v0.1 — Public preview

Name: Liminal AI
Author: Vireon Dynamics

Liminal AI's first public release. One-command install, two UIs, the full tool catalog, and the persona system.

This is the first public release of Liminal AI. It's the foundation — everything below is shipped, tested, and runnable today.

Install

One-command install for macOS, Linux, WSL, and Windows.
New liminal CLI: setup, doctor, web, tui, update, path.
Setup wizard writes .env, runs npm install / npm run build, smoke-tests the provider, and launches the web UI with persona bootstrap.
Non-interactive mode for CI install (--non-interactive, AGENT_API_KEY env).

ReAct loop with retries, provider circuit breakers, context compression, drift scoring.
Resumable streaming writes — large write_file / edit_file payloads stream through a staging file and can resume from a manifest after a length-truncated response.
Compensation ledger for partial-side-effect rollback.
Approval gates on destructive tools (run_shell, run_background).
Optional self-heal lint loop after edits (AGENT_SELF_HEAL_LINT).
Prompt caching with per-turn hit-rate logging on supported providers.

Files: read_file (chunked, with-imports), edit_file (replacements + fuzzy diff), write_file, multi_file_apply (atomic + rollback), rename_symbol (semantic project-wide TS/JS rename).
Code intelligence: ast_grep, symbol_index, find_references, grep_file.
Shell + process: run_shell, run_background, kill_process, run_command_with_pty, run_tests, run_lint.
Git: full suite including git_worktree for linked branches under .agent_worktrees/.
Web: web_search, web_fetch with readability extraction and 401/403 Firefox+Chrome retry.
Browser: full Playwright family with captcha_solve (2captcha / CapSolver).
Memory: remember, recall, recall_relevant (hybrid BM25 + vector), memory_graph, memory_consolidate.
Vault: Obsidian vault_read/write/search/links/graph with vault auto-discovery.
Reasoning: think, reason, plan, hypothesize, breakdown as observable tools.
Documents: full pipeline through to PPTX/DOCX/PDF with quality gate.
Orchestration: spawn_agent, wait_for_agents, sub-agent contract verification.
Meta: create_tool, edit_tool, list_dynamic_tools — agent-authored tools persisted to disk.

First-run persona bootstrap modal in TUI and web.
Generated PersonaUiTheme (V1 or V2) — palette, motion, shell, density, typography.
Web shell router: TerminalShell, HudShell, StudioShell, MinimalShell.
Streaming persona generation — artifacts stream live during bootstrap.

Per-turn intent classification on the fast model.
Reasoning budget inferred per turn (AGENT_REASONING_BUDGET).
External reasoning surface — model reasons via think() + reason() tools rather than a hidden native stream.
Two-tier routing: main model for ReAct, fast model for intent / distill / rewrite / critic / safety judge / auto-dream.
Effort learning: outcome scoring tunes reasoning budget per intent class.

Ink TUI: tool cards, think/reason/plan/breakdown cards, approval modal, persona bootstrap modal, memory strip, streaming text.
Web UI: SSE-streamed chat with reconnect via last-event-id, Settings modal exposing ~190 typed knobs, persona generation workbench, streaming write previews.
Both surfaces render every tool call inline.

Append-only JSONL session traces under .agent_sessions/.
self_telemetry tool surfaces the harness's own failure / rule / recipe / effort stats.
Recipe library — successful multi-tool turns recorded by (intent class, tool-phase shape).

22 scenario packs in packages/eval: basic, reliability, memory, retrieval, harness quality, harness capability, epistemic, multi-hop, contradiction, context-rot, approval-correctness, web research, long-horizon, tool-lazy-load, large-file-write, reasoning-budget, browser-local, document quality, and more.
JSON sink → .agent_eval_runs/.

Everything above is intentional. v0.1 is the foundation. The next phase is everything that makes Liminal feel like infrastructure.

— Vireon Dynamics