llm_mock_anthropic

When you test Ruby code that calls the Anthropic API, you usually don't want it hitting the network. One clean way to avoid that is to stub your Anthropic client and return a canned response object — but that runs into a wall: constructing a realistic Anthropic SDK response by hand is genuinely painful. A real Anthropic::Message needs id, model, role, type, usage, stop_reason, and typed content blocks; tool calls are nested; and streaming responses have no simple object to fake at all.

llm_mock_anthropic gives you small, ergonomic stand-ins for exactly those response shapes — messages, text blocks, tool_use blocks, and streams — so your stub can return something your code happily consumes:

allow(client.messages).to receive(:create).and_return(
  LlmMock::Anthropic.message([
    LlmMock::Anthropic.tool_use(id: "tu_1", name: "save_summary", input: {"text" => "…"}),
  ])
)

The common community approach is to stub at the HTTP layer (VCR/WebMock): record real HTTP and let the SDK deserialize it. That's a great fit when you can make a real call once. Stub at the object layer — what this gem is for — when that's awkward, most often because:

• Streaming. Replaying SSE streams through VCR is fiddly; returning a Stream double is trivial.
• You want to script the model's behavior deterministically (e.g. "this turn calls the complete tool") without recording anything.

If you want to record real calls once and replay them rather than hand-script responses, see deja — it builds on this gem.

# Gemfile
group :test do
  gem "llm_mock_anthropic"
end

Response value objects (duck-typed to the SDK's response surface — .content, block fields, .text, .accumulated_message):

The structs are also available directly (LlmMock::Anthropic::Message.new(...)) if you prefer.

Say the code under test streams a reply and finishes when the model calls a complete tool — it calls messages.stream, renders the incremental text, then inspects the final message:

class TutorTurn
  def run(client, conversation)
    stream = client.messages.stream(
      model: "claude-sonnet-4-5",
      max_tokens: 1024,
      messages: conversation,
      tools: [ complete_tool ],
    )

    stream.text.each {|chunk| broadcast(chunk) } # render text as it arrives

    stream.accumulated_message.content.each do |block|
      finish! if block.type == :tool_use && block.name == "complete"
    end
  end
end

In a test, return a fake stream so that code runs without the network — one text chunk plus a final message that includes the complete tool call:

fake = LlmMock::Anthropic.stream(
  text_chunks: [ "Here's the core idea. " ],
  message: LlmMock::Anthropic.message([
    LlmMock::Anthropic.text("Here's the core idea. "),
    LlmMock::Anthropic.tool_use(id: "tu_done", name: "complete", input: {}),
  ]),
)
allow(client.messages).to receive(:stream).and_return(fake)

stream.text yields the chunks and stream.accumulated_message.content is the final block list — exactly the surface TutorTurn#run reads from a real stream.

LlmMock::Anthropic::Provider implements the llm_mock contract — it builds a stub client routed through a responder, invokes the real client, and serializes/deserializes responses to/from plain hashes. That's how deja uses this gem to record and replay Anthropic calls. You don't need any of that to use the builders above.

MIT — see LICENSE.