Liminal — Features

Feature Status

Launch labels are governed by feature-claim-ledger-2026-05-06.md. A feature can be implemented while still needing live launch proof.

Feature	Status	Details
12 Creative Domains	Live-covered; release-rerun required	Current final QA proof covers p5, SVG, GLSL, Three.js, Hydra, Strudel, Tone.js, Revideo, HyperFrames, ASCII, Kinetic, and TextGen. Public launch copy must use a fresh, current-commit live receipt before calling release output proven.
Multi-Agent Critique	Implemented	3-agent board: Minimalist, Expressionist, Technician. Product launch claims require the claim ledger proof row.
Compost Mill	Experimental	Digest → Shred → Collide → Score → Promote pipeline; public value claims stay experimental until live proof is current.
Self-Improving Harness	Manual-memory mode	Failure logging and pattern detection record manual adaptation advice; automatic HarnessUpdater runtime fixes are not active launch behavior.
Music Theory Engine	Implemented	Euclidean rhythms, Markov chains, scales, chord progressions; claim ledger maps the current proof boundary.
Audio Pipeline	Implemented	Audio analysis, pitch detection, visual parameter mapping; launch proof is feature-specific.
Aesthetic Guardrails	Experimental	Color harmony, layout, typography, sound quality critics; not a blanket guarantee of visual quality.
Model-Agnostic LLM	Setup caveated	ProviderFactory supports multiple providers, but setup diagnostics and live proof are provider-specific.
Circuit Breaker	Stream-failure guarded	Fallback stream errors now surface as failures instead of empty successful assistant responses. Broad automatic failover reliability claims still need provider-specific live proof.
Bubble Tea TUI	Smoke-proven	Go-based TUI with HTTP/SSE bridge, provider switching, and deterministic user-surface proofs.
Task Ledger	Implemented	Self-hosting corpus, runner, verifier (`src/ledger/`); examples and provenance proof remain tracked separately.
LiminalFS	Implemented	ProjectStore, EventStore, AssetStore (`src/fs/`).
Guardrails M1-M18	Proof-limited	Core, extended, and compliance layers exist; launch wording must reference the current command proof per guardrail surface.
Security Hardening	Audit pending	SSRF protection, rate limiting, sandbox, CSRF, and helmet exist, but release security claims remain gated by final QA findings.
Thinking-Trace Loop	Active	Meta-learning from LLM reasoning traces
Plugin System	Active	PluginManager, extensible generator registry

Natural Language Interface

Chat and TUI interpret natural language input and route to the right handler.

Intent Detection

Pattern	Routes To	Example
Generation request	RalphLoop	"Create a particle system"
Evaluation request	Evaluator	"Rate this sketch"
Fix request	Auto-fix (LLM)	"Fix the validation in main.ts"
Status query	Status reporter	"What's the test status?"
Knowledge query	Knowledge base	"Tell me about p5.js noise"
Slash command	Direct handler	/fix, /status, /provider

Implemented in src/tui/NaturalInterface.ts. Works in both liminal chat and the Bubble Tea TUI.

LLM Mode Agent

An agent that can use tools to read, write, build, and test code autonomously.

Capabilities

Multi-step reasoning — Plans and executes complex repair workflows
21 tools — readFile, writeFile, applyEdit, runBuild, runTests, executeSkill, search, searchCode, searchDocs, listDir, typeCheck, npm, runLint, runFocusedTests, lsp, astValidate, importGuard, searchTools, gitStatus, createBackup, restoreBackup
Rate limits — 10 reads/min, 10 writes/min, 12 builds/min, 12 tests/min
Retry logic — Up to 3 retries with circuit breaker protection
Safety — Shell-free execFileSync in verifier, metacharacter guards, prefix whitelist

Source: src/harness/agent/LLMModeAgent.ts. Configurable max steps (default: 15).

Ralph-Wiggum Loop

The core generation cycle.

1. Prompt Enhancement
   Brain injects artistic knowledge, compost seeds, archive examples
        ↓
2. Generation
   LLM generates creative code in the target domain
        ↓
3. Evaluation
   Multi-dimensional scoring (technical + aesthetic)
   3-agent critique board deliberates
        ↓
4. Stagnation Detection
   Compares against recent iterations, detects loops
        ↓
5. Strategy Adaptation
   Adjusts temperature, prompt framing, or switches approach
        ↓
6. Quality Gate
   Stops when threshold met or max iterations reached

Aesthetic gate: When useAestheticGuardrails: true, the RalphLoop runs output through the AestheticCritic pipeline (color harmony, layout, typography, sound quality) before accepting.

Guardrail Details (M1-M18)

All 18 guardrails are implemented and active.

Core (M1-M8)

ID	Name	Source
M1	Code Validation	correctness/
M2	Type Safety	correctness/TypeCheckGuardrail
M3	Test Verification	correctness/TestVerificationGuardrail
M4	Schema Validation	validation/SchemaValidator
M5	Resource Limits	core/ResourceLimiter
M6	Code Style	hygiene/CodeStyleGuardrail
M7	Catastrophic Failure	rules/CatastrophicGuardrails
M8	Registry	core/GuardrailRegistry

Extended (M9-M11)

ID	Name	Source
M9	Semantic Validator	SemanticValidator.ts
M10	Runtime Health	RuntimeHealthMonitor.ts
M11	Accessibility	AccessibilityGuardrails.ts

Compliance (M12-M18)

ID	Name	Source
M12	Privacy	compliance/PrivacyGuardrail
M13	Fairness	compliance/FairnessGuardrail
M14	Supply Chain	compliance/SupplyChainGuardrail
M15	Resilience	compliance/ResilienceGuardrail
M16	Audit Trail	compliance/AuditGuardrail
M17	Explainability	compliance/ExplainabilityGuardrail
M18	Injection Prevention	compliance/InjectionGuardrail

Harness Task System

Structured self-improvement via task files in src/ledger/.

Task Corpus

Structured task files defining bug fixes and improvements. Each task specifies inputs, expected outputs, and verification criteria.

TaskRunner

Executes tasks with sandbox safety. Applies rate limits and validates tool usage patterns.

TaskVerifier

Validates results using shell-free execFileSync with metacharacter guards and a prefix whitelist for allowed commands.

Task Milestones

Milestone	Focus	Status
M1-M8	Core correctness and safety	Complete
M9-M11	Semantic validation, runtime health, accessibility	Complete
M12-M18	Compliance: privacy, fairness, supply chain, resilience, audit, explainability, injection	Complete

CLI: liminal ledger list, liminal ledger run <id>, liminal ledger verify <id>

Generation Modes

Mode	Flag	Description
Single	`default`	One model generates, evaluates, iterates
Swarm	`--use-swarm`	Five default personas (Kai, Nova, Rex, Sam, Max) generate in parallel and vote on best
Deep Collab	`--routing-mode`	Dual-model routing (fast + powerful)
Live AV	`--mode live-music`	Generate Strudel music + Hydra video-synth code

Swarm strategies: competitive, hybrid, ring, mesh. Controlled via --swarm-mode flag.

Features & Guardrails