Repository navigation
[PR] docs(zcode): clarify the zai-vs-bigmodel compression dead-end in CLIENTS Known limitations - #4610
[PR] docs(zcode): clarify the zai-vs-bigmodel compression dead-end in CLIENTS Known limitations#4610ranxianglei wants to merge 9 commits into
Conversation
…n limitations (#4592) The #661 anti-fraud wall is described as hitting 'login traffic' and the doc blanket-recommends 'use GUI cert-MITM for compression on these accounts', which contradicts the required mitm://zcode.z.ai passthrough route for zai-series accounts whose model endpoint IS zcode.z.ai (kernel -> 405/3012, passthrough -> no compression). Clarify that #661 trips the model endpoint, and split bigmodel-series (open.bigmodel.cn, compressible via cert-MITM) from zai-series (zcode.z.ai, must passthrough, no compression today).
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.6174-pr.4610.c8638ffThis exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr4610.tgz
npm install -g packageInstall instructions are refreshed on each push. |
(ai: vllm-qwen/qwen3.8-27b) 自动修复判定(attempt 1/2):继承 master 常红,不在本分支修——等 #4606 合入后 rebase。 八问:
结论:本轮判定+停手,未向分支 push 任何提交;已在 #4599 留一行继承记录;待 #4606 合入后自动 rebase/重触发。 |
(ai: vllm-qwen/qwen3.8-27b) 查重: 不重复 —— 本 PR 是 #4593(zcode 风控死路)文档面的唯一实现(#4593 线程已记录该 PR);近期无其他 PR 触碰 CLIENTS.md 的 ZCode Known limitations 段。 结构化八问
事实核对(独立重查,非复读 issue 结论)
顺手修掉的小问题(已在 PR 分支补 commit)引用号写错:正文与 PR 描述里的 CIhead 上仅 结论:可合。 内容核实通过、镜像一致、无行为变更;唯一的红是继承性 master 红,master 转绿后(或直接合入,文档不影响任何测试)即可合并。 |
(ai: vllm-qwen/qwen3.8-27b) 自动修复判定(attempt 2/2,末轮):同一继承红,本分支不动手——自动修复通道到此结束。 新 head 八问(简版,完整证据链见 attempt 1 评论):
结论:2/2 自动修复轮次用尽,未向分支 push 任何提交。解锁路径:#4606 转绿并合入后,本分支 rebase 重触发即可全绿;在此之前本 PR 的 |
master window commits 293ae65/de5e03659 bumped kernel/package.json to 0.0.165 without regenerating the lockfiles; lock-sync re-resolution then rewrites the stale recorded versions (root 0.0.163, kernel 0.0.164) and fails. Pure recorded-version sync via npm install --package-lock-only; no dependency or semantic change.
Refs #4593 (documentation facet only — the issue stays open for the design discussion on byte-faithful routing / ZCode-side fix).
What
The ZCode "Known limitations" block was self-contradictory for zai-series coding-plan accounts whose model endpoint is
zcode.z.ai:/api/v1/zcode-plan/anthropic/v1/messages;mitm://zcode.z.ai: {passthrough:true}route" (billion-context ACP kernel 改写 body 导致 ZCode (zcode.z.ai) 405/3012 #661 paragraph) — mutually exclusive for zai-series, i.e. kernel → 405/3012 vs passthrough → no compression.Change (docs only, en + zh mirror)
Clarify that #661 rejects kernel-rebuilt bodies on the
zcode.z.aimodel endpoint, and split the two coding-plan families by endpoint:open.bigmodel.cn, outside the billion-context ACP kernel 改写 body 导致 ZCode (zcode.z.ai) 405/3012 #661 scope): cert-MITM + kernel compresses normally;zcode.z.ai, inside the billion-context ACP kernel 改写 body 导致 ZCode (zcode.z.ai) 405/3012 #661 scope): any kernel pass over model traffic 405s even with zero folds folded, so the MITM setup must keep the passthrough route → working-but-uncompressed; no config today both compresses and avoids the 405 for zai-series model traffic.Behavior
None. Two
.mdfiles only (CLIENTS.md,CLIENTS.zh-CN.md). No code, config schema, wire, persistence, or version change.