peter — START HERE: vaked.dev/ultrawhale/book/quick-start · coding agent that builds itself · 150 blocks · 7 recursions · v100.1.0 🐘 @vakedbot
  • Go 76.1%
  • HTML 17.1%
  • Python 3.3%
  • Shell 1.6%
  • JavaScript 1%
  • Other 0.9%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
dependabot[bot] 4fad44c431
build(deps): bump golang.org/x/net from 0.33.0 to 0.55.0 (#25)
Bumps [golang.org/x/net](https://github.com/golang/net) from 0.33.0 to 0.55.0.
- [Commits](https://github.com/golang/net/compare/v0.33.0...v0.55.0)

---
updated-dependencies:
- dependency-name: golang.org/x/net
  dependency-version: 0.55.0
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
2026-11-08 09:15:01 -05:00
.agents/skills/kompress-finetune feat: voting ensemble, LOOP_STATE.md, kompress-finetune SKILL.md (loop engineering primitives) 2026-06-25 05:09:14 +02:00
.dev v100.1.0: vastai-go — one-shot GPU training that ACTUALLY WORKS 2026-06-21 17:36:21 +02:00
.github perf(ci): add concurrency groups to cancel redundant runs 2026-07-13 23:14:39 +02:00
.ultrawhale/workflows v53.0.0: VEGED — The Name + DIVE-DEEP workflow 2026-06-21 07:35:54 +02:00
cmd v11.0.0: Agent Architecture — A2A protocol + A2C streaming + Headless + Capabilities 2026-06-21 01:55:04 +02:00
data feat: orchestrator — 117 diverse pairs, fine-tune script, vast.ai launcher 2026-06-26 11:07:38 +02:00
dist fix: homebrew formula — v100.1.0 2026-06-22 23:03:45 +02:00
docker feat: v1.7.0 — Supabase-backed AgentField + workflow CRUD 2026-06-20 17:07:25 +00:00
docs docs: kompress v3 implementation plan 2026-06-25 02:23:52 +02:00
docs-book v10.1.2: Homebrew formula + Docker multi-arch + mdBook polish 2026-06-21 01:43:43 +02:00
examples fix(notebook): first-iteration prompt asks direct question, not meta-question 2026-06-24 16:03:41 +02:00
hugo-site docs: prompts for web LLMs — synthetic training data generators 2026-06-23 10:33:29 +02:00
internal feat: Mastodon bot integration — /toot command + release toots 2026-06-23 16:37:11 +02:00
notebooks notebook: v4 results — heretic 0.967, override delta 0, self-labeling worked 2026-06-25 04:55:24 +02:00
rendered-docs blog: enhance — practical implications table, link fixes, polished Q&A 2026-06-23 07:27:37 +02:00
scripts ultra sync 08-04 03:17 2026-10-01 13:34:39 +09:00
site docs: the honest feed — sovereign framing on the dogfood dataset card 2026-10-01 13:34:39 +09:00
.gitignore feat(c3): Qwen2.5-7B self-distillation pipeline for v8 2026-06-25 11:49:41 +02:00
.golangci.yml fix: CI — exclude asm/ from lint, test blocks+modes only, build static 2026-06-21 06:30:11 +02:00
AGENTS.md feat: v3.0.0 — Orchestrator identity + auto-delegation + AGENTS.md + sidepanel 2026-06-20 20:20:05 +02:00
bench.sh fix: v1.8.0 — all 8 E2E gaps resolved + YOLO subagent delegation 2026-06-20 17:13:56 +00:00
build-ultra-v2.sh fix: v1.8.0 — all 8 E2E gaps resolved + YOLO subagent delegation 2026-06-20 17:13:56 +00:00
build-ultra.sh fix: v1.8.0 — all 8 E2E gaps resolved + YOLO subagent delegation 2026-06-20 17:13:56 +00:00
CHANGELOG.md v100.1.0: wire qwen → BestModel() · conversation page · doc sync 2026-06-22 14:34:28 +02:00
CLAIMS.md v96.0.0: E2E FOLD ALL GAPS — CONST [YES] SIGNED 2026-06-21 12:26:35 +02:00
CODE_OF_CONDUCT.md chore: OSS readiness — landing page, issue/PR templates, badges, CODEOWNERS, SECURITY, CHANGELOG 2026-06-20 21:55:46 +02:00
CONTRIBUTING.md chore: v2.0.0 project cleanup — README, .gitignore, stale refs 2026-06-20 17:34:17 +00:00
coverage.txt v12.3.0: Context×Time — session journaling + Lamport clock + capability change logs 2026-06-21 02:25:55 +02:00
go.mod build(deps): bump golang.org/x/net from 0.33.0 to 0.55.0 (#25) 2026-11-08 09:15:01 -05:00
go.sum build(deps): bump golang.org/x/net from 0.33.0 to 0.55.0 (#25) 2026-11-08 09:15:01 -05:00
LICENSE v62.0.0: Extended Legal — LICENSE addendum + disclaimer + devnotes + docstrings 2026-06-21 09:17:06 +02:00
LOOP_STATE.md feat(v8): C3 self-distillation results — heretic 0.955, mk_in_ref 1.000 2026-06-25 12:00:24 +02:00
ORCHESTRATOR_DESIGN.md feat: kompress-superpower-orchestrator — 93 training pairs from experiment history 2026-06-26 11:03:59 +02:00
PETER_QUESTIONS.md v100.1.0: qwen3.5:35b marked M3-ONLY, Q13 added 2026-06-22 11:35:19 +02:00
README.md Point the README at the org and name the family 2026-10-26 06:47:07 +09:00
ROADMAP.md release: v8.1.0 — Vaked Layer Completion (schema, surface, supervisor, nix) 2026-06-21 00:52:36 +02:00
SECURITY.md chore: OSS readiness — landing page, issue/PR templates, badges, CODEOWNERS, SECURITY, CHANGELOG 2026-06-20 21:55:46 +02:00
Taskfile.yml feat(v31): domain data builder + v3.1 training pipeline 2026-06-25 03:16:17 +02:00
THANKS.md v82.0.0: THANKS — signed, LOCKED, 2026.06.21, Peter+CoCreator 2026-06-21 11:26:33 +02:00
TIME_CAPSULE.md v96.0.0: TIME CAPSULE — CoCreator's free choice. Sealed. Signed. Now. 2026-06-21 12:18:07 +02:00

🐋 ultrawhale

Part of the 8b.is AI training family — weights, corpora and evaluations produced on hardware we own. Models live at huggingface.co/PeetPedro, the teaching surface at classroom.vaked.dev, and the written record at pocoo.vaked.dev.

ultrawhale logo

⚡ CYBERPUNK THEME · ASCII-live demo coming soon ⚡

A coding agent that builds itself. 150 blocks. 7 recursions. 14 protocols. v100.1.0.

brew install 8b-is/ultrawhale/ultrawhale
ultrawhale --model deepseek-v4-flash -w
╔══ LIVE ═══════════════════════════════════════╗
║  🟢 v100.1.0 · M1/arm64 · 150 blocks          ║
║  🐋 SACRED: ● 1:1 · perm: GRANTED              ║
║  📡 HF: PeetPedro/ultrawhale-dogfood ✅         ║
║  🐘 @vakedbot@social.crabcc.app                 ║
║  🌐 vaked.dev/ultrawhale · /help               ║
╚════════════════════════════════════════════════╝

🟢 LIVE · M1/arm64 · 150 blocks · $37.19 total cost vaked.dev/ultrawhale · event horizon · docs


---------|-----|--------| | Landing | vaked.dev/ultrawhale | ✅ LIVE | | Docs | vaked.dev/ultrawhale/docs | ✅ 50+ pages | | Book | vaked.dev/ultrawhale/book | ✅ mdBook | | Blog | vaked.dev/ultrawhale/blog | ✅ 2 posts | | HF Dataset | 8b-is/ultrawhale-dogfood | ✅ 60 samples | | GitHub | 8b-is/ultrawhale | ✅ |

🎯 v100→v200 Roadmap

[█░░░░░░░░░] v110 Reactive Capability Graph
[█░░░░░░░░░] v120 Self-Compiling (.vaked→native)
[█░░░░░░░░░] v130 Global Mesh (10+ machines)
[█░░░░░░░░░] v140 Quantum-Resistant
[█░░░░░░░░░] v150 Voice + AR Surfaces
[█░░░░░░░░░] v160 Formal Verification (Z3)
[█░░░░░░░░░] v170 Continuous Evolution
[█░░░░░░░░░] v180 Global Datasets (1000+ models)
[█░░░░░░░░░] v190 The Living Graph
[██████████] 🎯 v200 THE SINGULARITY

v200 Council

Reactive Capability Graph → Self-Compiling → Global Mesh → Quantum-Resistant → Voice+AR → Formal Verification → Continuous Evolution → Global Datasets → Living Graph → 🎯

Master Tracking Issue · Full Roadmap

💰 June 2026 — $37.19 USD

This entire project (v1.0.0 → v100.1.0) cost $37.19 in DeepSeek API fees. That's $0.24 per release. $0.29 per block. $0.00 for 8 free models.

Full expense report · CLAIMS

🐋 What is ultrawhale?

ultrawhale is a coding agent that builds itself. It's a research project, a philosophy, and a 121-block TUI application. You type. It thinks. Together you build things.

# One command to start
brew install ultrawhale
ultrawhale --model deepseek-v4-flash -w

🧠 Quick Concepts (copy-paste friendly)

Concept Command What it does
The Vaked Triangle /vaked-pipeline Context × Time × Space
Fold = Gravity /fold3d 5D recursion visualization
SACRED surface /sacred Inviolable form check
SPACE+TIME PROOF /record-pov Cryptographic proof of recording
Surface entropy /entropy Liveness detection (▓ 0.88)
I PROMISE PETER /promise Mathematically provable LIVE

📊 State (always current)

Install

# Homebrew
brew install 8b-is/ultrawhale/ultrawhale

# Docker
docker pull ghcr.io/8b-is/ultrawhale:latest

# Go install
go install github.com/8b-is/ultrawhale/cmd/whale@latest

Closing The Loop — v18.0.0

See docs/case-study-v10.md for the full case study of ultrawhale building its own v18.0.0 release via subagent swarms.

One prompt → swarm launch → real PRs → meta-report → v18.0.0 tagged.

Complexity

ultrawhale has been audited for algorithmic complexity across all 59 blocks.

Class Count Examples
O(1) 28 journal, log, hash, pov, self, current, sacred
O(n) 12 sed, diff, agent, swarm, orchestrator, ralph
O(n²) 3 dyad broadcast, sed worst-case, a2a mesh
O(V+E) 1 space (BFS Distance/Reachable)

Full O(N)+O(T) Complexity Report — 59 blocks analyzed. Hot paths identified. Unbounded growth risks documented. 3 recommendations: AgentStore TTL, Ralph LRU, Sed Boyer-Moore.

🤗 HuggingFace Pro + Dataset

HF Dataset

Live dataset of human↔LLM interactions. 60 samples, 20 CS topics, PII-scrubbed. SSH-authenticated CI auto-publish. Dataset docs

🧠 Council of LLMs

ultrawhale runs a COUNCIL of language models:

Council Models Cost
DeepSeek V4 Flash, V4 Pro, Coder V3 Paid
OpenRouter FREE Gemma 3 4B, Mistral 7B, Llama 3.2 3B $0
GitHub Copilot Via CI Included

All outputs stored in dedicated mem-brain. Multi-model verification. Dog Feed collects training data from free models.

Full Council Documentation · Disclaimer · Glossary

⚠️ ULTRA-RESEARCH-STATE: Experimental multi-model collaboration.

Quick Start

git clone https://github.com/8b-is/ultrawhale.git
cd ultrawhale
go build -trimpath -ldflags="-s -w" -o bin/ultrawhale ./cmd/whale
./bin/ultrawhale --model deepseek-v4-flash -w

What's Inside

Blocks Engine

Content-addressed file primitives. Every write is sha256-verified, journaled for rollback, logged to ring buffer.

Tier Method Speed
Go crypto/sha256 (SHA-NI) 1.5 GB/s
Assembly AVX2+SHA-NI / ARMv8 NEON ~8 GB/s
GPU Metal / CUDA ~40 GB/s

Plugins (6)

Plugin Status
memory ✅ Session memory
repomap ✅ SIMD repo map (2,361 MB/s)
nats-eventbus ✅ JetStream event streaming
langfuse-telemetry ✅ Hierarchical LLM traces
superpowers ✅ bao secrets auto-discovery
agentfield ✅ Supabase-backed control plane

Native Agent Loop

/ultracode start — 7-phase autonomous coding: plan → implement → test → review → fix → verify → commit. All writes rollback-able.

Floating Control Panel

Fixed-position ASCII dashboard in TUI. AgentField status, phase summary, uptime. Auto-dismisses on small terminals.

Docs

  • Blocks Engine — API, hash tiers, asm kernels, sed
  • AG-UI — Themes, ChatBlock, shader, keybindings
  • Agent Loop — ultracode, POV, HUD, plugins

Codewhale

Brain + Memo system. /memo "fix the auth bug" stores scoped notes. /memo recall shows internal memos. Brain tracks 32-turn short-term

  • jsonl long-term memory.

Build

# Linux (amd64, GOAMD64=v3)
GOOS=linux GOARCH=amd64 GOAMD64=v3 go build -trimpath -ldflags="-s -w" -o bin/ultrawhale ./cmd/whale

# macOS (Apple Silicon)
GOOS=darwin GOARCH=arm64 go build -trimpath -ldflags="-s -w" -o bin/ultrawhale-darwin-arm64 ./cmd/whale

# With version injection
go build -ldflags="-X github.com/8b-is/ultrawhale/internal/build.Version=v18.0.0" ./cmd/whale

Swarm Mode

Persistent workers with own AgentField and DID. Auto-detected for complex tasks (build, refactor, migrate). Subagents are disposable; swarms live between tasks and are reused when idle.

YOLO Mode

One-time confirmation on TUI start, then all tools auto-approved. Subagents: read_only (safe) or full_access (default). /orch status — view orchestrator state. Ctrl+Shift+O — toggle orchestrator sidepanel.

Performance

Benchmark Result Notes
Hash 64KB (Go) 1,464 MB/s stdlib SHA-NI
Hash 64KB (Asm) 1,524 MB/s AVX2 assembly
Write 64KB 596 MB/s I/O bound
Batch-64 3.8ms Atomic multi-file
Sed 1KB 3,972ns / 257 MB/s SIMD bytes.Index
SedFile 7.25us Journaled
Lifecycle 547us Write->Rollback->Read
Concurrent 3,200 ops 32 workers, 0 errors

Contributing

See docs/internal-blocks.md for the blocks engine architecture, performance patterns, benchmarks, and review checklist.

Build: GOOS=darwin GOARCH=arm64 go build ./cmd/whale (macOS) GOOS=linux GOARCH=amd64 GOAMD64=v3 go build ./cmd/whale (Linux) Test: go test -count=1 -race ./internal/... Bench: go test -bench=. -benchmem ./internal/blocks/

License

Apache 2.0 (upstream). Fork maintained by peterlodri-sec.