Added support for external hooks to observe model registry events using the `ModelRegistryHook` interface. Implemented thread-safe, non-blocking execution of hooks with panic recovery. Comprehensive tests added to verify hook behavior during registration, unregistration, blocking, and panic scenarios.
- GLM-4.7: Uses extra_body={"thinking": {"type": "enabled"}, "clear_thinking": false}
- MiniMax-M2.1: Uses reasoning_split=true for OpenAI-style reasoning separation
- Added preserveReasoningContentInMessages() to support re-injection of reasoning
content in assistant message history for multi-turn conversations
- Added ThinkingSupport to MiniMax-M2.1 model definition
Per Google's official documentation, Gemini 3 models should use
thinkingLevel (string) instead of thinkingBudget (number) for
optimal performance.
From Google's Gemini Thinking docs:
> Use the thinkingLevel parameter with Gemini 3 models. While
> thinkingBudget is accepted for backwards compatibility, using
> it with Gemini 3 Pro may result in suboptimal performance.
Changes:
- Add model family detection functions (IsGemini3Model, IsGemini25Model,
IsGemini3ProModel, IsGemini3FlashModel)
- Add ApplyGeminiThinkingLevel and ApplyGeminiCLIThinkingLevel functions
for applying thinkingLevel config
- Add ValidateGemini3ThinkingLevel for model-specific level validation
- Add ThinkingBudgetToGemini3Level for backward compatibility conversion
- Update NormalizeGeminiThinkingBudget to convert budget to level for
Gemini 3 models
- Update ApplyDefaultThinkingIfNeeded to not set a default level for
Gemini 3 (lets API use its dynamic default "high")
- Update ConvertThinkingLevelToBudget to preserve thinkingLevel for
Gemini 3 models
- Add Levels field to all Gemini 3 model definitions:
- Gemini 3 Pro: ["low", "high"]
- Gemini 3 Flash: ["minimal", "low", "medium", "high"]
Backward compatibility:
- Gemini 2.5 models continue to use thinkingBudget as before
- If thinkingBudget is provided for Gemini 3, it's converted to the
appropriate thinkingLevel
- Existing configurations continue to work
NormalizeThinkingModel now checks ModelSupportsThinking before removing
"-thinking" or "-thinking-<ver>", avoiding accidental parsing of model
names where the suffix is part of the official id (e.g., kimi-k2-thinking,
qwen3-235b-a22b-thinking-2507).
The registry adds ThinkingSupport metadata for several models and
propagates it via ModelInfo (e.g., kimi-k2-thinking, deepseek-r1,
qwen3-235b-a22b-thinking-2507, minimax-m2), enabling accurate detection
of thinking-capable models and correcting base model inference.
- Added support for parsing and normalizing dynamic thinking model suffixes.
- Centralized budget resolution across executors and payload helpers.
- Retired legacy Gemini-specific thinking handlers in favor of unified logic.
- Updated executors to use metadata-based thinking configuration.
- Added `ResolveOriginalModel` utility for resolving normalized upstream models using request metadata.
- Updated executors (Gemini, Codex, iFlow, OpenAI, Qwen) to incorporate upstream model resolution and substitute model values in payloads and request URLs.
- Ensured fallbacks handle cases with missing or malformed metadata to derive models robustly.
- Refactored upstream model resolution to dynamically incorporate metadata for selecting and normalizing models.
- Improved handling of thinking configurations and model overrides in executors.
- Removed hardcoded thinking model entries and migrated logic to metadata-based resolution.
- Updated payload mutations to always include the resolved model.
- Added `ContextLength` field with a value of 200,000 to all applicable Claude model definitions.
- Standardized `MaxCompletionTokens` values across models for consistency and alignment.
Fixes an issue where Claude thinking models would return 400 errors when
the thinking.budget_tokens was greater than or equal to max_tokens.
Changes:
- Add MaxCompletionTokens: 128000 to all Claude thinking model definitions
- Add ensureMaxTokensForThinking() function in claude_executor.go that:
- Checks if thinking is enabled with a budget_tokens value
- Looks up the model's MaxCompletionTokens from the registry
- Ensures max_tokens is set to at least the model's MaxCompletionTokens
- Falls back to budget_tokens + 4000 buffer if registry lookup fails
This ensures Anthropic API constraint (max_tokens > thinking.budget_tokens)
is always satisfied when using extended thinking features.
Fixes: #339🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Added new `Gemini 3 Pro Image Preview` model with detailed metadata and configuration.
- Removed outdated `Claude Sonnet 4.5 Thinking` model definition for cleanup and relevance.
- Add Claude thinking model definitions (sonnet-4-5-thinking, opus-4-5-thinking variants)
- Add Thinking support for antigravity models with -thinking suffix
- Add injectThinkingConfig() for automatic thinking budget based on model suffix
- Add resolveUpstreamModel() mappings for thinking variants to actual Claude models
- Add extractAndRemoveBetas() to convert betas array to anthropic-beta header
- Update applyClaudeHeaders() to merge custom betas from request body
Closes#324
- Added `shouldLogRequest` helper to simplify path-based request logging logic.
- Updated middleware to skip management endpoints for improved security.
- Introduced an explicit `nil` logger check for minimal overhead.
- Updated dependencies in `go.mod`.
**feat(auth): add handling for 404 response with retry logic**
- Introduced support for 404 `not_found` status with a 12-hour backoff period.
- Updated `manager.go` to align state and status messages for 404 scenarios.
**refactor(translator): comment out debug logging in Gemini responses request**
Add gemini-3-pro-preview model to GetGeminiCLIModels() to make it
available for OAuth-based Gemini CLI users, matching the model
already available in AI Studio provider.
Model spec:
- ID: gemini-3-pro-preview
- Version: 3.0
- Input: 1M tokens
- Output: 64K tokens
- Thinking: 128-32K tokens (dynamic)
- Introduced `gpt-5.1-codex-max` variants to model definitions (`low`, `medium`, `high`, `xhigh`).
- Updated executor logic to map effort levels for Codex Max models.
- Added `lastCodexMaxPrompt` processing for `gpt-5.1-codex-max` prompts.
- Defined instructions for `gpt-5.1-codex-max` in a new file: `codex_instructions/gpt-5.1-codex-max_prompt.md`.
Extract reasoning effort mapping into a reusable function `setReasoningEffortByAlias` to reduce redundancy and improve maintainability. Introduce support for the "gpt-5.1-none" variant in the registry and runtime executor.
Stop advertising and mapping the unsupported gpt-5.1-minimal variant in the model registry and Codex executor, and align bare gpt-5.1 requests to use medium reasoning effort like Codex CLI while preserving minimal for gpt-5.
