Today's LlmClient.generate(prompt: string, options) takes a single string. The assistant flattens conversation history into that string with User: / Assistant: role labels (packages/assistant/src/assistant-service.ts:formatHistoryAsPrompt), which works but is a workaround.

Scope

• Refactor LlmClient.generate to accept either a string (back-compat) or a messages: ChatMessage[] array (new path).
• Each provider in packages/llm-client/src/providers/{anthropic,openai,google,ollama}.ts translates the messages array into its native chat-completions format.
• Update AssistantService to pass messages directly; delete formatHistoryAsPrompt and the User: / Assistant: workaround.
• Update tests across @skytwin/llm-client and @skytwin/assistant.

Out of scope

• Tool-use / function-calling — different refactor (track separately).
• Vision / multi-modal inputs — different refactor.

Why

Native multi-turn is more accurate (the model knows turn boundaries explicitly), more efficient (some providers cache prompts at message-array boundaries), and removes a comment-laden workaround.

Should land before phase 2c (action routing) so action intent detection works on real turn boundaries instead of the flattened string.