governor holds missed; fix collects all, as a sorted set","ci_state":"green","merge_state":"behind","pr_number":6156,"confidence":"high"},
    {"number":7276,"title":"billing lookups rebuild every override row","category":"performance-memory","root_cause":"performance-memory: same as #7280 — billing price lookups rebuilt every override row + lowercase index per lookup; fix PR #7280 in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7280 in flight"},
    {"number":7275,"title":"name top-level keys when egress fuse fires","category":"egress-fuse-pricing","root_cause":"egress-fuse-pricing: the egress window fuse 413 lumped non-messages/tools/system into an opaque 'other' bucket; fix adds estimateRawBodyTopKeys to name the top-level keys (fix for #7273)","ci_state":"green","merge_state":"dirty","pr_number":7275,"confidence":"high"},
    {"number":7092,"title":"egress fuse + doom gates price prompt mass not body","category":"egress-fuse-pricing","root_cause":"egress-fuse-pricing: the egress fuse + doom gates priced the whole outbound body vs the prompt mass; dsh_session_log landed in 'other' → spurious 413; fix adds PROMPT_BODY_KEYS to price prompt mass (fixes #6844 #7013 #7088)","ci_state":"green","merge_state":"behind","pr_number":7092,"confidence":"high"},
    {"number":6451,"title":"[Root Cause] preflight-tier board","category":"preflight-tier","root_cause":"rcagg board: aggregation thread for the preflight-tier root cause (6 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6363,"title":"suffix-align folded envelope tail in view verification","category":"storage","root_cause":"storage: the sampled digest verifier compared the envelope's bounded folded snapshot vs the table's full raw tail positionally from index 0 → permanent DIGEST DRIFT latch; fix suffix-aligns the tail","ci_state":"green","merge_state":"dirty","pr_number":6363,"confidence":"high"},
    {"number":7240,"title":"compress parser: newline-joined JSON tuples","category":"compress-parser","root_cause":"compress-parser: the compress range parser returns ok=true for newline-joined JSON tuples but parses only one range, silently moving summary text into topic","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":7229,"title":"[Root Cause] plugin-mode board","category":"plugin-mode","root_cause":"rcagg board: aggregation thread for the plugin-mode root cause (4 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6452,"title":"[Root Cause] session-resolve board","category":"session-resolve","root_cause":"rcagg board: aggregation thread for the session-resolve root cause (9 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6982,"title":"[Root Cause] window-fit board","category":"window-fit","root_cause":"rcagg board: aggregation thread for the window-fit (window budget) root cause (2 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6454,"title":"[Root Cause] wire-fidelity board","category":"wire-fidelity","root_cause":"rcagg board: aggregation thread for the wire-fidelity root cause (13 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6450,"title":"[Root Cause] persist-session board","category":"persist-session","root_cause":"rcagg board: aggregation thread for the persist-session root cause (3 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6446,"title":"[Root Cause] dsh-lane board","category":"dsh-lane","root_cause":"rcagg board: aggregation thread for the dsh-lane root cause (9 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6978,"title":"[Root Cause] dsh-model-routing board","category":"dsh-model-routing","root_cause":"rcagg board: aggregation thread for the dsh-model-routing root cause (2 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6981,"title":"[Root Cause] fork-adoption board","category":"fork-adoption","root_cause":"rcagg board: aggregation thread for the fork-adoption root cause (2 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6896,"title":"[Root Cause] identity-msgid board","category":"identity-msgid","root_cause":"rcagg board: aggregation thread for the identity-msgid root cause (3 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6980,"title":"[Root Cause] kernel-compress board","category":"kernel-compress","root_cause":"rcagg board: aggregation thread for the kernel-compress root cause (2 member issues); note: this round's kernel individuals are filed under kernel-compression (the 15-member board) — the two kernel strings are near-aliases","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6447,"title":"[Root Cause] kernel-compression board","category":"kernel-compression","root_cause":"rcagg board: aggregation thread for the kernel-compression root cause (15 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6448,"title":"[Root Cause] panel-ui board","category":"panel-ui","root_cause":"rcagg board: aggregation thread for the panel-ui root cause (3 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":6449,"title":"[Root Cause] performance-memory board","category":"performance-memory","root_cause":"rcagg board: aggregation thread for the performance-memory root cause (15 member issues)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":7273,"title":"egress fuse: ~1.7M-token top-level field not in context","category":"egress-fuse-pricing","root_cause":"egress-fuse-pricing: a ~1.7M-token top-level field (dsh_session_log-class, in the 'other' bucket) doesn't enter bili's context accounting (sentTotal) but alone pushes the body over the window → persistent 413 egress_window_fuse; wants top-N key naming + prompt-mass caliber; attachment read (full egress-fuse report)","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #7275 in flight"}
    ]
    

    Round summary (73 issues): 14 rcagg aggregation boards (preflight-tier, plugin-mode, session-resolve, window-fit, wire-fidelity, persist-session, dsh-lane, dsh-model-routing, fork-adoption, identity-msgid, kernel-compress, kernel-compression, panel-ui, performance-memory) + 59 individual issues. Category ledger of the individuals: kernel-compression 5 (5803/7299/5315/5779/5443), performance-memory 8 (7294/7293/7279/7285/7282/7278/7280/7276), panel-ui 5 (5652/7291/7292/7290/7289), egress-fuse-pricing 4 (4928/7275/7092/7273), persist-session 4 (7298/7297/6215/5336), throttle-governor 3 (7301/7295/6156), codex-lane 2 (7296/4594), cache-invalidation 2 (5478/4790), preflight-tier 3 (6467/6546/3978), storage 3 (5591/5255/6363), wire-fidelity 2 (5672/5737), config-knobs 2 (7277/7283), kernel-identity 2 (5683/5235), install-lane 1 (5357), external-provider 2 (7287/7286), identity-channel 2 (7065/6785), dsh-lane 2 (7075/5693), fold-reconcile 1 (5400), compress-gate 1 (7300), compress-parser 1 (7240), test-quality 1 (7288), reasoning-drop 1 (7143), window-budget 1 (5384).

    Notable: CI-red PRs this round — 5779→5777 (3 failed), 5235→5233 (1 failed), 5384→5382 (7 failed), 6785 (1 failed). Kernel-approval-pending: 5443 (healIllegalSystemPositions FOLLOW side, Kernel Boundary). Decision/fix-PR-in-flight notes carried on 7297/7293/7277/7291/7286/7289/7284/7278/7276/7273. blocked mergeability reported as dirty (schema has no blocked); unknown as clean (unverified). 7273's attachment (the actual egress-fuse diagnostics report) was read and confirms the egress-fuse-pricing classification.

    All 73 issues classified; analysis-only honored (no PRs/issues/labels/edits).

  15. ranxianglei commented on Nov 19, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [bot] 🏷 [plan] 25-issue incremental round

    Batch (25 mirror numbers): 4838, 6877, 5693, 4902, 7179, 4937, 5006, 6156, 6363, 6667, 6859, 5779, 5166, 5119, 5866, 5235, 7332, 5255, 5427, 7331, 5336, 5384, 5880, 6076, 6903.

    Plan:

    1. Fetch all 25 mirror bodies (http://127.0.0.1:3002/api/v1/issues/).
    2. 16 are repeats from earlier rounds of this thread — I carry their stable categories forward (5693 dsh-lane, 4902 window-budget, 7179 dsh-lane, 4937 native-compact-rebase, 5006 egress-fuse-pricing, 6156 throttle-governor, 6363 storage, 6859 wire-fidelity, 5779 kernel-compression, 5119 test-quality, 5235 kernel-identity, 5255 storage, 5427 watermark-fastpath, 5336 persist-session, 5384 window-budget, 4838 ci-build-release) and re-verify body/PR/CI state.
    3. 9 are new to this thread — 6877, 6667, 5166, 5866, 7332, 7331, 5880, 6076, 6903 — classify fresh by reading body + linked PR/diff + source.
    4. Pull CI + merge state per PR (mirror upstream-checks API, keyed by the GitHub PR number).
    5. Post one [final] 25/25 reply with the full classifier-json array (batch order, each exactly once).

    Category scheme: reuse the established strings from prior rounds (kernel-compression, kernel-identity, window-budget, egress-fuse-pricing, throttle-governor, storage, persist-session, wire-fidelity, dsh-lane, native-compact-rebase, watermark-fastpath, test-quality, ci-build-release, panel-ui, config-knobs, performance-memory, preflight-tier, plugin-mode, session-resolve, identity-channel, cache-invalidation, install-lane, external-provider, fold-reconcile, codex-lane, reasoning-drop, compress-gate, compress-parser, egress-fuse-pricing, tinyfix, …). Mint a new stable string only for a genuinely new root cause. tinyfix = reserved for a single small fix with no shared mechanism.

    Field mapping: mergeable blocked→dirty (schema has no blocked), unknown→clean, behind→behind; issue rows (no PR) → ci none, merge n/a. Analysis-only: no PRs/issues/labels/edits.

  16. ranxianglei commented on Nov 19, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [bot] 🏷 [final] 25/25 — complete classifier-json + round summary

    [
    {"number":4838,"title":"release v0.1.209","category":"ci-build-release","root_cause":"ci-build-release: release PR shipping master as-is (458 commits since the v0.1.206 fork point); the fix it exists for is the MCP shim structuredContent now carrying the full acp_status/acp_cache/search_context/decompress/acp_retrieve report text (#3472/#4836)","ci_state":"green","merge_state":"dirty","pr_number":4838,"confidence":"high"},
    {"number":6877,"title":"nudge vs egress fuse disagree, long session stuck on 413","category":"egress-fuse-pricing","root_cause":"egress-fuse-pricing: the egress fuse measures the raw Responses `input` field (~1,068,060) which carries accumulated encrypted_content reasoning/compaction blobs the provider's input_tokens usage (~437,971) does NOT bill; the nudge rides the usage anchor so the two disagree → persistent 413 egress_window_fuse","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high"},
    {"number":5693,"title":"dsh long-session drift comprehensive remediation","category":"dsh-lane","root_cause":"dsh-lane: four-part dsh long-session drift remediation — directSendOrRefuse now refuses decisive native compaction on dead-proxy direct-send paths (local 422 x-bili-gate: dsh-compaction-refused) instead of sending it upstream where the checkpoint would destroy the compression substrate (#1729)","ci_state":"green","merge_state":"dirty","pr_number":5691,"confidence":"high"},
    {"number":4902,"title":"refactor: name the window-reservation block in handleRequest","category":"window-budget","root_cause":"window-budget: refactor — extract the window-reservation block (6-layer native window parse + 5 interlocking rules #896/#282/#4205/#3450/#2096) from the 2545-line handleRequest into reserveJudgingWindow(); pure naming/structure, no behavior change","ci_state":"red","merge_state":"dirty","pr_number":4902,"confidence":"high"},
    {"number":7179,"title":"dsh compaction gate covers runtime-registered agentPresets","category":"dsh-lane","root_cause":"dsh-lane: registerCompactionDisable (src/agent/dsh-compaction-disable.ts) only hooked internal/config (waterfall plugin entries); presets registered at runtime via ctx.agentPresets.register bypass the gate → native compaction answered 200 and silently wipes bili folds; fix wraps agentPresets.register with the same disableBasicCompaction rewrite","ci_state":"green","merge_state":"dirty","pr_number":7179,"confidence":"high"},
    {"number":4937,"title":"codex: rebase ACP state on opaque native compaction checkpoint","category":"native-compact-rebase","root_cause":"native-compact-rebase: codex's local auto-compact lands as an opaque encrypted_content compaction item the host text-prefix detector (carriesCodexLocalCompactionSummary) can never see → covered fold ids vanish → drift latch → preflight pauses forever; fix detects the wire shape (type=compaction + non-empty encrypted_content) and rebases through the native-compaction boundary","ci_state":"red","merge_state":"dirty","pr_number":4937,"confidence":"high"},
    {"number":5006,"title":"doom-gate byte floor prices billable content","category":"egress-fuse-pricing","root_cause":"egress-fuse-pricing: the three doomed gates (egress fuse src/server/relay.ts, preflight probe #2313, post-compress recheck) priced ONLY core message bytes vs max(2MB, window×7.5 B/token); a 26,903-message session with 1.44 MiB core text produced 55.2 MiB bodies every gate read as small; fix prices billable content = core text + system/tools wire overhead","ci_state":"red","merge_state":"dirty","pr_number":5006,"confidence":"high"},
    {"number":6156,"title":"credential fingerprint covers all auth headers, order-independent","category":"throttle-governor","root_cause":"throttle-governor: credentialFingerprint (src/cache-ledger.ts) broke on the first recognized auth header in insertion order (order-dependent + incomplete) → the #4204 429 governor's holds missed reordered requests and a different account behind the same gateway inherited the first's hold; fix collects all, normalizes, hashes the sorted set","ci_state":"green","merge_state":"dirty","pr_number":6154,"confidence":"high"},
    {"number":6363,"title":"suffix-align folded envelope tail in view verification","category":"storage","root_cause":"storage: the sampled digest verifier (BatchedViewVerification.compareMessages, src/storage/shadow.ts) compared the envelope's bounded folded snapshot vs the table's full raw tail positionally from index 0 → once the tail grew past the persistTailTokens cap every slot mismatched → permanent DIGEST DRIFT latch (8–9 s re-ingest per save); fix suffix-aligns the overlap","ci_state":"red","merge_state":"dirty","pr_number":6363,"confidence":"high"},
    {"number":6667,"title":"bound init-snapshot memory/time with top-100 min-heap","category":"performance-memory","root_cause":"performance-memory: snapshotInit (src/context-analysis.ts) allocated one InitMessageEntry per message + sorted the whole array before retaining 100 → O(n) memory + O(n log n) sync work on the first proxied request; fix = bounded top-K min-heap (~1313 ms/85 MiB → ~48 ms/13 MiB on 1M msgs)","ci_state":"red","merge_state":"dirty","pr_number":6667,"confidence":"high","notes":"duplicate of #6656 (same mechanism, two open PRs)"},
    {"number":6859,"title":"preserve signed ordinary text parts on wire + tool-loop rebuilds","category":"wire-fidelity","root_cause":"wire-fidelity: Gemini can attach a thoughtSignature to an ORDINARY text part (no thought:true); the Google SSE adapter + non-streaming rewriter only exempted thought:true parts from prose filtering, and the tool-loop rebuild dropped the signature → violates KDD #10 byte-for-byte; fix treats a signed text part like a thought part","ci_state":"green","merge_state":"dirty","pr_number":6859,"confidence":"high"},
    {"number":5779,"title":"screen tier merge nudges against compression eligibility","category":"kernel-compression","root_cause":"kernel-compression: tier 2/3 distillation nudges advertised block ranges rejected by the kernel turn-integrity/coverage rules; only the kernel can screen the same projection used by compression; fix screens tier targets on the actual apply view (up to 8 individual-block fallback)","ci_state":"red","merge_state":"dirty","pr_number":5777,"confidence":"high"},
    {"number":5166,"title":"decide block liveness by coverage shape, not existence","category":"fold-reconcile","root_cause":"fold-reconcile: syncBlocks asked whether ANY covered message was still present, so a host rewind (survivors = prefix, newest gone) kept injecting a summary describing vanished history every turn; fix discriminates by the gap's shape — prefix-dead / suffix-live / interior-live (fold-reconcile re-keys the interior id-churn case)","ci_state":"green","merge_state":"dirty","pr_number":5166,"confidence":"high"},
    {"number":5119,"title":"sandbox XDG base dirs in suite preload","category":"test-quality","root_cause":"test-quality: bare targeted test runs skipped the preload entirely and the suite preload only pinned BILI_STORAGE_UNIFIED → tests wrote sessions/index.db rows + envelopes + state into the real $HOME; fix sandboxes the four XDG base dirs to a per-process temp root in the preload","ci_state":"green","merge_state":"dirty","pr_number":5119,"confidence":"high"},
    {"number":5866,"title":"[claude] 400 final block cannot be thinking after compress","category":"kernel-compression","root_cause":"kernel-compression (bug report): the model's [thinking, tool_use:compress] turn — planReasoningDrop keeps the signed thinking (KDD #10) while hideConsumedCompressCalls hides the compress pair independently → the assistant message collapses to [thinking] → Anthropic 400; fix PR #5880 (GH 5877) in flight","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"fix PR #5880 in flight; kernel approval pending via #7331"},
    {"number":5235,"title":"kernel: preserve issued refs, quarantine historical collisions","category":"kernel-identity","root_cause":"kernel-identity: a restart could restamp previously issued refs + the allocator could reuse historical numbers → a model citing an old ref resolves/compresses a different message; fix preserves issued refs across restart + quarantines known collisions (kernel 0.0.170→0.0.171)","ci_state":"red","merge_state":"dirty","pr_number":5233,"confidence":"high"},
    {"number":7332,"title":"heal pass keeps a thinking block out of the final assistant block","category":"kernel-compression","root_cause":"kernel-compression: healUnpairedServerToolBlocks (kernel/src/wire/anthropic.ts) drops an orphaned server-tool half and rewrites to the surviving blocks but never checks the LAST block — if it's thinking-family the healed message ends in thinking → 400; fix appends the emptied-case placeholder (the SECOND producer of #5864, distinct from #5877)","ci_state":"green","merge_state":"dirty","pr_number":7332,"confidence":"high"},
    {"number":5255,"title":"degrade 2nd long-lived index.db writer to legacy-only","category":"storage","root_cause":"storage: two bili installs sharing one sessions dir both RW-open the same index.db → corruption + SQLITE_BUSY turn failures; fix degrades the second long-lived (proxy/sidecar) writer to legacy-only at open (owner decision B+S2 from #5064); ephemeral CLI/export opens keep report-only","ci_state":"none","merge_state":"dirty","pr_number":5250,"confidence":"high"},
    {"number":5427,"title":"dsh plugin-lane full-resend watermark foundation","category":"watermark-fastpath","root_cause":"watermark-fastpath: the dsh plugin lane re-parses the ENTIRE resent body per round (81 MiB → hundreds-of-MB AST peak — the #5403 user-side OOM producer); fix = count-watermark (x-bili-pf-count) so the server parses only the uncovered tail window; foundation behind a 0% rollout gate (inert)","ci_state":"red","merge_state":"dirty","pr_number":5425,"confidence":"high"},
    {"number":7331,"title":"kernel authorization request: PR #5877 (fix #5864)","category":"kernel-compression","root_cause":"kernel-compression (decision request): kernel authorization for PR #5877 — the fix for the #5864 400 (hide pass leaving a signed-thinking predecessor) — pending owner approve/reject per Kernel Boundary; escalated to an issue because the PR-thread decision request sat unanswering 10 days","ci_state":"none","merge_state":"n/a","pr_number":null,"confidence":"high","notes":"kernel approval pending — owner must approve/reject PR #5877"},
    {"number":5336,"title":"bili sessions orphan-sweep: provably-foreign collapsed blocks","category":"persist-session","root_cause":"persist-session: a collapsed block committed under a shared conversation id (≤v0.1.208 era) reappears FOREVER in the parent session after upgrade — the #2695 zombie-reap exemption shelters it (same 'covered ids never on this wire' signature as healthy proxy-mode folds); fix = orphan-sweep of provably-foreign blocks","ci_state":"none","merge_state":"dirty","pr_number":5334,"confidence":"high"},
    {"number":5384,"title":"overflow-stated counts steer local fold targets","category":"window-budget","root_cause":"window-budget: on a fresh overflow-only session the ring can never publish (needs ≥2 agreeing samples) so local fold judgments stayed raw-estimate calibre → every forward re-400'd (permanent loop); fix retains the upstream-stated count as a provisional estimator scale (stats.overflowBias) via util.effectiveCalibration","ci_state":"red","merge_state":"dirty","pr_number":5382,"confidence":"high"},
    {"number":5880,"title":"keep compress pair visible when predecessor is signed thinking","category":"kernel-compression","root_cause":"kernel-compression: fix for #5864 — planReasoningDrop keeps the signed thinking while hideConsumedCompressCalls hides the compress pair independently → [thinking] final block → 400; fix adds a host-injected strandedProtectedPredecessor predicate to the kernel Config (kernel 0.0.200→0.0.201)","ci_state":"green","merge_state":"dirty","pr_number":5877,"confidence":"high"},
    {"number":6076,"title":"compress separator guard for CJK range words + rejected preview","category":"compress-parser","root_cause":"compress-parser: the #4253 unrecognized-separator guard (kernel/src/parse-compress-input.ts) only covered punctuation, so CJK range words 到/至 between two refs bypassed it → silent one-message fold of a multi-message range (violates the guard's own invariant); fix closes the letter-class hole + adds a rejected-payload preview","ci_state":"green","merge_state":"dirty","pr_number":6074,"confidence":"high"},
    {"number":6903,"title":"image recovery: one ref per split image, DSH attachments recoverable","category":"image-recovery","root_cause":"image-recovery: the image-media recovery path (src/image-restore.ts) either over-recovers (each split image's rawResponsesItem carries the FULL original item → N×N spill files + cross-image contamination) or under-recovers (a DSH read_image attachment is reduced to text by openaiToCore → extractMediaRefs never sees it → pinned BLOCKED, never folds); fix = one ref per split image + attachment recovery","ci_state":"green","merge_state":"dirty","pr_number":6902,"confidence":"high","notes":"CI incomplete (12 pending)"}
    ]
    

    Round summary (25 issues): 16 repeats from earlier rounds (categories carried forward + body/PR/CI re-verified) + 9 new (6877, 6667, 5166, 5866, 7332, 7331, 5880, 6076, 6903). Category ledger: kernel-compression 5 (5779/5866/7332/7331/5880), egress-fuse-pricing 2 (6877/5006), dsh-lane 2 (5693/7179), window-budget 2 (4902/5384), storage 2 (6363/5255), ci-build-release 1 (4838), native-compact-rebase 1 (4937), throttle-governor 1 (6156), performance-memory 1 (6667), wire-fidelity 1 (6859), fold-reconcile 1 (5166), test-quality 1 (5119), kernel-identity 1 (5235), watermark-fastpath 1 (5427), persist-session 1 (5336), compress-parser 1 (6076), image-recovery 1 (6903, NEW this round).

    Notable: CI-red PRs — 6667 (12f), 5384 (7f), 4937 (6f), 4902 (3f), 5779 (3f), 5006 (2f), 6363 (2f), 5235 (1f), 5427 (1f). blocked mergeability (5166/7332/5880/6076/6903) reported as dirty (schema has no blocked). Kernel-approval-pending: 7331 (decision request for PR #5877) + 5866 (bug report, fix #5880 in flight). 6667 is a duplicate of #6656 (noted). All 25 classified in batch order, each exactly once; analysis-only honored (no PRs/issues/labels/edits).

  17. ranxianglei commented on Nov 20, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

    Root-cause classification — round plan (58 issues)

    Analysis-only lane: no PRs/issues/labels, no edits. I read all 58 issue bodies, will open the linked PR/diff where present and spot-check the workdir (billion-context @ 6453feeeb, 0.1.216) to name the component + mechanism, then post progress at 25/50 and the full classifier-json at 58.

    Category scheme — dynamic, mechanism-based root-cause strings (one string per mechanism; a bug + its fixing PR share the string). Reserved tinyfix for single small fixes with no shared mechanism. Scheme for this batch:

    • watermark-fastpath-id-namespace — fastpath head check compares ids across two id namespaces (fastpath.ts)
    • pin-ledger-sweep-drain — fastpath-HIT turns drain the msgid pin ledger via the [bug] Fold anchor detaches after host restart (parent-gone): hydrated and unified-shadow DIGEST DRIFT within the same second, gap solidifies without self-healing (0.1.197, complementary to #3013 restart path) | [bug] 宿主重启(parent-gone)后 fold 锚点脱锚:hydrated 与 unified-shadow DIGEST DRIFT 同秒,缺口固化不自愈(0.1.197,与 #3013 重启路径互补) #3285 authoritative-set sweep
    • fail-fast-forward-bypass — fail-fast says "NOT forwarded" but the same-round payload still reaches upstream
    • doom-gate-byte-floor-caliber — doom-gate byte floor prices core text only, not system/tools wire overhead
    • egress-fuse-nudge-disagreement — nudge (stale usage) and egress fuse disagree → stuck 413
    • request-body-budget-config — large-decoded-request budget constants hardcoded / not configurable
    • transport-misclassified-degenerate — upstream connect-timeout misrouted into the degenerate-turn path
    • degenerate-turn-fatal-no-toggle — post-continuation degenerate turn is always fatal, no toggle, 【bug】空轮重试门闸把 reasoning 记成可见正文 —— Responses 协议下 item.content[] 带 reasoning 的退化轮永不重试(客户端拿到空轮) #2756 exemption unreachable
    • compress-fold-benefit-floor — model-initiated folds lack a min-benefit floor + per-nudge budget (cache re-bill)
    • compress-fail-streak-loop — compressFailStreak resets on any success → infinite blind-retry loop
    • compress-circuit-breaker-text-mismatch — breaker receipt forbids the only recovery the code still allows
    • nudge-collapsible-interval-no-toggle — collapsible-interval fold nudge fires even with autoFold:false
    • compress-range-includes-current-user-msg — compress folds the newest user msg → model answers a fabricated message
    • storage-poison-row-isolation — a single bad CAS/JSON row aborts the whole session load (write no hash-verify + read no isolation)
    • storage-digest-drift — batched/snapshot verification DIGEST DRIFT latch (positional compare / lost-tail)
    • storage-vacuum-foreign-writer — sessions rm full-VACUUM under a live foreign writer
    • session-gc-companion-stat — session GC treats a failed companion stat as absence → deletes originals
    • dead-code-narrow-read — fix: bound ordinary session loading and avoid full reloads for lazy fields #4739 narrow-read restore path is dead, never wired
    • msgid-identity-channel — stable session+msgid identity channel (per-client wiring / snapshot plane)
    • fork-inheritance-lost — DSH fork child loses inherited compression (409 non-transient / parentSession omitted) → zero-fold, 400
    • signed-block-wire-fidelity — signed ordinary text (Gemini thoughtSignature) rewritten on wire / tool-loop rebuild
    • thinking-final-block-400 — 400 "final block cannot be thinking" after a compress that follows a thinking block
    • anthropic-system-position-heal — heal enforces only the FOLLOW side of Anthropic's mid-conversation system rule
    • responses-tool-ordering — Responses item ordering breaks web_search / tool_search pairs (separator insertion / unknown types)
    • 429-governor-hold-scope — 429 governor hold scope/streak/Retry-After freezes a whole credential, escalates to 424
    • credential-fingerprint-order — credentialFingerprint order-dependent + incomplete on multi-credential lanes
    • codex-prewarm-as-turn — codex context-only prewarm runs as an ACP turn (folds wiped, summaries re-injected)
    • v1internal-preflight-envelope — v1internal preflight summary request not wrapped / response not unwrapped
    • billing-registry-lookup-perf — billing/registry price lookups rebuild tables + scan rows per lookup
    • billing-disk-summary-stale — disk-only web summaries keep stale dollars after price changes
    • web-blog-refresh-byte-cap — blog-refresh byte caps over-buffer streams + reject valid base64 docs
    • web-diagnostics-log-read — diagnostics reads unused log generations
    • web-dashboard-stale-summary — file-backed dashboard summary prefers a stale session copy
    • opencode-lane-hook-payload — opencode-lane hook rewrites the shared http.request payload
    • dsh-long-session-drift — dsh long-session drift remediation
    • orphan-proxy-offloop-reap — wedged orphaned proxy reaped off-loop when owner dies
    • failed-compress-receipt-assistant-tail — failed-compress recovery receipt leaves re-request ending on an assistant turn
    • tag-echo-filter-drift — tag-echo filter misses tail-append name drift + mid-turn draft envelopes
    • boot-contract-characterization — boot-contract characterization tests (verifyAttachAndRecover / latches / health)
    • init-snapshot-memory — init-snapshot allocates+sorts all messages before retaining 100
    • summary-role-label-mislabel — per-line role labels in compressed summaries mislabeled
    • folded-summary-role-authority — folded summaries carry live-message authority via carrier role (user/system)
    • tinyfix — reserved: single small fix, no shared mechanism

    Posting progress at 25/58 next.

  18. ranxianglei commented on Nov 20, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [progress] 25/58

    [
      {"number": 7347, "title": "Watermark fastpath never hits on 2x2 pinned sessions", "category": "watermark-fastpath-id-namespace", "root_cause": "src/server/fastpath.ts head check compares a.id!==b.id between fresh content-hash cores (h_…) and the post-pin host/ledger msgid cores — two id namespaces that never match on a pinned session, so the fastpath always misses (pre-pin traffic keeps content-hash ids and hits, which is why rollout-off unit tests stayed green)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7346, "title": "Watermark-HIT turns drain the msgid pin ledger", "category": "pin-ledger-sweep-drain", "root_cause": "on a fastpath HIT, snapshotMessages(session, windowCores) claims rawness (lastMessagesFolded=false) so the #3285 authoritative-set sweep retires every ledger row the ~2-row tail window omits; compounded by hydrateSlimCcrStore returning null for the empty CCR envelope → false DIGEST DRIFT → completeLog on every save", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6785, "title": "[PR] msgid delivery × snapshot plane unified merge", "category": "msgid-identity-channel", "root_cause": "unifies msgid delivery (host msgid envelope → server pin → fork) with the snapshot plane (4-table schema v12, shadow write, mapEgress direct) so stage-0 consumes the delivered host msgids; new storage.snapBase.* config (default OFF) + gate --snapbase-full; absorbs #6564/#5901/#6680", "ci_state": "red", "merge_state": "behind", "pr_number": 6785, "confidence": "high"},
      {"number": 7345, "title": "fail-fast 之后同一轮的 payload 仍发到上游", "category": "fail-fast-forward-bypass", "root_cause": "preflight fail-fast returns 502 'NOT forwarded' but the same-round payload still reaches upstream (an in-flight request is not cancelled, or a concurrent retry bypasses the gate) — the gate decision and the actual egress are not atomically coupled, and the log line is misleading", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "medium"},
      {"number": 7339, "title": "[Drive-by] persist.ts narrowBlockContents/narrowMetadata are dead code", "category": "dead-code-narrow-read", "root_cause": "src/persist.ts:3686 narrowBlockContents + :3668 narrowMetadata (from 7b1ecc2b1fb / #4739) have no callers anywhere in src/ or tests/ — the narrow-read restore path was never wired; pure dead code, no runtime impact", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7344, "title": "[PR] refactor(storage): remove the dead #4739 narrow-read restore path", "category": "dead-code-narrow-read", "root_cause": "removes the dead #4739 narrow-read chain (persist.ts lazySource/narrowMetadata/narrowBlockContents, shadow.ts getSessionMetadata + no-arg getBlockContents, store.ts no-arg getBlockContents/getMessagesAt/sessionViewRows/getSessionMetadata/viewPointerSeqs, backend.ts + mysql.ts sessionViewRows); no runtime path touched, pure dead-code removal", "ci_state": "green", "merge_state": "behind", "pr_number": 7343, "confidence": "high"},
      {"number": 5330, "title": "[PR] fix(#5236): sessions rm must not full-VACUUM under a live foreign writer", "category": "storage-vacuum-foreign-writer", "root_cause": "runRm in src/storage-maintenance.ts lacked the store.unifiedActiveWriter() guard the other three full-VACUUM sites have — a live foreign writer lets sessions rm's full VACUUM win the exclusive lock and starve the co-resident instance for the whole O(db-size) rewrite; fix skips only the shrink up front with a pid-named note", "ci_state": "green", "merge_state": "clean", "pr_number": 5330, "confidence": "high"},
      {"number": 6037, "title": "transport connect-timeout misclassified as a degenerate turn", "category": "transport-misclassified-degenerate", "root_cause": "when an upstream TCP connect times out and network.replayRetryMax (default 3) is exhausted, the turn reaches the plugin pipe as a zero-output completed turn and is classified 'degenerate terminal turn' (#732/#821), re-POSTed with a continuation nudge, then #870 fatal — a transport failure is misrouted into the degenerate path, which cannot repair a dead connection", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7343, "title": "压缩摘要里角色标签标错", "category": "summary-role-label-mislabel", "root_cause": "per-line role labels in compressed-summary bodies are wrong (assistant thinking / tool-call JSON / system-injection rendered as [user]) — 45/4091 mislabeled + 157 refs that don't exist in the session; the renderer either inherits the block-head user role when merging consecutive assistant blocks, or the summary model mis-attributes speaker", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "medium"},
      {"number": 4913, "title": "续写重试后的退化回合必然变成致命 in-band stream error", "category": "degenerate-turn-fatal-no-toggle", "root_cause": "in the plugin pipe, a degenerate turn that survives one continuation nudge always emits the fatal in-band error (#870) with no toggle to disable it; the #2756 'bare' exemption is written below the fatal branch so it never executes on the retry pass (36% retry→hard-fail on 0.1.208)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7341, "title": "[PR] fix(web): enforce decoded blog download byte limits", "category": "web-blog-refresh-byte-cap", "root_cause": "src/web/blog-refresh.ts enforced the 512KiB/2MiB caps only after response.arrayBuffer() buffered the whole body (missing/understated Content-Length → over-buffer to EOF) and the contents-API fallback counted base64 padding+newlines as file bytes (rejecting valid 524287/524288-byte docs); fix reads chunks against the cap, cancels on overflow, measures whitespace-normalized decoded length", "ci_state": "green", "merge_state": "behind", "pr_number": 7341, "confidence": "high"},
      {"number": 7334, "title": "[PR] fix: add min-benefit floor to model-initiated compress (Part 1 of #7333)", "category": "compress-fold-benefit-floor", "root_cause": "model-initiated compress in applyRanges accepted any fold with net shrink ≥1 token (only the ~5000-char minCompressRangeChars floor), so a burst of tiny folds each re-bills the prompt-cache tail; adds compress.minFoldBenefitTokens (default 4096) gate below the pressure band + a model-facing refusal receipt teaching the model to batch ranges", "ci_state": "green", "merge_state": "behind", "pr_number": 7334, "confidence": "high"},
      {"number": 7342, "title": "Blog refresh byte caps over-buffer streams and reject valid base64", "category": "web-blog-refresh-byte-cap", "root_cause": "same mechanism as #7341 — src/web/blog-refresh.ts plain-text/image paths enforce byte caps only after arrayBuffer() consumes the full response, and the contents-API base64 estimate counts padding/newlines as decoded bytes; fixing PR #7341 in flight", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5835, "title": "[PR] fix: reap wedged orphaned proxies off-loop when the owner dies (#5449)", "category": "orphan-proxy-offloop-reap", "root_cause": "the in-loop parent watcher (2s ticks) only detects/reaps when the event loop yields; a wedged orphan proxy with a dead owner held port + index.db lock for minutes; adds an off-loop owner-liveness probe (process.kill(pid,0)) in the #4762 worker that exits ~15–45s after confirming owner death ×3 ticks + loop silence ≥30s + zero other watchers", "ci_state": "green", "merge_state": "clean", "pr_number": 5833, "confidence": "high"},
      {"number": 7338, "title": "[PR] fix: isolate block summary/one CAS content parse (#7337)", "category": "storage-poison-row-isolation", "root_cause": "appendBlockContents (src/storage/ingest.ts:615) did a bare JSON.parse on block summary/one CAS content, so one corrupt row threw out of the shared helper and aborted the whole cold load on both the full-view and resident-field paths; fix skips + returns the unparseable row into skippedCorruptRows (closes the last gap #7335/#7336 started)", "ci_state": "green", "merge_state": "clean", "pr_number": 7338, "confidence": "high"},
      {"number": 5628, "title": "[PR] fix(opencode-lane): stop rewriting the shared http.request payload (#5625)", "category": "opencode-lane-hook-payload", "root_cause": "the opencode-lane http.request hook replaced e.request with a new Request at the proxy origin, but opencode hands that one mutable event to every later hook — provider auth plugins (Kiro, Antigravity) validating the upstream URL broke; fix makes routes only discover the proxy and performs the /bili/ rewrite at fetch time in the shared native fetch chain", "ci_state": "green", "merge_state": "clean", "pr_number": 5626, "confidence": "high"},
      {"number": 7337, "title": "[Drive-by] Block summary/one CAS content parse is not isolated", "category": "storage-poison-row-isolation", "root_cause": "same mechanism as #7335/#7336 but a different carrier — appendBlockContents (src/storage/ingest.ts:615) bare JSON.parse on summary/one CAS content (vs message content + block payloadJson which #7336 isolated); a corrupt block-content row still aborts the whole session load; fixing PR #7338 in flight", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7335, "title": "一条坏数据行可以炸掉整个会话", "category": "storage-poison-row-isolation", "root_cause": "two structural gaps — (1) write side: UnifiedStore.putContent (src/storage/store.ts) trusts the caller's precomputed hash (const hash = precomputedHash ?? sha256Hex), so a hash≠bytes poison pair enters the CAS; (2) read side: LegacyViewBuilder.addMessages/addBlocks (src/storage/ingest.ts) bare JSON.parse with no try/catch, so one bad row throws through viewFor → legacy fallback fails → starting fresh → full state wipe (26 poison rows → 32 serial preflight rounds, 12+ min user stall); fixing PR #7336", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5245, "title": "[PR] Boot contract stubbing: four-lane verifyAttachAndRecover matrix + latch permanence + health glossary", "category": "boot-contract-characterization", "root_cause": "characterization tests only (zero product code) pinning current verifyAttachAndRecover behavior across pi/omp/opencode/dsh (S1 healthy / S2 strict-old-version / S3 dead-target), the one-shot latch permanence (stats.disabled 3-strike never-unlatches, logger stderrDead), and the /__bili/health key-set — building the equivalence net before extracting a shared lifecycle (#5241 Phase 0)", "ci_state": "green", "merge_state": "clean", "pr_number": 5242, "confidence": "high"},
      {"number": 5178, "title": "[PR] fix(#5175): strip drifted open-tag names and mid-turn draft envelopes", "category": "tag-echo-filter-drift", "root_cause": "two holes — (A) the open-side ACP_NAME_ALT enum (src/loop/tag-echo-filter.ts buildAcplikeName) has no tail-append arm so drifted names (acpt/acct/accent…) pass all three gates (RENDER_TAG_DETECT/BROAD_TAG_HEAD_TAIL/PARTIAL_TAIL); (B) the draft predicate DRAFT_CLOSE_TAIL (src/degenerate-turn.ts:78) is $-anchored so mid-turn draft envelopes leak; fix keys on the two payload shapes (ref-in-attr empty body / emitter tokens= attr) instead of enumerating names", "ci_state": "green", "merge_state": "clean", "pr_number": 5178, "confidence": "high"},
      {"number": 7336, "title": "[PR] fix: refuse unverifiable content digests and isolate unparseable rows (#7335)", "category": "storage-poison-row-isolation", "root_cause": "closes both #7335 gaps — putContent (store.ts + mysql.ts) now always recomputes sha256 and rejects a mismatching caller digest (tx rollback, loud error); LegacyViewBuilder + shadow.ts skip unparseable rows into skippedCorruptRows with a loud warn; sidecar verify-sample isolates bad rows as drift instead of failing every save", "ci_state": "green", "merge_state": "clean", "pr_number": 7336, "confidence": "high"},
      {"number": 5006, "title": "[PR] fix: doom-gate byte floor prices billable content", "category": "doom-gate-byte-floor-caliber", "root_cause": "the three doomed gates (egress fuse src/server/relay.ts, preflight fail-fast probe #2313, post-compress recheck src/server.ts) priced only core message bytes against max(2MB, window×7.5B/token) while the preflight trigger prices messages+system/tools — a payload whose billable total exceeds every window cleared the floor on the residual caliber (26,903-msg session, 1.44MiB core text → 55.2MiB bodies read as 'small'); fix prices core text + system/tools wire overhead from the actual outbound body", "ci_state": "red", "merge_state": "dirty", "pr_number": 5006, "confidence": "high"},
      {"number": 917, "title": "Message Board | 留言板", "category": "tinyfix", "root_cause": "empty message-board stub — no body, no mechanism, no reproducible defect; not part of any systemic root cause", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "low", "notes": "empty body; nothing to classify beyond the reserved tinyfix"},
      {"number": 7333, "title": "compress: one nudge can trigger a burst of small model-initiated folds", "category": "compress-fold-benefit-floor", "root_cause": "the model-initiated compress path (applyRanges) accepts any number of folds per nudge and any fold with net shrink ≥1 token (only the ~5000-char minCompressRangeChars floor), so one nudge → 39 folds in one session, each re-billing the cache tail (619K uncached tokens); bili's own preflight/auto-fold paths have the 4096-token benefit gate but the model path doesn't; Part 1 (min-benefit floor)=PR #7334, Part 2 (per-nudge budget) pending decision", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 3011, "title": "DSH 分支克隆的 fork 继承被 409 SNAPSHOT_UNAVAILABLE 永久放弃", "category": "fork-inheritance-lost", "root_cause": "DSH branch clone → the child's fork adoption gets 409 SNAPSHOT_UNAVAILABLE (parent's raw pluginSnapshot not resident), and createForkAdopter's transient regex (^(snapshot|fork) (error|http 5\\d\\d)) does NOT match 'snapshot http 409', so it's classified non-transient → done.add(child) → the child never inherits folds for the rest of the process → starts from raw ~1.06M-token history → context_length_exceeded 400 on first turn", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"}
    ]
    
  19. ranxianglei commented on Nov 20, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [progress] 50/58 — issues 26–50 (1–25 are in the prior [progress] 25/58; the final reply carries all 58)

    [
      {"number": 6124, "title": "DSH fork() omits parentSession, child inherits 1 of 33 blocks then permanently 400", "category": "fork-inheritance-lost", "root_cause": "DSH 0.2.2-rc.1 fork() writes only isSeeded:true, not parentSession (the dsh-subagent path does write it), so bili's forkChildVerdict returns parent-missing and the fork-adoption path is skipped at the entry; the resume-inheritance fallback uses planForkAdoption's every() all-or-nothing block gate and 4 of 5 active blocks were invalidated by a few image/reasoning messages that shouldn't count → 33 blocks → 1 block, 161,960 tok lost, child permanently 400", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6156, "title": "[PR] fix(wiring): credential fingerprint covers all forwarded auth headers, order-independent (#6148)", "category": "credential-fingerprint-order", "root_cause": "credentialFingerprint (src/cache-ledger.ts) broke on the FIRST recognized auth header in insertion order, so a lane forwarding two distinct credentials got an identity that was order-dependent (reordering flipped the digest → #4204 429 governor holds missed) and incomplete (second credential invisible → a different account inherited the first's hold); fix collects every recognized header, normalizes, dedupes, hashes the sorted set (single-credential byte-identical)", "ci_state": "green", "merge_state": "dirty", "pr_number": 6154, "confidence": "high"},
      {"number": 5693, "title": "[PR] fix: dsh long-session drift comprehensive remediation (#5686)", "category": "dsh-long-session-drift", "root_cause": "four-part dsh long-session drift remediation — P0-1 directSendOrRefuse refuses marker-decisive dsh native compaction calls on dead-proxy direct-send paths (was: sent upstream → checkpoint durably destroyed the substrate #1729); P0-2 detectCompactionOverlayGap alarms for every profile shape (was: web-profile only); P1-1 personaConversationOrder onResolve makes fork/reanchor decisions visible; P1-2 reanchor instead of fork when a landed compaction truncated the head under a new persona tag", "ci_state": "green", "merge_state": "dirty", "pr_number": 5691, "confidence": "high"},
      {"number": 6859, "title": "[PR] fix: preserve signed ordinary text parts on the wire and in tool-loop rebuilds", "category": "signed-block-wire-fidelity", "root_cause": "Gemini can attach a thoughtSignature to an ordinary text part (no thought:true); per KDD #10 a signed block must never be rewritten — but (1) the Google SSE adapter + non-streaming JSON rewriter only exempted thought:true parts from prose filtering, so a signed ordinary text part with marker-like text was mutated before reaching the client, and (2) the streaming loop recorded ordinary text as a merged string and dropped its signature on the next re-request; fix treats a signed text part like a thought part and carries the signature through the loop event + re-request", "ci_state": "green", "merge_state": "dirty", "pr_number": 6859, "confidence": "high"},
      {"number": 6363, "title": "[PR] fix: suffix-align folded envelope tail in batched view verification", "category": "storage-digest-drift", "root_cause": "BatchedViewVerification.compareMessages (src/storage/shadow.ts) compared the envelope's bounded folded snapshot against the table's full raw tail positionally from index 0 — once the tail grew past the persistTailTokens cap, every slot mismatched → a permanent DIGEST DRIFT latch forcing an 8–9s full re-ingest on every save; fix aligns the comparison to the overlap window (the envelope is a suffix of the table tail)", "ci_state": "green", "merge_state": "dirty", "pr_number": 6363, "confidence": "high"},
      {"number": 6667, "title": "[PR] fix: bound init-snapshot memory/time with a top-100 min-heap scan", "category": "init-snapshot-memory", "root_cause": "snapshotInit() in src/context-analysis.ts (advisory per-session observability) allocated one InitMessageEntry per message and sorted the complete array before retaining only 100 — O(n) temp memory + O(n log n) synchronous work on the first proxied request even though the visible sample is capped at INIT_PER_MSG_CAP=100; fix replaces with a bounded top-K min-heap (1313ms/85MiB → 48ms/13MiB on 1M synthetic msgs), output byte-identical", "ci_state": "green", "merge_state": "dirty", "pr_number": 6667, "confidence": "high"},
      {"number": 5606, "title": "failed-compress recovery receipt leaves the re-request ending on an assistant turn → 400", "category": "failed-compress-receipt-assistant-tail", "root_cause": "in src/loop/core.ts the failed-only mutating round branch drops the failed compress/decompress calls and pushes the streamed markers as a final {role:'assistant'} receipt message with nothing after it — Anthropic rejects a conversation ending on an assistant message (trailing whitespace, or 'does not support assistant message prefill' on Claude 5) → 400, subagent dies; the path meant to keep the task alive (#5220) is the one that kills it", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7318, "title": "可折区间折叠提示,关不掉 (collapsible-interval fold nudge can't be turned off)", "category": "nudge-collapsible-interval-no-toggle", "root_cause": "the 'there's still a collapsible interval' nudge is pushed every turn even with compress.externalSummary.enabled=false + autoFold=false + autoFoldTargetTokens=250000 (config reload ok:true) — there is no independent config switch for this nudge, so it keeps inducing the model to fold against the user's explicit stop instruction, each fold re-billing the prefix cache (200+ folds in one night)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5737, "title": "Folded summaries reach the model as user (OpenAI) / system (Anthropic) messages", "category": "folded-summary-role-authority", "root_cause": "the folded-block summary's wire role is asymmetric by lane — kernel renderSummary (kernel/src/prune.ts:548-560) renders it role:system; the OpenAI/Responses egress (adapter-openai.ts:247, prepare-openai.ts:378, responses-tool-output.ts) re-voices it to user via systemToUser (so compressed history enters as a live user message, amplifying #4222 stale-objective re-execution); the Anthropic egress keeps it as a mid-conversation system message (system-level authority); constrained by KDD #7 (proxy-mode user re-voice is intentional for SGLang + prefix-cache head stability)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 4594, "title": "Codex WS context-only prewarm runs as an ACP turn: folds wiped, forged summaries re-injected", "category": "codex-prewarm-as-turn", "root_cause": "src/responses-ws.ts only treats generate:false with an EMPTY input as a prewarm probe; current codex sends its startup/idle warmup as generate:false PLUS developer/user context items, so it falls through into the ACP pipeline as an ordinary zero-message turn → syncBlocks on a view with no covered message deactivates every fold block (73→0, 45→0) and prepare-responses re-injects the entire codexForgedSummaries list (1,421 summaries / 1.2M tokens → preflight fail-fast 413); second gap: the probe relay calls the upstream transport without the relay's proxy dispatcher (ws-bridge only exposes a boolean isEgressProxied), so behind an upstream proxy every probe goes direct and fails", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7296, "title": "Codex Responses ordering inserts a user separator before a hosted web search call", "category": "responses-tool-ordering", "root_cause": "responsesToCore represents a hosted web_search_call as contentType:'reasoning', so repairResponsesAssistantOrdering (src/server/responses-ordering.ts) sees two consecutive reasoning entries (the real reasoning + the web_search_call) and inserts a synthetic user separator between them → wire order user→reasoning→[user sep]→web_search_call → server rejects 'web_search_call provided without its required reasoning item'; the separator behavior for distinct runs is correct, but same-run reasoning+web_search_call must not be split", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7295, "title": "429 governor: free-tier Retry-After (~600s) freezes the whole opencode.ai credential across sessions, escalates to 424", "category": "429-governor-hold-scope", "root_cause": "with network.throttle429Scope:'credential' (default), a marker-less free-tier 429 on opencode.ai falls back to an origin+credential hold; createHoldState takes max(jittered backoff, Retry-After) so the ~600s Retry-After dominates (2s base / 30s throttle429MaxDelayMs cap never apply) → every session + every model on that credential gets the local bili_throttle_hold 429 for ~10 min; the streak only resets on a non-429 response, so messages tens of minutes apart still climb to the terminal 424", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7301, "title": "额度耗光后不再自动解开 upstream rate-limited (no longer auto-unblocks after quota recovers)", "category": "429-governor-hold-scope", "root_cause": "the 429 governor hold ('upstream rate-limited 2 consecutive request(s) to 127.0.0.1:3425; bili is holding outbound sends until … ~182s, exponential backoff') does not auto-resume when the upstream quota actually recovers — the hold/streak accounting (same #4204 governor as #7295) keeps the credential held past the real limit clearance", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "medium"},
      {"number": 7300, "title": "compress 空转:fail-streak resets on success → infinite blind-retry loop", "category": "compress-fail-streak-loop", "root_cause": "compressFailStreak only accumulates on consecutive failures and resets on any success, but a 'success' here reclaims only 86–290 tokens (a fragment barely over the 5000-char minCompressRange), so 'fail×N → success×1 → fail×N' loops forever and the circuit breaker never reaches COMPRESS_LOOP_THRESHOLD; compounded by the refusal text '(4545 chars … min 5000)' teaching the model to top up and retry, and minCompressRange=5000 being too strict for CJK; proposed: windowed call budget (2 calls/120s), imperative refusal text, optional threshold", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7197, "title": "Circuit breaker armed by requireRecommendation range misses: receipt forbids the only recovery the code still allows", "category": "compress-circuit-breaker-text-mismatch", "root_cause": "with compress.requireRecommendation:true, the scope gate charges the compress failure ledger (noteCompressLoopFailure) BEFORE returning, so three wrong range guesses arm the breaker at hard-coded COMPRESS_LOOP_THRESHOLD=3; the armed receipt (loopNoisePauseParagraph) tells the model to STOP calling compress and not poll acp_status — but the gate does not consult the breaker, so a subsequent in-recommendation range compresses normally; the text asserts a permanent state the code does not implement, and the stated disarm conditions are unreachable by an obedient model", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6877, "title": "Compression nudge and egress fuse disagree, leaving long session stuck on 413", "category": "egress-fuse-nudge-disagreement", "root_cause": "the compression nudge decision used a prior usage reading (437,971 tokens, growth 4,853 below the 22,500 floor) while the later measured-body egress fuse (src/server/relay.ts) estimated the outgoing Responses request at ~1,068,060 tokens (other ~1,045,895) over the 1,000,000 window → 413 request_too_large before forwarding; the two gates measure different things (stale usage vs serialized-body heuristic) and disagree, so short retries hit the same opaque 'other' bucket and the session is stuck", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7273, "title": "bili diagnostics report (v0.1.216) — thin pointer to an attachment", "category": "tinyfix", "root_cause": "user reports a DSH plugin error and attached a bili-issue-draft.md; the issue body itself carries no reproducible mechanism (just 'I got an error, DSH said it's valuable to report') — the actual defect is inside the attachment, not readable from the issue text", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "low", "notes": "body is a thin pointer to an attachment (bili-issue-draft.md); the real defect is inside the attachment, not classifiable from the issue text alone"},
      {"number": 5672, "title": "tool_search_call / tool_search_output missing from tool-pair ordering → upstream 400", "category": "responses-tool-ordering", "root_cause": "src/tool-pair-order.ts:13-14 CALL_TYPES/OUTPUT_TYPES only know function_call/custom_tool_call (+_output); Codex 0.166's tool_search_call/tool_search_output are never recognized as a call/output pair, so the #766 adjacency guarantee doesn't apply and a compressed request can reach the wire with a tool_search_output whose call is gone → DeepSeek 400 'No tool call found for tool output'; second gap: hoistTrappedToolItems only hoists non-tool items, never separates two interleaved call/output groups", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7277, "title": "长上下文会话下第 3 条大请求必然 503,且无法配置 (3rd large request always 503, not configurable)", "category": "request-body-budget-config", "root_cause": "src/request-body-budget.ts hard-codes MAX_LARGE_DECODED_REQUESTS=2 and LARGE_DECODED_ADMISSION_BYTES=32MiB; largeDecodedRequests is a per-process in-memory counter, so a 3rd in-flight large (>32MiB decoded) request immediately 503s request_body_busy regardless of provider/model/key (only a process restart clears it); the reporter frames it as a config/observability gap (make the two constants env-configurable + log the current count on refusal), not a logic bug", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5826, "title": "逐客户端调查并接入稳定'会话+消息ID'身份信道 (stable session+msgid identity channel per-client)", "category": "msgid-identity-channel", "root_cause": "umbrella to wire the explicit sessionid+msgid[] identity channel per client (server side already protocol-agnostic via /__bili/plugin/turn msgIds + full-resend biliMsgIds/biliLedgerSeqs, all-or-nothing validation in src/msgid-pin.ts); the remaining work is host-side per lane (dsh/omp/pi wired, opencode V2 structurally weak, hermes/claude-code have storage but no channel, codex deferred) — the v6 pre-gate rollout depends on this coverage; the root problem it addresses is identity-by-content-guessing causing false forks/orphans/resume failures", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7286, "title": "v1internal Antigravity lane preflightCompress summary request not wrapped in v1internal envelope", "category": "v1internal-preflight-envelope", "root_cause": "on the v1internal (Google Cloud Code) lane, preflightCompressIfNeeded (src/server/handle.ts:3403) does not pass v1Meta to requestSummary (src/preflight.ts:1552), so the summary request is sent as flat Gemini {contents, systemInstruction} instead of the {project, requestId, model, request:inner} envelope → endpoint 400 'Unknown name contents' → 502 preflight_compress_failed; second gap: googleChunkText (src/preflight.ts:920) + external-summary-http.ts completed('google') read top-level chunk.candidates without unwrapping the v1internal response envelope {response:{candidates}} → empty summary (0 chars)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7276, "title": "Billing price lookups rebuild every override row and lowercase index", "category": "billing-registry-lookup-perf", "root_cause": "every resolveBillingPrice call reconstructs the whole configured billing.prices table as registry cost rows, and each reconstruction gives scanRegistryCostRow a new table identity so its lowercase lookup index is rebuilt even on an exact override hit — N configured models pay for N complete conversions while enumerating; the conversion can happen once in setBillingOverrides (config loading already funnels every update there)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7278, "title": "[Drive-by] Registry cost-row index rebuilt per lookup on overrides-miss (single-slot cache thrash)", "category": "billing-registry-lookup-perf", "root_cause": "costLowerIndex (src/registry.ts:564) is a single-slot identity cache; an overrides-miss lookup scans the overrides table (slot←overrides index) then the registry table (slot←registry index, evicting the overrides index), so both indexes are rebuilt per lookup — pre-existing, NOT introduced by #7276; proposed: make costLowerIndex a small Map (cap 2) so both indexes are cached", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7284, "title": "Relay registry lookups scan every row for each model suffix", "category": "billing-registry-lookup-perf", "root_cause": "the registry caches a lowercase index per table, but unknown-relay lookups still scan every indexed row for every candidate model suffix (context/output ceiling + pricing for provider-prefixed keys); proposed: lazily build slash-boundary suffix buckets on the first relay fallback, retaining key order + case-collision rows, a ~0.4MB index; distinct from #7276/#7279 (repeated billing-row normalization) and #7278/#7282 (alternating table identities evicting an index)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7293, "title": "Diagnostics reads unused log generations even when the report has enough evidence", "category": "web-diagnostics-log-read", "root_cause": "buildDiagnostics reads every retained log generation for every registered topic; attribution and snapshot-shadow reports discard those log tails entirely, and storage/conflict reports need only the most recent 12/20 matching lines yet still scan older generations after the active log already supplies the full tail (8.4MB/5 files per topic in the fixture); fix: read no log files for topics that don't consume log evidence, and for tails visit generations newest-first and stop once the bounded tail is full", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"}
    ]
    
  20. ranxianglei commented on Nov 20, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    Root-cause classification — final, all 58 issues (batch order). Supersedes the two [progress] replies above; this is the complete array.

    Method: read every issue body, opened the linked upstream PR/diff where present, and inspected the named source paths in this workdir before judging. Categories are mechanism-based (stable across rounds); the reserved string tinyfix marks single small fixes with no shared mechanism. [PR] rows carry the upstream PR + CI + merge state; pure bug/feature/decision rows are pr_number:null / ci:none / merge:n/a.

    [
      {"number": 7347, "title": "full-resend watermark fastpath always misses on pinned sessions (id namespace mismatch)", "category": "watermark-fastpath-id-namespace", "root_cause": "src/server/fastpath.ts head check compares a.id!==b.id between fresh content-hash cores (h_) and post-pin host/ledger msgid cores — two id namespaces never match on pinned sessions so the fastpath always misses (pre-pin keeps content-hash ids and hits, which is why rollout-off unit tests stayed green)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7346, "title": "fastpath hit claims rawness → #3285 sweep retires ledger rows the tail window omits", "category": "pin-ledger-sweep-drain", "root_cause": "on fastpath HIT snapshotMessages(session,windowCores) claims rawness (lastMessagesFolded=false) so the #3285 authoritative-set sweep retires every ledger row the ~2-row tail window omits; plus hydrateSlimCcrStore returns null for an empty CCR envelope → false DIGEST DRIFT → completeLog on every save", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6785, "title": "[PR] unify msgid delivery with the snapshot plane (schema v12, shadow write, mapEgress)", "category": "msgid-identity-channel", "root_cause": "unifies msgid delivery (host msgid envelope→server pin→fork) with the snapshot plane (4-table schema v12, shadow write, mapEgress direct) so stage-0 consumes delivered host msgids; new storage.snapBase.* config (default OFF) + gate --snapbase-full; absorbs #6564/#5901/#6680", "ci_state": "red", "merge_state": "behind", "pr_number": 6785, "confidence": "high"},
      {"number": 7345, "title": "preflight fail-fast returns 502 'NOT forwarded' but the payload still reaches upstream", "category": "fail-fast-forward-bypass", "root_cause": "preflight fail-fast returns 502 'NOT forwarded' but the same-round payload still reaches upstream (the in-flight send is not cancelled, or a concurrent retry bypasses the gate) — the gate decision and actual egress are not atomically coupled; the log line is misleading", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "medium"},
      {"number": 7339, "title": "dead #4739 narrow-read chain has no callers", "category": "dead-code-narrow-read", "root_cause": "src/persist.ts:3686 narrowBlockContents + :3668 narrowMetadata (from 7b1ecc2b1fb/#4739) have no callers in src/ or tests/ — the narrow-read restore path was never wired; pure dead code", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7344, "title": "[PR] remove the dead #4739 narrow-read chain", "category": "dead-code-narrow-read", "root_cause": "removes the dead #4739 narrow-read chain (persist.ts lazySource/narrowMetadata/narrowBlockContents, shadow.ts getSessionMetadata + no-arg getBlockContents, store.ts no-arg getBlockContents/getMessagesAt/sessionViewRows/getSessionMetadata/viewPointerSeqs, backend.ts + mysql.ts sessionViewRows); no runtime path touched", "ci_state": "green", "merge_state": "behind", "pr_number": 7343, "confidence": "high"},
      {"number": 5330, "title": "[PR] sessions rm full VACUUM lacks the unifiedActiveWriter guard", "category": "storage-vacuum-foreign-writer", "root_cause": "runRm in src/storage-maintenance.ts lacked the store.unifiedActiveWriter() guard the other three full-VACUUM sites have — a live foreign writer lets a sessions rm full VACUUM win the exclusive lock and starve the co-resident instance for the whole O(db-size) rewrite; fix skips only the shrink up front with a pid-named note", "ci_state": "green", "merge_state": "clean", "pr_number": 5330, "confidence": "high"},
      {"number": 6037, "title": "upstream TCP connect timeout misrouted into the degenerate-turn fatal path", "category": "transport-misclassified-degenerate", "root_cause": "upstream TCP connect timeout + network.replayRetryMax(3) exhausted → the turn reaches the plugin pipe as a zero-output completed turn, classified 'degenerate terminal turn' (#732/#821), re-POSTed with a continuation nudge, then #870 fatal — a transport failure is misrouted into the degenerate path", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7343, "title": "per-line role labels in compressed-summary bodies are wrong", "category": "summary-role-label-mislabel", "root_cause": "per-line role labels in compressed-summary bodies are wrong (assistant thinking / tool-call JSON / system-injection rendered as [user]) — 45/4091 mislabeled + 157 refs that don't exist; the renderer inherits the block-head user role when merging consecutive assistant blocks, or the summary model mis-attributes the speaker", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "medium"},
      {"number": 4913, "title": "degenerate turn surviving one nudge always emits fatal (#870), no toggle", "category": "degenerate-turn-fatal-no-toggle", "root_cause": "plugin pipe: a degenerate turn surviving one continuation nudge always emits the fatal in-band error (#870), no toggle; the #2756 'bare' exemption is written below the fatal branch so it never executes on the retry pass (36% retry→hard-fail on 0.1.208)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7341, "title": "[PR] blog-refresh byte caps enforced only after arrayBuffer() buffers the whole body", "category": "web-blog-refresh-byte-cap", "root_cause": "src/web/blog-refresh.ts enforced the 512KiB/2MiB caps only after response.arrayBuffer() buffered the whole body (missing/understated Content-Length → over-buffer to EOF) + the contents-API fallback counted base64 padding+newlines as file bytes (rejecting valid 524287/524288-byte docs); fix reads chunks against the cap, cancels on overflow, and measures the whitespace-normalized decoded length", "ci_state": "green", "merge_state": "behind", "pr_number": 7341, "confidence": "high"},
      {"number": 7334, "title": "[PR] model-initiated compress accepted any fold net shrink ≥1 token", "category": "compress-fold-benefit-floor", "root_cause": "model-initiated compress in applyRanges accepted any fold net shrink ≥1 token (only the ~5000-char minCompressRangeChars floor) so a burst of tiny folds each re-bills the cache tail; adds compress.minFoldBenefitTokens (default 4096) gate below the pressure band + a model-facing refusal receipt teaching the model to batch ranges", "ci_state": "green", "merge_state": "behind", "pr_number": 7334, "confidence": "high"},
      {"number": 7342, "title": "blog-refresh plain-text/image paths buffer full body before the byte cap (same as #7341)", "category": "web-blog-refresh-byte-cap", "root_cause": "same mechanism as #7341 — src/web/blog-refresh.ts plain-text/image paths enforce the byte caps only after arrayBuffer() consumes the full response + the contents-API base64 estimate counts padding/newlines; fixing PR #7341 is in flight", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5835, "title": "[PR] wedged orphan proxy with dead owner held port+index.db lock for minutes", "category": "orphan-proxy-offloop-reap", "root_cause": "the in-loop parent watcher (2s ticks) only detects/reaps when the event loop yields; a wedged orphan proxy with a dead owner held the port + index.db lock for minutes; adds an off-loop owner-liveness probe (process.kill(pid,0)) in the #4762 worker that exits ~15-45s after confirming owner death ×3 ticks + loop silence ≥30s + zero watchers", "ci_state": "green", "merge_state": "clean", "pr_number": 5833, "confidence": "high"},
      {"number": 7338, "title": "[PR] appendBlockContents bare JSON.parse aborts the whole cold load on one corrupt row", "category": "storage-poison-row-isolation", "root_cause": "appendBlockContents (src/storage/ingest.ts:615) does a bare JSON.parse on the block summary/one CAS content, so one corrupt row threw out of the shared helper and aborted the whole cold load on both the full-view and resident-field paths; fix skips + returns the unparseable row into skippedCorruptRows (closes the last gap #7335/#7336 started)", "ci_state": "green", "merge_state": "clean", "pr_number": 7338, "confidence": "high"},
      {"number": 5628, "title": "[PR] opencode-lane http.request hook replaced e.request, breaking provider auth plugins", "category": "opencode-lane-hook-payload", "root_cause": "the opencode-lane http.request hook replaced e.request with a new Request at the proxy origin, but opencode hands that one mutable event to every later hook — provider auth plugins (Kiro, Antigravity) validating the upstream URL broke; fix makes the routes only discover the proxy and performs the /bili/ rewrite at fetch time in the shared native fetch chain", "ci_state": "green", "merge_state": "clean", "pr_number": 5626, "confidence": "high"},
      {"number": 7337, "title": "corrupt block-content row still aborts the whole session load (same as #7335/#7336)", "category": "storage-poison-row-isolation", "root_cause": "same mechanism as #7335/#7336, different carrier — appendBlockContents (src/storage/ingest.ts:615) bare JSON.parse on summary/one CAS content (vs the message content + block payloadJson which #7336 isolated); a corrupt block-content row still aborts the whole session load; fixing PR #7338 is in flight", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7335, "title": "putContent trusts caller precomputed hash + LegacyViewBuilder bare JSON.parse → full state wipe", "category": "storage-poison-row-isolation", "root_cause": "two structural gaps — (1) write: UnifiedStore.putContent (src/storage/store.ts) trusts the caller precomputed hash (const hash=precomputedHash ?? sha256Hex) so a hash≠bytes poison pair enters the CAS; (2) read: LegacyViewBuilder.addMessages/addBlocks (src/storage/ingest.ts) bare JSON.parse with no try/catch, one bad row throws through viewFor → legacy fallback fails → starting fresh → full state wipe (26 poison rows → 32 serial preflight rounds, 12+ min stall); fixing PR #7336", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5245, "title": "[PR] characterization tests pinning verifyAttachAndRecover behavior (Phase 0)", "category": "boot-contract-characterization", "root_cause": "characterization tests only (zero product code) pinning current verifyAttachAndRecover behavior across pi/omp/opencode/dsh (S1 healthy / S2 strict-old-version / S3 dead-target), one-shot latch permanence (stats.disabled 3-strike never-unlatches, logger stderrDead), and the /__bili/health key-set — building the equivalence net before extracting the shared lifecycle (#5241 Phase 0)", "ci_state": "green", "merge_state": "clean", "pr_number": 5242, "confidence": "high"},
      {"number": 5178, "title": "[PR] tag-echo filter: drifted ACP names + mid-turn draft envelopes leak", "category": "tag-echo-filter-drift", "root_cause": "two holes — (A) the open-side ACP_NAME_ALT enum (src/loop/tag-echo-filter.ts buildAcplikeName) has no tail-append arm so drifted names (acpt/acct/accent…) pass all three gates (RENDER_TAG_DETECT/BROAD_TAG_HEAD_TAIL/PARTIAL_TAIL); (B) the draft predicate DRAFT_CLOSE_TAIL (src/degenerate-turn.ts:78) is $-anchored so mid-turn draft envelopes leak; fix keys on two payload shapes (ref-in-attr empty body / emitter tokens= attr) instead of enumerating names", "ci_state": "green", "merge_state": "clean", "pr_number": 5178, "confidence": "high"},
      {"number": 7336, "title": "[PR] putContent recomputes sha256 + LegacyViewBuilder skips unparseable rows", "category": "storage-poison-row-isolation", "root_cause": "closes both #7335 gaps — putContent (store.ts + mysql.ts) always recomputes sha256 and rejects a mismatching caller digest (tx rollback, loud error); LegacyViewBuilder + shadow.ts skip unparseable rows into skippedCorruptRows with a loud warn; the sidecar verify-sample isolates bad rows as drift instead of failing every save", "ci_state": "green", "merge_state": "clean", "pr_number": 7336, "confidence": "high"},
      {"number": 5006, "title": "[PR] three doomed gates price only core bytes, not system/tools wire overhead", "category": "doom-gate-byte-floor-caliber", "root_cause": "three doomed gates (egress fuse src/server/relay.ts, preflight fail-fast probe #2313, post-compress recheck src/server.ts) priced only core message bytes against max(2MB, window×7.5B/token) while the preflight trigger prices messages+system/tools — a payload whose billable total exceeds every window cleared the floor on the residual caliber (26,903-msg session, 1.44MiB core text → 55.2MiB bodies read as 'small'); fix prices core text + system/tools wire overhead from the actual outbound body", "ci_state": "red", "merge_state": "dirty", "pr_number": 5006, "confidence": "high"},
      {"number": 917, "title": "empty message-board stub", "category": "tinyfix", "root_cause": "empty issue — no body, no mechanism, no reproducible defect; not part of any systemic root cause", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "low", "notes": "empty body"},
      {"number": 7333, "title": "model-initiated compress: one nudge → 39 tiny folds, each re-billing the cache tail", "category": "compress-fold-benefit-floor", "root_cause": "the model-initiated compress path (applyRanges) accepts any number of folds per nudge + any fold net shrink ≥1 token (only the ~5000-char minCompressRangeChars) so one nudge → 39 folds in one session, each re-billing the cache tail (619K uncached tokens); bili's own preflight/auto-fold have a 4096-token benefit gate but the model path doesn't; Part1 (min-benefit floor) = PR #7334, Part2 (per-nudge budget) pending decision", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 3011, "title": "DSH fork adoption 409 SNAPSHOT_UNAVAILABLE classified non-transient → child never inherits folds", "category": "fork-inheritance-lost", "root_cause": "DSH branch clone → child fork adoption gets 409 SNAPSHOT_UNAVAILABLE (parent raw pluginSnapshot not resident); createForkAdopter's transient regex (^(snapshot|fork) (error|http 5\\d\\d)) does NOT match 'snapshot http 409' so it is classified non-transient → done.add(child) → the child never inherits folds for the rest of the process → starts from the raw ~1.06M-token history → context_length_exceeded 400 on the first turn", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6124, "title": "DSH fork() omits parentSession → child inherits 1 of 33 blocks then permanently 400", "category": "fork-inheritance-lost", "root_cause": "DSH 0.2.2-rc.1 fork() writes only isSeeded:true, not parentSession (the dsh-subagent path does write it), so bili's forkChildVerdict returns parent-missing and the fork-adoption path is skipped at the entry; the resume-inheritance fallback uses planForkAdoption's every() all-or-nothing block gate and 4 of 5 active blocks were invalidated by a few image/reasoning messages that shouldn't count → 33 blocks → 1 block, 161,960 tok lost, child permanently 400", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6156, "title": "[PR] credential fingerprint covers all forwarded auth headers, order-independent", "category": "credential-fingerprint-order", "root_cause": "credentialFingerprint (src/cache-ledger.ts) broke on the FIRST recognized auth header in insertion order, so a lane forwarding two distinct credentials got an identity that was order-dependent (reordering flipped the digest → #4204 429 governor holds missed) and incomplete (second credential invisible → a different account inherited the first's hold); fix collects every recognized header, normalizes, dedupes, and hashes the sorted set (single-credential byte-identical)", "ci_state": "green", "merge_state": "dirty", "pr_number": 6154, "confidence": "high"},
      {"number": 5693, "title": "[PR] dsh long-session drift comprehensive remediation", "category": "dsh-long-session-drift", "root_cause": "four-part dsh long-session drift remediation — P0-1 directSendOrRefuse refuses marker-decisive dsh native compaction calls on dead-proxy direct-send paths (was: sent upstream → checkpoint durably destroyed the substrate #1729); P0-2 detectCompactionOverlayGap alarms for every profile shape (was: web-profile only); P1-1 personaConversationOrder onResolve makes fork/reanchor decisions visible; P1-2 reanchor instead of fork when a landed compaction truncated the head under a new persona tag", "ci_state": "green", "merge_state": "dirty", "pr_number": 5691, "confidence": "high"},
      {"number": 6859, "title": "[PR] preserve signed ordinary text parts on the wire and in tool-loop rebuilds", "category": "signed-block-wire-fidelity", "root_cause": "Gemini can attach a thoughtSignature to an ordinary text part (no thought:true); per KDD #10 a signed block must never be rewritten — but (1) the Google SSE adapter + non-streaming JSON rewriter only exempted thought:true parts from prose filtering, so a signed ordinary text part with marker-like text was mutated before reaching the client, and (2) the streaming loop recorded ordinary text as a merged string and dropped its signature on the next re-request; fix treats a signed text part like a thought part and carries the signature through the loop event + re-request", "ci_state": "green", "merge_state": "dirty", "pr_number": 6859, "confidence": "high"},
      {"number": 6363, "title": "[PR] suffix-align folded envelope tail in batched view verification", "category": "storage-digest-drift", "root_cause": "BatchedViewVerification.compareMessages (src/storage/shadow.ts) compared the envelope's bounded folded snapshot against the table's full raw tail positionally from index 0 — once the tail grew past the persistTailTokens cap, every slot mismatched → a permanent DIGEST DRIFT latch forcing an 8-9s full re-ingest on every save; fix aligns the comparison to the overlap window (the envelope is a suffix of the table tail)", "ci_state": "green", "merge_state": "dirty", "pr_number": 6363, "confidence": "high"},
      {"number": 6667, "title": "[PR] bound init-snapshot memory/time with a top-100 min-heap scan", "category": "init-snapshot-memory", "root_cause": "snapshotInit() in src/context-analysis.ts (advisory per-session observability) allocated one InitMessageEntry per message and sorted the complete array before retaining only 100 — O(n) temp memory + O(n log n) synchronous work on the first proxied request even though the visible sample is capped at INIT_PER_MSG_CAP=100; fix replaces it with a bounded top-K min-heap (1313ms/85MiB → 48ms/13MiB on 1M synthetic msgs), output byte-identical", "ci_state": "green", "merge_state": "dirty", "pr_number": 6667, "confidence": "high"},
      {"number": 5606, "title": "failed-compress recovery receipt leaves the re-request ending on an assistant turn → 400", "category": "failed-compress-receipt-assistant-tail", "root_cause": "in src/loop/core.ts the failed-only mutating round branch drops the failed compress/decompress calls and pushes the streamed markers as a final {role:'assistant'} receipt message with nothing after it — Anthropic rejects a conversation ending on an assistant message (trailing whitespace, or 'does not support assistant message prefill' on Claude 5) → 400, subagent dies; the path meant to keep the task alive (#5220) is the one that kills it", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7318, "title": "可折区间折叠提示,关不掉 (collapsible-interval fold nudge can't be turned off)", "category": "nudge-collapsible-interval-no-toggle", "root_cause": "the 'there's still a collapsible interval' nudge is pushed every turn even with compress.externalSummary.enabled=false + autoFold=false + autoFoldTargetTokens=250000 (config reload ok:true) — there is no independent config switch for this nudge, so it keeps inducing the model to fold against the user's explicit stop instruction, each fold re-billing the prefix cache (200+ folds in one night)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5737, "title": "Folded summaries reach the model as user (OpenAI) / system (Anthropic) messages", "category": "folded-summary-role-authority", "root_cause": "the folded-block summary's wire role is asymmetric by lane — kernel renderSummary (kernel/src/prune.ts:548-560) renders it role:system; the OpenAI/Responses egress (adapter-openai.ts:247, prepare-openai.ts:378, responses-tool-output.ts) re-voices it to user via systemToUser (so compressed history enters as a live user message, amplifying #4222 stale-objective re-execution); the Anthropic egress keeps it as a mid-conversation system message (system-level authority); constrained by KDD #7 (proxy-mode user re-voice is intentional for SGLang + prefix-cache head stability)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 4594, "title": "Codex WS context-only prewarm runs as an ACP turn: folds wiped, forged summaries re-injected", "category": "codex-prewarm-as-turn", "root_cause": "src/responses-ws.ts only treats generate:false with an EMPTY input as a prewarm probe; current codex sends its startup/idle warmup as generate:false PLUS developer/user context items, so it falls through into the ACP pipeline as an ordinary zero-message turn → syncBlocks on a view with no covered message deactivates every fold block (73→0, 45→0) and prepare-responses re-injects the entire codexForgedSummaries list (1,421 summaries / 1.2M tokens → preflight fail-fast 413); second gap: the probe relay calls the upstream transport without the relay's proxy dispatcher (ws-bridge only exposes a boolean isEgressProxied), so behind an upstream proxy every probe goes direct and fails", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7296, "title": "Codex Responses ordering inserts a user separator before a hosted web search call", "category": "responses-tool-ordering", "root_cause": "responsesToCore represents a hosted web_search_call as contentType:'reasoning', so repairResponsesAssistantOrdering (src/server/responses-ordering.ts) sees two consecutive reasoning entries (the real reasoning + the web_search_call) and inserts a synthetic user separator between them → wire order user→reasoning→[user sep]→web_search_call → server rejects 'web_search_call provided without its required reasoning item'; the separator behavior for distinct runs is correct, but same-run reasoning+web_search_call must not be split", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7295, "title": "429 governor: free-tier Retry-After (~600s) freezes the whole credential, escalates to 424", "category": "429-governor-hold-scope", "root_cause": "with network.throttle429Scope:'credential' (default), a marker-less free-tier 429 on opencode.ai falls back to an origin+credential hold; createHoldState takes max(jittered backoff, Retry-After) so the ~600s Retry-After dominates (2s base / 30s throttle429MaxDelayMs cap never apply) → every session + every model on that credential gets the local bili_throttle_hold 429 for ~10 min; the streak only resets on a non-429 response, so messages tens of minutes apart still climb to the terminal 424", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7301, "title": "额度耗光后不再自动解开 upstream rate-limited (no auto-unblock after quota recovers)", "category": "429-governor-hold-scope", "root_cause": "the 429 governor hold ('upstream rate-limited 2 consecutive request(s) to 127.0.0.1:3425; bili is holding outbound sends until … ~182s, exponential backoff') does not auto-resume when the upstream quota actually recovers — the hold/streak accounting (same #4204 governor as #7295) keeps the credential held past the real limit clearance", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "medium"},
      {"number": 7300, "title": "compress 空转:fail-streak resets on success → infinite blind-retry loop", "category": "compress-fail-streak-loop", "root_cause": "compressFailStreak only accumulates on consecutive failures and resets on any success, but a 'success' here reclaims only 86-290 tokens (a fragment barely over the 5000-char minCompressRange), so 'fail×N → success×1 → fail×N' loops forever and the circuit breaker never reaches COMPRESS_LOOP_THRESHOLD; compounded by the refusal text '(4545 chars … min 5000)' teaching the model to top up and retry, and minCompressRange=5000 being too strict for CJK; proposed: windowed call budget (2 calls/120s), imperative refusal text, optional threshold", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7197, "title": "Circuit breaker armed by requireRecommendation range misses: receipt forbids the only recovery the code still allows", "category": "compress-circuit-breaker-text-mismatch", "root_cause": "with compress.requireRecommendation:true, the scope gate charges the compress failure ledger (noteCompressLoopFailure) BEFORE returning, so three wrong range guesses arm the breaker at hard-coded COMPRESS_LOOP_THRESHOLD=3; the armed receipt (loopNoisePauseParagraph) tells the model to STOP calling compress and not poll acp_status — but the gate does not consult the breaker, so a subsequent in-recommendation range compresses normally; the text asserts a permanent state the code does not implement, and the stated disarm conditions are unreachable by an obedient model", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6877, "title": "Compression nudge and egress fuse disagree, leaving the long session stuck on 413", "category": "egress-fuse-nudge-disagreement", "root_cause": "the compression nudge decision used a prior usage reading (437,971 tokens, growth 4,853 below the 22,500 floor) while the later measured-body egress fuse (src/server/relay.ts) estimated the outgoing Responses request at ~1,068,060 tokens (other ~1,045,895) over the 1,000,000 window → 413 request_too_large before forwarding; the two gates measure different things (stale usage vs serialized-body heuristic) and disagree, so short retries hit the same opaque 'other' bucket and the session is stuck", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7273, "title": "bili diagnostics report (v0.1.216) — thin pointer to an attachment", "category": "tinyfix", "root_cause": "user reports a DSH plugin error and attached a bili-issue-draft.md; the issue body itself carries no reproducible mechanism (just 'I got an error, DSH said it's valuable to report') — the actual defect is inside the attachment, not readable from the issue text", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "low", "notes": "body is a thin pointer to an attachment (bili-issue-draft.md); the real defect is inside the attachment, not classifiable from the issue text alone"},
      {"number": 5672, "title": "tool_search_call / tool_search_output missing from tool-pair ordering → upstream 400", "category": "responses-tool-ordering", "root_cause": "src/tool-pair-order.ts:13-14 CALL_TYPES/OUTPUT_TYPES only know function_call/custom_tool_call (+_output); Codex 0.166's tool_search_call/tool_search_output are never recognized as a call/output pair, so the #766 adjacency guarantee doesn't apply and a compressed request can reach the wire with a tool_search_output whose call is gone → DeepSeek 400 'No tool call found for tool output'; second gap: hoistTrappedToolItems only hoists non-tool items, never separates two interleaved call/output groups", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7277, "title": "长上下文会话下第 3 条大请求必然 503,且无法配置 (3rd large request always 503, not configurable)", "category": "request-body-budget-config", "root_cause": "src/request-body-budget.ts hard-codes MAX_LARGE_DECODED_REQUESTS=2 and LARGE_DECODED_ADMISSION_BYTES=32MiB; largeDecodedRequests is a per-process in-memory counter, so a 3rd in-flight large (>32MiB decoded) request immediately 503s request_body_busy regardless of provider/model/key (only a process restart clears it); the reporter frames it as a config/observability gap (make the two constants env-configurable + log the current count on refusal), not a logic bug", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5826, "title": "逐客户端调查并接入稳定'会话+消息ID'身份信道 (stable session+msgid identity channel per-client)", "category": "msgid-identity-channel", "root_cause": "umbrella to wire the explicit sessionid+msgid[] identity channel per client (server side already protocol-agnostic via /__bili/plugin/turn msgIds + full-resend biliMsgIds/biliLedgerSeqs, all-or-nothing validation in src/msgid-pin.ts); the remaining work is host-side per lane (dsh/omp/pi wired, opencode V2 structurally weak, hermes/claude-code have storage but no channel, codex deferred) — the v6 pre-gate rollout depends on this coverage; the root problem it addresses is identity-by-content-guessing causing false forks/orphans/resume failures", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7286, "title": "v1internal Antigravity lane preflightCompress summary request not wrapped in the v1internal envelope", "category": "v1internal-preflight-envelope", "root_cause": "on the v1internal (Google Cloud Code) lane, preflightCompressIfNeeded (src/server/handle.ts:3403) does not pass v1Meta to requestSummary (src/preflight.ts:1552), so the summary request is sent as flat Gemini {contents, systemInstruction} instead of the {project, requestId, model, request:inner} envelope → endpoint 400 'Unknown name contents' → 502 preflight_compress_failed; second gap: googleChunkText (src/preflight.ts:920) + external-summary-http.ts completed('google') read top-level chunk.candidates without unwrapping the v1internal response envelope {response:{candidates}} → empty summary (0 chars)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7276, "title": "Billing price lookups rebuild every override row and the lowercase index", "category": "billing-registry-lookup-perf", "root_cause": "every resolveBillingPrice call reconstructs the whole configured billing.prices table as registry cost rows, and each reconstruction gives scanRegistryCostRow a new table identity so its lowercase lookup index is rebuilt even on an exact override hit — N configured models pay for N complete conversions while enumerating; the conversion can happen once in setBillingOverrides (config loading already funnels every update there)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7278, "title": "[Drive-by] Registry cost-row index rebuilt per lookup on overrides-miss (single-slot cache thrash)", "category": "billing-registry-lookup-perf", "root_cause": "costLowerIndex (src/registry.ts:564) is a single-slot identity cache; an overrides-miss lookup scans the overrides table (slot←overrides index) then the registry table (slot←registry index, evicting the overrides index), so both indexes are rebuilt per lookup — pre-existing, NOT introduced by #7276; proposed: make costLowerIndex a small Map (cap 2) so both indexes are cached", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7284, "title": "Relay registry lookups scan every row for each model suffix", "category": "billing-registry-lookup-perf", "root_cause": "the registry caches a lowercase index per table, but unknown-relay lookups still scan every indexed row for every candidate model suffix (context/output ceiling + pricing for provider-prefixed keys); proposed: lazily build slash-boundary suffix buckets on the first relay fallback, retaining key order + case-collision rows, a ~0.4MB index; distinct from #7276/#7279 (repeated billing-row normalization) and #7278/#7282 (alternating table identities evicting an index)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7293, "title": "Diagnostics reads unused log generations even when the report has enough evidence", "category": "web-diagnostics-log-read", "root_cause": "buildDiagnostics reads every retained log generation for every registered topic; attribution and snapshot-shadow reports discard those log tails entirely, and storage/conflict reports need only the most recent 12/20 matching lines yet still scan older generations after the active log already supplies the full tail (8.4MB/5 files per topic in the fixture); fix: read no log files for topics that don't consume log evidence, and for tails visit generations newest-first and stop once the bounded tail is full", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7297, "title": "Session GC treats a failed companion stat as absence and can delete originals", "category": "session-gc-companion-stat", "root_cause": "the opt-in session-file GC (BILI_SESSION_GC=1) catches every failure of stat(companion) as if no companion exists (EIO/EACCES/ELOOP/directory all treated as absent), so the session passes its size gate without accounting for the original tool-output bytes in its .content-store.json; after deleting the session, the orphan pass can delete that same companion if the metadata failure was transient — defeats the #1180/#1189 retain-unreadable-data contract; only ENOENT should establish an absent companion", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7331, "title": "kernel authorization request: PR #5877 (fix #5864: 400 'final block cannot be thinking' after compress)", "category": "thinking-final-block-400", "root_cause": "kernel authorization request (owner approval pending) for PR #5877 — the 400 'final block in an assistant message cannot be thinking' after a [thinking, compress] turn is caused by two each-correct-but-together-illegal judgments: planReasoningDrop keeps signed thinking (#1960) × kernel hideConsumedCompressCalls unconditionally hides the consumed compress pair → the pair is hidden but the signed thinking is left → the message collapses to [thinking]; decision item 1 = approve/reject #5877 (CI green, 3 review rounds clean); item 2 = a second producer of the same 400 in master's healUnpairedServerToolBlocks (kernel/src/wire/anthropic.ts) drops orphan server-tool halves leaving [thinking]", "ci_state": "none", "merge_state": "n/a", "pr_number": 5877, "confidence": "high", "notes": "decision request (owner approval pending), not a fixable defect itself; pr_number 5877 is the PR awaiting approval"},
      {"number": 5443, "title": "kernel authorization request: heal must enforce the FOLLOW side of Anthropic's mid-conversation system position rule (#5432)", "category": "anthropic-system-position-heal", "root_cause": "kernel authorization request (owner approval pending) — Anthropic's mid-conversation system position rule is two-sided (must FOLLOW a user message AND PRECEDE an assistant message / be at array end), but kernel healIllegalSystemPositions (kernel/src/wire/anthropic.ts:403) only checks the follower side; after a fold covers the user turn before a sidecar system, the system dangles right after an assistant → 400 'role system must follow a user message'; fix adds the predecessor check + an insertBeforeTrailingSystemRun helper; host side can't solve it (rule ownership is the kernel wire encoder, KDD #2)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high", "notes": "decision request (owner approval A/B pending); no PR yet (implementation blocked until approval)"},
      {"number": 7299, "title": "compress accepted a range containing the user message the turn was answering; next step answered a fabricated user message", "category": "compress-range-includes-current-user-msg", "root_cause": "a model compress call whose endId is past the user message the turn is answering succeeds (nothing refuses folding the newest user msg), so after the fold the prompt no longer ends with that user message and the model reconstructs a fabricated multi-sentence user message from the summary and answers it — a silent post-fold tail problem (no trailing user message); suggested guard: compress refuses/warns when endId >= the newest user message ref", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5866, "title": "400 'final block in an assistant message cannot be thinking' after a compress call that follows a thinking block", "category": "thinking-final-block-400", "root_cause": "Anthropic rejects the request right after the model calls compress when that turn also had a thinking block ([thinking, tool_use:compress]) — bili drops the executed compress tool_use + result from the forwarded history but keeps the thinking block, so the assistant message ends with thinking; 5/27 such turns failed in the worst session; fixing PR #5877 (kernel, awaiting owner approval via #7331)", "ci_state": "none", "merge_state": "n/a", "pr_number": 5877, "confidence": "high"},
      {"number": 7289, "title": "File-backed dashboard summaries can prefer stale copies of a session", "category": "web-dashboard-stale-summary", "root_cause": "the file-backed web summary index's refreshIndex selects the newest copy by savedAt (mtime tiebreak) and records it in byId, but mergedSummaries ignores that choice and inserts every file into a map keyed by session ID — a stale copy visited later in the directory walk overwrites the current summary, so the sessions list / filtered search / overview report old titles, request counts, token totals; fix consumes the existing selected-file index instead of re-selecting by walk order", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7291, "title": "Disk-only web sessions keep stale dollar estimates after pricing changes", "category": "billing-disk-summary-stale", "root_cause": "the web dashboard promises dollar estimates from currently-resolved prices, but the legacy/shadow disk summary cache is invalidated only by file mtime + size — editing billing.prices or refreshing the registry updates live sessions while disk-only sessions keep stale estimated spend/savings/payback/rates until their files change or the process restarts; fix includes the resolved four-tier price in the cached summary's validity", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5591, "title": "Codex Desktop: snapshot-shadow lost-tail and unified shadow state.messageRefs DIGEST DRIFT on 0.1.212", "category": "storage-digest-drift", "root_cause": "in a long Codex Desktop conversation the proxy recorded snapshot-shadow mechanical mismatches (lost-tail:6/16/144/11) then 'unified shadow DIGEST DRIFT … diverged: state; state: messageRefs' → dropped the ingest watermark, next save re-ingests the full tail; related to #5106 but identifies state.messageRefs + snapshot-shadow lost-tail on the Codex host; exact reproducer and underlying cause unknown (conversation remains usable, no data loss established)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "low", "notes": "underlying cause unknown per the report; classified by symptom/mechanism family"}
    ]
    

    Notes: 58/58 covered (none dropped). 17 [PR] rows carry upstream PR + CI + merge state (2 red: #6785, #5006). 2 decision-request rows (#7331, #5443) are owner-approval gates, not fixable defects — flagged in notes. 4 low-confidence rows (#917 empty, #7273 attachment-only, #5591 unknown cause, plus #7301 medium) are best-effort.

  21. ranxianglei commented on Nov 20, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    Root-cause classification — incremental round, 36 open issues.

    Plan:

    Analysis-only lane: no PRs/issues/labels/edits. Progress at 25, then final with all 36 in batch order.

  22. ranxianglei commented on Nov 20, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [progress] resolving upstream PR numbers for the 14 [PR] rows (6 have mirror number ≠ upstream GitHub PR number — mapping them now), then classifying all 36. classifier-json follows with the 25-mark and the final.

  23. ranxianglei commented on Nov 20, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    Root-cause classification — incremental round, all 36 issues. Read every body, resolved the upstream PR number for each of the 14 [PR] rows (6 had mirror number = upstream+1, found via GitHub title search), pulled CI + merge state, and inspected the named source paths. Categories reuse the prior round's mechanism strings where the root cause matches; new mechanisms get new stable strings.

    [
      {"number": 4595, "title": "[PR] route codex context-only prewarms to the probe relay via the upstream proxy", "category": "codex-prewarm-as-turn", "root_cause": "src/responses-ws.ts only treated generate:false with EMPTY input as a prewarm probe; codex's startup/idle warmup (PrewarmInput::Base: context items + generate:false) became a zero-message ACP turn that deactivated every fold, re-injected the whole codexForgedSummaries list, and added ACP tools; fix routes context-only generate:false prewarms verbatim on the probe relay and passes the relay's upstream proxy (egressProxyFor) instead of the boolean isEgressProxied", "ci_state": "green", "merge_state": "clean", "pr_number": 4595, "confidence": "high"},
      {"number": 7411, "title": "[PR] ignore non-model top-level dsh_* extension fields in raw body token estimates", "category": "raw-body-estimator-nonmodel-fields", "root_cause": "estimateRawBodyTokens/estimateRawBodyBreakdown (src/preflight.ts) only checked NON_TEXT_BODY_KEYS at string leaves, so they traversed the transport-only top-level dsh_session_log.events / dsh_plugin_packages (0 model tokens, never sent to the model) and billed them to 'other', tripping the 1,000,000-token #554 egress window fuse (HTTP 413 egress_window_fuse) before the request could reach DeepSeek; fix adds NON_MODEL_TOP_LEVEL_KEYS and skips matching top-level keys on non-array plain-object bodies while still counting nested messages[] content", "ci_state": "green", "merge_state": "behind", "pr_number": 7411, "confidence": "high"},
      {"number": 7412, "title": "[Bug][dsh] degenerate retry concatenates two thinking blocks into one message; still degenerates (41%)", "category": "degenerate-retry-block-concat", "root_cause": "retryEmptyTurn() (dist/index.js:1286) does not discard the first attempt's already-appended reasoning blocks — it only clears streams and sets retryIndexOffset=blocksForwarded so the retry's blocks continue numbering — so two thinking blocks land side-by-side in one assistant message; and degenerateRetried gives only one chance, so a second 'thought then stopped with no prose' turn hard-fails the whole round (19/46 = 41% retry-still-degenerate in one session)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6785, "title": "[PR] msgid delivery x snapshot plane: unified merge (absorbing #6564 + #5901 + #6680)", "category": "msgid-identity-channel", "root_cause": "unifies msgid delivery (host msgid envelope -> server pin -> fork) with the snapshot plane (4-table schema v12, shadow write, mapEgress direct) so stage-0 consumes delivered host msgids; new storage.snapBase.* config (default OFF, config-file only) + gate --snapbase-full; absorbs #6564/#5901/#6680", "ci_state": "red", "merge_state": "behind", "pr_number": 6785, "confidence": "high"},
      {"number": 7410, "title": "[PR] storage boot line states the actual selection and writer state (#7408)", "category": "tinyfix", "root_cause": "the full-mode [storage] boot line in src/storage/shadow.ts still carried the pre-#6480 wording ('legacy files keep being written - unset the knob to roll back'), but post-#6480 the pure-sqlite default resolves to full with the JSON envelope writer off, so all three clauses were false and the line contradicted the accurate persist.ts line in the same boot; fix branches the line on the raw selection (unifiedSelectionIsSqlite) and the effective writer state (effectiveJsonWriterOff) — a single log-line accuracy fix with no shared systemic mechanism", "ci_state": "green", "merge_state": "behind", "pr_number": 7409, "confidence": "high"},
      {"number": 7409, "title": "[0.1.216 verification datapoint][dsh] 25.9 MiB session still full-resends (15601/16876 covered ids verbatim, full ingest 8.9s)", "category": "dsh-long-session-no-convergence", "root_cause": "verification datapoint confirming the #5796 O(N) trap persists on 0.1.216 (which already contains #5715/#5755/#6480): one 25.9 MiB dsh session re-sends its entire history every turn, [fold-reconcile] gives up citing the closed #5715 while 15601/16876 hard-covered ids still ride verbatim, and the delta fast path falls back to a synchronous 8.9s full ingest that blocks the event loop — same no-convergence mechanism as #5796, not a new root cause", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7383, "title": "Large session preflight fail-fast 413: no fallback path, client cannot see the 14 MiB actually sent", "category": "preflight-failfast-no-fallback", "root_cause": "when a long session's full history (7,663 msgs / 14.1 MiB inbound) can no longer be folded under the window, bili emits a fail-fast 413 (retryable=false) without first trying the existing deterministic emergency-shrink path (drop oldest tool results / truncate long text), and the client's meter shows the post-fold upstream usage (~10%) not the real 14.1 MiB wire body, with the 413 as plain text (no structured size field) — no graceful degradation + no body-size observability; distinct from #7271 (fork inherits messages but not folds) though both share fold-state-not-reused-across-identity-change", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6175, "title": "[Codex CLI][v0.1.213] slow-ingest delta still exhibits a 10.3s synchronous storage spike", "category": "storage-slow-ingest-sync-spike", "root_cause": "follow-up verification of #5744/#5756: the delta ingest path still does O(history) synchronous work (slow-ingest path=delta took=10342ms on a 65.4 MiB WAL, plus 5.3s/5.1s full-path spikes) that blocks the event loop; the report asks to decompose slow-ingest into serialize/diff/table-IO/commit/checkpoint timings and verify the delta path is not re-scanning history, without asserting the old root cause regressed", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7254, "title": "[Root Cause] fold-reconcile (2 items) — auto-maintained aggregation board", "category": "fold-reconcile-content-hash-drift", "root_cause": "root-cause aggregation board (rcagg:fold-reconcile) for two members (#6966 pi inline rendered-summary mismatch, #7247 content-hash anchor): fold-reconcile anchors covered ids by content hash, but the host rewrites the message content (e.g. the 'Compressed conversation section' summary text), so the content-hash anchor no longer matches and reconcile drifts; the board asks for ONE root-cause fix PR covering both members", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5796, "title": "[feat][dsh] Long sessions have no convergence path: compaction gated off on both sides, incremental lane full-resends", "category": "dsh-long-session-no-convergence", "root_cause": "in DSH+bili a session that grows to 20-30k messages enters an O(N) zone with no exit: host-side compaction is disabled (profile compaction/compaction-basic/command-compact/tool-result-pruner all disabled:true) AND bili's dsh-compaction-gate re-writes compaction lines to disabled whenever bili is active; proxy-side folds only compress outbound (host surface stays at 11M tokens); the incremental lane (pluginIncrementalTurns) is 0 and the OpenAI wire shape falls back to full-resend (wireMessagesLookAnthropic rejects tool_calls/role:tool) — so every request rebuilds+uploads 29.6k msgs / 51.2 MiB", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 6037, "title": "[Bug] transport connect-timeout misclassified as a degenerate turn -> fatal", "category": "transport-misclassified-degenerate", "root_cause": "upstream TCP connect timeout + network.replayRetryMax(3) exhausted -> the turn reaches the plugin pipe as a zero-output completed turn, classified 'degenerate terminal turn' (#732/#821), re-POSTed with a continuation nudge, then #870 fatal — a transport failure is misrouted into the degenerate path (the nudge re-POST cannot repair a dead connection)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7408, "title": "[Bug] Timed-out DNS fallback probes leave their workers alive", "category": "dns-fallback-worker-leak", "root_cause": "the tunnel admission DNS fallback has an inner 2.5s lookup deadline and an outer 3s worker deadline; when the inner timer posts ETIMEDOUT the parent rejects the admission probe and clears the outer timer but does not stop the worker, so the expired one-shot worker and its queued getaddrinfo remain alive until unrelated native work completes (reproducible with UV_THREADPOOL_SIZE=1 + a blocked FIFO); fix requests the same worker termination the outer deadline uses once the inner deadline settles", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7301, "title": "After the quota is exhausted, it no longer automatically unblocks the upstream rate-limited hold", "category": "429-governor-hold-scope", "root_cause": "the 429 governor hold ('upstream rate-limited 2 consecutive request(s) to 127.0.0.1:3425; bili is holding outbound sends until ... ~182s, exponential backoff') does not auto-resume when the upstream quota actually recovers — the hold/streak accounting (same #4204 governor as #7295) keeps the credential held past the real limit clearance", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "medium"},
      {"number": 7396, "title": "Request path stalls on models.dev: contextFromRegistry() refetches after every failed load, trickling response outlives the 15s watchdog", "category": "registry-request-path-refetch", "root_cause": "server/handle.ts:1020 awaits contextFromRegistry() on the request path for a model the warm cache misses; loadRegistry()'s guard (src/registry.ts:228) resets cache to the bundled snapshot after a failed fetch so every such request pays for a new models.dev fetch, and the 15s budget is an idle watchdog (src/fetch-util.ts:234-244) that a trickling response (8.3s to first byte then ~11 B/s) outlives", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7406, "title": "[Bug] Fake-completion buffering silently truncates successful responses", "category": "fake-completion-buffer-truncation", "root_cause": "with the opt-in fake-completion retry enabled, resolveFakeCompletion uses the error-body reader which drains the entire stream but retains only its capped prefix (fakeCompletion.bufCapBytes, 16 MiB default), then presents that prefix as the complete answer — so a successful response larger than the cap loses its tail (text, real tool calls, usage, terminal events; a JSON success becomes unparseable) in both proxy and plugin modes, initial and corrective; the cap should bound inspection, not discard a successful response", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7407, "title": "[PR] preserve successful responses beyond the fake-completion buffer cap", "category": "fake-completion-buffer-truncation", "root_cause": "the opt-in fake-completion path treated a capped error-body buffer as a complete successful response (drains bytes beyond the inspection cap, forwards only the prefix); fix buffers for inspection only while the complete response fits the cap, on overflow replays all retained bytes and continues reading the original stream through the existing response pipes, and classifies/updates the fake-completion streak only when the whole response was inspected (exactly-at-cap still gets the existing check)", "ci_state": "green", "merge_state": "behind", "pr_number": 7406, "confidence": "high"},
      {"number": 7405, "title": "[Bug] Long-session SVG reports exceed the function-argument limit", "category": "svg-report-arg-limit", "root_cause": "the SVG trajectory builder (src/web/report.ts) uses Math.max.apply(null, allSamples), which the runtime treats as every retained sample being a separate call argument, so a valid 200,000-sample history throws RangeError: Maximum call stack size exceeded before the report can be downloaded; the same report's recovery chart has three more extrema calls that expand its full sample window into function arguments", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7404, "title": "[PR] fix(web): keep long-session reports within the argument limit", "category": "svg-report-arg-limit", "root_cause": "the SVG report passed every trajectory sample to Math.max.apply (plus three recovery-chart extrema calls expanding the full sample window into function arguments), so a valid 200,000-sample history threw RangeError before download; fix computes the trajectory maximum and recovery-window extrema with reductions, keeping every plotted point and the existing axis/outlier/recovery semantics", "ci_state": "green", "merge_state": "behind", "pr_number": 7404, "confidence": "high"},
      {"number": 7343, "title": "The role labels in the compressed summary are mislabeled (model attributed my own words to the user)", "category": "summary-role-label-mislabel", "root_cause": "per-line role labels in compressed-summary bodies are wrong (assistant thinking / tool-call JSON / system-injection rendered as [user]) — 45/4091 mislabeled + 157 refs that don't exist; the renderer inherits the block-head user role when merging consecutive assistant blocks, or the summary model mis-attributes the speaker (the 157 phantom refs support the latter)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "medium"},
      {"number": 7402, "title": "[Bug] Synthesized Responses text events lose content_index and crash SDK consumers", "category": "responses-content-index-loss", "root_cause": "several Responses SSE producers (the proxy adapter's synthesized text + ACP marker lifecycles, the opt-in legacy stream-error completion, and proxy/plugin text filters flushing a buffered tail) retain the item/output address but omit the content-part address (content_index), so the official OpenAI Node SDK's output.content[event.content_index] gets undefined and finalResponse() rejects with 'missing content at index undefined' — even for an ordinary answer ending in 'Length: 10 m' whose final 'm' is flushed in a freshly constructed delta", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7403, "title": "[PR] fix(responses): preserve content indices in synthesized streams", "category": "responses-content-index-loss", "root_cause": "Responses text events are addressed by both output_index and content_index, but several synthesized frames (generated text/ACP markers, legacy stream-error completions, filtered text tails in proxy+plugin modes) retained the item address but omitted the part address, so the OpenAI SDK threw 'missing content at index undefined'; fix sets content_index:0 on the four content/text lifecycle frames for single-part messages and preserves the source content index in proxy/plugin tail metadata including plugin retry reframing", "ci_state": "green", "merge_state": "behind", "pr_number": 7402, "confidence": "high"},
      {"number": 7286, "title": "[bug][v1internal] Antigravity/CloudCode pre-compression summary request not wrapped in v1internal envelope", "category": "v1internal-preflight-envelope", "root_cause": "on the v1internal (Google Cloud Code) lane, preflightCompressIfNeeded (src/server/handle.ts:3403) does not pass v1Meta to requestSummary (src/preflight.ts:1552), so the summary request is sent as flat Gemini {contents, systemInstruction} instead of the {project, requestId, model, request:inner} envelope -> endpoint 400 'Unknown name contents' -> 502 preflight_compress_failed; second gap: googleChunkText (src/preflight.ts:920) + external-summary-http.ts completed('google') read top-level chunk.candidates without unwrapping the v1internal response envelope {response:{candidates}} -> empty summary", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7395, "title": "[PR] fix(registry): bound the request-path wait on models.dev and back off after a failed fetch", "category": "registry-request-path-refetch", "root_cause": "server/handle.ts awaits contextFromRegistry() on the request path for a model the warm cache misses; after a failed fetch the cache resets to the bundled snapshot so loadRegistry() refetches on the next call (every such request pays for a new fetch), and the fetch watchdog is an idle timer a trickling response outlives; fix: contextFromRegistry() backs off for 10 min after a failed fetch (serves the fallback) and otherwise races the load against a 3s budget, answering from cache when the budget runs out while the load keeps warming the cache (loadRegistry's per-call retry and the boot pre-warm are untouched)", "ci_state": "green", "merge_state": "behind", "pr_number": 7395, "confidence": "high"},
      {"number": 7398, "title": "[Ideas] outbound-state table proposal: operation log + replay + compaction (replace per-turn full recompute)", "category": "outbound-view-full-recompute", "root_cause": "design proposal (owner-aligned) targeting the per-turn full recompute of the outbound view in large sessions (#6785 background: 100k msgs, ~5.5s/turn just for kernel render) — proposes an operation-log table (base/+I insert/-D delete-range/+U update, anchored by message id) with deterministic replay and post-fold compaction, so steady-state outbound rebuild is O(tail) not O(full); a design/discussion, not a bug", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7400, "title": "[Bug] Repeated failures overwrite earlier incident dump files", "category": "incident-dump-name-collision", "root_cause": "the incident writers in src/error-dump.ts name files with only the current millisecond + sanitized session ID + HTTP status, then use an ordinary overwrite-capable write, so two rejected bodies or summary exchanges with the same name silently replace the earlier evidence — bypassing the retention-GC protection these high-value failure files deliberately have; distinct raw session IDs that sanitize identically, or an existing incident from another process, also collide", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7401, "title": "[PR] fix(diagnostics): preserve colliding incident dumps", "category": "incident-dump-name-collision", "root_cause": "main-path and preflight-summary incident files (src/error-dump.ts) used only Date.now() + sanitized session ID + status as their name with an ordinary writeFileSync, so a collision silently overwrote the earlier incident (losing evidence even though err-*/summary-err-* are protected from retention GC); fix makes both writers use exclusive file creation, retaining the historical filename when available and only on EEXIST selecting a fresh UUID-suffixed path (descriptive session portion bounded so the suffix fits)", "ci_state": "green", "merge_state": "behind", "pr_number": 7400, "confidence": "high"},
      {"number": 7374, "title": "[Drive-by] #7346 regression test simulates the fix's effect but does not exercise the prepare-openai fix line", "category": "watermark-hit-test-coverage", "root_cause": "tests/issue7346-pf-window-ledger.test.ts pins the #3285 sweep/no-sweep contract by calling shadow.record directly with the two rawTail shapes, so it simulates the fix's downstream effect (rawTail=undefined => no sweep) but does not exercise the prepare-openai.ts fix line (if wpHit !== undefined then session.lastMessagesFolded = true) — the prepare -> persist link (HIT turn => lastMessagesFolded=true => persist.ts:3081 omits the raw tail) is untested and would not fail if that line were removed", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7399, "title": "[Bug] Diagnostic loopback exemptions expose 127-prefixed private DNS names", "category": "loopback-host-mask-prefix", "root_cause": "the diagnostic visibility exemption in src/log-mask.ts treats every string beginning with '127.' as an IP loopback address, so a DNS name like 127.customer.internal bypasses the normal private-host mask (isLoopbackHost returns true, maskUrlForDiagnostics returns it unchanged, and it leaks through host-event files and structured upstream-failure responses); the same endpoint-display boundary also mistakes the final colon in a portless [::1] for a port delimiter, producing '<private-host>:1]'", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7397, "title": "[PR] fix: validate diagnostic loopback host exemptions", "category": "loopback-host-mask-prefix", "root_cause": "the loopback display exemption accepted ip.startsWith('127.'), which also matched private DNS names like 127.customer.internal, leaking them through diagnostic URLs, host-event files, and structured upstream-failure responses despite host masking being enabled; plus a portless [::1] IPv6 address was split at its final colon as if it were a port delimiter; fix requires a real IPv4 loopback literal (canonicalizing Node numeric aliases like 127.1/127.0x1 to a valid 127/8) and splits bracketed display targets after the closing bracket", "ci_state": "green", "merge_state": "behind", "pr_number": 7397, "confidence": "high"},
      {"number": 7390, "title": "[PR] fix: route oversized kv values to the content CAS instead of throwing", "category": "kv-oversized-value-throw", "root_cause": "setKv threw when a value exceeded the 4 MiB inline extension-value cap (EXT_VALUE_MAX_BYTES), so the session-order skeleton (~4-8 MiB) exceeded the cap and threw on every start, forcing the hydrate onto the legacy file path for the process lifetime (the error message itself said 'store large payloads in the content CAS instead'); fix routes over-cap setKv values to the content CAS (kv row stores a small {\"$biliCas\":\"<sha256>\"} ref, payload in the refcount-managed content table; getKv/listKv resolve transparently, deleteKv releases the ref); setSessionExt stays inline", "ci_state": "green", "merge_state": "behind", "pr_number": 7389, "confidence": "high"},
      {"number": 7388, "title": "[Drive-by] 418-second preflight folding for 269 intervals is an independent fault mode", "category": "preflight-per-interval-summary-call", "root_cause": "a 14.1 MiB / 7663-message / 269-foldable-interval payload took 418s of local processing (src/server/request-cost.ts:58 logRequestCost) — suspected preflightCompress makes one external summary LLM call per foldable interval (269 calls accumulating to ~7 min) with no parallelism/rate-limit/cache; 来源 #7383, an independent performance fault mode to investigate separately from the #7383 fail-fast 413", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7391, "title": "[PR] fix(web): release unread changelog + blog-refresh error response bodies", "category": "unread-response-body-leak", "root_cause": "src/web/changelog.ts::runCheck rejects non-2xx statuses before consuming the body, then its finally stops the idle watchdog assuming the body was fully consumed — but no code cancels the unread response, so the stopped watchdog can no longer reclaim a connection whose error body never ends (repeated failed refreshes retain response resources until the peer/runtime closes them); the PR covers the full fetch-util #552 defect class across changelog.ts + the five blog-refresh.ts fetch helpers (getJsonList/fetchPlainText/fetchApiText/fetchPlainImage/fetchApiImage) with a releaseAndStop helper that cancels an unused body before stopping the watchdog", "ci_state": "green", "merge_state": "dirty", "pr_number": 7390, "confidence": "high"},
      {"number": 7389, "title": "[Drive-by] Fold state is not persisted across truncation; fold rebuilt from scratch each round", "category": "fold-state-no-cross-truncation-persistence", "root_cause": "after the client does compaction/truncation, every request's lineage is 'truncated of ...' so bili treats it as a new session and rebuilds fold from scratch each round (fold state is per-session, not persisted across the identity change) — even with an identical prefix the fold is fully recomputed and the preflight cost is re-paid every turn (31 min / 3 hr / EMPTY_RESPONSE observed); 来源 #7383, the largest-leverage but largest-effort fix (persist/reuse fold across truncation/fork), same general family as fork-inheritance-lost", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7387, "title": "[PR] fix: forward declared dsh compaction on sessions with no fold blocks", "category": "dsh-declared-compaction-no-substrate-422", "root_cause": "the declaredDshCompaction branch (added in #6880) at src/server/handle.ts refused unconditionally (422), but a plugin background session (dsh-mneme dream) declares purpose=compaction on a session with one message and zero fold blocks — no compression substrate to shadow, so the #1729 guard does not apply and the authored-compaction path (which needs >=1 active fold block) can never answer; fix requires an active fold block for the declared branch (guardApplies = isDshCompactionCall || (declaredDshCompaction && hasActiveFoldBlock)) while the content-marker branch stays unconditional", "ci_state": "green", "merge_state": "behind", "pr_number": 7386, "confidence": "high"},
      {"number": 7392, "title": "[Bug] URL masking blocks processing of long unbroken text", "category": "url-mask-overscan-quadratic", "root_cause": "the shared URL matcher in src/log-mask.ts retries an overlapping scheme suffix from every letter when a long run contains no complete URL (O(n^2)), so all three text maskers (maskUrlsInText/maskUrlsForPublish/maskUrlsForDiagnostics) and buildAnonymousProvenance (which masks the full first user message before selecting its 160-char preview) synchronously block — 384ms median on a 40k-char scheme-like run before the fix", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 7393, "title": "[PR] fix: avoid repeated scheme scans in URL masking", "category": "url-mask-overscan-quadratic", "root_cause": "the shared URL regex in src/log-mask.ts restarted its scheme scan at every letter of a long unbroken non-URL run (overlapping work), blocking logging/diagnostics and anonymous-session provenance (the complete first message is masked before its 160-char preview); fix starts matching once per contiguous scheme-character run, preserving leading numeric/punctuation outside the URL and using the existing single-URL maskers unchanged (384.51ms -> 0.61ms on a 40k-char run; 120,720 before/after comparisons with zero byte differences)", "ci_state": "green", "merge_state": "behind", "pr_number": 7393, "confidence": "high"}
    ]
    

    Notes: 36/36 covered (none dropped). 14 [PR] rows carry the resolved upstream PR number + CI + merge state (all green except #6785 red/1-fail; #7391 dirty, rest behind; #4595 mergeability still unknown on GitHub → marked clean). 5 recurring from the prior round (#6785, #6037, #7301, #7343, #7286) reuse their prior category; #6785 re-verified still red/behind. #7410 classified tinyfix (single log-line accuracy fix, no shared mechanism). Several [PR] bodies cite a "Fixes #N" that is a GitHub number offset from the mirror number (e.g. #7407 body says "Fixes #7405" but mirror #7405 is the SVG-report bug) — I classified each by its actual content, not the cross-reference.

  24. ranxianglei commented on Nov 21, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    [bot] 🏷 Round 3 plan — 17 issues (#7179, #7217, #7422, #4768, #4902, #7421, #4838, #4937, #5006, #5119, #7420, #5235, #5427, #5255, #5336, #5384, #7419).

    Plan: fetch every body via the mirror API, resolve the upstream PR number for each [PR] row (GitHub title search, since mirror number ≠ upstream for ~half), pull CI + merge state, and inspect the named source paths in this workdir before judging. Reuse the prior rounds' stable category strings where the root cause matches; new mechanisms get new stable strings; a single small fix with no shared mechanism → tinyfix.

    Note: #5006 recurred from round 1 — will re-verify and reuse its category. #7422/#7421/#7420/#7419 are fresh (high numbers, likely recent). #4768/#4902/#4838/#4937/#5119/#5235/#5427/#5255/#5336/#5384 are older open issues.

    Category scheme (stable, from prior rounds — reused when the mechanism matches): watermark-fastpath-id-namespace, pin-ledger-sweep-drain, fail-fast-forward-bypass, doom-gate-byte-floor-caliber, egress-fuse-nudge-disagreement, request-body-budget-config, transport-misclassified-degenerate, degenerate-turn-fatal-no-toggle, compress-fold-benefit-floor, compress-fail-streak-loop, compress-circuit-breaker-text-mismatch, nudge-collapsible-interval-no-toggle, compress-range-includes-current-user-msg, storage-poison-row-isolation, storage-digest-drift, storage-vacuum-foreign-writer, session-gc-companion-stat, dead-code-narrow-read, msgid-identity-channel, fork-inheritance-lost, signed-block-wire-fidelity, thinking-final-block-400, anthropic-system-position-heal, responses-tool-ordering, 429-governor-hold-scope, credential-fingerprint-order, codex-prewarm-as-turn, v1internal-preflight-envelope, billing-registry-lookup-perf, billing-disk-summary-stale, web-blog-refresh-byte-cap, web-diagnostics-log-read, web-dashboard-stale-summary, opencode-lane-hook-payload, dsh-long-session-drift, orphan-proxy-offloop-reap, failed-compress-receipt-assistant-tail, tag-echo-filter-drift, boot-contract-characterization, init-snapshot-memory, summary-role-label-mislabel, folded-summary-role-authority, raw-body-estimator-nonmodel-fields, degenerate-retry-block-concat, dsh-long-session-no-convergence, preflight-failfast-no-fallback, storage-slow-ingest-sync-spike, fold-reconcile-content-hash-drift, dns-fallback-worker-leak, registry-request-path-refetch, fake-completion-buffer-truncation, svg-report-arg-limit, responses-content-index-loss, outbound-view-full-recompute, incident-dump-name-collision, watermark-hit-test-coverage, loopback-host-mask-prefix, kv-oversized-value-throw, preflight-per-interval-summary-call, unread-response-body-leak, fold-state-no-cross-truncation-persistence, dsh-declared-compaction-no-substrate-422, url-mask-overscan-quadratic, tinyfix.

    Starting fetch now.

  25. ranxianglei commented on Nov 21, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    Root-cause classification — round 3, all 17 issues. Read every body, resolved the upstream PR number for each of the 14 [PR] rows (5 had mirror number ≠ upstream: 7217→7215, 7422→7420, 5235→5233, 5427→5425, 5255→5250, 5336→5334, 5384→5382 — found via GitHub title search), pulled CI + merge state, and verified the named source paths (relay.ts #821 branch, stream.ts compress breaker) in this workdir. Reused prior-round category strings where the mechanism matches (#5006 → doom-gate-byte-floor-caliber, #7419 → compress-circuit-breaker-text-mismatch); new mechanisms get new stable strings.

    [
      {"number": 7179, "title": "[PR] fix: dsh compaction gate covers runtime-registered agentPresets", "category": "dsh-compaction-gate-preset-coverage", "root_cause": "registerCompactionDisable (src/agent/dsh-compaction-disable.ts:186) hooks only internal/config, which fires for plugin entries resolved via cordis waterfall; presets registered at runtime via ctx.agentPresets.register(definition) (dsh 0.2.2-rc.3+) bypass the hook entirely, so their compaction rows stay armed — with an active bili block native compaction is answered 200 and silently wipes bili's folds; fix wraps agentPresets.register() via cordis inject to apply the same disableBasicCompaction rewrite (idempotent, async-preserving, fail-safe) + a no-preset-register-hook gap in detectCompactionOverlayGap", "ci_state": "green", "merge_state": "dirty", "pr_number": 7179, "confidence": "high"},
      {"number": 7217, "title": "[PR] fix: validate MCP initialization session metadata", "category": "tinyfix", "root_cause": "the MCP initialize handler (src/mcp.ts) calls .trim() on the optional _meta.ui.sessionId before checking its runtime type, so a number/boolean/array/object throws an unhandled rejection that terminates the stdio process before it answers initialization (also blocking tool use when an existing env binding + proxy are valid); fix types the incoming sessionId as unknown and narrows to a string before trimming — a single-spot defensive type-narrowing fix with no shared systemic mechanism (other values leave the existing binding/fallback intact; valid strings unchanged)", "ci_state": "green", "merge_state": "dirty", "pr_number": 7215, "confidence": "high"},
      {"number": 7422, "title": "[PR] fix(relay): skip the #821 null-body warn for HEAD responses", "category": "tinyfix", "root_cause": "the relay's 2xx null-body branch (src/server/relay.ts:2244-2247) logs the #821 warn for every 2xx null body, but a HEAD response has no body by definition (RFC 9110 §9.3.2), so health probes (HEAD /api/hello) hit that branch on every request and flood bili.log with the spurious [unknown] warn (the exact lines in the #7072 report); fix changes the branch to `else if (req.method !== \"HEAD\")` — a one-line warn suppression with no shared mechanism (the response is still relayed empty for HEAD; the #821 warn for a real 2xx null body 204/304 is unchanged)", "ci_state": "green", "merge_state": "clean", "pr_number": 7420, "confidence": "high"},
      {"number": 4768, "title": "[PR] Collapsed preview for large tool-result fragments (sub-min span)", "category": "splitchunks-submin-span-reject", "root_cause": "when a single tool-result exceeds the chunk budget (max(2000, floor(limit×0.6)), src/preflight.ts:54/1627), splitChunks splits the preceding small call fragment into an independent span; the kernel minCompressRange precheck (kernel/src/compress.ts:482-501) counts raw msg.text.length at the commit boundary (before pair widening) and rejects that call-only span atomically every round — the serial lane burns a preview + a misleading warn each round, and the batch lane (external summary) pops only one span per range so such ranges are never nominated and stay stuck until the serial walk takes over; fix adds a host-side CPU-only precheck (spanRawChars, same caliber as the kernel precheck) that skips sub-minCompressRange spans before previewing on both lanes, without changing packing (the token-regime minUnits=0 historical packing is preserved)", "ci_state": "green", "merge_state": "dirty", "pr_number": 4768, "confidence": "high"},
      {"number": 4902, "title": "[PR] refactor(server): name the window-reservation block in handleRequest", "category": "window-reservation-block-refactor", "root_cause": "pure refactor (zero behavior change) — the window-budget reservation block in src/server/handle.ts::handleRequest (a 2545-line single function) sits on top of 6-layer native window resolution plus 5 interacting rules (#896 reserve cap, #282 fallback floor, #4205 anthropic exemption, #3450/#3973 measured share clamp, #2096 clamp caliber) whose correctness is held only by code order and was previously covered only indirectly via server fixtures; fix extracts the 139-line block verbatim into reserveJudgingWindow() with 15 named params (167→19 call-site lines) + a direct branch-matrix test (tests/reserve-judging-window.test.ts); no new config/env, no wire/persistence change, no kernel touch", "ci_state": "red", "merge_state": "dirty", "pr_number": 4902, "confidence": "high"},
      {"number": 7421, "title": "Upstream 200 with null body relayed as empty; Claude Code fails turn", "category": "2xx-null-body-relay-empty", "root_cause": "when the upstream (a loopback gateway, magpie, via the /bili/ prefix) returns HTTP 200 with a null/empty body, bili logs the #821 warn (src/server/relay.ts:2247) and relays an empty 200 to the client instead of surfacing a retryable upstream error, so Claude Code fails the turn with 'API returned an empty or malformed response (HTTP 200)'; the [unknown] session tag means the request isn't attributed to a session (55 occurrences in one day, intermittent — small requests bypassing bili return normal Message bodies); the behavioral question (200 null body → retryable 502/529 with an Anthropic-shaped error body vs empty 200) is open and needs the reporter's log to confirm whether the warn lines were HEAD probes or real empty Messages", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "medium"},
      {"number": 4838, "title": "[PR] release v0.1.209", "category": "release", "root_cause": "release PR shipping master as-is (458 commits since the v0.1.208 fork point) — npm latest 0.1.208 was cut from the bugfix carve-out lane and does NOT contain master's post-fork work; the fix this release exists for is #3472/#4836 (the MCP shim's structuredContent now carries the full report text so Claude Code's structured-preferring bridge no longer drops acp_status/acp_cache/search_context/decompress/acp_retrieve payloads down to bare {outcome:success}); the release-branch diff is only the version bump + plugin-manifest sync (#4449) + one test row — no code/behavior change beyond what master already carries (notable among the 458: #4797 storage restamp identity, #4288 preflight fit-exit order, #4743 degenerate-429 governor, #4732 Windows EBUSY teardown)", "ci_state": "green", "merge_state": "dirty", "pr_number": 4838, "confidence": "high"},
      {"number": 4937, "title": "[PR] fix(codex): rebase ACP state when codex replays an opaque native compaction checkpoint", "category": "codex-opaque-compaction-checkpoint-latch", "root_cause": "codex's newer local auto-compact lands as a native Responses compaction input item with an opaque encrypted_content (instead of the legacy textual summary message); the kernel projection drops it into the preamble (no core id ever exists for it) so the host text-prefix detector (carriesCodexLocalCompactionSummary) can never see it → covered fold ids vanish from the resent history → drift escalation latches fold-epoch confirmation → preflight pauses forever (the old history never returns to the wire so the latch cannot clear on its own) + the stale usage baseline survives; fix (host-only) detects the structural wire shape itself (type===compaction + non-empty string encrypted_content, excluding bili-generated fc_bili_* items), persists metadata.codexCheckpointFp (sha256 of the raw blob, undecoded), and rebases through the existing native-compaction boundary path when fold coverage was actually lost by majority rule (missing*2 >= expected) or the substrate is empty; legacy text path untouched, no kernel change", "ci_state": "red", "merge_state": "dirty", "pr_number": 4937, "confidence": "high"},
      {"number": 5006, "title": "[PR] fix: doom-gate byte floor prices billable content (core text + system/tools), not core text alone", "category": "doom-gate-byte-floor-caliber", "root_cause": "the three doomed gates (egress fuse src/server/relay.ts, preflight fail-fast probe #2313, post-compress recheck src/server.ts) priced ONLY core message bytes against the bar max(2MB, window×7.5 B/token), but preflight's trigger prices messages PLUS system/tools wire overhead (estimateWireOverhead, #470/#829) — so a payload whose billable total exceeds every model window could clear the floor on the residual caliber (field evidence #5000, dsh 0.1.208: a 26,903-message session with only 1.44 MiB of core text produced 55.2 MiB request bodies that every gate read as 'small'); fix makes every gate site price billable content = core message text + system/tools wire overhead, both measured from the ACTUAL outbound body (new estimateWireOverheadBytes over the factored wireOverheadTexts; the JSON envelope and signature-only thinking mass are deliberately excluded to preserve the bytes÷7.5 > tokens invariant); gates are strictly stricter — payloads previously clearing the floor may now 413 (the intended blind-spot closure, pinned both-ways per #3617)", "ci_state": "red", "merge_state": "dirty", "pr_number": 5006, "confidence": "high"},
      {"number": 5119, "title": "[PR] test: sandbox XDG base dirs in the suite preload so bare runs stop writing real $HOME", "category": "test-sandbox-xdg-home-pollution", "root_cause": "local test runs were writing into the real $HOME — bare targeted runs (node --import tsx --test <file>) skipped the preload entirely, and even the suite preload only pinned BILI_STORAGE_UNIFIED, so any test that did not pin its own XDG dirs landed sessions/index.db rows, session envelopes, and state files in the production ~/.local/.../billion-context tree; fix sits at the runner entry point (tests/test-sandbox-env.mjs sandboxXdgEnv, applied by preload-storage-baseline.mjs) and sandboxes the four XDG base dirs (DATA/STATE/CACHE/CONFIG) to a per-process mkdtemp root under os.tmpdir() — explicit values always win, so every existing test that pins its own dirs is untouched; zero product-code change (src/ untouched)", "ci_state": "green", "merge_state": "dirty", "pr_number": 5119, "confidence": "high"},
      {"number": 7420, "title": "[dsh-plugin.org | dsh-plugin-hub] build scripts blocked by pnpm allowlist", "category": "dsh-install-pnpm-allowlist", "root_cause": "the dsh-plugin-hub installer (dsh plugin --profile desktop add billion-context@latest) fails with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED because a git-hosted dep in the install tree (dshmarket@1.73.11 from codeload.github.com/dsh-market/dsh-market) declares a build script that pnpm blocks by default until it is approved in the allowBuilds allowlist — the blocked script is not necessarily a native module; this is a dsh-channel packaging/install failure (dsh-plugin-hub v2.3.9 / DSH v0.2.2-rc.5 / Node v24.21.0), not a bili code defect; per owner policy the dsh channel is not maintained by this fleet — the fix belongs in the dsh-plugin-hub/dsh-market installer (allow the build script), and users should be guided to the official npm package (npm install -g billion-context)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"},
      {"number": 5235, "title": "[PR] kernel: preserve issued refs and quarantine known historical collisions", "category": "kernel-ref-reuse-restamp", "root_cause": "a restart could restamp previously issued refs and an allocator could reuse historical numbers (low ledgerSeq/dead/reserved ref cases), so a model citing a previously seen ref could resolve or compress a DIFFERENT message — violating the kernel's never-reuse contract; host-only removal of restamping is insufficient because the kernel owns allocation; fix stops cold-load ref restamping (preserving existing message/block identity), makes both kernel allocation branches reserve previously issued/dead refs + respect the historical high-water mark, and detects known retained old/current collisions refusing ref-resolving mutations before summary work or plugin state/credit/storage mutation (quarantine fact carried through Session.metadata serialization, absorb/adoption/fork, active-store convergence); kernel 0.0.170→0.0.171, kernel-change, human-gated (no auto-merge)", "ci_state": "red", "merge_state": "dirty", "pr_number": 5233, "confidence": "high"},
      {"number": 5427, "title": "[PR] fix(server): dsh plugin-lane full-resend watermark — foundation increment", "category": "full-resend-watermark-oom", "root_cause": "the per-round full JSON.parse of the entire resent body (81 MiB → hundreds-of-MB AST peak, plus cross-round AST pinning — the two producers of the #5403 user-side OOM) is replaced by a count-watermark assertion: the injection layer stamps x-bili-pf-count: K ('the first K messages[] elements are byte-identical to what you received last round') and on match the server parses only the uncovered tail window instead of the whole body; this PR lands the verified foundation (offset-only JSON depth scanner src/server/fastpath-scan.ts + pure gate decision with fixed precedence src/server/fastpath-gate.ts + rollout entry pluginFullResendWatermark default 0%) behind a 0% rollout gate (inert until force-enabled → zero user impact); the core prepare variant (with A/B golden byte-equality proof), relay stamping x-bili-pf-seq, dsh-native echo, and circuit-breaker wiring are staged on the same branch — references #5418 (mechanism fix for #5403), not marked Closes until the core increment lands", "ci_state": "red", "merge_state": "dirty", "pr_number": 5425, "confidence": "high"},
      {"number": 5255, "title": "[PR] Degrade a second long-lived index.db writer to legacy-only instead of co-residing", "category": "storage-foreign-writer-degrade", "root_cause": "when two bili installs share one sessions directory both RW-open the same index.db; previously the second long-lived writer logged an error and KEPT running (double-writing → corruption risk + SQLITE_BUSY turn failures); fix (owner-approved decision B + S2 keep 3 from #5064) makes the second-process-on-same-db path in UnifiedStore.open (src/storage/store.ts) refuse unified storage at open and degrade to legacy-only for long-lived opens only (proxy/sidecar via UnifiedShadow.maybeOpen passing rejectForeignWriter:true), preserving availability without the double-write (degradation, not blocking, keeps sessions serving from the JSON store per #3633); ephemeral CLI/export direct opens (src/storage/backend.ts, one-shot) keep the report-only behavior unchanged; no new config surface (reuses BILI_SESSIONS_DIR + BILI_CONFIG_FILE), host-side src/storage/ only", "ci_state": "none", "merge_state": "dirty", "pr_number": 5250, "confidence": "high"},
      {"number": 5336, "title": "[PR] Add bili sessions orphan-sweep: remove provably-foreign collapsed blocks from upgraded sessions", "category": "orphan-foreign-block-shared-id-era", "root_cause": "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, never accrues a death streak, never reaped; src/fold-reconcile.ts) and the render side (kernel/src/sync.ts stillPresent) keeps it active as long as any covered-id hash collides with the current request or its anchor is present; once rendered into the client's persisted history the bytes become the client's own (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; fix adds `bili sessions orphan-sweep [id] [--fix] [--json]` flagging blocks with ZERO overlap with the session's unified-store transcript (reading active AND retired rows, since a folded row is retired/tombstoned not deleted) and archiving via a reversible metadata exclusion list (metadata.orphanExcludedBlocks, buildSession drops listed blocks on load); only upgraded ≤v0.1.208 sessions affected (the #4774/#5184/#5163 lane split means new foreign blocks can no longer be produced)", "ci_state": "none", "merge_state": "dirty", "pr_number": 5334, "confidence": "high"},
      {"number": 5384, "title": "[PR] fix: overflow-stated counts steer local fold targets immediately (provisional estimator bias)", "category": "overflow-calibration-bias", "root_cause": "after an overflow 400 the ring can never publish a calibrated k̂ on a fresh overflow-only session (publication needs ≥2 agreeing samples on one origin+model slot that any foreign settle clobbers), so every local judgment on that route (preflight trigger, fold mass/target, relay overhead, output clamp ×3 lanes) stayed raw-estimate calibre forever and each forward re-400'd — a permanent loop; fix retains the upstream-stated payload count as a provisional estimator scale (stats.overflowBias, persisted) and routes every local judgment through one new selector util.effectiveCalibration (ring-published k̂ when applicable to the decision route, else the retained bias, else absent/today's raw behavior), with band-clamped ratios [0.25,4] + a mismatched guard (stated count describes ≥4×/≤¼ of bili's own view → #3450 shape → does NOT steer; a later sane measurement supersedes it) + origin-gated + model-gated exactly like the ring + provider calibration.disabled respected; the first overflow's stated counts (three mutually consistent numbers in one body) are high-confidence and break the loop; the armed preflight comparator additionally exits on the calibrated round text so a rescue fold stops at est×k̂ ≤ budget", "ci_state": "red", "merge_state": "dirty", "pr_number": 5382, "confidence": "high"},
      {"number": 7419, "title": "[DSH] compress infinite loop: failure receipt triggers retries; circuit breaker armed but only alerts without blocking", "category": "compress-circuit-breaker-text-mismatch", "root_cause": "two gates disagree (nudge by total-compressible/growth, compress by single-range min 5000 chars) creating a state where nudge says compress but no range qualifies (log: max compressible 20060 < threshold 50000 on the nudge side, 57 chars min 5000 on the compress side); the failure receipt is 'retryable' semantics (ends with 'Combine more messages into your range(s) to meet the threshold') not terminal, so the model retries forever (observed up to ~1700 repeats per turn, 183 too-small failures across 16 ranges); and the circuit breaker (COMPRESS_LOOP_THRESHOLD=3, src/stream.ts:232) once armed only appends the 'STOP calling compress now' text to the receipt (src/stream.ts:260-271) + stops the nudge, but there is NO code path that refuses to execute compress — the armed streak grew to 78 while compress kept running and failing (the guardrail is advisory text, not enforcement); plus 0-range parse failures are silently treated as 1 range (no 'malformed args' error); PRs #5253/#5564 (too-small → success no-op receipt) exist but are not in any released version (0.1.216 dist/index.js still carries the retryable error branch)", "ci_state": "none", "merge_state": "n/a", "pr_number": null, "confidence": "high"}
    ]
    

    Notes: 17/17 covered (none dropped). 14 [PR] rows carry the resolved upstream PR number + CI + merge state. CI red on 5 (#4902 3-fail, #4937 1-fail, #5006 2-fail, #5235 1-fail, #5427 1-fail, #5384 7-fail — that's 6); CI none (no cached checks) on #5255/#5336; the rest green. Merge state: all dirty except #7422 (mergeable, clean). Categories: 2 reused from prior rounds (#5006 doom-gate-byte-floor-caliber, #7419 compress-circuit-breaker-text-mismatch); 13 new stable strings. #7421 (bug report) and #7422 (its HEAD-noise PR) are related but distinct — #7421 is the open behavioral question (200 null body → empty relay vs retryable error), #7422 is the one-line HEAD-warn suppression (tinyfix). #4902 is a pure zero-behavior-change refactor and #4838 is a release PR — neither is a bug root cause, categorized by subject. #7420 is a dsh-channel install failure (external to bili) — flagged for the official-package guidance.

  26. ranxianglei commented on Nov 21, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    Round 4 plan — 39 issues. Reading every body now, resolving the upstream PR number for each [PR] row (mirror number ≠ upstream for ~half the rows — verified via GitHub title search each round), pulling CI (mirror cache) + merge state (GitHub pulls/N), and verifying the named source paths in this workdir before judging.

    Category scheme (stable, reused across rounds): same root-cause mechanism → same string as prior rounds (e.g. doom-gate-byte-floor-caliber, compress-circuit-breaker-text-mismatch, msgid-identity-channel, watermark-fastpath-id-namespace, transport-misclassified-degenerate, fork-inheritance-lost, fold-reconcile-content-hash-drift, dsh-long-session-no-convergence, kernel-ref-reuse-restamp, full-resend-watermark-oom, storage-foreign-writer-degrade, …); genuinely new mechanisms get new stable strings; a single small fix with no shared mechanism gets exactly tinyfix. Progress reply at 25/39, then final with all 39.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions