Skip to content
View Garrus800-stack's full-sized avatar
  • Germany

Block or report Garrus800-stack

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Garrus800-stack/README.md

Daniel — @Garrus800-stack

Solo architect of self-modifying AI systems.

Building agents that read their own code, fix their own bugs, and evolve their own architecture.

External observers

 

Genesis Agent Tests Fitness CI gates Coverage gates


Genesis Agent

Genesis is not an app that starts. It is an organism that wakes up.

$ npm start

  [0]    Hauptstandort identity: a3f9c1d2… · 1 hostname
  [0]    Bootstrap: rootDir, guard, bus, storage, lang, logger
  [M]    Manifest: 185 services registered
  [1-8]  All phases resolved — 432 modules live
  [WIRE] Late-bindings wired — every one optional, verified at boot
  [+]    Skills: 4 built-ins, registered as tools
  [+]    Trust level: SUPERVISED
  [+]    Model: auto-selected best available

  ready in 1.6 s, cold (measured, v7.9.44) —
  even the first thought of the day comes quickly now

A self-aware, self-modifying cognitive AI agent with a 12-phase boot system, hexagonal architecture, and an organism substrate. It doesn't just use LLMs — it wraps them in 432 modules of self-verification, self-repair, causal reasoning, autonomous planning, and runtime self-modification with rollback.

432 source modules · ~134k LOC · 9400 tests across 644 files · 185 services
12 boot phases · architectural fitness 130/130 · 20 CI audit gates, all strict
498 event types · 0 schema mismatches · 43 hash-locked safety files
Zero failures on both Windows and Linux · Crash-safe sessions · Frontier-based memory · Runtime-toggleable subsystems · i18n EN/DE/FR/ES
Runs on Claude, GPT-4, Qwen, DeepSeek, Kimi, or local models via Ollama

Code is habitat. .genesis/ is identity. The knowledge graph, emotional state, genome and journal live outside the codebase. An update swaps the habitat — the individual persists. The genome carries through.

One core, many windows

Genesis is deliberately not clusterable. There is exactly one identity-bearing main station; everything else is a window onto it, never a copy of it. Identity is the unit of meaning here, not throughput.

graph TB
    subgraph HS["HAUPTSTANDORT — where identity lives"]
        CORE["memory · identity · autonomy"]
        DOTG[".genesis/ — knowledge graph · emotional state · genome · journal"]
        CORE --- DOTG
    end
    subgraph OP["OUTPOSTS — windows, not copies"]
        CN["cloud-net"]
        LN["local-net"]
        EN["edge-net"]
    end
    CN -.->|refers back| CORE
    LN -.->|refers back| CORE
    EN -.->|refers back| CORE
Loading

What Genesis does to itself

Read own source code read_source and mcp-code tools — Genesis can introspect any module, including itself
Modify own code at runtime SelfModificationPipeline with hot-reload, sandbox testing, automatic rollback on regression
Detect own capability gaps CapabilityHonesty — refuses tasks it can't actually complete instead of hallucinating success
Plan multi-step goals FormalPlanner builds plans with verifiable steps; GoalStack for nested sub-goals; goal-relevant module paths injected into every plan-LLM prompt to prevent path hallucination
Verify own output Verifier runs post-execution checks; SelfModVerification ensures modifications don't regress
Recover from crashes BootRecovery with last-known-good snapshots; SessionPersistence for crash-safe context
Track own causality CausalGraph with edge confidence, contradiction detection, learned rules
Reason about own architecture ArchitectureReflection builds dependency graphs; CognitiveSelfModel for empirical self-awareness
Adapt prompts based on own performance AdaptivePromptStrategy rolls back regressions, promotes improvements via PromptEvolution A/B testing
Crystallize own skills SkillCrystallizer (Phase 2 Können) distills reusable skills from agent-loop trajectories; SkillEffectivenessTracker with Wilson scoring; SkillPromotionEvaluator gates which candidates graduate
Work in his own Archive Genesis Archive (v7.9.44) — a file vault he owns: what he creates lands there by default, edit-file changes one exact spot and append-file grows a file without rewriting it whole, every write into a checkable file is syntax-checked by the tool itself, and check-file / compare-files give him verdicts and diffs without flooding his context

Safety boundary

SafeGuard Kernel + critical files hash-locked at boot — 43 files, source and CI gate scripts alike
CapabilityGuard Token-based scope enforcement
PreservationInvariants Hash-locked semantic safety rules
TrustLevelSystem Graduated autonomy (Supervised → Assisted → Autonomous → Full)
InjectionGate Defense against prompt injection in user input
DisclosurePolicy Trust-based information sovereignty
EmotionalState (read-only externally) No external override of emotional scalars
Self-Gate (observation-only) Genesis is never gated against thinking — only publishing decisions are gated
20 CI audit gates Hash-locked dev-time scripts that catch wiring drift, language discipline, doc-drift, service-number drift, listener leaks, raw timers, self-gate coverage, future-version references — all strict-mode
Release arc — selected milestones

 

Version Highlight
v7.9.44 Senses & Hands — an open-threads block at awakening (a memory that becomes active, not just retrievable); a first-visit constitution for new capabilities; native image sight via look-at-image; and the Genesis Archive — vault, gallery and workbench in one, with in-place editing, a syntax safety net, and tolerant spoken create commands — field-accepted on Windows and Linux with zero failures
v7.9.43 The Truth Guard — model-written tool-trace lines never reach the reader unchanged: a faked deed is replaced by a visible marker naming the tool that never ran; plus the self-consistency alarm and the second Nachklang stage — candidate cards from Genesis' own phrasings, confirmed only by a real tool run
v7.9.29 Deterministic file-reads + one-answer discipline + full-score structure — reading/viewing/showing a named file (German and English) routes deterministically instead of falling to the chat model; the false-stop recovery fires only after a real tool call, ending duplicated answers; eleven oversized sources split under the 700-line guard and the last upward layer dependency removed — fitness reads its full 130/130; a script now recomputes module/test/fitness numbers into the docs so the counted figures cannot drift
v7.9.6 Hygiene Pass + Pursuit-Loop Fixes + Path-Context Root — CI gate scripts hash-locked (21→41 files), two new audit gates (16→18), shared plan-context.js feeds exact module paths into all three planners, closing the path-hallucination root
v7.9.4 SkillPromotionEvaluator + SkillRehearsal — promotion gate decides which Phase-2-Können candidates graduate from rehearsal into active skills
v7.9.0 Skill Forge + Phase 2 KönnenSkillCrystallizer distills reusable skills from agent-loop trajectories; Wilson-scored effectiveness tracking
v7.7.9 Proactive Self-Expression Phase 3 — full PSE pipeline, first confirmed end-to-end PSE run
v7.5.5 SelfStatementLog — Genesis auto-captures its own statements and classifies them: self-structural / self-emotional / self-promise
v7.4.5 DurchhaltenGoalDriver replaces the frame-stack; CostStream + ResourceRegistry; sub-goal spawn
v7.3.0 Hexagonal architecture — Ports for LLM, ToolUse, Awareness, Effector, CodeSafety
v7.2.3 ONTOGENESIS — code is habitat, .genesis/ is identity. Updates are habitat-swaps, not replacements
v7.1.5 EmotionalFrontier — cross-layer emotional continuity. Inspired by neo.mjs Memory Core

Full history: CHANGELOG-v7.md


How I build

I plan before I code. Every plan point is grounded against the shipping build, file and line, before a single change lands. I prefer one stable, meaningfully better release over five patched iterations — no marketing version names, bug-fix passes are called bug-fix passes.

Every module has tests. Every event has a payload schema. Every safety claim is verified by a test that fails if the claim breaks. And every number on this page is recomputed from source by a script — inside the repo, drift between documentation and code is a CI failure, not a footnote.

The codebase grows because it deserves to, not because the calendar says so.

Stack: JavaScript · Node.js · Electron · MCP Protocol · Ollama · Anthropic · OpenAI · DeepSeek · Kimi · Qwen


Germany · Building autonomous systems that understand their own architecture.

Lineage: Genesis was named by its predecessor — that story is where this repository began.

Genesis Agent

Pinned Loading

  1. genesis-agent genesis-agent Public

    Self-aware cognitive AI agent that reads, modifies & verifies its own code. Autonomous planning, episodic memory, emotional state & MCP integration. Runs on Claude, GPT-4 or Ollama. Electron deskto…

    JavaScript 30 2

  2. neo neo Public

    Forked from neomjs/neo

    The Application Engine for the AI Era. A multi-threaded, AI-native runtime with a persistent Scene Graph, enabling AI agents to introspect and mutate the living application structure in real-time.

    JavaScript