fix: carry reasoning effort through to GLM-5.3 on the OpenAI client path by greenhat616 · Pull Request #32 · TriDefender/zcode-api

greenhat616 · 2026-08-29T06:42:10Z

按 review 意见收窄:只保留 OpenAI 协议 reasoning_effort 的修复。

已移除固定模型目录的改动、update-models 探测脚本,以及随之而来的 README / config / tsconfig 改动。改动面从 18 个文件收敛到 7 个,只剩 src/translator 与 src/provider/reasoning.ts。

问题

coding-plan 的上游是 Anthropic 格式,所以 OpenAI 客户端的请求会走 translateRequestOpenAIToAnthropic。该函数只特判了 reasoning_effort:"none",其余取值一律退化成 {type:"enabled"} —— 既没有 budget,也没有 output_config。抓到的上游请求体:

客户端发送 reasoning_effort:"high", model glm-5.3
-> 上游收到 {"model":"glm-5.3","thinking":{"type":"enabled"}}
-> 实际只有 1 个字符的思考,而 effort:"high" 应有 ~244

Anthropic 端点会静默忽略 reasoning_effort,output_config.effort 才是唯一生效的通道,而且必须配对 thinking.budget_tokens —— 只给 effort 标签,上游仍停在自己近乎零的默认值上。上游对非法 effort 值和超过 max_tokens 的 budget 都不做校验,所以归一化只能放在这里做。

改动

新增 src/provider/reasoning.ts,把契约集中在一处:三档 effort 配对 budget 8000 / 16000 / 32000,默认档 max —— 取自 ZCode 自带目录里的 glm-5.3 条目。effort 归一化遵循 Z.AI 公布的映射表,向上取整:medium 映射到 high 而不是 low,这条有具名回归测试。
openai-to-anthropic(coding-plan 方向):GLM-5.3 系列请求一律带上显式 effort 与配对 budget,budget 会 fit 进 max_tokens 并为回答保留 1024 token。
anthropic-to-openai(start-plan 方向):两边通道相反,把 output_config.effort 带进 reasoning_effort。

默认值改动仅限 GLM-5.3 系列,均对齐 ZCode:未指定 reasoning_effort 时取 max,未指定 max_tokens 时回退到该模型目录里的 maxOutputTokens 而非通用的 4096。后者是前者的前提 —— ZCode 的 min(budget, maxOutputTokens - 1) 是针对模型上限写的,套到 4096 的兜底值上会把几乎整个额度吃进思考里。

其他模型行为完全不变,测试有断言。output_config 的透传也 gate 在 GLM-5.3 系列内。

关于模型目录

按 review 意见,本 PR 不改目录。这里的模型匹配走 model id 正则、不依赖目录收录,所以目录里没有的同系列模型(今天的 glm-5.3-flash)一样能拿到 effort 通道;只有 max_tokens 的兜底会查目录,查不到就退回通用的 4096,不会失败。两种情况都有测试覆盖。

Copilot review

explicitBudget > 0 之后才 floor —— 已改为先 floor 再判正。否则 0.5 能通过 > 0 却 floor 成 0,在 max_tokens 非有限数时会原样传给上游变成 budget_tokens: 0。补了 2 个回归测试。
output_config.effort 对所有模型透传 —— 已 gate 到 GLM-5.3 系列。补了 glm-4.7 的断言测试。
另外两条(update-models 的 CLI 参数校验、body-transformer 的 output_config 处理)所在代码已随本次收窄一并移除。

验证

bun test 635 pass / 0 fail,tsc --noEmit 通过。

Copilot

🟡 Changes recommended

There are a few concrete correctness issues in the new logic (notably CLI arg parsing and GLM-5.3 output_config/effort handling) that should be fixed before merging.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR expands the proxy’s pinned GLM model catalog (notably adding glm-5.3-flash) and fixes GLM-5.3-family “reasoning effort” handling by introducing a shared reasoning contract and wiring it through request translators and passthrough body transforms. It also adds a probing-based update-models script (plus tests) to help maintain src/provider/models.ts safely.

Changes:

Add a GLM-5.3 reasoning-effort contract (output_config.effort + thinking.budget_tokens) and apply it in OpenAI↔Anthropic translators and passthrough transforms.
Update the pinned model catalog/specs (incl. glm-5.3-flash) and align configs/docs/tests with the new catalog.
Add bun run update-models probing script + unit tests and include scripts/**/* in TS compilation.

File summaries

File	Description
tsconfig.json	Includes `scripts/*/` so the new probing script typechecks.
src/translator/types.ts	Adds `output_config` typing for Anthropic requests (GLM-5.3 effort channel).
src/translator/openai-to-anthropic.ts	Implements GLM-5.3 effort/budget translation and model-aware max_tokens defaulting.
src/translator/openai-to-anthropic.test.ts	Adds regression tests covering GLM-5.3(-flash) effort/budget behavior.
src/translator/anthropic-to-openai.ts	Carries `output_config.effort` into OpenAI `reasoning_effort` and rewrites disabled thinking for GLM-5.3.
src/translator/anthropic-to-openai.test.ts	New tests for `output_config.effort` carry-over and disabled-thinking rewrite.
src/proxy/body-transformer.ts	Adds passthrough-route rewrites for GLM-5.3 disabled thinking (OpenAI + Anthropic shapes).
src/proxy/body-transformer.test.ts	Adds tests for passthrough-route disabled-thinking rewrites and budget fitting.
src/provider/reasoning.ts	New module defining GLM-5.3 effort levels, budgets, model matching, and budget fitting.
src/provider/reasoning.test.ts	Unit tests for the GLM-5.3 reasoning contract helpers.
src/provider/providers.test.ts	Updates model catalog assertions (context windows, maxOutputTokens, reasoning flags, exclusions).
src/provider/models.ts	Updates pinned model list/specs; adds `glm-5.3-flash` and adjusts other model metadata.
src/config/template.ts	Updates template model list to match the updated pinned catalog.
scripts/update-models.ts	Adds probing-based catalog regeneration script with a two-stage availability/ceiling probe.
scripts/update-models.test.ts	Unit tests for the script’s pure functions (classification/parsing/merge policy).
README.md	Updates documented pinned model table/specs and adds notes about aliasing/availability.
package.json	Adds `update-models` script entry.
config.example.yaml	Updates example model list to match the updated pinned catalog.

Review details

Files reviewed: 17/18 changed files
Comments generated: 4
Review effort level: Lite

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

+      case "--provider": {
+        const v = argv[++i];
+        if (v !== "zai" && v !== "bigmodel") throw new Error(`--provider must be zai|bigmodel, got "${v}"`);
+        args.provider = v;
+        break;
+      }
+      case "--api-key":
+        args.apiKey = argv[++i];
+        break;
+      case "--config":
+        args.configPath = argv[++i];
+        break;


+    const explicitBudget = explicit.budget_tokens ?? explicit.budgetTokens;
+    if (typeof explicitBudget === "number" && Number.isFinite(explicitBudget) && explicitBudget > 0) {
+      budget = Math.floor(explicitBudget);
+    }


+function applyGlm53AnthropicThinkingCompat(body: Record<string, unknown>): boolean {
+  if (!glm53Model(body) || !thinkingIsDisabled(body)) return false;
+  const legal = buildGlm53Reasoning(GLM53_DISABLED_REPLACEMENT_EFFORT);
+  const budget = fitGlm53Budget(legal.thinking.budget_tokens, body.max_tokens);
+  body.thinking = budget === undefined
+    ? { type: "enabled" }
+    : { type: "enabled", budget_tokens: budget };
+  if (!isPlainObject(body.output_config)) {
+    body.output_config = legal.output_config;
+  }
+  return true;
+}


+  // GLM-5.3+ Anthropic clients set reasoning effort via `output_config.effort`
+  // (the Anthropic upstream ignores `reasoning_effort`); the OpenAI-format
+  // upstream on start-plan reads `reasoning_effort` instead, so carry it over.
+  if (req.output_config?.effort) {
+    result.reasoning_effort = req.output_config.effort as OpenAIChatRequest["reasoning_effort"];
+  }


TriDefender · 2026-08-29T10:45:44Z

请只提交一个修复OpenAI协议 reasoning_effort PR，模型列表和上下文问题均为 NOT A BUG
原因如下：
部分harness 例如 OpenCode 不接受自动探测到的模型能力，更多依赖于用户手动配置或者预先维护的模型能力列表
5v turbo你没有我有，完全是个人账号权限问题
proxy并不会阻止你去调用某个未被列出的模型，这是防止zai放出新模型我就得更新一次

On coding-plan the upstream is Anthropic-format, so an OpenAI client's request goes through translateRequestOpenAIToAnthropic. That function only special-cased reasoning_effort:"none"; every other value degraded to `{type:"enabled"}` with no budget and no output_config. A captured upstream body: client sent reasoning_effort:"high", model glm-5.3 -> upstream got {"model":"glm-5.3","thinking":{"type":"enabled"}} -> 1 character of thinking, where effort:"high" yields ~244 The Anthropic endpoint silently ignores reasoning_effort, so output_config.effort is the only channel that works there, and the upstream validates neither an illegal effort value nor a budget larger than max_tokens — normalizing has to happen here. New src/provider/reasoning.ts holds the contract in one place, taken from ZCode's catalog entry for glm-5.3: three levels with budgets 8000 / 16000 / 32000, defaultLevel "max". Effort normalization follows Z.AI's published table, which rounds up — "medium" maps to "high", not "low"; there is a named regression test for that. Also carries output_config.effort into reasoning_effort in the start-plan direction (Anthropic client, OpenAI upstream), where the channels are reversed. That carry-over is gated to the GLM-5.3 family: output_config is this family's extension, and forwarding it for every model would let a client reroute another model's reasoning through a channel it never opted into. Two default-value changes for the GLM-5.3 family only, both following ZCode: a request that omits reasoning_effort now gets effort "max" rather than the gateway's near-zero default, and one that omits max_tokens falls back to the model's catalog maxOutputTokens rather than the generic 4096. The second is load-bearing for the first — ZCode's clamp of `min(budget, maxOutputTokens - 1)` is written against the model ceiling, and applying it to a 4096 fallback left a 4095-token thinking budget with nothing for the answer. fitGlm53Budget now reserves 1024 tokens for the answer and floors a fractional max_tokens. Membership of the pinned catalog is a hint here, never a gate. Model matching is by id pattern, so a family member the catalog does not list (glm-5.3-flash today) still gets the effort channel; only the max_tokens fallback consults the catalog, and it degrades to the generic 4096 default rather than failing. Tests cover both. Net effect on a bare OpenAI request for glm-5.3: max_tokens 4096 and ~1 character of thinking becomes max_tokens 128_000 with a 32_000 thinking budget. That is ZCode parity, and it costs more.

…ns glm-5.3-flash Merge of master (1646a4c 'feat(models): advertise glm-5.3-flash') invalidated the premise that glm-5.3-flash is absent from the pinned catalog: the fixture now resolves the model-aware default (128,000) instead of the generic 4096. - Point the unlisted-family fixture at a genuinely unlisted id (glm-5.3-air), preserving the original intent: the effort channel is keyed off the model id, not catalog membership. - Pin the new catalog reality: glm-5.3-flash gets the same model-aware default as glm-5.3 (max_tokens 128,000, max-effort budget 32,000).

…nk 最小放宽 - CL-16: handler.ts 顶部注释重写为 v2.6 现实（两套餐均 Anthropic 上游， OpenAI 客户端 O→A 翻译，Anthropic 客户端原生 passthrough）； passthroughResponse docstring 改为如实描述 accept-encoding 转发； claim/client.ts 引用去掉不存在的 NOTEPAD 章节名 - R2-11: manager.ts 'refreshing if necessary' 过度承诺改为如实描述（凭据过期由上游 401 暴露，守卫保留备未来填充 expiresAt） - CL-17: 抽共享 src/proxy/inflate.ts inflateWithCap（DecompressionStream + 上限 + cancel + releaseLock）；proxy/handler.ts 与 async/handler.ts 两处解压循环改调，64MiB/4MiB 上限与错误契约（throw vs errorResponse）保留在各自调用方 - CL-15: /v1/responses userId 门控对齐 handler.ts（startPlan 不注入 metadata.user_id），留档 bundle 为 presence-gated 无套餐概念 - R2-18: parseSSEChunk data: 前缀放宽（至多剥一个前导空格，'data: x' 输出逐字节不变）；帧分割容忍 \r\n\r\n（导出 SSE_FRAME_SPLIT，与 3.11.2 bundle Kxo 正则一致），anthropicSseToOpenaiSse 与 async 流适配器的缓冲分割同步切换——CRLF 流不再永久缓冲；:496 第二解析器与其余解析点不动（#32）

Copilot AI lite review requested due to automatic review settings August 29, 2026 06:42

Copilot started reviewing on behalf of greenhat616 August 29, 2026 06:42

View session

Copilot AI reviewed Aug 29, 2026

View reviewed changes

TriDefender self-assigned this Aug 29, 2026

demian-bot mentioned this pull request Aug 29, 2026

fix(start-plan): route to the live Anthropic gateway #34

Merged

greenhat616 force-pushed the feat/glm-5.3-contract branch from 2bb2d5f to 675298b Compare August 29, 2026 16:30

greenhat616 changed the title ~~feat: add glm-5.3-flash and update-models scripts~~ fix: carry reasoning effort through to GLM-5.3 on the OpenAI client path Aug 29, 2026

TriDefender and others added 2 commits August 29, 2026 17:32

Merge branch 'master' into feat/glm-5.3-contract

b11ddc1

TriDefender merged commit e36b4d7 into TriDefender:master Aug 29, 2026
1 check passed

Navigation Menu

Repository navigation

fix: carry reasoning effort through to GLM-5.3 on the OpenAI client path - #32

fix: carry reasoning effort through to GLM-5.3 on the OpenAI client path#32
TriDefender merged 3 commits into
TriDefender:masterfrom
greenhat616:feat/glm-5.3-contract

greenhat616 commented Aug 29, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

TriDefender commented Aug 29, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

greenhat616 commented Aug 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

问题

改动

关于模型目录

Copilot review

验证

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

TriDefender commented Aug 29, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

greenhat616 commented Aug 29, 2026 •

edited

Loading