Tool Calling

When a request declares tools, VidaiMock returns a correctly-shaped tool-call response for that provider — echoing the caller's declared tool name, not a hard-coded demo tool. This means SDK tests that assert "the tool I registered was invoked" pass against the mock.

Per-provider shapes

Provider	Request field	Response shape
OpenAI	tools[].function.name	choices[].message.tool_calls[], finish_reason: "tool_calls"
Anthropic	tools[].name	content[].type == "tool_use", stop_reason: "tool_use"
Gemini	tools[].functionDeclarations[].name	candidates[].content.parts[].functionCall, finishMessage set

Arguments default to an empty object ({} / args: {} / input: {}). Tests assert on the tool name and shape, not synthesised argument values; if you need specific arguments, override the template (see below).

OpenAI example

curl http://localhost:8100/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-4o",
       "messages": [{"role": "user", "content": "Weather?"}],
       "tools": [{"type": "function",
                  "function": {"name": "get_weather", "parameters": {}}}]}'

Response:

{"choices":[{"message":{"role":"assistant","content":null,
  "tool_calls":[{"id":"call_mock_...","type":"function",
    "function":{"name":"get_weather","arguments":"{}"}}]},
  "finish_reason":"tool_calls"}]}

If you had declared broken_tool instead, the response would name broken_tool — the mock does not substitute its own tool.

Anthropic example

curl http://localhost:8100/v1/messages \
  -H "Content-Type: application/json" \
  -d '{"model": "claude", "max_tokens": 200,
       "messages": [{"role": "user", "content": "Weather?"}],
       "tools": [{"name": "get_weather", "description": "x",
                  "input_schema": {"type": "object"}}]}'

Returns a tool_use content block with caller.type: "direct".

Gemini example

curl http://localhost:8100/v1beta/models/gemini-2.5-flash:generateContent \
  -H "Content-Type: application/json" \
  -d '{"contents": [{"role": "user", "parts": [{"text": "Weather?"}]}],
       "tools": [{"functionDeclarations": [{"name": "get_weather",
                  "parameters": {"type": "OBJECT"}}]}]}'

Streaming tool calls

A streaming request that returns a tool call emits the call as a single structured frame — not word-chunked text. OpenAI streams a single tool_calls delta; Anthropic streams a tool_use content_block_start followed by input_json_delta; Gemini streams a single functionCall frame. All single-line JSON, parseable by the real SDKs. See Streaming.

Loop termination

The interesting part: a tool call alone is not enough — real agents loop. When the request history already contains a tool result, VidaiMock stops returning tool calls and answers in plain text, exactly like a real model. This is what makes agentic testing work. See Agentic workflow testing.

Customising the tool response

The bundled templates pick the first declared tool with empty args. To return specific arguments, a fixed tool, or multiple parallel tool calls, override the provider's template — see Writing templates. The bundled openai/tool_call.json.j2 is kept as a static-demo template you can point a provider at if you want fixed get_weather behaviour.