feat: adaptor reason effort convert

[zijiren233]

No description provided.

[Copilot]

Pull request overview

This PR introduces a unified “thinking / reasoning” compatibility layer across request modes (OpenAI Chat/Completions, OpenAI Responses, Gemini, Claude/Anthropic) and updates multiple adaptors to normalize and convert reasoning parameters consistently (with origin-model-first matching), backed by expanded documentation and tests.

Changes:

• Add normalized reasoning model + parsing/application helpers, and wire them into OpenAI/Gemini/Claude conversion paths (including Responses reasoning.effort).
• Update several adaptors (Ali/Doubao/DeepSeek/Zhipu/VertexAI/AWS/Azure/Baidu/Moonshot, etc.) to use origin-first model matching and to apply reasoning conversions/hooks.
• Add comprehensive docs (EN/ZH) describing the compatibility rules and conversion examples.

Reviewed changes

Copilot reviewed 50 out of 50 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
docs/REASONING_COMPATIBILITY.md	New English spec for reasoning/thinking normalization + conversion rules and examples.
docs/REASONING_COMPATIBILITY.zh.md	New Chinese spec mirroring the compatibility behavior and examples.
core/relay/utils/reasoning.go	Core implementation: parse/normalize/apply reasoning across OpenAI/Responses/Gemini/Claude + model-name resolution helpers.
core/relay/model/reasoning.go	Adds NormalizedReasoning + ReasoningEffort enum + normalization function.
core/relay/model/response.go	Adds reasoning to CreateResponseRequest for Responses API output.
core/relay/model/gemini.go	Extends Gemini thinking config to include thinkingLevel.
core/relay/model/completions.go	Adds OpenAI-compatible reasoning fields (reasoning_effort) and Ali fields to GeneralOpenAIRequest.
core/relay/model/claude.go	Adds reasoning fields and Claude adaptive/output_config structures.
core/relay/adaptor/openai/gemini.go	Adds hook support + maps Gemini thinkingConfig into OpenAI reasoning fields (and Responses).
core/relay/adaptor/openai/gemini_test.go	Adds tests verifying Gemini thinking -> OpenAI reasoning_effort.
core/relay/adaptor/openai/claude.go	Adds hook support + maps Claude thinking/output_config into OpenAI reasoning (and Responses).
core/relay/adaptor/openai/chat.go	Ensures Chat -> Responses conversion writes reasoning.effort.
core/relay/adaptor/openai/constants.go	Adds IsResponsesOnlyModelAny (origin-first, actual-fallback).
core/relay/adaptor/openai/chat_test.go	Tests for IsResponsesOnlyModelAny.
core/relay/adaptor/openai/adaptor.go	Uses IsResponsesOnlyModelAny across routing/conversion/response handling.
core/relay/adaptor/azure/main.go	Uses origin-first IsResponsesOnlyModelAny for Azure Responses routing.
core/relay/adaptor/zhipu/adaptor.go	Applies reasoning conversion into Zhipu’s simplified thinking.type.
core/relay/adaptor/zhipu/adaptor_test.go	Tests Chat/Gemini reasoning -> Zhipu thinking.
core/relay/adaptor/deepseek/adaptor.go	Adds Chat/Gemini reasoning conversion into DeepSeek thinking.type.
core/relay/adaptor/deepseek/adaptor_test.go	Tests reasoning conversions for DeepSeek.
core/relay/adaptor/doubao/main.go	Adds reasoning hook for Gemini/Anthropic -> Doubao simplified thinking + origin-first routing helper.
core/relay/adaptor/doubao/chat.go	Converts Chat reasoning_effort -> thinking + origin/actual deepseek-reasoner detection via AST callbacks.
core/relay/adaptor/doubao/main_test.go	Tests origin-first routing and reasoning conversions + deepseek-reasoner prompt injection.
core/relay/adaptor/ali/chat.go	Replaces legacy thinking patch with reasoning-based conversion; adds Gemini->Ali conversion with model-specific hooks.
core/relay/adaptor/ali/adaptor.go	Routes Gemini mode through Ali’s new Gemini conversion function.
core/relay/adaptor/ali/adaptor_test.go	Adds tests for Gemini->Ali reasoning and budget behavior.
core/relay/adaptor/gemini/openai.go	Applies normalized reasoning to Gemini thinkingConfig; origin-first feature model resolution.
core/relay/adaptor/gemini/adaptor.go	Uses origin-first logic for deciding Gemini API version selection.
core/relay/adaptor/gemini/openai_test.go	Adds tests for reasoning_effort -> Gemini thinkingConfig behavior.
core/relay/adaptor/vertexai/adaptor.go	Uses origin-first “feature model” to pick inner adaptor and route publishers/paths.
core/relay/adaptor/vertexai/adaptor_test.go	Tests origin-first routing decisions for VertexAI publisher endpoints.
core/relay/adaptor/vertexai/claude/adapter.go	Uses origin-first resolved model name when fixing betas.
core/relay/adaptor/vertexai/claude/adapter_test.go	Tests native-thinking preservation behavior.
core/relay/adaptor/aws/adaptor.go	Uses origin-first model selection with fallback for AWS inner adaptor selection.
core/relay/adaptor/aws/claude/adapter.go	Uses origin-first resolved model name for betas, context mgmt edits, and model ID selection.
core/relay/adaptor/aws/claude/adapter_test.go	Tests native-thinking preservation behavior.
core/relay/adaptor/antling/adaptor.go	Uses origin-first resolved model name when fixing Anthropic betas.
core/relay/adaptor/baidu/adaptor.go	Uses origin-first model selection with endpoint-map fallback.
core/relay/adaptor/baidu/adaptor_test.go	Tests origin-alias fallback to actual model for endpoint routing.
core/relay/adaptor/moonshot/adaptor.go	Explicitly rejects Responses modes while delegating other modes to OpenAI-compatible adaptor.
core/relay/adaptor/moonshot/adaptor_test.go	Tests Responses-mode rejection + supported modes.
core/relay/adaptor/anthropic/utils.go	Adds origin-first model resolver + adaptive-thinking detection and normalization logic.
core/relay/adaptor/anthropic/openai.go	Switches OpenAI->Claude conversion to normalized reasoning + adaptive normalization.
core/relay/adaptor/anthropic/main.go	Uses resolved model for default max_tokens; simplifies thinking handling.
core/relay/adaptor/anthropic/gemini.go	Applies Gemini thinkingConfig via normalized reasoning when converting Gemini->Claude.
core/relay/adaptor/anthropic/gemini_test.go	Tests adaptive normalization for Gemini->Claude path.
core/relay/adaptor/anthropic/adaptor.go	Uses resolved model name when fixing betas.
core/relay/adaptor/anthropic/openai_test.go	Adds coverage for adaptive-thinking model behavior and origin-first matching.
README.md	Adds link to new reasoning/thinking compatibility docs; removes contributing section.
README.zh.md	Adds link to new reasoning/thinking compatibility docs; removes contributing section.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

These test cases use new(16384), new(0), etc. In Go the built-in new takes a type, not a value, so this won’t compile. Use a small helper like intPtr(v int) *int (or assign to a local and take &local) when you need *int literals in the table.

[Copilot]

This test uses new("low") / new("medium") to create *string values. The built-in new takes a type (e.g. new(string)), so new("...") won’t compile. Prefer a helper like stringPtr(v string) *string (or a local variable + &local) for pointer literals.