fix(openai): send native tool definitions by default

[hannesrudolph]

Summary

Ensure OpenAI (openai-compatible) requests include native tool definitions by default for custom model info configurations.

Changes

• Merge native tool defaults into OpenAI model info so native tools are considered supported
• Add unit test asserting tools are passed through to the OpenAI request

---

Important

OpenAiHandler now defaults to including native tool definitions in requests by merging NATIVE_TOOL_DEFAULTS into model info, with a new test verifying this behavior.

• Behavior:
  • OpenAiHandler.getModel() now merges NATIVE_TOOL_DEFAULTS into model info, ensuring native tool support by default.
  • Requests include tool definitions if supportsNativeTools is true.
• Testing:
  • Added unit test in openai-native-tools.spec.ts to verify tools are included in requests when supportsNativeTools is not specified in custom model info.

^{This description was created by for a6aba35. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Oroocle   See task on Roo Cloud

Review complete.

• Update openai-native-tools.spec.ts to assert the real regression path (tools omitted when openAiCustomModelInfo lacks supportsNativeTools), so the test fails without the getModel() merge.

Previous reviews

• c8cfe10: Review #1

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}

[roomote]

This test will pass even without the getModel() change because OpenAiHandler.createMessage() always forwards tools when metadata.tools is present. The real regression fixed in this PR is earlier in Task.attemptApiRequest(), where metadata.tools is omitted when modelInfo.supportsNativeTools is missing from openAiCustomModelInfo. Consider setting openAiCustomModelInfo in this test without supportsNativeTools and only providing metadata.tools when handler.getModel().info.supportsNativeTools is true, so the test fails without this PR.

_{Fix it with Roo Code or mention @roomote and request a fix.}

[roomote]

Fixaroo   See task on Roo Cloud

Fixed the test to properly assert the regression path. The test now sets openAiCustomModelInfo without supportsNativeTools and only provides tools when handler.getModel().info.supportsNativeTools is true, mimicking Task.attemptApiRequest() behavior.

_{View commit | Revert commit}

[hem951740-max]

@roomote 你可以把原生协议作为openai的默认协议，但是为什么要删除xml协议呢，导致我现在的有些自定义模型无法使用了？？？
Date/time: 2025-12-24T05:18:08.508Z
Extension version: 3.37.1
Provider: openai
Model: claude-sonnet-4-5

400
OpenAI completion error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"..custom.input_schema: JSON schema is invalid. It must match JSON Schema draft 2020-12 (https://.org//*** Learn more about tool use at https://.com///

[roomote]

Fixaroo   See task on Roo Cloud

Thank you for reporting this issue. This PR has already been merged, so I cannot modify it directly.

Understanding the Problem:

You're using an OpenAI-compatible endpoint that routes to an Anthropic-style backend (based on your model name claude-sonnet-4-5). The PR change enables native OpenAI tool definitions by default, but your backend expects different tool schema formatting, causing the invalid_request_error.

Workaround:

You can override this default by explicitly setting supportsNativeTools: false in your custom model info configuration. This will prevent native tool definitions from being sent and should restore the previous behavior.

In your settings, configure openAiCustomModelInfo to include:

{
  "supportsNativeTools": false,
  "maxTokens": 4096,
  "contextWindow": 200000
}

For a Permanent Fix:

If the workaround doesn't resolve your issue, please open a new GitHub issue describing your use case in detail. The team can then evaluate whether the default behavior should be adjusted or if additional configuration options are needed for OpenAI-compatible endpoints that proxy to non-OpenAI backends.

[hem951740-max]

Fixaroo 查看 Roo Cloud 上的任务   

感谢您报告此问题。此 PR 已合并，因此我无法直接修改。

理解问题：

您使用的是一个与 OpenAI 兼容的端点，该端点会路由到 Anthropic 风格的后端（基于您的模型名称claude-sonnet-4-5）。PR 更改默认启用原生 OpenAI 工具定义，但您的后端需要不同的工具模式格式，从而导致此问题invalid_request_error。

变通方法：

您可以通过在自定义模型信息配置中显式设置来覆盖此默认设置supportsNativeTools: false。这将阻止发送原生工具定义，并应能恢复之前的行为。

在设置中，配置openAiCustomModelInfo为包含以下内容：

{
  "supportsNativeTools": false,
  "maxTokens": 4096,
  "contextWindow": 200000
}

永久解决方案：

如果此方法无法解决您的问题，请在 GitHub 上提交一个新的 issue，详细描述您的使用场景。团队随后可以评估是否需要调整默认行为，或者是否需要为代理到非 OpenAI 后端的 OpenAI 兼容端点添加额外的配置选项。

请告知我详细的设置方法，

是这样操作吗

[hem951740-max]

@roomote 请回复

[roomote]

Fixaroo   See task on Roo Cloud

This PR has already been merged and all review feedback was addressed. The review item to update the test to assert the real regression path was completed in commit a6aba35.

No further action is needed.