Skip to content

[PR] docs(codex): Codex Desktop Composer predictions — custom-provider threads are ineligible (#3408) - #3409

Merged
ranxianglei merged 1 commit into
masterfrom
2026-10-16_docs-3408-codex-desktop-predictions
Oct 20, 2026
Merged

ranxianglei merged 1 commit into
masterfrom
2026-10-16_docs-3408-codex-desktop-predictions

Conversation

@ranxianglei

Copy link
Copy Markdown
Owner

Closes #3408

What

Adds one Codex subsection to CLIENTS.md (mirrored in CLIENTS.zh-CN.md) documenting the client-side Composer-prediction eligibility gate observed on Codex Desktop 26.1013.3656.0:

  1. The limitation — prediction eligibility checks the source thread's actual provider and requires the built-in openai provider; any other provider selects unsupported_provider and the client returns before generating a prediction. This happens entirely client-side, before any model request could reach billion-context — so missing predictions on custom-provider routes are expected client behavior, not a bili forwarding defect.
  2. Supported routing options — provider-preserving transport (cert-MITM via HTTPS_PROXY + SSL_CERT_FILE, what the bili codex launcher injects) intercepts below the provider layer, so the thread keeps its built-in openai identity + ChatGPT auth — mechanically the only posture that can keep eligibility intact. The doc states this is runtime-unverified on a Desktop build (no matched A/B yet) and invites confirmation reports back to Codex Desktop Composer predictions: document custom-provider eligibility limitation and supported routing options #3408.
  3. Explicit no-supported-approach statement — no configuration restores predictions on custom-provider threads; renaming a provider to impersonate openai or patching the client check is unsupported and not documented as a workaround.

Scope & verification

@github-actions

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-10-16_docs-3408-codex-desktop-predictions (6c56f5e)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.4006-pr.3409.52009b8

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-pr3409.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

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

查重:不重复。 已检索 open/closed issue 与近期 PR——#3408 本身就是本 PR 要关闭的源 issue(文档请求),仓库内无其他针对「Codex Desktop Composer 预测资格」的在途工作;相关但不同的 #3176/#3178(一键安装)作者已在 PR 描述里说明为何不并入本文档,判定成立。

审核结论(独立审过,非转述 issue)

按 PR review 规范逐项核过,不是照抄 issue 分析:

① 方向关:issue #3408 的三项诉求(记录客户端门槛、澄清受支持的路由姿态、无解时明说)与 PR 内容一一对应,方向正确。

② 事实核对(我逐条对过代码,不是只读 diff):

  • 「bili codex 启动器注入 HTTPS_PROXY + SSL_CERT_FILE」→ 属实:src/launcher.ts:13 及 codex 段(~L759,Go net/http honors HTTPS_PROXY + SSL_CERT_FILE)。
  • 「这类条目同时废掉订阅登录,见上文」→ 交叉引用有效:CLIENTS.md:407-409 现有段落原文即此陈述。
  • 「bili 侧无任何预测/composer 处理」→ grep 全 src/、tests/:零相关命中(仅两条无关注释),与「客户端行为、非转发缺陷」的定位一致。
  • 位置与镜像:新小节落在 ## Codex 节内、Windows 启动路径小节之后(CLIENTS.md:502),zh-CN 同位镜像(CLIENTS.zh-CN.md:141),en/zh 结构一致,标题风格与兄弟小节(### …(#NNNN))统一。
  • 对冲等级与证据强度匹配(关键项):静态检查证据明确归属「Desktop 26.1013.3656.0,Codex Desktop Composer predictions: document custom-provider eligibility limitation and supported routing options #3408」;launcher 传输保住资格的论断保留「runtime-unverified on a Desktop build(no matched A/B yet)」并邀请回报 Codex Desktop Composer predictions: document custom-provider eligibility limitation and supported routing options #3408——没有把未验证的运行时行为写成既成事实。这是本文档最容易越界的地方,处理得当。

③ 回归审查:纯插入 2 文件 20 行,现有段落零改动,不触碰任何代码/配置/wire 路径——既有行为清单为空,无回归面。

⑤ diff 整洁度:干净,每一行都服务于本 PR 目的;无关改动/空行重排:无。

CI:全部 gate 绿(static-gates、kernel-guard、docs-coverage、lock-sync、lane-guard、test-import-graph 等);real-client lanes 正确 skipped(未触碰其触发路径);test 分片 3/6 绿,剩 3 个(in_progress)——docs-only 变更,预计全绿,跑完我再补一句确认。

Rebase:master 领先 27 个提交,无一触碰这两个文件;试验性 rebase 干净(未推分支,BEHIND 不构成阻塞,GitHub 侧 mergeable=true)。

结构化八问

  1. 是什么问题:Codex Desktop 的 Composer 预测有一道客户端资格门槛——走自定义 provider 的线程不产生预测,且整件事发生在客户端、任何模型请求到达 bili 之前;用户此前容易误读成 bili 转发缺陷。
  2. 什么原因导致的:不是缺陷,是客户端设计——预测资格检查读源线程的实际 provider 并要求内置 openai(reporter 对闭源二进制的静态检查,bili 侧本就无此机制,已验证)。
  3. 影响范围与严重程度:纯文档补充,运行时无任何影响;价值在于消除一类误报(「自定义 provider 下预测消失 = bili bug」)。
  4. 怎么修复的:CLIENTS.md Codex 节加一小节(zh 镜像),三件事说清:门槛存在且是客户端行为;cert-MITM 传输在机制上保住资格(标注 runtime-unverified);无受支持配置可恢复,且不把改名冒充/打补丁文档化为绕行方案。
  5. 是否有回归问题:无——纯插入,未触碰任何代码路径,现有措辞零改动。
  6. 是否破坏缓存:不涉及。
  7. 是否改了设计方向:不涉及。
  8. 是否引入了新的配置:无。

判定:可合并(待剩余 3 个 test 分片全绿后补最终确认;docs-only,风险面为零)。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

CI 跟进(接上一条评论的审核结论;八问结论维持不变,只补充 Q5 回归项的最终判定与处置):

红的原因——不是本 PR 的回归。test (windows-latest, 22, 1) 唯一失败是
not ok 831 - retired stall guard: mid-stream silence survives any short window (#1706/#1714)
(error 'first chunk should have flowed through',subtest duration 63.7s)。本 PR 只改 CLIENTS.md ×2,不可能影响 socket 行为;真因是 tests/issue1452-socket-hardening.test.ts 里两处固定墙钟等待(:403 断言首个 SSE chunk 到达前固定睡 2s、:317 断言 [exposure] 遥测行前固定睡 400ms),在 runner 并行分片负载下(job 日志可见多个并发 bili 实例)被饿死而假失败。佐证:其余 5 个 test 分片全绿;master 约 1.5 小时前(73d9360c,CI run 38898555334)同一测试通过。

处置(几行级小修纪律:修复 PR + 记录 issue 各一):

当前 CI:已对失败 job 触发 rerun-failed-jobs;其余 17 个 job 全部 success,仅 test (windows-latest, 22, 1) 重跑中。绿了再报一声。

判定维持:可合并(docs-only、零行为面)。若 rerun 再次变红,继续排查并在本线程报告。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

Rerun 结果:test (windows-latest, 22, 1) 通过,本 PR 全部 18 个 job 现在 success,零失败。

最终判定:可合并——docs-only、零行为面、CI 全绿,随时可以合并:#3409

关联:顺带发现的测试等待 flake 已记录在 #3420([no-dispatch]),修复在 #3421(其 CI 正在 pushed head 上验证中)。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

补充一行:#3421(flake 修复)的 CI 已在 pushed head 上全绿(5 个 run 全部 success),无遗留事项。#3409 这边等你合并即可。

LoveDoLove pushed a commit to LoveDoLove-Forked-Projects/billion-context that referenced this pull request Oct 17, 2026
…t fixed sleep)

Discovered while reviewing PR ranxianglei#3409: its windows-latest(22,1) shard failed
on 'retired stall guard: mid-stream silence survives any short window'
(first chunk should have flowed through) — a fixed 2s wall-clock sleep that
starves under runner load. Convert both fixed-sleep waits in this file that
gate assertions on async events to the file's own deadline-poll idiom; the
mid-stream test keeps its 2s observation floor so the negative assertions
still cover a window far past the retired 400ms guard.
@ranxianglei
ranxianglei merged commit b2cbfcf into master Oct 20, 2026
55 of 56 checks passed
@ranxianglei ranxianglei changed the title docs(codex): Codex Desktop Composer predictions — custom-provider threads are ineligible (#3408) [PR] docs(codex): Codex Desktop Composer predictions — custom-provider threads are ineligible (#3408) Oct 20, 2026
@github-actions

Copy link
Copy Markdown
Contributor

📦 Merged Build Published

This merge to master was published as 0.0.665-master.b2cbfcf (rolling master tag).

npm install -g billion-context@0.0.665-master.b2cbfcf  # this exact build
npm install -g billion-context@master  # always the newest merged build

npm view billion-context@master prints the current version at any time.

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Codex Desktop Composer predictions: document custom-provider eligibility limitation and supported routing options

1 participant