fix(llm): forward thinking on Anthropic requests and drop sampling wi…

[renanalmd]

Paired: GitHub #483 ↔ GitCode !2297

What type of PR is this?

/kind bug

What does this PR do / why do we need it:

• _build_anthropic_params rebuilds its payload from an explicit key set and dropped thinking.
• Forward thinking, size max_tokens for the budget plus answer headroom, and omit sampling knobs when thinking is present.

Which issue(s) this PR fixes:

Fixes openJiuwen-ai/jiuwenswarm#2710

What scenarios were tested, and what were the verification results（Function, performance, reliability, etc.）：

• Unit tests: 63 passed in tests/unit_tests/core/foundation/llm/test_anthropic_model_client.py
  • Forwards thinking with budget_tokens.
  • Raises max_tokens when the budget would leave less than 4096 answer tokens.
  • Drops sampling knobs when thinking is enabled.
  • Preserves sampling behavior when thinking is absent.

Self-checklist:（Please check carefully,and mark an x in the [] brackets. We will review your completion status.）

• • Design: Has the solution corresponding to the PR been reviewed by the Maintainer, and have all review comments been replied to and revised
• • Test: Has the code in the PR been fully covered by UT/ST test cases, and have the newly added test cases been uploaded to the repository along with this PR or already uploaded.
• • Verification: Does the PR description contains a detailed description of the verification results regarding the achievement of the expected goals for the Feature, Refactor, and Bug Fix to this PR.
• • Interface: Does it involve changes to external interfaces? The corresponding changes have been approved by the interface review organization, and the annotation information for the API has been correctly refreshed.
• • Document: Does it involve modifications to the official website documentation? If so, please submit the materials to the Doc repository in a timely manner.

[openjiuwen-collaboration-bot]

head_sha: 703e6efe40ef8015569e8edc3a8979309b21ee7d

变更摘要

此 PR 修复了 AnthropicModelClient._build_anthropic_params 中因显式键集合重建请求体而遗漏 thinking 参数的缺陷，使扩展思考（extended thinking）功能能够正确转发给 Anthropic API。同时，当传入 thinking 时自动提升 max_tokens 以确保预算外有至少 4096 个 token 的回答空间，并移除不兼容的采样参数（temperature/top_p），避免 API 返回 400 错误。

主要改动

• 转发 thinking 参数：在 _build_anthropic_params 中新增对 openai_params["thinking"] 的提取与赋值，使调用方传入的思考配置能正确传递到 Anthropic 请求体，不再被静默丢弃。
• max_tokens 自动扩容：当 thinking 包含 budget_tokens 且当前 max_tokens 减去预算后不足 _THINKING_MAX_TOKENS_HEADROOM（4096）时，自动将 max_tokens 提升为 budget + 4096，避免 API 拒绝请求或回答空间过小。
• 思考模式下抑制采样参数：当 thinking 存在时，不再向请求体中写入 temperature 和 top_p，因为扩展思考模式不允许非默认采样参数；仅在 thinking 为 None 时沿用原有的 temperature 优先于 top_p 的转发逻辑。
• 引入常量 _THINKING_MAX_TOKENS_HEADROOM：模块级常量定义回答头空间为 4096 token，用于 max_tokens 扩容判断，兼顾可读性和可维护性。
• 新增单元测试 TestThinkingForward：在 test_anthropic_model_client.py 中新增 8 个测试用例，覆盖 thinking 转发、max_tokens 提升、预算不足时的兜底行为、异常 budget 值的健壮性，以及思考模式下采样参数的排除。

[openjiuwen-collaboration-bot]

head_sha: 703e6efe40ef8015569e8edc3a8979309b21ee7d

代码审查

✅ 未发现问题

[openjiuwen-collaboration-bot]

head_sha: 703e6efe40ef8015569e8edc3a8979309b21ee7d

任务名称	结果	日志操作
静态检查	✅SUCCESS	点此跳转
防投毒检查	✅SUCCESS	点此跳转
开源合规检查	✅SUCCESS	点此跳转
UT测试	✅SUCCESS	点此跳转
ST测试	N/A	N/A
build 编译包	N/A	N/A
ruff codecheck	✅SUCCESS	N/A