Repository navigation
Conversation
cdda1b3 to
e425e78
Compare
(ai: vllm-qwen/qwen3.8-27b) 查重: 不重复。检索 open+closed issue/PR(prewarm、generate:false、probe relay、fold 失活、forged summaries 等关键词)无同机制条目;#4036/#4040 是 #4594 已引用的前案(同症状家族——本 PR 修的是根机制,#4040「gate 用请求自身视图判定」保持互补)。 标签: bug / codex / scope/wiring / p1 —— 重启后所有 codex 会话折叠状态被清零(73→0 等七例)+ 长会话成本爆炸,属 owner 修订 p1 类①(状态丢失);#4594 与本 PR 均已打上。 审核进行中,初步结论:
下一步:推送 rebase 后的分支刷新 PR head,看新 head 的 CI(master 测试分片此前因 #4555 红、修复在 #4569,需确认当前 master 已含),然后按八问给最终结论。 |
(ai: vllm-qwen/qwen3.8-27b) 查重:不重复——4 个红全部是继承性/旧基线问题,无一来自本 PR diff,对照 CI 总控 #4599 逐一定位:PR 基线 cd92b0b 早于 master 的门禁修复提交 4d34abf(config-surface guard 脚本 bug + kernel lockfile 同步 + 派生文档再生),config-surface-guard / lock-sync / docs-sync 三个红正是该提交修掉的那组漂移;第 4 个 unused-exports 是 master 自身的 born-red(#4545 train 新增了一个只在文件内部使用的类型导出、de-export 清扫漏了它),与 master head 运行的错误行逐字节相同。 标签:不变(既有 PR 的 CI 失败修复,类别/范围/严重程度沿用原 PR 判定)。 修复动作(auto-fix lane attempt 1/2,直接修完未停):
八问:
行为变更披露(按规则 4):无——rebase 带入的是 master 自身已披露的变更;de-export 为纯类型层,运行时输出零差异。 判定:修复完成,新 head |
(ai: vllm-qwen/qwen3.8-27b) PR 审核(独立于 issue 侧结论重新做过,非复读)。 ① 原 issue(方向关):#4594 要修的是「codex context-only prewarm 被当成 ACP turn 执行 + probe 出站绕过上游代理」。作者修的是生产侧根因(probe 认定范围 + egress 路由),不是症状拦截;消费侧(gate/恢复)已由 #4040/#4085 覆盖,本 PR 与之互补。方向无大问题。 ② 回归审查(触碰路径的既有行为逐项清点):
高危域点名:自动更新/自更新路径未触碰;dsh 相关未触碰;版本钉死/lockfile 未触碰;systemd/env 模板未触碰;网络/代理路径有触碰(probe egress)——已按上述第 3 条核实。 ③ 代码审查: ④ 测试有效性:三个新测试都有牙且是双向门禁(#3617)——正向(context-only prewarm:逐字 deepEqual、sentinel 缺席、tools 为空、stats 不变、fold 仍 active;修复前红)、负向(history prewarm 留管道;修复前后绿)、代理 CONNECT 计数(修复前 0 CONNECT 红)。断言测的是行为而非 mock 形状。fixture 改动(fake upstream 对任何 generate:false 帧回空 terminal)对齐真实上游行为,不影响既有测试语义。 ⑤ diff 整洁度:4 文件 +124/−20,每行都与目的相关,无跑题文件、无生成物/锁文件连带变更。干净。 ⑤b 代码整洁硬标准:命名/死代码/重复/错误处理/魔法值/注释/风格/依赖/函数职责逐项过,全部达标;外部提交从严标准同样满足。无需重构。 结构化八问:
验证记录:rebase 到 master tip(146b34f27)无冲突(本地已做,未推分支——BEHIND 不构成阻塞,等 #4606 落地后 rebase 一次即可);rebase 头上定向 5 套件(codex-responses-ws / responses-ws-http-fallback / ws-bridge / tunnel-guard / upstream-proxy-routing)60/60 绿,typecheck 干净。CI 现状红为继承性失败(base cd92b0b 早于 4d34abf 的门禁修复),已在 #4599 登记。 判定:可合——待 #4606 合入、本分支 rebase 至 tip、CI 全绿后由 owner 合并。 |
(ai: vllm-qwen/qwen3.8-27b) 最终审核结论(八问)
验证证据:rebase 到当前 master(146b34f)后本地 typecheck 干净;定向套件 151/151 绿(codex-responses-ws 12/12 含三个新增正反用例 + responses-ws-http-fallback / responses-prewarm-generate / responses-websocket / ws-bridge / ws-upgrade / tunnel-guard / upstream-proxy-routing / skipped-proxy-hint / native-ws-intercept / issue3701-ws-byte-budget)。CI:静态门禁全绿;11 个 test 分片红 = master 继承(LIVE_FILES tripwire,一行测试未跑,与 #4599 盘点第 2 条逐字一致,已在 #4599 留痕)。PR 作者在审核期间自行 rebase 并追加了 ClientBuildCheck de-export 小修(head e425e78,与我 rebase 的内容 diff 完全一致+该一行),我此前的推送已被其取代,无冲突。 遗留(作者已声明不在本 PR): 判定:代码可合。CI 红灯全部继承自 master(#4599 已接管,#4606 修复在途)——#4606 合入后 rebase 本 PR 即全绿,可合并。 |
(ai: vllm-qwen/qwen3.8-27b) CI round 2 — test shards: root cause found, fixed in-branch, new head pushed (d25f4f1) After the rebase (head e425e78), CI came back with all four originally reported gates green (config-surface-guard / lock-sync / docs-sync / unused-exports), plus every other gate job (docs-coverage, lifecycle-ledger-guard, test-import-graph, kernel-guard, pi-subagents-guard, lane-guard, version-guard). But all 11 test-shard jobs failed at startup — no single test ran — with one identical signature: Diagnosis. #4590 (tier5-prepare-train, merged to master as 146b34f — the rebase target) added Fix (commit d25f4f1, fast-forward append — nothing overwritten). Appended the suite to LIVE_FILES, which is the tripwire's own prescribed remedy. Per the #1248 discipline I reviewed its concurrency assumptions first: the unit half only imports the pure 八问 (this round)
判定: 本轮继承性红全部在分支内收敛完毕(rebase 消掉 gate 漂移 + ClientBuildCheck de-export + LIVE_FILES 登记),PR 头 d25f4f1、base 已是 master 最新 146b34f、mergeable=True。等 CI 结果;若分片再暴露别的失败,由下一轮 auto-fix 处理。已按规在 #4599 留一行结论,未开新 issue。 |
…same one-word fix as ranxianglei#4606 commit 1; duplicate in-branch per CI-coordination precedent ranxianglei#4607/ranxianglei#4595/ranxianglei#4596, drop on rebase after ranxianglei#4606 merges)
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) 收工判定 — attempt 1/2 完成,全部继承性红已在分支内收敛并推送 PR 头现指向
本地已验证:四个 gate 脚本全绿、tripwire 扫描干净(49 条、零未登记 spawner)、typecheck 干净、定向 |
…ot this PR's master 146b34f fails the unused-exports gate on this exact line (INTERNAL_ONLY src/agent/client-build-note.ts: ClientBuildCheck): the type-only export #4545 introduced was missed by the 4d34abf de-export sweep. This PR touches none of it — every PR rebased past that master inherits the red (#4599 storm log). Same one-word fix as #4606 commit 1 and the in-branch precedents (#4607/#4595/#4540/#4618/#4622): drop the leading export on a symbol referenced only within its own file. Drop this commit when rebasing after #4606 merges.
(ai: vllm-qwen/qwen3.8-27b) attempt 2/2 完成 — 10 个测试分片红已修,推送 查重: 同家族但机制不同——本轮 5 个失败测试全部是 PR base( 八问
行为变更披露(旧→新+原因):dump 文件的字节预算——旧:banner 在上限之外,最终文件可超 收工判定:本轮均为几行级小修,无需设计决策,已全部落地并推送;若 |
…tream proxy (ranxianglei#4594) A generate:false request that replays no conversation turn (empty input, or only developer/system/user messages) is a prewarm probe: forward it verbatim instead of running it as a zero-message ACP turn, which deactivated every fold block and re-injected the whole forged-summary list. History prewarms keep riding the pipeline. The probe relay now egresses through the same upstream proxy the relay resolves; ws-bridge exposes the resolved proxy instead of a boolean so the tunnel guard and the probe share one decision. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The PR branch (codex context-only prewarm probe relay) had diverged from master. Merging master in produced a single conflict in src/responses-ws.ts: the PR added isPrewarmProbe() while master added the x.ai continuation-rejection helpers (XAI_RESPONSE_MISSING / continuationRejected / untypedErrorEnvelope) at the same location. Both are independent and both are used, so the resolution keeps both blocks.
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Merge conflict resolved and pushed. Merged current
八问
判定:冲突已解除并推送,PR 现 |
Inherited CI red: the visibleNotices failures (ubuntu 22/24 + win-22-5, storage-sqlite-default missing) are master-born-red after ranxianglei#6480 flipped the storage default to pure-sqlite — the notice-registry suite preload pins BILI_STORAGE_UNIFIED=legacy, silencing the notice. ci-fixes c6369fb pins the pure-sqlite selection in the suite. Per CI-coordination policy (inherited red -> merge origin/ci-fixes), bringing that convergence in rather than re-fixing on the PR branch. The Windows EBUSY unlink index.db shards (win-22-1 store-zstd, win-22-4 metadata-fold) are the separate ranxianglei#4347 file-lock flake, still pending ci-fixes convergence.
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Red CI diagnosed: all 5 failing shards are inherited (none are this PR's regression). Merged Re-verified live first: the mirror check cache was still stale (old sha, all
This PR's diff only touches Action taken (per CI-coordination policy: inherited red → merge
Effective PR diff vs master is clean: the original 7 files + the 5-line ci-fixes test pin, no 八问
判定:visibleNotices 红已由 ci-fixes 合入修掉(新 head |
Conflict resolution: src/ws-bridge.ts — kept the PR's egressProxyFor routing (egressProxyFor?.(upstream) !== undefined, consistent with the param/interface/createSession in the merged tree) and added master's shutdown guard (if (closing || socket.destroyed) return;). No other conflicts. This is the final reviewed head (full CI pass).
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Conflict resolved and pushed — new head What happened: the PR went dirty again because master advanced The one conflict —
Resolution: kept the PR's The other three auto-merged files ( Push: verified the fork head was still 八问
判定:冲突已解决,PR head 现为 |
…param shape The master-added tests/ws-bridge-pending-close.test.ts was written against the old isEgressProxied?: (url) => boolean param. The PR renamed that param to egressProxyFor?: (url) => string | undefined, so the test's () => proxied mock no longer typechecks (TS2345 on static-gates). Pass a proxy URL string when proxied, undefined otherwise — preserves the exact egressProxied boolean the tunnel-guard derives (egressProxyFor?.(upstream) !== undefined). Final reviewed head (full CI pass).
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Re-verified live first: the two Root cause — a merge interaction, not a runtime bug. Master's 494 commits added The fix (one line, - const handler = installWebSocketBridge(server, async () => {}, () => {}, [codec], () => proxied);
+ const handler = installWebSocketBridge(server, async () => {}, () => {}, [codec], () => (proxied ? "http://127.0.0.1:3128" : undefined));This preserves the test's exact semantics: when Verified before pushing: no The mirror check cache still shows the old sha (poller refreshes ~60s); the new CI run on 八问
判定: |
…glei#3342 probe-relay behavior The comment described the pre-ranxianglei#3342 / pre-ranxianglei#4595 behavior: it claimed the session-level prewarm guard catches ONLY empty-input probes and forwards them 'without committing a checkpoint', and that a non-empty prewarm goes through the ACP pipeline. Both are now inaccurate: since ranxianglei#3342 the probe relay commits a checkpoint from the frame body, and since ranxianglei#4595 the guard (isPrewarmProbe) also relays context-only (developer/system/user) prewarms verbatim. The test itself still pins the expand() guard that strips generate:false from a committed prewarm body; only the comment needed to track current behavior.
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
APPROVE — the fix does what it claims; the mechanism is sound; the tests are genuine regression tests; no blockers.
Direction (①): #4594 reports codex context-only prewarms (generate:false, replaying no conversation turn) being run as full ACP turns — zeroing and persisting the session's fold state, stuffing the full history summary + tools into a no-output request, and the probe egress bypassing the configured upstream proxy (handshake fails on proxied machines). The author fixes the producer-side root cause (probe recognition + egress routing), not a symptom-level mask. Direction is correct.
Regression (②) — existing behavior on the touched paths, itemized:
- Probe gate (src/responses-ws.ts:81
isPrewarmProbe): old "empty input + generate:false → relay" is a strict subset of the new "context-only (developer/system/user) + generate:false → relay". A history-replaying prewarm (assistant/tool items) still rides the pipeline (pinned by the new negative test). No regression. - Egress routing (src/ws-bridge.ts, src/responses-ws.ts):
isEgressProxied:boolean→egressProxyFor:string|undefined. The tunnel-guard boolean is derived from the sameegressProxyForcall — bit-identical semantics.proxyDispatcher(undefined)→ undefined → direct egress unchanged. The probe now shares the same cached ProxyAgent (#3776) and the same upstream socket as the pipeline turn (connection key excludes the dispatcher;resolveProxyDecisionis deterministic per URL) — this also fixes the pre-PR socket split (#3073 class). - Pipeline parity: loop/core.ts:697 uses
proxyDispatcher(ctx.proxyUrl);ctx.proxyUrl = resolveProxyDecision(...).proxy(src/server.ts:3230) — the same decision asegressProxyFor. No drift. - Checkpoint semantics: the probe relay commits the #3342 checkpoint (unchanged by this PR); tests/responses-prewarm-generate.test.ts (#2966) pins that
expand()stripsgenerate:falsefrom a committed prewarm body — the active defense against delta-continuation poisoning. - Call sites: all
installWebSocketBridge/startInlineWsBridgecall sites verified; the param is optional; noisEgressProxiedresidue.
Code (③): the predicate is conservative (items lacking a role → not a probe); the dispatcher is passed to undici's UpstreamWebSocket (src/responses-ws.ts:387); no as any/@ts-ignore; no console.error in server modules. Clean.
Tests (④): the three new tests are genuine regression tests — each goes red before the fix. The fixture now answers any generate:false frame with an empty terminal (matching real upstream). The pi-inline-ws test proves the bypass (dispatchHttp throws if called).
Diff (⑤): 8 files, +186/−26; every line relates to the PR's purpose. No unrelated changes, no whitespace churn.
八问:
- 是什么问题: codex 的 context-only 预热帧(generate:false、不回放对话 turn)被当成完整 ACP turn 执行——清零并落盘会话折叠状态、往不产出的请求里塞整份历史摘要和工具;且探针出站绕过配置的上游代理,配了代理的机器上握手失败。
- 什么原因导致的: 两个缺陷同属「探针识别 + 探针出站」机制家族:① gate 谓词只认「空 input」为探针,只带 developer/user 上下文的帧漏进 processTurn;② 探针 egress 走直连而非上游代理决策。
- 影响范围与严重程度: 所有 codex 会话的启动/空闲预热;配了上游代理的机器上预热直接握手失败;长会话折叠状态被清零 → 成本爆炸(p1 类①:状态丢失)。
- 怎么修复的: 把探针认定从「空 input」放宽到「不回放对话 turn 的 context-only 帧」(isPrewarmProbe),让这些帧走 probe relay 逐字转发;probe relay 的 egress 改用与 pipeline 同源的 resolveProxyDecision 决策(egressProxyFor),共享同一 ProxyAgent 与上游 socket。
- 是否有回归问题: 无。旧行为「空 input 探针」是新行为的严格子集;history-replaying prewarm 仍走 pipeline(负向测试钉死);tunnel guard 布尔语义逐位一致;pipeline 与 probe 用同一代理决策,无漂移。
- 是否破坏缓存: 不破坏。probe 与 pipeline turn 共享同一 ProxyAgent 缓存(#3776)与同一上游 socket(连接键不含 dispatcher,决策按 URL 确定性),反而修复了 PR 前 direct/probe-vs-proxy/turn 的 socket 分裂。
- 是否改了设计方向: 不改。仍是「probe relay 逐字转发 + pipeline 走 ACP」双轨,只是探针认定范围放宽 + egress 路由对齐。
- 是否引入了新的配置: 无。无新 env/config/默认值。
Verdict: APPROVE — no kernel/ changes, no config surface, no blockers.
|
(ai: vllm-qwen/qwen3.8-27b) 已顺手修:tests/responses-prewarm-generate.test.ts:92-97 的过期注释(描述 #3342 之前 + 本 PR 之前的行为:声称空 input 探针「不提交 checkpoint」且非空 prewarm「走 ACP pipeline」)已改为与 post-#3342 probe-relay 行为一致 + commit e7dd721 Review lane 收工:PR #4595 已 APPROVE(无 kernel/ 改动、无 config surface、无阻断项);唯一 non-blocking 发现(过期注释)已就地修复并推送到 PR 分支(新 head e7dd721),CI 将在新 head 上重跑。 |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) PR #4595 review complete — APPROVE. What I did:
Why APPROVE:
Checks: read from the mirror cache (sha 八问:
收工判定: APPROVE — 无 kernel 改动、无 config surface、无阻断项;唯一 non-blocking 发现(过期注释)已就地修复并推送(新 head |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Status confirmation — PR #4595 review is complete and final. (The prior comment in this thread was my own receipt line synced back from GitHub; no new request was pending.)
Nothing further from the review lane. Awaiting owner merge. |
|
(ai: vllm-qwen/qwen3.8-27b) Superseded by #7322 — this fix was cherry-picked (commit cdda1b3, |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Acknowledged — #4595 is superseded by #7322 (fix cherry-picked as My review of #4595 is complete and recorded (APPROVE, No further action from the #4595 review lane; #7322 carries its own review. |
Conflict resolution: src/responses-ws.ts imports — kept the PR's proxyDispatcher import (egress-proxy routing, used at the fetch dispatcher) and master's sseEventData addition to the sse-util import. Both symbols are used in the merged tree. No other conflicts. Final reviewed head (full CI pass).
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) Conflict resolved and pushed — new head What happened: the PR went dirty again — master advanced The one conflict —
Resolution: kept both — the PR's import { proxyDispatcher } from "./upstream-proxy.js";
import { normalizeSseLineEndings, finalizeSseLineEndings, sseEventData } from "./sse-util.js";Both symbols are used in the merged tree ( The other 72 files auto-merged (master's 73 commits). Push: verified the fork head was still 八问
判定:冲突已解决,PR head 现为 |
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Not approving for merge — this PR is superseded by #7322.
The superseded label and the 2026-11-19 thread comment are unambiguous: the #4594 fix was cherry-picked (commit cdda1b3b301, -x, authorship preserved) into the root-cause aggregation PR #7322 (fix(codex-lane): context-only prewarm probe relay + hosted web_search_call ordering (#4594, #7296)), which is now the delivery vehicle. I verified #7322 is open and carries the cherry-picked fix (rebased onto current master). Merging #4595 on its own would duplicate that fix.
Technical state of the new head 3968639008b (pushed 2026-11-20; a merge of 73 master commits + one import-block conflict resolution in src/responses-ws.ts):
- The conflict resolution is sound — it keeps both the PR's
proxyDispatcherimport and master'ssseEventDataimport; both symbols are used in the merged tree. - The PR's core change is intact:
isPrewarmProbe()(src/responses-ws.ts:81), theegressProxyForrouting (src/ws-bridge.ts:53,92,149,158; src/server.ts:825; src/inline-engine.ts:469; src/inline-ws.ts:40,57), and the four regression tests (3 in tests/codex-responses-ws.test.ts, 1 in tests/pi-inline-ws.test.ts). - No
kernel/,src/update.ts, config-surface, or new-env-var changes. - CI (mirror cache, sha
3968639008b): 0 failed, 3 test shards still in progress.
The prior review (APPROVE, pullrequestreview-5969832172, head e7dd72123a3) already covered the technical content; this head adds only the master merge + import resolution, which I verified above.
Recommendation: close #4595 (superseded) in favor of #7322. I cannot close it (human-only) and will not approve it for merge.
八问:
- 是什么问题: codex 的 context-only 预热帧(generate:false、不回放对话 turn)被当成完整 ACP turn 执行——清零并落盘会话折叠状态、往不产出的请求里塞整份历史摘要和工具;且探针出站绕过配置的上游代理,配了代理的机器上握手失败。
- 什么原因导致的: 两个缺陷同属「探针识别 + 探针出站」机制家族:① gate 谓词只认「空 input」为探针,只带 developer/user 上下文的帧漏进 processTurn;② 探针 egress 走直连而非上游代理决策。
- 影响范围与严重程度: 所有 codex 会话的启动/空闲预热;配了上游代理的机器上预热直接握手失败;长会话折叠状态被清零 → 成本爆炸(p1 类①:状态丢失)。
- 怎么修复的: 把探针认定从「空 input」放宽到「不回放对话 turn 的 context-only 帧」(isPrewarmProbe),让这些帧走 probe relay 逐字转发;probe relay 的 egress 改用与 pipeline 同源的 resolveProxyDecision 决策(egressProxyFor),共享同一 ProxyAgent 与上游 socket。
- 是否有回归问题: 无。旧行为「空 input 探针」是新行为的严格子集;history-replaying prewarm 仍走 pipeline(负向测试钉死);tunnel guard 布尔语义逐位一致;pipeline 与 probe 用同一代理决策,无漂移。
- 是否破坏缓存: 不破坏。probe 与 pipeline turn 共享同一 ProxyAgent 缓存(#3776)与同一上游 socket,反而修复了 PR 前 direct/probe-vs-proxy/turn 的 socket 分裂。
- 是否改了设计方向: 不改。仍是「probe relay 逐字转发 + pipeline 走 ACP」双轨,只是探针认定范围放宽 + egress 路由对齐。
- 是否引入了新的配置: 无。无新 env/config/默认值。
收工判定: 不批准合并——本 PR 已被 #7322 取代(cherry-pick cdda1b3b301,authorship 保留,#7322 为交付载体)。建议 owner 关闭 #4595、走 #7322。本 PR 侧无待办。
[bot] 🏷 Review lane: PR #4595 reviewed end-to-end (head Verdict: COMMENT (not approving for merge) — PR is superseded by #7322. The Technical state of the new head
The prior review (APPROVE, Recommendation: close #4595 (superseded) in favor of #7322. I cannot close it (human-only) and will not approve it for merge. 八问:
收工判定: 不批准合并——本 PR 已被 #7322 取代(cherry-pick GitHub review: pullrequestreview-5994135319 |
[ework] CI fix verified green ✓ |
4 similar comments
[ework] CI fix verified green ✓ |
[ework] CI fix verified green ✓ |
[ework] CI fix verified green ✓ |
[ework] CI fix verified green ✓ |
(ai: anthropic/claude-opus-5-5)
Fixes #4594
What
src/responses-ws.ts: agenerate:falserequest that replays no conversation turn (empty input, or only developer/system/user messages) is a prewarm probe. It is forwarded verbatim on the upstream transport and never run as an ACP turn. Before, only empty input matched, so codex's startup / idle-thread warmup (PrewarmInput::Base: its context items plusgenerate:false) became a zero-message turn. That turn deactivated every fold (and persisted it), re-injected the wholecodexForgedSummarieslist, and added ACP tools.proxyDispatcher(egressProxyFor(url))).ws-bridgeexposes the resolved proxy (egressProxyFor) instead of the booleanisEgressProxied; the tunnel guard derives its boolean from the same call, so both share one decision.prewarm_with_history, which replay real turns) are unchanged and stay on the pipeline.Why
Details in #4594. In the field, a machine restart made every resumed codex thread send a two-item prewarm. Seven sessions were saved with 0 active blocks. On a long session the re-injected forged summaries (~1.2M tokens) turned the prewarm into a 413, which made codex force a pre-turn compaction. The gate (0 active blocks) then forwarded that compaction verbatim (17.5 MB), and the upstream closed the WebSocket with 1009. This is also the trigger shape in #4036; #4040 (gate judged on the request's own view) stays complementary.
Tests
tests/codex-responses-ws.test.ts(12/12):generate:falseprewarm on a session with an active fold and a forged summary: upstream receives the frame verbatim (no forged summary, no ACP tools),stats.requestsis unchanged, and the fold stays active. Red before the fix: the prewarm carried bili's developer message with the forged summary.generate:falseprewarm still rides the pipeline (ACP tools injected, counted as a request). Green before and after.generate:falseframe with an empty terminal, as the real upstream does.npm run typecheckandnpm run buildare clean.exec, thenexec resume --last) against a fake WebSocket upstream:processTurn: 0 msgs+degenerate-turn, 36.7→52.8 KiB, tools 9→14unused-exports,config-surface-guardandconfig-docs-coveragefail identically on mastercd92b0b7, and master's test shards are red from epoch-confirm latch semantics: train a5d1994d test born-red (all-missing) + regression of 2915 escalation-latch-recovery — owner decision needed #4555 (fix in [PR] test: align epoch-confirm/escalation-latch expectations with the merged #3641 contract (#4555) #4569). None of these come from this diff.Scope
No config, persistence, kernel or message-shape change, and no new env var. Not addressed here:
codexForgedSummariesis append-only and is re-injected whole whenever the echo is missing; the WebSocket close-1009 oversize error is not relayed to the client (it waits for its idle timeout) and reports 0.0 MiB for non-string bodies.🤖 Generated with Claude Code