mirror of
https://github.com/JackChen-me/open-multi-agent.git
synced 2026-10-22 02:53:26 -05:00
Jack Chen
990594aa01
Add `parallelToolCalls?: boolean` to AgentConfig, CoordinatorConfig, LLMChatOptions, and RunnerOptions. The OpenAI adapter forwards it as `parallel_tool_calls` in both chat() and stream() payloads. Anthropic adapter ignores; default `undefined` keeps cloud OpenAI behavior unchanged. Some local OpenAI-compatible servers (vLLM, llama-server, certain quantized setups) truncate concurrent tool_call deltas during streaming or return malformed tool_calls when more than one is emitted. Setting `parallelToolCalls: false` forces serial tool calling on those servers without regressing cloud OpenAI users (whose server default of `true` applies when the field is omitted). Field placement mirrors topK / minP / frequencyPenalty: spread before extraBody so users can override via the escape hatch, but after the transport-level fields so model/messages/tools/stream stay protected. Closes #168