Development Status

Here you can find the current development status and recent updates for llm-connector.

Current Version: 0.7.2

[0.7.2] - 2026-02-24

📝 Documentation

• Added comprehensive changelog and development status to website
• Updated website configuration with new development section

[0.7.1] - 2026-02-23

🔧 Maintenance

• Resolved clippy warnings and stabilized tests
• Fixed cargo fmt issues in CI

[0.7.0] - 2026-02-23

Added

• Per-Request Overrides (Multi-Tenant / Gateway)
  • ChatRequest now supports api_key, base_url, and extra_headers for per-request overrides
  • with_api_key(), with_base_url(), with_header(), with_extra_headers() builder methods
  • Supports multi-tenant routing without creating a new client per tenant
  • Custom headers (e.g. X-Trace-Id, anthropic-version) override default provider headers
  • Works with all GenericProvider-based providers (OpenAI, Anthropic, DeepSeek, Moonshot, Volcengine, etc.)

[0.6.1] - 2026-02-20

🔧 Build / Compatibility

• Rust 2024 edition (MSRV: Rust 1.85+)
• Reqwest uses rustls-tls by default (better Android cross-compilation compatibility)

[0.6.0] - 2026-02-15

🚀 New Features

• Rust 2024 edition — MSRV is now Rust 1.85+

• Builder Pattern for LlmClient — LlmClient::builder() provides a fluent API for client construction with optional base_url(), timeout(), proxy() configuration. Supports all 12+ providers.

  let client = LlmClient::builder()
      .deepseek("sk-...")
      .timeout(60)
      .build()?;

• Zhipu Multimodal Support — Zhipu protocol now supports image URLs and base64 images via MessageBlock::image_url() and MessageBlock::image_base64(). Works with glm-4v-flash and other vision models.

⚡ Breaking Changes (minor)

• Streaming now enabled by default — streaming feature is included in default features, so chat_stream(), ChatStream, StreamingResponse etc. are available without extra configuration. Downstream libraries no longer need features = ["streaming"].

Recent History

[0.5.17] - 2026-02-14

🚀 New Features

• Mock Client for Testing
  • New MockProvider for unit testing without real API calls
  • MockProviderBuilder with fluent API for fine-grained control
  • LlmClient::mock("content") one-liner for simple cases
  • Sequential response mode for multi-turn test scenarios
  • Error simulation support for testing error handling paths
  • Request tracking via as_mock().request_count() / get_requests()
  • Tool call simulation via MockProviderBuilder::with_tool_calls()

[0.5.16] - 2026-02-14

🚀 New Features

• Enhanced Tool Calling (P0)

  • ChatResponse: Added tool_calls(), is_tool_call(), finish_reason() convenience methods
  • ToolCall: Added parse_arguments<T>() for typed deserialization, arguments_value() for generic JSON
  • Message: Added assistant_with_tool_calls() constructor for multi-turn tool use
  • Tool: Added Tool::function() convenience constructor

• Structured Outputs (P1)

  • ResponseFormat: Extended to support json_schema mode (OpenAI Structured Outputs)
  • New JsonSchemaSpec type with name, description, schema, strict fields
  • Convenience constructors: ResponseFormat::text(), json_object(), json_schema(), json_schema_with_desc()
  • OpenAI protocol now correctly passes response_format to API requests

• Error Type Refinement (P2)

  • Added ContextLengthExceeded error variant
  • is_retryable() now includes ServerError, TimeoutError, ConnectionError
  • New helper methods: should_reduce_context(), is_auth_error(), is_rate_limited()
  • Context length detection in OpenAI, Anthropic, Aliyun, Zhipu map_error implementations

• Token Usage (P2)

  • Usage now derives Default for easier construction

📝 Documentation

• Updated README with new Tool Calling and Structured Output examples
• Updated docs/TOOLS.md with convenience API usage
• Added integration test example: examples/test_wishlist.rs

For more details, please refer to the Full Changelog.