v0.0.1 is day one on main: a runnable model → tools → UI loop from a clone, before lazy loading, vault tools, or installer scripts existed.

Shipped

Monorepo

  • packages/core, tools, tui, web, eval workspaces
  • Build order: core → tools → interfaces
  • CLAUDE.md architecture guide for contributors

Harness (packages/core)

  • AgentHarness ReAct loop with tool dispatch, retries, and context manager
  • Approval path for destructive tools; think() pre-flight for danger gates
  • Initial harness env wiring and session lifecycle

Tools (packages/tools)

  • File read/write/edit, shell, git, memory (remember / recall), web search/fetch
  • Reasoning tools: think, plan, reason (external, observable)
  • System prompt v1 (PROTOCOL_CORE)

Interfaces

  • Ink TUI — streaming assistant text and tool call cards
  • Web — Express API + Vite React client with SSE chat
  • Eval — first behavioral integration scenarios

Project

  • Rebrand from internal codename to Liminal
  • Default model experiments (OpenRouter / minimax free tier)

Not yet

  • Lazy tool families, Obsidian vault, persona bootstrap UI, sub-agent DAG, browser automation, liminal CLI, MIT license file on repo
Next: v0.0.2 →Dynamic persona generation and child harness wiring.