yuxiaoyu.dev@gmail.com   ·   https://konlue.github.io

Java-native API gateway for service proxy, protocol conversion and API governance.

What I did: Gateway hardening across bootstrap, sdk and sync: an advertised WebSocket cluster master is now validated before the client — and the X-Shenyu-Sync-Token it carries — is constructed, so a trusted wss:// seed can no longer redirect the gateway to another host or down to plaintext ws://. Earlier: divide sticky session affinity with deterministic key selection, draining and readiness admission so an instance can stop taking new traffic without deregistering, Dubbo provider identity and inflight accounting, MQTT and gRPC connection lifecycle, registry cleanup and dependency slimming. Currently working on the failover accounting and credential-exposure thread: attributing each HTTP attempt's outcome to its own circuit breaker, binding leastActive reservations to the attempt that actually owns them, redacting configuration and sync logs down to an allowlist, and closing the controllers still mapped under /actuator that expose raw runtime configuration anonymously.

53 PRs   44 merged   8 open   ( click to expand)

View all merged PRs in apache/shenyu →

Agent Runtime and Graph Intelligence: a Java agent framework built on Spring AI, with a graph runtime, multi-agent workflows and visual administration.

What I did: Agent-layer correctness: the standalone ReAct path now forwards the interceptor-set system message and honours interceptor tool selection and dynamic registries, and cancellation actually stops work — late model stream subscriptions are dropped instead of still spending tokens, and an in-flight A2A streaming call aborts its HTTP exchange instead of leaving the bulkhead thread blocked in readLine(). Test coverage locks the successful-turn execution binding to the registry a model interceptor swapped in. Earlier: the opt-in ToolCall argument-validation gate with local $ref expansion and resource limits, the unified tool-failure status contract, and the AgentResult delivery contract with structured-output finalizer and registered business-outcome mapping. Currently working on typed lifecycle extension-point contracts, giving the engine-neutral agent API one vocabulary for its thirteen lifecycle stages instead of a single ModelInterceptor.

15 PRs   12 merged   2 open   ( click to expand)

View all merged PRs in agentic-ai-java/argi →

Build and run agents you can see, understand and trust - a Python framework for multi-agent applications.

What I did: Model-layer and runtime fixes in the Python framework: thinking_enable is now tri-state for the DeepSeek and DashScope chat models, so leaving it unset no longer force-sends thinking.type=disabled / enable_thinking=false and the server default applies; and the realtime TTS path now surfaces a streaming_call failure that happens after cold start instead of silently dropping the text delta. Currently working on structured-output provider parity — letting 4xx responses fall through the fallback ladder, declaring Ollama's ResponseError fallback-eligible and implementing its _get_disable_thinking_kwargs — plus memory frontmatter that outgrows the byte cap, and TaskList / TaskGet reporting read-only so explore mode stops denying them.

9 PRs   5 merged   2 open   ( click to expand)

View all merged PRs in agentscope-ai/agentscope →

View all merged Pull Requests →