SYSTEM DEFINITION

One core protocol / Multiple native surfaces

Hypertaks establishes a single source of truth across all AI development platforms through thin host adapters, evidence ledgers, and deterministic intake framing.

SYSTEM CORE

01. INTAKE User request, project state, constraints, permissions, evidence ledger

↓

02. HYPERTAKS CORE PROTOCOL INTAKE → FRAME → ALLOCATE → BIND → EXECUTE → VERIFY

↓

01. CANONICAL

Instructions & Skills

Define the operating method and the public command surface across hosts.

02. DELIVERY

Host Adapters

Expose the core through host-native plugin, extension, skill, or instruction mechanism.

03. EXECUTION

Tools & Python

Filesystem, GitHub, browser, shell, database, connectors, and MCP tools perform real actions.

04. EVIDENCE

Verification

Tests, citations, ledgers, checkpoints, and reconciliation prove what happened.

DECIDE

Objective over weak method

Separate the real objective from a proposed method that would harm quality, cash, trust, or capacity.

EXECUTE

Sized specialist work

Allocate only the roles and verified capabilities the task needs. No padded teams.

REMEMBER

Evidence-backed context

Preserve founder context with source, freshness, and authority boundaries. Memory never approves action.

PROVE

Reconcile before done

Tests, artifacts, Git state, handoffs, and acceptance criteria must agree before a completion claim.

02

OPERATING MODEL

Fixed loop. Proportional weight.

The founder loop stays the same from Nano to Hyper. Only the team size, evidence depth, and approval weight change with the task.

PHASE 0 TO 5

Boss objectiveT1 authority
Safety kerneltask contract
Founder loopPhase 0 to 5
Specialist worksized by tier
Evidence packtests and artifacts
Integrated resultfounder-grade

Founder Brain
evidence-backed memory

Graph and retrieval
need-routed intelligence

Continuity
checkpoint, resume, handoff

Optional systems
Graphify, Obsidian, MCP when verified

Phase	Founder responsibility
0. Intake and Verify	Confirm the real objective, deliverable, scope, evidence, permissions, and definition of done.
1. Frame	Test feasibility, business impact, risk, and strategic fit.
2. Pick Roles	Select distinct specialist perspectives without padding the team.
3. Equip	Bind the smallest sufficient verified capabilities and frameworks.
4. Produce	Execute inside the contract, dependency order, and permission boundary.
5. Integrate and Deliver	Reconcile conflicts, validate evidence, disclose limits, and deliver one result.

Nano answers directly. Lite uses one Founder. Standard uses three roles. Prime uses five. Hyper or Omega splits a large program into distinct workstreams with human gates.

03

RELEASE / PACKAGE STATUS

v4.5.8 Founder Anti-Slop, Subagent Tiering & 88/88 Behavioral Certification

Hypertaks v4.5.8 embeds the Founder Anti-Slop Mandate, deterministic subagent model tiering discipline, mandatory multi-modal visual delivery, and achieves 88/88 behavioral eval PASS with release threshold 80, while preserving the five canonical skills and four read-only MCP tools.

SYNCED MANIFEST

RUNTIME STATUS

Product 4.5.8

Health badge: STALE BASELINE

Source: baseline 4.5.8

Build revision: unknown

PUBLIC MANIFEST https://hypertaks.crimsonriftstudio.com/manifest

Machine-readable product boundary for websites and operators. No secrets. Read-only.

OPEN MANIFEST JSON

GITHUB RELEASE v4.5.8

Release notes, tag, and repository evidence stay on GitHub.

OPEN RELEASE

Surface	Status
Public product	Five canonical Hypertaks skills
Remote MCP surface	Four read-only tools; no mutating remote tool is exposed
Founder runtime	Existing runtime behavior with synchronized release metadata
Expansion Lab	Non-production research, prototypes, fixtures, and evidence
Behavioral certification	Documented host routes are compatibility records, not universal live host certification

Release boundary: prototypes do not widen the five public skills or the four read-only remote MCP tools. Production integration needs a later explicit contract.

02

INSTRUCTIONS / PUBLIC INTERFACE

Exactly five canonical skills

Public commands expose Hypertaks capabilities to users and agents across all registered hosts without duplicating logic.

5 SKILL CORE

#	Command	Name	Scope & Operating Purpose
01	`/hypertaks`	Founder Operating System	Complete founder loop for analysis, decisions, builds, and integrated delivery.
02	`/hypertaks-verify`	Verification Kernel	Environment, project, brain, storage, Graphify, and Obsidian verification.
03	`/hypertaks-brain`	Founder Memory	Evidence-backed founder facts, decisions, preferences, corrections, and memory promotion.
04	`/hypertaks-graph`	Relationship Graph	Relationship queries, change-impact analysis, Graphify routing, or direct-search fallback.
05	`/hypertaks-continuity`	Continuity Ledger	Checkpoint, resume, handoff, reconciliation, and proof of done.

The skills instruct the host agent. They do not manufacture capabilities that the current host has not exposed.

03

PLUGIN / ADAPTER LAYER

Thin adapters preserve one source of truth

Host adapters translate native plugin manifest structures, extension APIs, and prompt schemas directly into Hypertaks core skills.

ADAPTER MATRIX

Adapter Type	Target Hosts	Integration Method & Purpose
NATIVE PLUGIN	Claude Code, Codex, Cursor, Kimi, OpenCode	Marketplace or plugin-manager metadata, commands, rules, and skill discovery.
HOST EXTENSION	Pi, GitHub Copilot, Open WebUI, LibreChat, Hermes, OpenClaw	Host APIs, git plugin routes, or extension surfaces register the canonical Hypertaks package.
MANAGED SKILL	Antigravity, Cline, Windsurf, Kilo, Goose, OpenHands	Generated or managed packages expose the same five skills without duplicating the product core.
PROJECT INSTRUCTIONS	Aider, Claude.ai, Gemini App	Hosts without a full native plugin route consume the canonical instruction bundle.
CHATGPT APP ADAPTER	ChatGPT App	HTTPS MCP transport connects to a hosted MCP endpoint.
PACKAGE POLICY	All generated packages	Canonical skills + runtime policies + thin adapters. Generated packages must not drift from skill core.

04

MCP / TOOLS / CAPABILITIES

Plugin is the package / MCP is a transport / Tools do the work

Distinguishing distribution packages, remote communication protocols, and underlying executable capabilities ensures exact host capability binding.

MCP TRANSPORT

Term	Role	Description
PLUGIN	Distribution	Skills, commands, rules, manifests, metadata, adapters, and lifecycle instructions.
MCP	Protocol	Connects the host to external tools or data. Required for ChatGPT transport; optional elsewhere.
TOOL	Execution	Read, create, update, delete, execute, communicate, browse, query, deploy, or validate.
CONNECTOR	Boundary	GitHub, email, calendar, storage, database, browser, and authenticated systems.

MCP CONNECTION / RUNTIME

Connect with POST. Inspect with JSON.

Configure an MCP-compatible client with the HTTPS endpoint. Browser GET on /mcp returns 405 by protocol design. That is not an outage.

MCP ENDPOINT https://hypertaks.crimsonriftstudio.com/mcp

JSON-RPC Streamable HTTP. Use this URL in client configuration, not as a normal web page.

HEALTH CHECK https://hypertaks.crimsonriftstudio.com/healthz

Human and machine health JSON. Not a ChatGPT connection URL.

OPEN HEALTH JSON

PUBLIC MANIFEST https://hypertaks.crimsonriftstudio.com/manifest

Version, skills, package status, and read-only MCP tool inventory.

OPEN MANIFEST JSON

Setup notes

Copy the MCP endpoint into your MCP-compatible client.
Prefer protocol version 2025-06-18 or another supported 2025 Streamable HTTP version.
Call initialize, then tools/list.
Use hypertaks_route before hypertaks_get_skill for ambiguous requests.
Do not claim local filesystem mutation from this remote adapter.

Read-only MCP tools (4)

hypertaks_manifest
hypertaks_get_skill
hypertaks_route
hypertaks_verify_installation

Remote MCP cannot replace host filesystem, browser, database, or deployment tools.

05

TOOLS & EXECUTABLE ACTIONS

Bounded execution with verifiable output

Hypertaks orchestrates real tools provided by the host environment, enforcing safety boundaries, single-commit rules, and empirical verification.

TOOL ENGINE

Operation	Capability Domain	Execution Purpose & Safety Boundary
`READ`	Filesystem / Repository / Web	Collect evidence and inspect state. Read operations require no interactive prompt.
`CREATE & UPDATE`	Files / Code / Configuration	Mutating operations remain permission-bound and limited to explicit task paths.
`EXECUTE`	Shell / Tests / Python / TypeScript	Run reproducible calculations, builds, validators, and runtime checks.
`COMMUNICATE`	Email / Issues / Deployment	External effects require preview, approval, single commit, and reconciliation.

06

PROCESS & AI SERVER DISPATCHER

Interactive Hypertaks Skill & Adapter Dispatcher

Generate a configuration preview for a selected Hypertaks skill, host adapter, and runner. This interface does not execute anything.

INTERACTIVE RUNNER

1. SELECT HYPERTAKS SKILL

2. SELECT HOST ADAPTER

3. SELECT TOOL / RUNNER REFERENCE

4. USER INSTRUCTION / COMMAND TO AI AGENT

HYPERTAKS CONFIGURATION PREVIEW

STATUS: READY

Hypertaks configuration preview ready.
Select a skill, host adapter, and runner reference to generate a configuration preview.
Ready to generate a preview.

07

VISUALS / VISUAL DELIVERY

Evidence discipline and visual validation

Visual outputs, diffs, diagrams, and screenshots are verified through empirical evidence rather than speculative descriptions.

EVIDENCE GATE

GATE 01

INTAKE GATE

Resolve objective, feasibility, permissions, evidence needs, and definition of done before writing code.

GATE 02

EVIDENCE ROUTING

Bind only verified skills, tools, MCP capabilities, connectors, and permissions.

GATE 03

EXTERNAL EFFECTS

PREPARE → PREVIEW → APPROVAL → COMMIT ONCE → RECONCILE

GATE 04

VISUAL VALIDATION

Screenshots, diffs, and rendered outputs must be empirically verifiable, not described.

The connected MCP server determines the tools and access available to ChatGPT. It does not replace filesystem, browser, database, or deployment tools.

08

PROCESS / WORKFLOW

Six phases with explicit allocation tiers

Tasks progress through six structured phases, matching agent team size and allocation depth to measured risk and scope.

6-PHASE FLOW

PHASE 00

Intake & Verify

Resolve objective, feasibility, permissions, evidence needs, and definition of done.

PHASE 01

Frame

Separate the desired outcome from the proposed method and detect contradictions early.

PHASE 02

Allocate

Choose Nano, Lite, Standard, Prime, or Hyper from real scope and risk.

PHASE 03

Bind

Select only verified, relevant skills, tools, MCP capabilities, connectors, and permissions.

PHASE 04

Execute

Run distinct specialist roles or synthesize them honestly when parallel agents are unavailable.

PHASE 05

Integrate & Prove

Reconcile outputs, test claims, disclose limitations, and record evidence.

Tier Name	Max Concurrent Subagents	Target Task Scope & Complexity
NANO	0 Agents	Direct response, trivial single-file edit or query
LITE	1 Agent	Focused task, single component bug fix or refactor
STANDARD	3 Agents	Moderate scope, feature addition with test coverage
PRIME	5 Agents	Cross-domain architecture work, multi-package refactor
HYPER	6–10+ Agents	Multi-workstream program, major codebase transformation

09

KNOWLEDGE BASE / METHODS

Retrieve, evaluate, and pack evidence before generation

Hybrid retrieval algorithms combine exact code symbols, semantic vectors, Graphify relationship graphs, and Obsidian notes into packed evidence prompts.

RAG KERNEL

NEED→ SCOPE→ ROUTE→ RETRIEVE→ FUSE→ RERANK→ EVALUATE→ PACK EVIDENCE

Retrieval Route	Methodology	Use Case & Targeted Assets
EXACT / KEYWORD	Identifiers and literal vocabulary	Codes, filenames, proper names, quoted phrases, and exact terminology.
VECTOR / SEMANTIC	Meaning and paraphrase	Conceptual similarity, multilingual meaning, and vocabulary mismatch.
HYBRID / FUSION	Mixed query populations	Combine independent routes only when both materially contribute.
RERANKING	Ambiguous candidate sets	Use a verified reranker only when accuracy gain justifies cost and latency.
FOUNDER BRAIN	Evidence-backed memory	Persistent context remains attributable, scoped, and explicit about uncertainty.
GRAPHIFY / GRAPH	Relationships and impact	Connect decisions, dependencies, entities, files, and downstream effects.
OBSIDIAN	Human-readable continuity	Checkpoints, operating records, and project memory remain reviewable outside model.

10

PROGRAMS / REPOSITORY REFERENCE

Runtime code and references support the instruction core

Tracked repository modules implement Python installers, TypeScript memory runners, distribution manifests, and automated evaluation suites.

REPOSITORY MAP

Repository Path	Role	Architectural Purpose
`skills/`	Five canonical public skills	The source of product behavior and instruction policy.
`skills/hypertaks/references/`	Operating references	Security kernel, state transactions, retrieval, professional execution, visual delivery.
`runtime/`	TypeScript routing & brain	Strict reference implementations, state logic, tests, and ChatGPT transport.
`scripts/`	Python utilities	Installer, updater, retrieval evaluation, charts, validators, distribution builders.
`distribution/`	Host compatibility records	Registry, package templates, capability matrix, and evidence classifications.
`evals/`	Verification corpus	Security, quantitative, workflow, retrieval, host, and cross-host conformance cases.

HOST INSTALLATION REFERENCE

git clone https://github.com/aabrur/hypertaks-agent.git
cd hypertaks-agent
python scripts/installer.py list-hosts --json

Clone the canonical repository, then use the host-specific installer or plugin manager documented for your environment.

The tracked catalog documents 22 host routes. A compatibility PASS means an official route and matching adapter were documented. It does not mean every account-bound host executed live in this release. Full matrix: repository README.

11

REFERENCES / SECURITY KERNEL

State integrity and safety policies

Strict security kernels prevent unauthorized file deletion, unverified external API mutations, and unbacked memory claims.

SECURITY KERNEL

Reference Module	Security Principle	Enforcement Mechanism
`security_kernel.md`	Permission-bounded execution	Explicit user approval required before external mutations or destructive shell commands.
`state_transactions.md`	Transactional continuity	Checkpoint state before complex edits; verify tests before committing changes.
`retrieval_intelligence.md`	Evidence attribution	All claims must cite authoritative local files or verified web sources.
`professional_execution.md`	Surgical code modifications	Touch only what is requested; clean up orphans created by your changes.

Turn a request into agoverned result