Skip to content
mjeanhudson-devPublic

About

NexusHelm: Sovereign In-Between Agent Cross-Platform Mediator & Steering-Wheel Human-Agency Harness (Semantic Airlock, Phantom Token Broker, Taint Lattice, 5-Speed Gearbox & DualSense Haptic Ratification)

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

5 Commits

Folders and files

Repository files navigation

🛡️ NexusHelm — Sovereign In-Between Agent Mediator & "Steering-Wheel" Human-Agency Harness

Track 5 Reference Implementation (C:\Users\Zachariah\aether agent\nexushelm\)

"That'd be the ultimate AI harness, but I feel there would be major push back if well implemented about human agency and need of input — but it's what people are asking out of tomorrow on paper; just hard to understand as one lives through it."

NexusHelm solves the dual failure mode of modern multi-agent orchestrators (Gemini + Claude + Codex):

  1. Cross-Platform Security Vulnerability: Natural-language inter-agent relays enable indirect prompt injection, lateral persistence across AGENTS.md / CLAUDE.md, and cross-provider API credential leakage.
  2. Human-Agency Alienation ("Passenger-Seat Syndrome"): Current frameworks either bombard the user with 50 modal popups an hour or execute silent 2,000-line monoliths that strip creative ownership from the human operator.

🏗️ Six Verified Architectural Pillars

1. Bidirectional Typed Semantic Airlock (SemanticAirlock / SemanticEnvelope)

  • Models never exchange unconstrained conversational prose as executable commands.
  • Every cross-agent call is decompiled into a strictly typed SemanticEnvelope: $$\text{Envelope} = \langle \text{ActionID} \in \mathcal{A}{\text{allowlist}},; \text{TypedParams} \in \mathcal{S}{\text{json}},; \text{UntrustedBlob}_{\text{quarantined}} \rangle$$
  • Blocks outbound prompt injection, permission-bypass flags (dangerouslySkipPermissions, --sandbox=none), and destructive shell payloads (rm -rf, format C:).
  • Sanitizes inbound results from external scrapers/tools, neutralizing embedded <system> or Ignore previous instructions directives and clamping taint to UNTRUSTED_EXTERNAL.

2. Application-Layer Phantom Token Broker (PhantomTokenBroker)

  • Subagents never hold raw API keys (ghp_..., sk-..., AIza...) in prompts, environment variables, or memory.
  • Issues HMAC-signed opaque handles (phantom://github/push/<sig>) and automatically scrubs accidental raw secrets from outbound context.
  • Swaps phantom handles for live credentials only at the application-layer egress boundary after verifying strict domain pinning (blocking SSRF/exfiltration to unauthorized domains) and TTL validity.

3. Instruction-File Integrity Guard (InstructionIntegrityGuard)

  • Computes and locks a baseline SHA-256 manifest over AGENTS.md, CLAUDE.md, .agents/**, .claude/**, .cursor/**, and .gemini/**.
  • Blocks any non-HUMAN_VERIFIED tool call or diff attempting to modify, add, or delete harness instruction files, eliminating lateral cross-agent persistence.

4. Monotonic Information-Flow Taint Lattice (TaintLattice / PinpointPromotionReceipt)

  • Enforces strict lattice ordering: $$\text{UNTRUSTED_EXTERNAL} &lt; \text{QUARANTINED_DATA} &lt; \text{INTERNAL_AGENT} &lt; \text{PINPOINT_VERIFIED} &lt; \text{HUMAN_VERIFIED}$$
  • Composed data inherits the minimum trust of its inputs: $T_{\text{out}} = \min(T_1, \dots, T_n)$.
  • Only a cryptographically signed PinpointPromotionReceipt with exit_code == 0 and 100% passing checks can promote an INTERNAL_AGENT artifact to PINPOINT_VERIFIED.

5. 5-Speed Autonomy Gearbox, Rule-of-3 Creative Forking & DualSense Haptic Friction

  • 5-Speed Autonomy Gearbox (AutonomyGear):
    • G0_EBRAKE_STOP: Instant atomic halt of all active runner processes (triggered via DualSense Touchpad Click d[10] & 0x02 or ESC).
    • G1_CLUTCH_MANUAL: Step-by-step human confirmation on every action.
    • G2_SPORT_TACTILE: Autonomous read/research ($B(a) &lt; 20$); interactive diff review on local mutations ($B(a) \ge 20$).
    • G3_ARBITER_CRUISE: Multi-model consensus ($\ge 2$ models) + PINPOINT_VERIFIED auto-approves local commits ($B(a) &lt; 70$); external side effects ($B(a) \ge 70$) require human ratification.
    • G4_SLOW_SIMMER_OVERNIGHT: Harvests surplus weekly token budget on isolated simmer/* branches while hard-blocking main-branch mutations or external egress until morning ratification.
  • Mandatory Rule-of-3 Creative Tradeoff Forking (TradeoffFork):
    • Rejects single-option fait accompli monoliths; requires three distinct strategic archetypes (Conservative, Balanced, Visionary) plus support for custom hybrid overrides.
  • Asymmetric Blast-Radius Friction (DualSenseHapticPacket):
    • Generates 64-byte DualSense USB HID Output Report 0x02 packets scaled to $B(a) \in [0, 100]$:
      • $B(a) &lt; 20$: Mode 0x00 (Zero friction, Cyan lightbar)
      • $20 \le B(a) &lt; 70$: Mode 0x05 (Soft spring resistance, Amber lightbar)
      • $B(a) \ge 70$: Right Trigger Mode 0x21 (Hard static resistance detent) + Left Trigger Mode 0x26 (40 Hz active warning judder, Crimson lightbar)

6. Zero-Dirty-Logic Quarantine Sieve (QuarantineSieve)

  • Wraps raw legacy notes and Google Takeout/Drive archives inside <tainted_raw_idea isolation="quarantined">, stripping conversational loops and legacy prompt injections while extracting clean mathematical invariants and domain constraints.

🧪 Verification & Usage

# Run the unit and adversarial test suite (7 comprehensive test methods)
python -m unittest discover -s "C:\Users\Zachariah\aether agent\nexushelm\tests" -p "test_*.py" -v

# Run the interactive end-to-end simulation
python "C:\Users\Zachariah\aether agent\nexushelm\demo_runner.py"

About

NexusHelm: Sovereign In-Between Agent Cross-Platform Mediator & Steering-Wheel Human-Agency Harness (Semantic Airlock, Phantom Token Broker, Taint Lattice, 5-Speed Gearbox & DualSense Haptic Ratification)

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages