- Add handling for message_start event to capture initial token usage
- Fix message_delta to use assignment (=) instead of addition (+=)
since Anthropic sends cumulative token counts, not incremental
- Add comprehensive tests for all providers (Google, OpenAI Completions,
OpenAI Responses, Anthropic)
- Document OpenAI limitation: token stats only available at stream end
Fixes issue where aborted streams had zero token counts despite
Anthropic sending input tokens in the initial message_start event.
Fixes issue where unpaired Unicode surrogates in tool results cause JSON serialization errors in API providers, particularly Anthropic.
- Add sanitizeSurrogates() utility function to remove unpaired surrogates
- Apply sanitization in all provider convertMessages() functions:
- User message text content (string and text blocks)
- Assistant message text and thinking blocks
- Tool result output
- System prompts
- Valid emoji (properly paired surrogates) are preserved
- Add comprehensive test suite covering all 8 providers
Previously only Google and Groq handled unpaired surrogates correctly.
Now all providers (Anthropic, OpenAI Completions/Responses, Google, xAI, Groq, Cerebras, zAI) sanitize text before API submission.
- Changed ToolRenderer return type from TemplateResult to ToolRenderResult
- ToolRenderResult = { content: TemplateResult, isCustom: boolean }
- isCustom: true = no card wrapper, false = wrap in card
- Updated all existing tool renderers to return new format
- Updated Messages.ts to handle custom rendering
This enables tools to render without default card chrome when needed.
- Add 'aborted' as a distinct stop reason separate from 'error'
- Change AssistantMessage.error to errorMessage for clarity
- Update error event to include reason field ('error' | 'aborted')
- Map provider-specific safety/refusal reasons to 'error' stop reason
- Reorganize utility functions into utils/ directory
- Rename agent.ts to agent-loop.ts for better clarity
- Fix error handling in all providers to properly distinguish abort from error
- Added partial-json package for parsing incomplete JSON during streaming
- Tool call arguments now contain partially parsed JSON during toolcall_delta events
- Enables progressive UI updates (e.g., showing file paths before content is complete)
- Arguments are always valid objects (minimum empty {}), never undefined
- Full validation still occurs at toolcall_end when arguments are complete
- Updated all providers (Anthropic, OpenAI Completions/Responses) to use parseStreamingJson
- Added comprehensive documentation and examples in README
- Added test to verify arguments are always defined during streaming
- Switch from Zod to TypeBox for tool parameter schemas
- TypeBox schemas can be serialized/deserialized as JSON
- Use AJV for runtime validation instead of Zod's parse
- Add StringEnum helper for Google API compatibility (avoids anyOf/const patterns)
- Export Type and Static from main package for convenience
- Update all tests and documentation to reflect TypeBox usage
- Replace JSON Schema with Zod schemas for tool parameter definitions
- Add runtime validation for all tool calls at provider level
- Create shared validation module with detailed error formatting
- Update Agent API with comprehensive event system
- Add agent tests with calculator tool for multi-turn execution
- Add abort test to verify proper handling of aborted requests
- Update documentation with detailed event flow examples
- Rename generate.ts to stream.ts for clarity
- Add 'zai' as a KnownProvider type
- Add ZAI_API_KEY environment variable mapping
- Generate 4 zAI models (glm-4.5-air, glm-4.5v, etc.) using anthropic-messages API
- Add comprehensive test coverage for zAI provider in generate.test.ts and empty.test.ts
- Models support reasoning/thinking capabilities and tool calling
- Anthropic API requires tool call IDs to match pattern ^[a-zA-Z0-9_-]+$
- OpenAI Responses API generates IDs with pipe character (|) which breaks Anthropic
- Added sanitizeToolCallId() to replace invalid characters with underscores
- Fixes cross-provider handoffs from OpenAI Responses to Anthropic
- Added test to verify the fix works
- Implement QueuedGenerateStream class that extends AsyncIterable with finalMessage() method
- Add new types: GenerateStream, GenerateOptions, GenerateOptionsUnified, GenerateFunction
- Create generateAnthropic function-based implementation replacing class-based approach
- Add comprehensive test suite for the new generate API
- Support streaming events with text, thinking, and tool call deltas
- Map ReasoningEffort to provider-specific options
- Include apiKey in options instead of constructor parameter
- Test handling of empty content arrays
- Test handling of empty string content
- Test handling of whitespace-only content
- All providers handle these edge cases gracefully
- Updated generate-models.ts to fetch these providers directly from models.dev API
- OpenRouter now only used for xAI and other third-party providers
- Fixed test model IDs to match new model names from models.dev
- Removed unused import from google.ts
- Add transformMessages utility to handle cross-provider compatibility
- Convert thinking blocks to <thinking> tagged text when switching providers
- Preserve native thinking blocks when staying with same provider/model
- Add comprehensive handoff tests verifying all provider combinations
- Fix OpenAI Completions to return partial results on abort
- Update tool call ID format for Anthropic compatibility
- Document cross-provider handoff capabilities in README
- Anthropic, Google, and OpenAI Responses providers now return partial results when aborted
- Restructured streaming to accumulate content blocks incrementally
- Prevents submission of thinking/toolCall blocks from aborted completions in multi-turn conversations
- Makes UI development easier by providing partial content even when requests are interrupted
- Emit start event with model and provider info after creating stream
- Add abort signal tests for all providers
- Update README abort signal section to reflect non-throwing API
- Fix model references in README examples
- Collect complete output items during streaming instead of building blocks incrementally
- Handle reasoning summary parts with proper newline separation
- Support refusal content in message outputs
- Preserve full reasoning items and message IDs for multi-turn resubmission
- Emit proper streaming events for text and thinking deltas
BREAKING CHANGE: AssistantMessage now uses content array instead of separate fields
- Changed AssistantMessage.content from string to array of content blocks
- Removed separate thinking, toolCalls, and signature fields
- Content blocks can be TextContent, ThinkingContent, or ToolCall types
- Updated streaming events to include start/end events for text and thinking
- Fixed multiTurn test to handle new content structure
Note: Currently only Anthropic provider is updated to work with new API
Other providers need to be updated to match the new interface
- Added contentSignature tracking for assistant messages
- Fixed message format in convertToResponsesFormat (output_text instead of input_text)
- Properly preserve message IDs for multi-turn conversations
- Added proper ResponseOutputMessage type satisfaction
- Updated tests to cover more providers and multi-turn scenarios
- Added image tests to OpenAI Completions (gpt-4o-mini)
- Added image tests to Anthropic (claude-sonnet-4-0)
- Added image tests to Google (gemini-2.5-flash)
- Tests verify models can process and describe the red circle test image
- LLM constructors now take Model objects instead of string IDs
- Added provider field to AssistantMessage interface
- Updated getModel function with type-safe model ID autocomplete
- Fixed Anthropic model ID mapping for proper API aliases
- Added baseUrl to Model interface for provider-specific endpoints
- Updated all tests to use getModel for model instantiation
- Removed deprecated models.json in favor of generated models
- Add proper Anthropic model aliases (claude-opus-4-1, claude-sonnet-4-0, etc.)
- Deduplicate models when same ID appears in both models.dev and OpenRouter
- models.dev takes priority over OpenRouter for duplicate IDs
- Fix test to use correct claude-3-5-haiku-latest alias
- Reduces Anthropic models from 11 to 10 (removed duplicate)
- Add Model interface to types.ts with normalized structure
- Create type-safe generic createLLM function with provider-specific model constraints
- Generate models from OpenRouter API and models.dev data
- Strip provider prefixes for direct providers (google, openai, anthropic, xai)
- Keep full model IDs for OpenRouter-proxied models
- Clean separation: types.ts (Model interface), models.ts (factory logic), models.generated.ts (data)
- Remove old model scripts and unused dependencies
- Rename GeminiLLM to GoogleLLM for consistency
- Add tests for new providers (xAI, Groq, Cerebras, OpenRouter)
- Support 181 tool-capable models across 7 providers with full type safety
- Switch from Node.js test runner to Vitest for better DX
- Add test suites for Grok, Groq, Cerebras, and OpenRouter providers
- Add Ollama test suite with automatic server lifecycle management
- Include thinking mode and multi-turn tests for all providers
- Remove example files (consolidated into test suite)
- Add VS Code test configuration
- Remove catch-all [key: string]: any from ModelInfo
- Make all required fields non-optional (attachment, reasoning, etc.)
- Add proper union types for modalities (text, image, audio, video, pdf)
- Mark only cost and knowledge fields as optional
- Export ModalityInput and ModalityOutput types
- Add models script to download latest model information
- Create models.ts module to query model capabilities
- Include models.json in package distribution
- Export utilities to check model features (reasoning, tools)
- Update build process to copy models.json to dist
- Add examples for Cerebras, Groq, Ollama, and OpenRouter
- Update OpenAI Completions provider to handle base URL properly
- Simplify README formatting
- All examples use the same OpenAICompletionsLLM provider with different base URLs
- Add multi-turn test to verify thinking and tool calling work together
- Test thinkingSignature handling for proper multi-turn context
- Fix Gemini provider to generate base64 thinkingSignature when needed
- Handle multiple rounds of tool calls in tests (Gemini behavior)
- Make thinking tests more robust for model-dependent behavior
- All 18 tests passing across 4 providers
- Added thinkingConfig with includeThoughts and thinkingBudget support
- Use part.thought boolean flag to detect thinking content per API docs
- Capture and preserve thought signatures for multi-turn function calling
- Added supportsThinking() check for Gemini 2.5 series models
- Updated example to demonstrate thinking configuration
- Handle SDK type limitations with proper type assertions
- Added GeminiLLM provider implementation with GoogleGenerativeAI SDK
- Supports streaming with text/thinking content and completion signals
- Handles Gemini's parts-based content system (text, thought, functionCall)
- Implements tool/function calling with proper format conversion
- Maps between unified types and Gemini-specific formats (model vs assistant role)
- Added test example matching other provider patterns
- Fixed typo in AssistantMessage type (stopResaon -> stopReason) across all providers
- Update LLMOptions interface to include completion boolean parameter
- Modify all providers to signal when text/thinking blocks are complete
- Update examples to handle the completion parameter
- Move documentation files to docs/ directory
- Implement OpenAICompletionsLLM for Chat Completions API with streaming
- Implement OpenAIResponsesLLM for Responses API with reasoning support
- Update types to use LLM/Context instead of AI/Request
- Add support for reasoning tokens, tool calls, and streaming
- Create test examples for both OpenAI providers
- Update Anthropic provider to match new interface
- Define clean API with complete() method and callbacks for streaming
- Add comprehensive type system for messages, tools, and usage
- Implement AnthropicAI provider with full feature support:
- Thinking/reasoning with signatures
- Tool calling with parallel execution
- Streaming via callbacks (onText, onThinking)
- Proper error handling and stop reasons
- Cache tracking for input/output tokens
- Add working test/example demonstrating tool execution flow
- Support for system prompts, temperature, max tokens
- Proper message role types: user, assistant, toolResult
