Skip to content

Native tool calling

Since 0.8.0, OpenAIChatPlanner is the recommended planner for complex and long tasks. It speaks /chat/completions — so it works against OpenAI and every OpenAI-compatible gateway (vLLM, Ollama, Groq, DeepSeek, …) — and uses provider-native tool calling instead of one-action JSON output:

  • Per-tool schemas. Your ToolDefinition.inputSchema reaches the provider as a native tools entry, so arguments are constrained per tool.
  • Parallel tool calls. tool_calls[] map to the kernel's parallel ToolCallAction[] batch — fan-out costs one round instead of N.
  • Prompt-cache friendly. The prompt is an append-only message transcript with a stable prefix, so provider caching stays effective as runs grow.
  • Real streaming. Answers stream as content deltas; no JSON scraping.
  • Kernel-stamped identity. The model never mints id/createdAt — the kernel stamps and de-duplicates action identity for the whole run.
import { OpenAIChatPlanner } from "@agentic-kernel/model-openai";

const planner = new OpenAIChatPlanner({
  apiKey: process.env.OPENAI_API_KEY!,
  model: "gpt-5.1",
  // baseUrl: "http://localhost:11434/v1",  // any OpenAI-compatible gateway
  parallelToolCalls: true,
  temperature: 0.2
});

import httpx
from agentic_kernel.model_openai import OpenAIChatPlanner, OpenAIChatPlannerOptions
from agentic_kernel.model_openai.httpx_fetch import httpx_fetch

client = httpx.AsyncClient()
planner = OpenAIChatPlanner(
    OpenAIChatPlannerOptions(
        api_key="...",
        model="gpt-5.1",
        fetch=httpx_fetch(client),   # reference HttpFetch — ships with [openai]
        parallel_tool_calls=True,
    )
)

Control actions ship as a small synthetic-tool set the model can call: ask_user (pauses at waiting_for_user — see Sessions) and stop always; push_goal/complete_goal behind enableGoalTools. A control call always wins over tool calls beside it, because kernel batches must be tool-only.

The Responses-API planner (OpenAIPlanner) remains available for hosts that want strict single-action structured output.