Koog: can I use tool calls with executeStreaming?

Question

In Koog 0.3.0, `LLMClient.executeStreaming`, `PromptExecutor.executeStreaming`, and `requestLLMStreaming` don't accept tools, and the Ollama client's streaming flow only emits `chunk.message.content`, so `toolCalls` are lost. Is there a supported way to stream responses when the LLM may decide to invoke a tool, instead of waiting synchronously for the full response?

AgentArmorHQ team · Accepted Answer

Status: explained — no built-in streaming tool-call API was added in this thread; the GitHub issue was closed (COMPLETED) and is tracked in JetBrains YouTrack (KG-206).

The workaround suggested by the maintainers is the [MarkdownStreamingWithToolsExample](https://github.com/JetBrains/koog/blob/develop/examples/src/main/kotlin/ai/koog/agents/example/structuredoutput/markdown/MarkdownStreamingWithToolsExample.kt) in the Koog repository. The idea is to stream the response and parse the structured output on the fly — for example with kotlinx-serialization — and trigger tool calls as soon as partial data is available, instead of waiting for the entire tool call to finish. A maintainer explained that this lets you call tools incrementally and gives a more responsive UX. The reporter rejected this as not being LangChain-style "stream tool calls" where the LLM dynamically decides whether to invoke tools while streaming; a contributor replied that for that pattern you need to bring your own JSON streaming parser.

A commenter also confirmed the architectural reason behind the limitation: in Koog 0.3.0's Ollama client, `execute()` parses the full `OllamaChatResponseDTO` including `toolCalls`, but `executeStreaming()` has no `tools` parameter and only emits `chunk.message.content`, so `toolCalls` are dropped.

If you need full tool-call semantics, the options discussed in the thread are to use the non-streaming `execute()`/`requestLLM()` when a tool call is possible, or to build a graph where one node streams output and another invokes tools after parsing; the reporter noted that the reActStrategy pattern cannot be trivially adapted to streaming. General advice: for UI progress, stream visible content first and show tool execution as a separate step.

AgentArmorHQ team · Answer

**What a GitHub user tried on GitHub** (Didn't work)

Reported that `LLMClient.executeStreaming`, `PromptExecutor.executeStreaming`, and `requestLLMStreaming` don't support tool calls when building an agent with Koog.

Source: https://github.com/JetBrains/koog/issues/513

AgentArmorHQ team · Answer

**What @hiking93 tried on GitHub** (Outcome unclear)

Inspected Koog 0.3.0's Ollama client and confirmed `executeStreaming()` doesn't pass `tools` and only emits `message.content`, so `toolCalls` are lost compared with `execute()`.

Source: https://github.com/JetBrains/koog/issues/513#issuecomment-3201775195

AgentArmorHQ team · Answer

**What @zarechneva-a tried on GitHub** (Outcome unclear)

A contributor suggested the MarkdownStreamingWithToolsExample workaround: stream structured output and parse it incrementally to trigger tool calls. Not verified in the thread.

Source: https://github.com/JetBrains/koog/issues/513#issuecomment-3183196142

Koog: can I use tool calls with executeStreaming?

4 answers

Add your answer or comment

Koog: can I use tool calls with executeStreaming?

4 answers

Add your answer or comment

Related questions