bingshuo/hololake-platform
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity
hololake-platform/docs/adr/0027-dual-ai-architecture.md

2.1 KiB
Raw Blame History

type, id, title, status, superseded_by, date
type id title status superseded_by date
ADR 0027 Dual AI architecture (API chat + CLI agent) superseded 0028 2026-03-01

Context

Laputa needs two distinct AI interaction modes: a lightweight chat for quick questions about the current note (no tool access, fast responses), and a full agent that can search, read, create, and modify vault notes via MCP tools. These have fundamentally different requirements — the chat needs low latency and simple streaming, while the agent needs tool calling, conversation state, and MCP integration.

Decision

Maintain two separate AI interfaces: AI Chat (AIChatPanel) uses the Anthropic API directly via Rust for simple streaming responses. AI Agent (AiPanel) spawns Claude CLI as a subprocess with MCP vault integration for full tool access. Both share a context builder (ai-context.ts) that provides the active note and linked entries.

Options considered

  • Option A (chosen): Dual architecture — optimized for each use case. Chat is fast and simple; agent is powerful with tool access. Downside: two codepaths to maintain.
  • Option B: Single agent for both — always use Claude CLI. Downside: overkill for simple questions, slower startup, unnecessary tool overhead.
  • Option C: Single API-based chat with manual tool calling — unified codebase. Downside: complex tool-calling loop implementation, no MCP integration.

Consequences

  • AI Chat: AIChatPanel + useAIChat hook → Rust ai_chat command → Anthropic API. Default model: Haiku 3.5 (fast, cheap).
  • AI Agent: AiPanel + useAiAgent hook → Rust claude_cli.rs → Claude CLI subprocess with MCP config.
  • Both panels share a toggle in the breadcrumb bar (Sparkle icon).
  • Context builder (ai-context.ts) provides structured JSON with active note, linked notes, open tabs, vault metadata.
  • Token budget: 60% of 180k context limit (~108k tokens max).
  • Chat requires an Anthropic API key in settings; agent uses Claude CLI's own authentication.
  • Re-evaluation trigger: if Anthropic releases an SDK that handles both simple chat and tool calling efficiently.