GoldBug is hunting for gold...
GoldBug is hunting for gold...
417+ Agents. 232 Skills. 9 Combat Modes. 40 Routes. 12 Task Types. 4-Layer Routing. A single command dispatches an entire legion — every agent, every skill, every engine orchestrated in perfect synchrony.
These commands are verified in production. Copy to your terminal, Legion deploys instantly.
232 skills — if you had to figure out which to use, how to combine them, and in what order every single time, they'd be useless. The Command Hub's value isn't the count. It's "in what situation, call whom, in what order, with what verification." One cut — the optimal path is already decided.
Regex + keyword matching. 12 task types cover EVERY possible user request. Zero ambiguity.
Complexity-gated. LOW skips overhead. HIGH adds wayfinding + tickets. Never over-engineer, never under-plan.
The assembled chain isn't just a list. Each link has built-in verification. If one link fails, the chain auto-corrects.
grill(clarify) → spec(lock) → implement(code) → tdd(test) → code-review(review) → verify(check) → commit → shipEvery link verifies the previous. TDD verifies code. Review verifies TDD. Verify verifies everything. Zero unverified output.
The Core Insight: Skills are useless without a commander.
232 skills scattered everywhere = noise. 232 skills called by ONE hub at the right time in the right order = invincible legion. That's why the Command Hub matters more than any single skill — it's the intelligence layer. You don't need to know 232 skills. You tell the hub what you want. The hub knows how to dispatch.
These are not theories. These are the exact chains our own Legion uses every day. Copy them. Use them. They work.
Why this works: grill prevents scope creep. spec locks decisions before code. tickets enable parallel work. implement's built-in tdd+review catches errors early. ocr catches what tdd misses. ship ensures clean release. Each step validates the previous.
Iron Rule: NEVER guess the fix. Always diagnose first. 80% of 'bug fixes' that fail are because the root cause was wrong. Tight feedback loop: reproduce → fix → verify in under 2 minutes.
Key: butterfly-hook extracts 12-dimension DNA from any website (layout/grid/typography/color/spacing/animation/interaction/component/iconography/imagery/tone/accessibility). Not just copying — understanding the design system.
Why multi-source: Single-source research = single-view bias. 3 parallel searches + adversarial verification catches contradictions. Cross-reference builds confidence. Zero unverified claims.
These 4 patterns cover 90% of our daily work.
SaaS building, bug fixing, UI design, deep research — these are the four most frequent task types of the Invincible Legion. Every chain is battle-tested and repeatedly optimized. You don't need to reinvent the workflow each time. The hub already knows the optimal path. You just say: 'What I want to do.'
Every request flows through four precision layers — classified, routed, and dispatched to the optimal agent chain.
12 task types. Instantly identifies what the user truly wants.
WRITE_CODE · FIX_BUG · REVIEW_CODE · IMPROVE_CODE · DESIGN_UI · ANIMATE · RESEARCH · DEPLOY · QA_TEST · SECURITY · DOCUMENT · TRIAGELOW → direct. MEDIUM → spec+implement. HIGH → wayfinder+spec+tickets per ticket.
Regex-driven complexity detection. 3-Tier model routing (WASM/Flash/Pro). 30-50% token savings.Frontend / Backend / Game / iOS / Data Science / Blockchain / GIS / Marketing / Security / Finance / Legal / Healthcare — 18+ domains with specialized agents.
22 Custom + 235 Agency + 49 GameStudios + 100 Built-in = 406+ specialized agents across every domain.Auto-assembles optimal skill chain: grill→spec→implement with tdd+review built-in. Conflict resolution matrix ensures best skill wins.
6 canonical pipelines. Conflict resolution: OCR review > MP code-review > built-in. Every task has one optimal path.Every task type has a complexity-graded routing chain. No guessing. No wandering. One optimal path.
WRITE_CODECore Pipelinegrill→spec→implement(tdd+review)FIX_BUGCore Pipelinediagnosing-bugs→tdd→code-reviewREVIEW_CODECore Pipelineocr:review (29 personas + debate)IMPROVE_CODECore Pipelineimprove-architecture→codebase-design→implementDESIGN_UIVisual Creationdesign-consultation→design-html→design-review / butterfly-hook(clone)ANIMATEVisual Creationfind-animation→GSAP 8-piece→review-animationsRESEARCHKnowledgeMP research + deep-research + wigolo + build-your-own-xDEPLOYOperationsship (test→version→CHANGELOG→PR→merge)QA_TESTOperationsverify + qa + self-qa + site-qa + e2e + benchmarkSECURITYOperationssecurity-review + api-security + cso + sandbox-guardDOCUMENTKnowledgedocument-generate + make-pdf + guizang-pptTRIAGEKnowledgetriage(3-stage)→wayfinder(decision maps)→to-spec→to-ticketsFrom solo scout to full legion assault — every mission has its optimal formation.
Explorer+Researcher → synthesize. Understand before acting.
Strategist→Architect→[Backend+Frontend]→Test→Review→Security→deliver. Full pipeline.
[Reviewer+Security+Perf] parallel → synthesize. Multi-angle verification.
BugDetective+Explorer→fix→Test verify. Root cause before fix.
Researcher×3 multi-angle → cross-verify → synthesize. No single-view bias.
8-gate stage pipeline. Each gate must pass before next. Zero tolerance.
3 agents parallel → comparison matrix → human decides. No single model bias.
4 topologies (hierarchical/mesh/hybrid/adaptive) + 5 consensus algorithms. Anti-drift.
5-stage closed loop. Mandatory post-mortem. Confidence-based routing auto-tuning.
"The future is not choosing a model — it's routing between models." 93% accuracy, 50× cost reduction.
<1ms · $0 · 6 deterministic transforms
var→const, add-types, add-error-handling, async-await, add-logging, remove-console
~500ms · $0.0002 · Simple tasks
Complexity <30%. Fixes, typos, comments, minor changes.
2-5s · $0.003-0.015 · Complex reasoning
Architecture, security, multi-file refactors. K3 for frontend, DSv4-Pro for code.
Base Reasoning · Code King
1.6T MoE · 49B active · $0.87/M output · MIT · 3206 Codeforces
Agent King · Frontend #1
2.8T · 104B active · 1679 Elo frontend · 91.2 BrowseComp · 49% non-hallucination
2nd Foundation · IndexShare
753B MoE · IndexShare 2.9× FLOPs · slime framework · Anti-Hacking
Mandatory post-mortem after every task. Confidence-based routing auto-tunes. The legion that never stops learning.
Look up routing table → select agents → execute.
Self-score 1-10. Which agent helped? What to improve?
Success → +0.02 confidence. Failure → -0.05 + try fallback.
After 5 similar tasks: best combos, retire/consolidate.
Every 10 tasks → evolution report: success rate, TOP5, new/eliminated routes.
Every task → mandatory post-mortem → score → record → optimize routes → stronger next time. 3 consecutive successes = +0.10 confidence boost. 2 consecutive failures = immediate demotion + alternate. The legion gets smarter with every battle.
5-step pipeline before every reply. 12 background workers. Triple redundancy midnight evolution. Never sleeps.
Only skipped for greetings/single-word confirmations. 100% execution otherwise.
Legion is the dispatch center. These skill engines are the execution muscles.
Garry Tan's 118K⭐ methodology. 53 skills embedded into 8 combat modes. Legion decides WHO, gstack decides HOW.
14+ distilled celebrity advisors (Paul Graham, Karpathy, Jobs, Musk, Munger, Taleb…). 4-phase distillation pipeline. Auto-activation by keywords.
29 distilled reviewer personas + structured adversarial debate protocol. 29 perspectives, not 29 domains — each reviews through a different lens of thinking.
17 Ruflow hooks merged with 5 Legion hooks. Security gating, routing, session persistence, collaboration, quality checks. Every lifecycle moment is intercepted.
4 topologies + 5 consensus algorithms. Byzantine fault tolerance for critical tasks. Auto-trigger on 3+ file changes. Anti-drift 4-layer protection.
WASM Agent Booster (6 deterministic transforms) → Flash/Haiku → Pro/Opus+DSv4-Pro. Complexity detection via regex. 30-50% token savings.
8 GSAP skills: core, timeline, scrolltrigger, react, flip, observer, scrolltoplugin, matchmedia. Full animation pipeline from discovery to review.
ComfyUI (116K⭐) node-graph visual editing. gpt-image-engine for text-to-image. Dual-provider, model-agnostic routing.
ViMax (Wan2.2/Cosmos3), OpenMontage, Seedance. 12 pipelines, 52 tools. Text-to-video, image-to-video, video editing.
LSP-based. 15+ MCP tools. Find symbol, find references, rename, replace, insert — all with semantic precision across the entire codebase.
7-level de-fat staircase: -80% to -94% code lines. Lite/Full/Ultra modes. Karpathy-style surgical precision. Lazy senior engineer philosophy.
3 autonomy levels (L1 report-only, L2 assist-fix, L3 unattended). 7 patterns. Safety denylist path matrix. ≤3 retries per fix.
The blade that never dulls. Self-evolving. Self-sharpening. Self-verifying.
Brain > Technique > Documentation. Every round: 'Am I enhancing the brain or writing docs?'
Blade Sharpness = PASS/(PASS+FAIL) per 5 rounds. Target: 100%.
Iron Page Rule: page must never break. Verify after every change.
Anti-adversarial self-audit: find counterexamples before submission. Zero counterexamples = didn't look hard enough.
Decision provenance: every key decision annotated with knowledge source.
Harness evolution: experience learning → tool creation → prompt distillation. Agent grows without weight changes.
Right-click → Save Image As… or click Download SVG · Hub-and-Spoke: Legion commands from center, everything radiates outward. Not static — continuously evolving.