[Bug] 自定义 OpenAI-compatible 模型将 context window 误用为 max_tokens，导致 Ollama GLM-5.2 400

提交前确认 · Pre-submission checklist

• 我已搜索过现有 issue，确认这不是重复 / I searched existing issues and confirmed this isn't a duplicate.
• 我已阅读 CONTRIBUTING.md / I've read CONTRIBUTING.md.

问题类别 · Category

模型设置 / 切换 · Model config

涉及的 Agent 框架 · Agent framework

ZCode Agent（自研）

严重程度 · Severity

阻塞使用 · Blocking

复现频率 · Reproducibility

必现 · Always

问题描述 · Description

在 ZCode Desktop 的 Model settings 中添加自定义 OpenAI chat completions (/v1/chat/completions) provider，接入本地 Ollama OpenAI-compatible endpoint 的 glm-5.2:cloud 时，如果按模型真实能力把 Context window 配成 200000 或 1000000，ZCode 会向 provider 发出过大的 max_tokens，导致请求被 Ollama 拒绝：

Provider rejected the model request.
Turn execution failed
Bad Request
provider=a9fb69dd-4c38-46b0-a9b4-74f10a9f3969 model=glm-5.2:cloud request=072b98c6-8e78-4d99-a6e4-d9fd03716c3f reason=invalid_request status=400 retryable=false

这看起来是模型设置里的 Context window 被用于输出上限 max_tokens 了。GLM-5.2 的上下文窗口约 976K/1M，但单次最大输出是 131072；二者不应共用同一个字段。

相关但不重复：#94 请求支持自定义 Payload。本 issue 更具体，指向现有模型设置字段到请求参数的映射 bug。

复现步骤 · Steps to reproduce

1. 打开 ZCode Desktop → Settings → Model settings。
2. Add provider：
   • Name: Ollama GLM
   • Base URL: http://localhost:11434/v1
   • API key: ollama
   • API format: OpenAI chat completions (/v1/chat/completions)
3. Add model：
   • Model ID: glm-5.2:cloud
   • Context window: 200000 或 1000000
4. 在 ZCode Agent 中选择该 custom provider/model，发送任意 prompt。
5. 看到 provider 400 / invalid_request。

期望表现 · Expected behavior

ZCode 应该区分：

{
  "limit": {
    "context": 1000000,
    "output": 120000
  }
}

或在 UI 中分别提供：

• Context window / input context limit：例如 1000000
• Max output tokens：例如 64000、120000，且不超过 provider/model 输出上限 131072

请求 /v1/chat/completions 时，max_tokens 应来自 output limit，而不是 context window。

如果使用 /v1/responses，也应使用 max_output_tokens 表示输出上限；context window 仍不应直接映射到 max_output_tokens。

实际表现 · Actual behavior

当 model 的 Context window 设置为 200000 时，底层 OpenAI-compatible endpoint 返回：

max_tokens (200000) exceeds model's maximum output tokens (131072) for model glm-5.2

当设置为 1000000 时同理：

max_tokens (1000000) exceeds model's maximum output tokens (131072) for model glm-5.2

本地将 ZCode 配置手动改为同时包含 context 和 output 后可绕过：

"glm-5.2:cloud": {
  "limit": {
    "context": 1000000,
    "output": 120000
  },
  "modalities": {
    "input": ["text"],
    "output": ["text"]
  }
}

ZCode 版本 · ZCode version

• ZCode Desktop: log shows current app initialized as 3.1.3; updater downloaded 3.2.5 during the same run.
• bundled ZCode CLI: 0.15.0

设备 / 系统 / 浏览器 · Device / OS / Browser

• macOS 26.3, Apple Silicon
• ZCode Desktop App
• Ollama OpenAI-compatible endpoint at http://localhost:11434/v1

截图 / 录屏 / 日志 · Screenshots / Recordings / Logs

ZCode log excerpt:

{
  "event": "model.request.failed",
  "baseURL": "http://localhost:11434/v1",
  "modelId": "glm-5.2:cloud",
  "providerId": "a9fb69dd-4c38-46b0-a9b4-74f10a9f3969",
  "providerKind": "openai-compatible",
  "requestId": "072b98c6-8e78-4d99-a6e4-d9fd03716c3f",
  "reason": "invalid_request",
  "statusCode": 400,
  "statusMessage": "Provider rejected the model request."
}

Local minimal reproduction against Ollama OpenAI-compatible endpoint:

POST http://localhost:11434/v1/chat/completions
model=glm-5.2:cloud
max_tokens=200000 -> 400
max_tokens=131072 -> 200
max_tokens=120000 -> 200

/v1/responses also works, but has the same output-token cap:

POST http://localhost:11434/v1/responses
model=glm-5.2:cloud
max_output_tokens=200000 -> 400
max_output_tokens=131072 -> 200
max_output_tokens=120000 -> 200

References:

• Ollama OpenAI compatibility docs: https://docs.ollama.com/api/openai-compatibility
  • /v1/chat/completions supports max_tokens.
  • /v1/responses supports max_output_tokens.
  • The OpenAI API does not have a way to set model context size directly.
• GLM-5.2 model page: https://ollama.com/library/glm-5.2
  • lists 976K context window / 1M long-context capability.

Suggested fix:

1. Add a separate output-token field in custom model settings, or infer a safe output cap from model metadata.
2. Store both limit.context and limit.output for custom provider models.
3. For OpenAI-compatible chat completions, map only limit.output to max_tokens.
4. For OpenAI Responses API, map only limit.output to max_output_tokens.
5. Keep limit.context for model/context budgeting and UI display, not as a request output limit.

Start at the ZCode Desktop Model settings and the OpenAI-compatible /v1/chat/completions and /v1/responses request paths described in the report. Trace how custom model limits are stored and mapped; done means context and output limits remain distinct and requests use only the output limit for max_tokens or max_output_tokens, with the reported Ollama reproduction passing.