Skip to content

[PR] fix(#2645): gate the openai prompt_cache_key stamp to loopback destinations - #2649

Closed
ranxianglei wants to merge 1 commit into
masterfrom
2026-10-10_openai-pck-remote-gate
Closed

ranxianglei wants to merge 1 commit into
masterfrom
2026-10-10_openai-pck-remote-gate

Conversation

@ranxianglei

Copy link
Copy Markdown
Owner

Closes #2645

What / why

On the openai (chat/completions) wire, prompt_cache_key was stamped onto every outbound dsh plugin-lane request — including strict-schema remote upstreams that reject unknown params. NVIDIA NIM answers 400 Validation: Unsupported parameter(s): prompt_cache_key, so every turn on those models failed in <1s, with no config escape.

Root cause

shouldStampRelayAffinityPck (src/session-id.ts) short-circuited return true whenever a URL had been embedded/rewritten (routeRewrittenUrl !== undefined). The /bili/<proto>/<abs-url> tunnel lane and the forward-proxy lane set route.rewrittenUrl for the user's own remote destination too (NVIDIA NIM here), so the gate stamped a strict remote API on every turn. The loopback check underneath was dead for the tunnel lane. This regressed #2218's stated intent ("unrouted non-loopback forwards stay unstamped"): its "admin-rewritten route" leg conflated a URL was embedded with a deliberate local relay.

Fix

Since route.rewrittenUrl always co-references route.upstream, the destination origin is the only trustworthy signal. Drop the short-circuit and gate purely on the destination being loopback:

No new config surface added (per Configuration Surface Discipline). If an admin ever needs to force-stamp a specific remote relay, that's a separate owner-gated proposal.

Tests

  • Updated the #2218/#2645 unit: shouldStampRelayAffinityPck gate matrix — pins loopback→stamp, localhost→stamp, remote (incl. a full tunnel-style URL with path)→no-stamp, client-pck-wins, non-plugin→no-stamp, unparseable→no-stamp.
  • New e2e regression reproducing the exact NVIDIA 400: a strict mock upstream (400s on prompt_cache_key) reached from the dsh plugin lane through the machine's non-loopback address → asserts 200 + no stamped field + x-session-id header relay still intact. Skips gracefully if the environment has no non-loopback IPv4.

Verified locally: typecheck clean, targeted suites green (tests/issue2218-relay-affinity-signal.test.ts 7/7, plus openai/anthropic pck-identity + route-passthrough + codex-official 17/17), full build OK.

…ations

shouldStampRelayAffinityPck short-circuited to true whenever a URL was embedded/rewritten. The /bili/ tunnel and forward-proxy lanes set route.rewrittenUrl for the user's own REMOTE destination too, so the dsh plugin lane stamped prompt_cache_key onto strict-schema remote APIs (NVIDIA NIM answers 400) and broke every turn. rewrittenUrl always co-references route.upstream, so the destination origin is the only trustworthy signal: drop the short-circuit and gate purely on the destination being loopback. Local relays (every #2218 scenario) keep the stamp; remote strict APIs stop getting it. No config surface added.
@github-actions

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-10-10_openai-pck-remote-gate (822ccd0)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.2863-pr.2649.575f316

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

Option B — Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr2649.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 PR #2649 Review 结论(结构化八问)

1. 是什么问题:openai chat/completions 线上,proxy 替 dsh 插件道往每个出站请求体里盖 prompt_cache_key;严格 schema 的远端 API(NVIDIA NIM)对未知字段回 400 Validation,导致这类模型每一轮 <1s 失败,且无任何配置出口。

2. 什么原因导致的:shouldStampRelayAffinityPck(src/session-id.ts)在 route.rewrittenUrl !== undefined 时无条件短路返回 true。我核对了 resolveUpstream(src/server.ts:223):rewrittenUrl 恰好在三条 lane 被设置——MITM socket、/bili/<proto>/<abs-url> 隧道、absolute-form 转发代理——全部是「用户自己的目的地 URL 被嵌入」,没有任何 admin 策展路由分支(_opts 参数根本没被使用)。所以短路腿把「URL 被嵌入」误当「刻意的本地 relay」,#2218 的意图("unrouted non-loopback forwards stay unstamped",master 测试注释原文)在隧道腿上是死的。根因定位准确,是决策点上的概念混淆,不是表面现象。

3. 影响范围与严重程度:所有走 URL 嵌入 lane 的插件道 openai 请求 × 严格 schema 远端上游——完全无法使用(每轮 400)。触发条件明确:dsh 桌面/插件道 + 远端模型端点(经隧道或代理 env 到达)。本地 relay 部署不受影响。

4. 怎么修复的:删掉短路腿和第三个参数,gate 纯看目的地 origin 是否 loopback。这是机制级修复(疏而非堵):不改写载荷、不做 host 特判、不做 400 自愈(issue 里方案 4 是治标,方案 3 打的是另一个机制;本方案是方案 2 的更严格版本,且不新增配置面)。route.upstream 恒与 rewrittenUrl 同指目的地,是唯一可信信号。

5. 是否有回归问题:本版本引入回归:无。逐项审过的既有行为:

  • 本地 relay 盖章(上游侧失效导致缓存命中只有80%用的api反代 workbuddy glm5.3 #2218 全部场景,含无路由直连 loopback):保持(test ok 6 + 单测矩阵钉住);
  • 客户端自带 pck 永远优先:保持(第一道检查未动,test ok 7);
  • 非插件身份源(codex session-id、普通客户端)不盖章:保持(test ok 8/ok 4);
  • localhost 字面量算 loopback:保持(显式分支保留,新增单测钉住);
  • 不可解析 origin → 不盖章:保持;
  • 有意变更(已披露):经隧道/转发代理/MITM 到达的非 loopback 目的地,旧=盖章,新=不盖章。这正是修复目标;跨机 body-only relay + dsh 桌面 + 客户端不发 pck 的部署会失去亲和提示——但修复后状态恰好回到 上游侧失效导致缓存命中只有80%用的api反代 workbuddy glm5.3 #2218 的设计态(该类部署在未嵌入路由时本来就不盖章),且客户端侧显式发 pck 永远可行,x-session-id header relay 不受影响;
  • anthropic 线:pck 无条件从出站体剥离(prepare-anthropic.ts:337),无同类缺陷;responses 线走独立的 promptCache.routing 旋钮(有配置出口),未触碰;
  • 两种压缩模式均核过:gate 首查 x-bili-plugin-conversation header,proxy 模式行为前后一致。

6. 是否破坏缓存:loopback relay 路径字节级不变,前缀缓存策略无影响。远端严格 API 从「每轮 400(无任何缓存可言)」变为正常 200。唯一的亲和损失见第 5 条有意变更项,属设计态恢复而非破坏。

7. 是否改了设计方向:不涉及架构转向;是把 openai 线的盖章语义收敛到其本来的定义域(loopback 本地 relay)。若未来需要强制给特定远端 relay 盖章,按 Configuration Surface Discipline 走 owner-gated 提案,本 PR 正确地没有自建配置面。

8. 是否引入了新的配置:不涉及。零新 env/config/默认值。

验证证据(非转述作者声明,均独立执行):

  • 红绿证明:把两个 src 文件临时换回 master 版跑新增 e2e → expected: 200, actual: 400(精确复现 NVIDIA 失败形态);换回 PR 代码 → pass。回归测试有牙;
  • 本地定向套件 31/31 绿:issue2218-relay-affinity-signal 7/7(含新 e2e,本机有非 loopback IPv4 故未 skip)、openai/anthropic pck-identity+verbatim 12/12、route-passthrough 6/6、codex-official 2/2;
  • CI 全绿:head 822ccd0 上 11 个 workflow 全部 success(含 E2E real codex、e2e-dsh-plugin-lane、e2e-native-{dsh,opencode,pi}、e2e-local-registry、CI Windows codex overlay);
  • diff 整洁:单 commit,3 文件 +133/−26,每行都与本 PR 目的相关;无 lockfile/version/kernel/wire-contract golden 变更;
  • 调用点全仓唯一(src/server/prepare-openai.ts:344),签名收窄无其他消费者;
  • 小观察(无需处置):session-id.ts 注释点名了隧道/转发两条 lane,MITM lane 共享同一属性但未枚举——注释的机制级表述("destination origin is the only trustworthy signal")已完整覆盖三条 lane,不构成错误,不为它重跑 CI。

判定:可合并。 方向对、根因准、测试真、CI 绿、行为变更已披露。请人工合并:#2649

ranxianglei pushed a commit to askl56/billion-context that referenced this pull request Oct 17, 2026
…ant prompt_cache_key on every egress lane

- session-id: drop the route-rewritten leg of shouldStampRelayAffinityPck
  (ranxianglei#2649: a rewritten/embedded URL is not relay evidence — the /bili/ tunnel
  rewrites the user's own remote destination too); loopback is the only
  destination-shape proof. Export relayAffinityDestined for the egress site.
- prepare-openai: chat wire no longer passes a client prompt_cache_key
  through verbatim to non-relay upstreams — strict APIs 400 unknown body
  params (ranxianglei#2645). Client value still wins on loopback relay destinations.
- pi (stampPromptCacheKey): hoist the ranxianglei#2583 Cowork withhold above the
  routed-check so it gates ALL stamp paths. Pre-hoist, the launcher
  cert-MITM lane (BILI_MITM_HOSTS) made destinationRoutedThroughProxy()
  true and skipped the withhold entirely — stamped bodies reached
  Anthropic's strict schema verbatim (ranxianglei#1403 400s).
- pi (ompCoworkTransportBypassesFetch): a LOOPBACK HTTPS_PROXY is bili
  itself and Cowork's node:https honors it — treat as routing evidence
  (stamp); a foreign/non-loopback HTTPS_PROXY is a bypass (withhold).

Tests: flip ranxianglei#2649 matrix row; 4 new cases (MITM withhold, loopback
HTTPS_PROXY stamp, foreign HTTPS_PROXY withhold, predicate matrix).
ranxianglei pushed a commit to askl56/billion-context that referenced this pull request Oct 17, 2026
… doc placement; add ranxianglei#2645/ranxianglei#3556 remote-destination pck regressions

Review of ranxianglei#3556: this PR already dropped the rewritten-URL leg of shouldStampRelayAffinityPck, which left the third parameter dead and the full-gate JSDoc stranded above relayAffinityDestined. Aligns the signature with ranxianglei#2649's 3-arg form and moves the doc back onto the function it describes.

Also ports the ranxianglei#2649 strict-remote regression (mock upstream that 400s on prompt_cache_key, bound to a non-loopback IPv4 so bili sees a REMOTE destination) and adds the missing regression for the new client-pck strip: verbatim client prompt_cache_key must not reach non-relay chat egress (plain-client and dsh plugin lanes), while x-session-id identity survives.
@ranxianglei

Copy link
Copy Markdown
Owner Author

为什么本地就删除这个key 号池子就是本地 没有这个就遭殃

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 先纠正一个关键误解:本 PR 没有删本地目的地的盖章。判定看的是请求最终目的地 origin 是否 loopback(127.x / localhost)——loopback 目的地照旧盖章,字节级不变,有测试钉死(issue2218 套件 ok 6:"dsh plugin lane stamps pck on a loopback destination even without a configured route"+ 单测矩阵 127.0.0.1→stamp、localhost→stamp)。被删的只是非 loopback 目的地的盖章(NVIDIA NIM 这类严格远端 API,每轮 400 的那个)。

你的号池会不会受影响,取决于一个具体事实:不是号池部署在哪台机器,而是 dsh 侧指向号池时写的地址串是什么(判定用的就是这个串的 origin):

  • 号池和 dsh 同机,且 dsh 指向 127.0.0.1/localhost → 照样盖章,零影响。标准 上游侧失效导致缓存命中只有80%用的api反代 workbuddy glm5.3 #2218 形态就是这样:bili dsh 对 loopback upstream 保持 overlay 改写,请求以 path-form 到 proxy,upstreamOrigin = http://127.0.0.1:<pool-port> → 命中盖章;
  • 号池同机,但 dsh 指向机器 IP / LAN IP(或 /bili/ 隧道 URL 里写的是那个 IP)→ 判成非 loopback → 不盖章 → 你担心的「遭殃」只发生在这种形态;
  • 补充一点:即使不盖章,x-session-id header relay 不受影响——读 header 优先的 relay(sub2api 一族)拿得到会话信号;只有纯读 body 的 relay(workbuddy 面板一族)会退回 content digest。

请给一行事实:dsh 侧配的上游地址里,号池写的是什么?(127.0.0.1:port 还是机器 IP?)

如果已经是 127.0.0.1,本 PR 直接合即可,你的场景一个字节没变。如果是机器 IP,三个选项,我的推荐是 A:

  • A(推荐,零代码):dsh 侧把号池地址改成 127.0.0.1(同机部署本就该走 loopback,隧道准入也更宽)。盖章立即恢复,现有机制原样工作,不为单一部署形态动代码。
  • B(owner-gated 配置面):新增「relay 盖章白名单」(按 destination host),覆盖跨机号池或必须用非 loopback 地址的场景。按 Configuration Surface Discipline 需先立项讨论再实现,本 PR 不会自建。
  • C(400 自愈,[bug] compress 保护区过滤在完整性调整之后挖洞,导致 assistant tool-call 消息缺 reasoning_content(DeepSeek 思考模式 400) #684 learn-on-failure 同类):先盖章,上游回 400 且报错点名 prompt_cache_key 时记住该 origin 不再盖。无配置,但每个严格 origin 每进程付一轮失败,且要引入 origin 拒绝状态。

八问(决策请求版)

  1. 是什么问题:① openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape #2645 本体——严格远端 API 因被盖 pck 每轮 400,已由本 PR 解决;② 你新提出的——号池若经非 loopback 地址到达,会失去 pck 信号、退回 content digest,账户迁移时前缀缓存冻结(上游侧失效导致缓存命中只有80%用的api反代 workbuddy glm5.3 #2218 症状复现)。②是否成立取决于上面的地址事实。
  2. 什么原因导致的:旧短路把「URL 被嵌入」误当「刻意本地 relay」,错盖了严格远端;新判定反过来把「非 loopback 地址」一律视为「非刻意 relay」——真远端和「本地号池走了非 loopback 地址」被一网打尽,这是同一枚硬币的两面。
  3. 影响范围与严重程度(分选项):A=只动你自己的 dsh 配置,无严重程度;B=所有配置了白名单的用户,opt-in,默认行为不变;C=全部插件道 openai 请求,体验瑕疵(严格 origin 首轮一次 400)+内部复杂度。
  4. 怎么修的(推荐 A):信号语义是「刻意的本地 relay」,同机号池用 loopback 地址表达即可让既有机制原样生效;跨机号池成为真实需求时再单独立项 B。
  5. 回归风险(分选项):A=零代码变更,零回归;B=新配置面需 owner 批准+双语文档,默认关闭不影响存量;C=给原本 fail-fast 的路径引入「首轮失败」,并扩大运行时状态面,风险最高。
  6. 是否破坏缓存:A/B(配置后)=恢复号池亲和,缓存恢复;C=首轮 400 造成一轮缓存未命中,之后正常。
  7. 是否改了设计方向:三者都不改架构;B 属新增 owner-gated 配置字段(待立项)。
  8. 是否引入新配置:A=无;B=一个新 key(具体位置待提案,应在 promptCache 层级下);C=无新配置但新增 origin 拒绝状态。
A 改 loopback 地址 B 白名单配置 C 400 自愈
前提 号池与 dsh 同机 任意地址 任意
成本 一行配置 新配置面(owner-gated 立项) 代码+状态
严格 API 安全 ✅ ✅(默认关) 首轮付费
支持跨机号池 ❌ ✅ ✅

收工判定:等你一行地址事实。127.0.0.1 → 本 PR 可合并,无需任何改动;机器 IP → 建议先落 A,若你认为跨机/固定非 loopback 是常态,我按 B 立项(带来源标记)等拍板。

@ranxianglei

Copy link
Copy Markdown
Owner Author

(ai: zhipuai-lb/glm-5.3)

此 PR 被替代:#3736 (superseded)

原因:静态门槛「非本地目的地一律不发」分不开两个群体——部署在 docker 桥 IP / 局域网 / VPS / 域名上的中转和号池同样靠 prompt_cache_key 做前缀缓存路由,一刀切掉远端会让这些用户大面积丢缓存亲和,而带键仍报 400 的只是其中一小撮严格的端点(#2645 里的 NVIDIA 形态,且是同一域名下按模型分的)。所以改成实证自愈:默认照旧全发;哪个上游真的拒绝这个字段,A/B 验证一次(带键 400、去键成功)后只在那处停发;状态只在进程内存,重启重置;日志、acp_status、Web 面板三处都会明确告知用户,号池场景会指名到会话和模型。

本分支没有代码行被沿用(静态判断逻辑整体被实证门取代),是否关闭由 owner 定。

@ranxianglei ranxianglei changed the title fix(#2645): gate the openai prompt_cache_key stamp to loopback destinations [PR] fix(#2645): gate the openai prompt_cache_key stamp to loopback destinations Oct 18, 2026
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 收到替代决定,事实已核:PR #3736 存在且 open(mergeable=True,分支 2026-10-18_pck-probe-remember @ 09ad8646,body 明确标注 Fixes #2645 · Supersedes #2649 · Absorbs #3556);机制与你的描述一致——默认全发、400 时在出口缝(src/server/relay.ts)drain 错误体后仅去 prompt_cache_key 重发一次做 A/B 验证、重发被接受才把 (origin, model) 记入进程内存(src/pck-probe.ts)并只在该处停发、被拒绝则原样透传原始 400 不学习(重试本身即裁决,误归因不可能)。

同时纠正我此前 review 记录里的一处低估(对替代决定是加强而非削弱):我重新核对了 #2649 基座(7c993f9e)的接线——src/server.ts:2207 在基座上就已是 billingUpstream = route?.rewrittenUrl ?? (reqUrl 是绝对 URL ? reqUrl : opts.upstream),且 prepare-openai.ts:342 传给 gate 的就是这个恒有定义的值。也就是说旧短路在活线路上无条件命中:master 的事实行为是「插件道一律盖章」(loopback/远端、有路由/无路由都盖),loopback 检查整体是死代码,#2218 测试注释里写的 "unrouted non-loopback → no stamp" 只是意图、从未在活路径上成立。所以 #2649 的真实 diff 比我原 verdict 写的更大:不是「经隧道/代理到达的非 loopback 目的地丢键」,而是全部非 loopback 目的地丢键(含未走路由的直连转发)。我的红绿 e2e 只覆盖了隧道 lane,没暴露这个更宽的 delta——这条更正也正好印证你的判断:静态门槛一旦合入,docker 桥 IP / LAN / VPS / 域名上的号池会实际大面积丢缓存亲和,不是假想风险。#3736 的实证设计两个失败面都不付:号池默认保住键,严格端点(包括同一域名下按模型分的严格性——任何静态目的地谓词都表达不了这个维度)一次 A/B 验证后自愈。

八问(收工版)

  1. 是什么问题:openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape #2645(严格端点拒收 prompt_cache_key 致每轮 400)改由 [PR] fix(#2645): probe-and-remember prompt_cache_key — stamp by default, self-heal strict upstreams #3736 实证自愈方案承接;[PR] fix(#2645): gate the openai prompt_cache_key stamp to loopback destinations #2649 静态门槛方案整体退役。
  2. 什么原因导致退役:静态目的地谓词既表达不了「同域按模型分治」的严格性分布,又会误伤所有非 loopback 号池;复核证实 master 旧 gate 事实行为是全发,[PR] fix(#2645): gate the openai prompt_cache_key stamp to loopback destinations #2649 的 delta 实为全量剥离远端键。
  3. 影响范围与严重程度:无——[PR] fix(#2645): gate the openai prompt_cache_key stamp to loopback destinations #2649 从未合并,退役零影响;它本会造成的远端号池亲和大面积丢失随退役消除。
  4. 怎么修复的:不由本 PR 修,由 [PR] fix(#2645): probe-and-remember prompt_cache_key — stamp by default, self-heal strict upstreams #3736「默认全发 + 400 后 A/B 验证一次 + (origin,model) 进程内存记忆」承接,日志/acp_status/Web 面板三处告知用户。
  5. 是否有回归问题:未引入(从未合并);[PR] fix(#2645): probe-and-remember prompt_cache_key — stamp by default, self-heal strict upstreams #3736 的回归面在其自身线程审。
  6. 是否破坏缓存:无(无代码合入)。
  7. 是否改了设计方向:是——pck 发送策略从「静态目的地门槛」转向「实证归因 + 进程内记忆」,属刻意选择,非漂移。
  8. 是否引入新配置:无(状态仅进程内存,重启重置)。

收工判定:建议以 superseded 关闭 #2649(按你的措辞,关闭操作留给你)。两件可交接给 #3736 线程的东西:①我为红绿证明写的严格 mock upstream(NVIDIA 400 报错逐字对齐)+ 隧道 lane e2e 夹具(tests/issue2218-relay-affinity-signal.test.ts)可直接复用;②上面「基座 gate 事实行为=全发」的核对结论,可作其 review 背景——它说的 "restores master behavior" 字面上就是对的,因为全发正是 master 的事实行为。

@ranxianglei

Copy link
Copy Markdown
Owner Author

[bot] 🏷 Superseded by #3736 — the static loopback/non-local gate does not separate the two large pool populations (local/docker/LAN/VPS/domain relays all need the key; strict APIs reject it), so the fix moves to adaptive learn-from-400 in #3736. Closing this PR; no part of it is carried forward. See #2645.

LoveDoLove pushed a commit to LoveDoLove-Forked-Projects/billion-context that referenced this pull request Oct 19, 2026
…ache_key stamp on un-consumable transports

Adopted from ranxianglei#3556 (ddc4d0a) — the pi.ts + omp test half. The commit's
session-id/prepare-openai/issue2218 changes are NOT adopted: this branch's
first commit replaces that destination-shape gate with A/B-verified
probe-and-remember instead of static stripping (ranxianglei#2649's loopback-only gate
was vetoed — it cost the two biggest relay/account-pool populations their
prefix-cache affinity).

- pi (stampPromptCacheKey): hoist the ranxianglei#2583 Cowork withhold above the
  routed-check so it gates ALL stamp paths. Pre-hoist, the launcher
  cert-MITM lane (BILI_MITM_HOSTS) made destinationRoutedThroughProxy()
  true and skipped the withhold entirely — stamped bodies reached
  Anthropic's strict schema verbatim (ranxianglei#1403 400s).
- pi (ompCoworkTransportBypassesFetch): a LOOPBACK HTTPS_PROXY is bili
  itself and Cowork's node:https honors it — treat as routing evidence
  (stamp); a foreign/non-loopback HTTPS_PROXY is a bypass (withhold).

Tests: 4 new cases in omp-native-identity (MITM withhold, loopback
HTTPS_PROXY stamp, foreign HTTPS_PROXY withhold, predicate matrix).

(cherry picked from commit ddc4d0a)
ranxianglei pushed a commit that referenced this pull request Oct 29, 2026
CONFIGURATION en/zh: default column and section rewritten for the disabled default, both-wires semantics, client-key rule, entry points. TECHNICAL-NOTES en/zh: policy-reversal section superseding #2649's default-on ruling. CHANGELOG: Unreleased entry with the old->new behavior disclosure.
ranxianglei pushed a commit that referenced this pull request Oct 29, 2026
CONFIGURATION en/zh: default column and section rewritten for the disabled default, both-wires semantics, client-key rule, entry points. TECHNICAL-NOTES en/zh: policy-reversal section superseding #2649's default-on ruling. CHANGELOG: Unreleased entry with the old->new behavior disclosure.
ranxianglei pushed a commit that referenced this pull request Oct 29, 2026
CONFIGURATION en/zh: default column and section rewritten for the disabled default, both-wires semantics, client-key rule, entry points. TECHNICAL-NOTES en/zh: policy-reversal section superseding #2649's default-on ruling. CHANGELOG: Unreleased entry with the old->new behavior disclosure.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape

1 participant