Hypertaks turns an AI agent into a founder-grade operating partner that frames the real objective, challenges harmful methods, coordinates specialist work, preserves verified founder context, and proves whether the work is done.
Execution profiles: CORE for smaller models and FULL through SKILL.md for frontier agents.
Hypertaks is a portable operating protocol for founder-shaped work across business strategy, software engineering, marketing, finance, ERP, supply chain, smart contracts, data, design, and IoT.
The Boss remains the final human authority. The Hypertaks Founder owns task framing, contract integrity, specialist coordination, capability relevance, evidence quality, founder continuity, risk disclosure, and final integration.
Hypertaks is not a hosted model, autonomous legal entity, mandatory memory service, or bundled agent runtime. It runs inside the AI agent and tools the user already selected.
Version 4.5.0 adds the Founder Brain and Continuity System.
A useful founder-grade agent must do more than answer one prompt well. It should remember approved decisions, preserve why they were made, distinguish facts from inference, resume from real repository state, pass work to another agent, and reject unsupported claims of completion.
Hypertaks 4.5.0 adds that capability without replacing its Founder Operating System identity with a generic memory layer.
- exactly five public Hypertaks skills;
- discovery and reuse of an existing main brain or shared agent memory;
- evidence-backed agent-private, project, and shared memory;
- optional Obsidian Vault storage owned by the user;
- optional Graphify relationship and impact analysis;
- Git-aware checkpoint, resume, and cross-agent handoff;
- proof of done based on current tests and acceptance evidence;
- approved-root path containment, atomic writes, and secret protection;
- EV-66 through EV-88 for founder brain and continuity behavior.
Hypertaks exposes exactly five public entry points.
| Skill | Purpose |
|---|---|
/hypertaks |
Main Founder Operating System flow for normal work |
/hypertaks-verify |
Verify project, brain, shared memory, storage, Graphify, and Obsidian configuration |
/hypertaks-brain |
Inspect, record, revalidate, correct, promote, demote, or archive founder memory |
/hypertaks-graph |
Query code relationships and impact through verified Graphify capabilities or direct search |
/hypertaks-continuity |
Checkpoint, resume, handoff, reconcile, status, and proof of done |
The main /hypertaks skill remains the default entry point. Users do not need
to memorize the other four commands. They can invoke them directly when focused
control is useful.
A repository validator rejects a missing command, duplicate command, or sixth public Hypertaks skill.
Every meaningful task follows the same six-phase operating loop:
Phase 0 Intake and verify
Phase 1 Frame the objective and feasibility
Phase 2 Pick the required specialist roles
Phase 3 Equip the smallest relevant capability set
Phase 4 Produce the work
Phase 5 Integrate, verify, and deliver
The loop scales by task complexity:
| Tier | Agents | Typical work |
|---|---|---|
| Nano | 0 | One factual answer with no build, decision, or durable state |
| Lite | 1 | Small focused correction or follow-up |
| Standard | 3 | Bounded multi-discipline deliverable |
| Prime | 5 | Founder-shaped cross-domain work |
| Hyper | 6-10+ | Multi-workstream program |
| Omega | 10+ | Long-running program with human go or no-go gates |
Harmless Nano work creates no brain, pointer, checkpoint, graph job, registry scan, or update check.
Hypertaks separates the Boss's objective from the proposed method.
It protects customer value, trust, cash flow, runway, margin, product quality, operational capacity, team health, compliance, reputation, strategy, and long-term growth. When the proposed method materially damages the objective, Hypertaks states the conflict, explains the consequence, and proposes a safer path.
The Boss remains the final authority unless system policy, permission, legal, security, or irreversible-action rules block execution.
Founder memory is structured evidence, not hidden authority.
AgentPrivate Temporary or agent-specific context
Project Verified project facts, decisions, risks, and preferences
Shared Cross-agent facts or Boss-approved decisions
Shared memory accepts only:
- a repository fact verified against the active repository, branch, commit, tracked path, and content hash; or
- a Boss-approved decision whose message and contract match a valid T1 approval proof.
Model inference remains INFERRED. Missing evidence remains UNVERIFIED.
Changed repository evidence becomes stale or invalidated rather than silently
remaining true.
Memory and Graphify output cannot approve actions, expand scope, grant permissions, or override current repository evidence.
/hypertaks-verify can configure:
- project-local storage;
- an explicit external local folder;
- an existing Obsidian Vault;
- a separate local Git repository;
- a verified MCP memory capability;
- session-only memory with no persistence.
Existing custom brains remain user-owned. Hypertaks references them through a project pointer and does not copy or reorganize the existing structure.
A new default namespace uses:
Brains/<agent-name>/
Creation occurs only after an explicit preview and approval.
Obsidian is optional.
Hypertaks 4.5.0 supports an approved Obsidian Vault as a user-owned filesystem
destination. It validates the Vault root, respects the selected folder, and
never modifies .obsidian/.
This is not a claim of Obsidian application-level integration. The
obsidianmd/obsidian-releases repository is a future distribution path for an
optional visual plugin, not a memory API.
Graphify is optional and is used only when relationships, architecture, dependencies, centrality, or change impact materially help the task.
Supported routes are:
- a verified local stdio MCP executor;
- a verified shared HTTPS MCP endpoint with an authentication handle and approved external boundary;
- a verified local command;
- direct repository search when Graphify is disabled or unavailable.
Hypertaks never reports Graphify success unless a real executor, command, or
direct search ran. Missing source metadata is UNVERIFIED. A branch or commit
mismatch is STALE.
Graphify results are evidence below the active Boss turn, workspace standards, approved contract, and current repository state.
A checkpoint records:
- objective and active contract;
- actual repository identity, branch, and commit;
- changed files;
- completed and pending work;
- blockers and exact next action;
- granted permissions;
- tests with command, exit code, timestamp, and tested commit;
- acceptance criteria with evidence.
Resume reads Git state internally and rejects repository, branch, or commit mismatches.
A handoff carries the smallest verified state another agent needs. It preserves permissions, blockers, risks, test status, and next action without dumping the raw transcript or raw secrets.
Proof of done returns DONE only when current tests and acceptance evidence
pass and no pending work or blocker remains. Otherwise it returns NOT_DONE
with exact reasons.
Hypertaks 4.5.0 enforces:
- source-bound authority;
- explicit permissions and contract activation;
- fail-closed external-system capability binding;
- canonical approved-root path containment;
- strict agent names and record identifiers;
- traversal, absolute-path, reserved-name, and symlink-escape rejection;
- runtime pointer and checkpoint validation;
- atomic persistence;
- full-artifact secret scanning;
- handoff redaction;
- real Git-state verification;
- no silent Graphify installation, server start, or remote upload.
External side effects follow:
PREPARE -> PREVIEW -> T1 APPROVAL -> COMMIT ONCE -> RECONCILE
A timeout is not proof of failure. Hypertaks reconciles before retrying.
Hypertaks still includes the v4.4 systems:
- Retrieval Intelligence Router for exact, semantic, mixed, structured, small-corpus, and unavailable queries;
- Capability Relevance Router for the smallest sufficient verified skill, tool, MCP capability, or connector;
- professional Python, Matplotlib, TypeScript, UI/UX, and image-generation execution profiles;
- Visual Necessity Router for required, recommended, optional, or unnecessary visuals.
Precise numbers use tables or charts before creative image generation. External systems fail closed unless the approved contract names the allowed boundary. Mutating operations require approval even when capability metadata is wrong.
/plugin marketplace add aabrur/hypertaks-agent
/plugin install hypertaks@hypertaks-marketplace
codex plugin marketplace add aabrur/hypertaks-agent --ref main
codex plugin add hypertaks@hypertaks-marketplace/add-plugin hypertaks
/plugins install https://github.com/aabrur/hypertaks-agent
pi install git:github.com/aabrur/hypertaks-agentUse the platform instructions under .opencode/, .openclaw/, and .hermes/.
A managed Git checkout exposed through a symlink or junction is preferred over
copying the skill directory.
Hypertaks, verify this project and connect my existing main brain.
/hypertaks-brain inspect
/hypertaks-graph impact runtime/router.ts
/hypertaks-continuity checkpoint
Hypertaks, resume the project and prove what remains unfinished.
Normal work can still begin with plain language:
Hypertaks, fix this bug.
Hypertaks, review this product decision.
Hypertaks, build the feature after the contract is approved.
Hypertaks, explain why the release is not done yet.
The GitHub Actions gate runs:
python3 scripts/validate_skill.py
python3 scripts/validate_public_skills.py
python3 scripts/run_evals.py --check
python3 scripts/run_evals.py --static
python3 -m unittest scripts.test_run_evals scripts.test_retrieval_eval -v
python3 scripts/retrieval_eval.py evals/fixtures/retrieval-sample.jsonl --output <report>
python3 scripts/plot_retrieval_eval.py <report> <output-base>
npm test
python3 -m compileall scripts
git diff --check origin/main...HEAD
The structural inventory contains 88 eval definitions. Static GREEN and local runtime success are not behavioral certification.
Hypertaks 4.5.0 is a Release Candidate after the final merge gate passes.
The historical 4.3.0 ledger remains 43/49 Behavioral PASS with 6 documented non-PASS harness cases. That evidence does not certify the changed v4.4 or v4.5 behavior.
Fresh independent behavioral runs are still required for EV-50 through EV-88, including real host skill discovery, Graphify capability behavior, approved Obsidian Vault behavior, and cross-agent continuity.
No tag, package publication, marketplace submission, deployment, guaranteed security claim, or stable behavioral certification is implied by the source merge.
hypertaks-agent/
├── skills/hypertaks/ # Main Founder Operating System
├── skills/hypertaks-verify/ # Environment and brain verification
├── skills/hypertaks-brain/ # Evidence-backed founder memory
├── skills/hypertaks-graph/ # Graphify and direct-search routing
├── skills/hypertaks-continuity/ # Checkpoint, resume, handoff, proof of done
├── runtime/ # Strict TypeScript reference runtime
├── evals/ # EV-01 through EV-88
├── scripts/ # Validators, evaluators, reports, updater
├── .claude-plugin/
├── .codex-plugin/
├── .cursor-plugin/
├── .kimi-plugin/
├── .agents/
└── LICENSE
MIT © abrur