F039: Cognition Protocol Stack (SSTP/CSTP/LSTP)

Status: Proposed Target: v1.0.0+ (Multi-Agent Cognition Network) Source: Cisco Outshift Internet of Cognition, README roadmap

Overview

Implement the three-layer protocol stack from Cisco Outshift’s Internet of Cognition architecture. Each layer optimizes for different trade-offs between human auditability, bandwidth efficiency, and inference fidelity.

Protocol Layers

SSTP - Semantic State Transfer Protocol

Layer: Semantic (human-auditable) Current status: Partially implemented as CSTP JSON-RPC

The highest-level protocol. Decisions, guardrails, and reasoning are expressed in natural language with structured metadata. Fully auditable by humans.

Use cases:

• Cross-vendor strategic coordination
• Policy governance and compliance
• Human-in-the-loop decision review
• Audit trails

Format: JSON-RPC with natural language fields (what we have today)

{
  "decision": "Adopt HSM architecture for long-context processing",
  "confidence": 0.85,
  "reasons": [{"type": "analysis", "text": "MIT research shows 81% improvement"}],
  "bridge": {
    "structure": "Architecture pattern selection",
    "function": "Optimize long-context inference performance"
  }
}

CSTP - Compressed State Transfer Protocol

Layer: Compressed (bandwidth-efficient)

Decisions and context compressed into abstract feature representations. Reduces payload by 10-50x while preserving semantic meaning. Useful for edge deployments, high-frequency decision streams, and WAN communication.

Use cases:

• Edge agents with limited bandwidth
• High-frequency trading/monitoring decisions
• WAN federation between distant agents
• Mobile/IoT agent decision sync

Format: Compressed feature vectors + minimal metadata

{
  "id": "dec_abc",
  "embedding": [0.12, -0.45, ...],
  "confidence": 0.85,
  "category_idx": 3,
  "outcome_idx": 1,
  "timestamp": 1707681600
}

LSTP - Latent State Transfer Protocol

Layer: Latent (highest fidelity)

Raw latent representations for maximum inference continuity. One agent’s internal state can be directly loaded by another agent with compatible architecture. Closest to “transferring a mind state.”

Use cases:

• Local cluster agents with shared model architecture
• High-fidelity reasoning continuity (no information loss)
• Agent cloning / forking
• Research and experimentation

Format: Serialized tensor state + model metadata

{
  "id": "state_abc",
  "model": "gemini-3-pro",
  "format": "safetensors",
  "layers": ["reasoning_head", "decision_context"],
  "compatible_models": ["gemini-3-pro", "gemini-3-flash"],
  "data": "<base64 tensor data>"
}

Protocol Selection

Factor	SSTP	CSTP	LSTP
Human auditability	Full	Partial	None
Bandwidth	High	Low	Very High
Fidelity	Good	Moderate	Perfect
Cross-vendor	Yes	Yes	No (model-dependent)
Latency	Medium	Low	High (transfer) / Low (resume)

Auto-Selection Logic

{
  "method": "cstp.negotiateProtocol",
  "params": {
    "peer": "cstp://peer.example.com",
    "constraints": {
      "auditRequired": true,
      "maxBandwidthKbps": 100,
      "modelCompatible": false
    }
  }
}

Implementation Priority

1. SSTP - Already partially implemented as current CSTP JSON-RPC. Formalize the spec.
2. CSTP (compressed) - Add embedding-only query mode and compressed bundle format.
3. LSTP - Research phase. Requires compatible model architectures.

API

cstp.negotiateProtocol

Auto-negotiate the best protocol layer for a given peer connection.

cstp.exportCompressed

Export decisions in CSTP compressed format.

cstp.importCompressed

Import compressed decision bundles.

Integration

• F035 State Transfer: Bundles can be serialized at any protocol layer
• F038 Federation: Protocol negotiation between federated peers
• F036 Reasoning Continuity: LSTP enables highest-fidelity thread resumption

Acceptance Criteria

• SSTP formalized (current JSON-RPC spec documented as SSTP)
• CSTP compressed format defined
• cstp.negotiateProtocol RPC method
• cstp.exportCompressed / cstp.importCompressed
• Embedding-only query mode for CSTP layer
• LSTP research document with feasibility analysis
• Protocol selection matrix in documentation
• MCP tools for protocol operations