What's New in v0.2.0
This release brings multimodal image support, a fully refactored LLM Responses protocol, improved reliability, and a richer developer toolset.
🖼️ Multimodal Image Input
- Image input support — Pass image data directly into the agent alongside text prompts.
- Orientation-adaptive dimension validation — Automatically validates image dimensions based on portrait/landscape orientation, ensuring API compliance without manual checks.
⚡ LLM & Responses Protocol Refactor
- Responses protocol upgrade — Rewrote the SSE streaming transport, stream parser, and type layer to align 1:1 with the upstream Grok Responses API.
- Prefix-continue on SSE idle timeout — The engine now automatically issues a prefix-continue request once when the SSE stream goes idle, recovering gracefully from temporary stalls.
- Fixed duplicated tool-call JSON — Resolved a bug where Responses tool-call payloads were being emitted twice per invocation.
- Omit placeholder args for hosted
web_search— Server-side hosted tools no longer send spurious empty arguments at turn start. - Aligned search behavior with Grok Build — Response search flow is now in sync with the upstream reference implementation.
🔄 Reliability & Retry
- Transient retry + multi-provider failover — The HTTP client retries on transient network errors and can automatically fail over to a secondary LLM provider, improving uptime in mobile network
conditions. - DuckDuckGo connectivity probe with cooldown — The web-search tool now probes DuckDuckGo availability before use and applies a cooldown on repeated failures, with a simplified user-facing fallback
notice. - Immediate turn completion for server-side tools — Pure server-side tool calls (e.g., hosted web search) no longer enter an unnecessary client-side loop, reducing latency.
- Improved
call_idparsing andweb_fetcherror handling — More robust parsing of Responsescall_idfields and cleaner error messages from the web fetch tool.
🛠️ Developer Experience
EngineConfigBuilderAPI — New fluent builder for constructing engine configuration without manually constructing JSON.- 1:1 client profile emulation — The engine can now mimic specific LLM client profiles for reproducible behavior across environments.
- Native tracing / logging callback (FFI) — Debug builds now export a native callback so iOS/Android host apps can receive structured log output directly via the C FFI layer.
- Expanded unit test suite — New tests covering the C FFI layer, busybox applets, the Boa JS script sandbox, and file tools.