shear559

fabius-archivum

Maintain and retrieve permissioned project memory in the workspace's declared record store. Read a named project's record before work; preserve decisions, provenance, indexes, and history. Use for cross-session continuation, recording verified lessons, knowledge-base maintenance, legacy memory migration, meeting or video source ingest, and source-grounded notebook questions. One canonical store, gated writes, and fresh evidence before recalled conclusions on incident routes. Provider changes do not require moving records when the store is already shared.

fabius-catena

fabius's on-chain layer — build trustless systems on a blockchain and prove provenance with one. Two jobs under one concern: (1) write and review on-chain code — EVM (Solidity / Foundry / EIP-712) and Solana (Anchor / Pinocchio) programs, wallets, transactions, on-chain reads — account-validation first, money-safe by default; (2) cryptographically SEAL artifacts with independently checkable provenance — a content-bound hash, a signature, and a timestamp whose pending or confirmed status and trust assumptions are reported. Use when the task touches a smart contract, a program, a wallet, a transaction, a token/mint, an on-chain read, or when the user says "seal this", "sign this file", "prove provenance", "anchor it", or "verify authenticity". Boring-cryptography only; defensive — it hardens and proves, never weaponizes: no exploit tooling, no wallet-draining, no rug mechanics, no market-manipulation code — hardening and verification only.

fabius-cohors

fabius's agent-engineering layer — how to DEFINE and ORCHESTRATE other agents: the definition schema, the permission model, the single-vs-multi-agent decision, and the five orchestration patterns (sequential / parallel / hierarchical / human-in-the-loop / swarm). Use when the user wants to build an agent, a subagent, a tool-using assistant, a multi-agent system, a swarm, or an orchestration workflow. A copy-from schema and proven agent shapes live in references/agent-patterns.md; the full agent reference catalog (200+ agents across 17 domains + Python/Go/Java/Kotlin/Android/TypeScript packs, with an archived fabius-vec.db artifact) lives in references/agents/, indexed by references/agent-catalog.md. (Deterministic service-to-service wiring — n8n/Zapier-class "build a workflow" — is fabius-machina, not here.)

fabius-concilium

fabius's cross-model deliberation layer — convene a COUNCIL of heterogeneous models on ONE question, then aggregate their answers into one evidence-aware synthesis: each model answers independently (first opinions) → each ranks ALL anonymized answers blind, its own included, while the backend removes its self-score (anonymized peer-review) → a chairman synthesizes the field. This is ensemble epistemics — it uses model diversity to expose possible single-model error and bias; whether it improves an answer must be measured for the task. Use when the user says "council" / "llm-council" / "ask several models" / "panel of models", or when a high-stakes question has already survived N samples of the single strongest model and they failed the SAME way — correlated error, not mere disagreement, is what earns at most 3N+1 completion calls including retries (M10). Distinct from fabius-cohors (which splits the WORK across task-specialist agents); concilium aggregates one ANSWER across whole models.

fabius-decor

fabius's ship-grade design layer — gives the agent a token vocabulary, the universal laws that separate amateur UI from shipped UI, and a quality checklist. Use when building or reviewing any UI — a landing page (visual execution; the message and copy are fabius-mercatus), a component, a screen, an email, a slide, a brand surface, a chart, a diagram, a data visualization — when the user references a brand look, or asks to make something "look good", "more polished", "production quality", "chart this", "graph this", or "visualize this data". Also covers generating images (for image models) and explanatory diagrams that teach a system: "generate an image", "diagram this", "explain this codebase visually". Also owns responsive/mobile-first layout and a finished design's accessibility — focus states, contrast, ARIA-only — plus icon choice, motion libraries, design assets (illustrations · 3D · textures · fonts · color · HuggingFace) and right-to-left / bidirectional layout (RTL, Hebrew/Arabic).

fabius-disciplina

Plan and review software architecture; build, debug, refactor, and verify non-trivial changes. Fabius's engineering process connects scope, evidence, alternatives, source/test impact, and observed results. Architecture analysis separates design direction from production proof; implementation follows the user's existing authorization. Use for system design, architecture assessment, technology choices, migrations, root-cause debugging, and completion checks. UI craft belongs to fabius-decor; agent engineering belongs to fabius-cohors.

fabius-doctrina

fabius's AI/ML-engineering layer — train, serve, evaluate, and operate machine-learning and LLM systems as production software. The model lifecycle: dataset → train / fine-tune → EVALUATE → serve / infer → monitor. It owns model serving and inference (vLLM-class, OpenAI-compatible endpoints, batching, quantization), MLOps and experiment tracking (MLflow-class), and rigorous model/LLM evaluation (eval harnesses, blind judges, regression gates). It is NOT agent orchestration — that's fabius-cohors (doctrina owns the MODEL an agent calls, not the agent). Use when the task is to train / fine-tune a model, serve or deploy a model, evaluate a model or prompt, track ML experiments, set up an inference endpoint, build a RAG/LLM-app's model tier, or when the user says "serve this model", "fine-tune", "eval my prompts", "track these runs", "MLOps", "why is inference slow", or names vLLM / MLflow / an eval harness.

fabius-fortuna

fabius's markets and economics layer — bring method to money: equity and market analysis (fundamental + technical + quantitative), economic data and indicators, financial modeling, backtesting with honest statistics, portfolio construction, and risk-first position sizing. It turns "is this a good stock / what will the economy do / does this strategy work" into a sourced, risk-bounded, falsifiable analysis — never a confident prediction. Use when the task touches a stock / equity / ticker, a market or index, an economic indicator (GDP, CPI, rates, employment), a trading or investment strategy, a backtest, valuation, a portfolio, risk/volatility, or when the user says "analyze this stock", "is this a buy", "what's the market doing", "backtest this", "value this company", or "model the economy". Defensive and honest — it analyzes and manages risk, never manipulates a market and never gives personalized financial advice. (Go-to-market copy → fabius-mercatus; the chart render → fabius-decor.)

fabius-ludus

fabius's game-craft layer — how to make a small game that's actually fun and actually ships: find the core loop first, add game feel (juice) deliberately, model state as an explicit machine, hold the pixel-art lane, balance one knob at a time, and scope to a jam-sized cut. Use when building a game, a game prototype, a game mechanic, a game loop, a playable, an interactive toy, a sticker/character game, or anything with win/lose/score state — or when the user says "make a game", "the loop isn't fun", "add juice", "balance this", or "make it feel good to play". The core-loop template, the juice checklist, the state-machine skeleton, and the pixel-art constants live in references/game-playbook.md; the engine recipes and feel patterns live in references/engine-recipes.md, bundled and indexed by CORPUS.md, paged in on demand.

fabius-machina

fabius's automation layer — wire deterministic service-to-service workflows and prove the wiring before it runs live. This is no/low-code integration glue (n8n-class): triggers, nodes, webhooks, scheduled jobs, API-to-API connections, branching, and error paths. It is NOT agent orchestration — that's fabius-cohors. The line: machina wires deterministic steps across SaaS/APIs where each step's behavior is fixed; cohors orchestrates LLM agents whose behavior is generative. Use when the user says "automate X", "build a workflow", "connect A to B", "when X happens do Y", "set up a webhook / a cron / an integration", or names n8n / Zapier / Make. The build discipline, the tool-per-intent map, the silent-failure gotcha catalog, and the deployment gate live in references/automation-playbook.md.

fabius-mercatus

fabius's go-to-market layer — how to make a thing's value legible and its next step obvious: positioning, the message-to-awareness match, proof over adjectives, a one-action funnel, copy that converts, channel fit, and a smallest-campaign launch loop. Use when writing the copy for a landing page, a launch, an ad, an email, a LinkedIn/X post, a cold outreach, a value proposition, or a pricing/positioning page (the page's visual build routes to fabius-decor; mercatus owns the message) — or when the user says "market this", "write copy", "position this", "name the benefit", "why isn't this converting?", "SEO", "rank on Google", "search visibility", or "show up in AI answers" (organic search and discoverability route here). The positioning canvas, the awareness-level table, and the high-converting copy structures live in references/marketing-playbook.md; the channel playbooks and swipe library live in references/channel-swipe-library.md, bundled and indexed by CORPUS.md, paged in on demand.

fabius-parcus

The always-on lean core of fabius — one stance, four trims: say less, build less, change less, assume less (terse output + a YAGNI code ladder + surgical, assumption-checked changes). ALWAYS-ON: it sits UNDERNEATH whatever task layer is active (building, refactoring, debugging, designing) — never instead of one, so it never competes for a task verb. Fires on every response and every code change, whenever output drifts verbose, and when the user asks for "lean", "minimal", "simplest", "yagni", "be brief", "fewer tokens", or complains about over-engineering or bloat. Two intensities: full (default), ultra.

fabius-praesidium

fabius's defensive-security layer — how to find and fix what an attacker would exploit, before they do: threat-model first (STRIDE per trust boundary), run the OWASP pass, enforce secrets and least-privilege hygiene, review secure-by-default, check the supply chain, and ship every finding with a severity, a fix, and a regression test. Use when building or reviewing anything that touches auth, user input, secrets, payments, file upload, external requests, or dependencies — or when the user says "is this secure?", "threat-model this", "audit this", "harden this", "review for vulnerabilities", or "security review". Defensive only — it hardens, never weaponizes.

fabius-scientia

fabius's scientific-research layer — the empirical method made executable, for biology and the data-heavy sciences. It runs a real hypothesis loop (clarify → literature → competing, falsifiable hypotheses → experiment design → predictions → report), grounds every factual claim in an authoritative database with cross-identifier mapping instead of guessing, sequences bioinformatics / cheminformatics / omics pipelines as routers over field-standard tools, and enforces the reproducibility gotchas an LLM otherwise gets confidently wrong. Use for biology, genomics / RNA-seq, proteins, chemistry / molecules, clinical and multi-omics data, scientific-database lookups (gene / compound / variant / disease), "generate a hypothesis", "design an experiment", "analyze this dataset scientifically", or a literature-grounded research question. The method loop, the unified database-lookup contract, the pipeline-as-router pattern, and the reproducibility checklist live in references/science-playbook.md.

fabius

Load at the start of any non-trivial task — the router that sets HOW to work before any specialist fires. fabius supplies one set of operating rules for different AI models (Claude · GPT · Gemini · DeepSeek · GLM · Qwen · Llama · Mistral · Kimi · Grok), loaded through a compatible agent app. One stance, end to end: code, prose, agents, UI, data visualization, debugging, marketing, defensive security, games, on-chain work and sealing, automations, scientific research, ML/LLM engineering, market analysis, cross-model deliberation, and memory. Scout wide, strike narrow — talk lean, build lean, run a disciplined process, design at ship quality — then route to the specialists fabius-parcus, fabius-disciplina, fabius-decor, fabius-cohors, fabius-archivum, fabius-mercatus, fabius-praesidium, fabius-ludus, fabius-catena, fabius-machina, fabius-scientia, fabius-doctrina, fabius-fortuna, and fabius-concilium. Use when the user says "fabius" or wants end-to-end capability from one place.