---

name: lingua-universale version: 0.1.0 description: Verify agent-to-agent communication against session type protocols. Mathematical proofs, not trust. author: CervellaSwarm homepage: https://github.com/rafapra3008/cervellaswarm tags: [protocol-verification, session-types, mcp, ai-agents, formal-methods] user-invocable: true metadata: clawdbot: emoji: "🔬" requires: bins: - uvx env: []

Lingua Universale - Protocol Verification Skill

What it does

Verifies that agent-to-agent messages follow a formally defined protocol using session types -- the same mathematical framework used in distributed systems research (Honda/Yoshida POPL 2008, Scribble, MPST).

No API keys required. No external services. Runs entirely locally.

When to use

• Before sending a message to another agent: confirm the message is expected
• After a conversation: validate the full transcript against the protocol
• During design: check safety properties (no deadlock, always terminates)
• When choosing a protocol pattern: browse 20 standard library templates

Tools provided

lu_load_protocol

Parse a .lu protocol definition and extract its structure.

Input: protocol_text (string) -- full content of a .lu file Output: JSON with protocol name, roles, steps, choices, and declared properties

lu_verify_message

Check if a message is valid in the context of an ongoing session.

Input:

• protocol_text (string) -- the protocol definition
• messages (list) -- already-sent messages: [{"sender": "a", "receiver": "b", "action": "ask"}]
• next_message (dict) -- message to validate: {"sender": "b", "receiver": "a", "action": "return"}

Output: {"valid": true, "next_expected": "..."} or {"valid": false, "violation": "...", "expected": "...", "got": "..."}

lu_check_properties

Verify the formal safety properties declared in a protocol.

Input: protocol_text (string) Output: JSON with per-property verdicts (PROVED / SATISFIED / VIOLATED / SKIPPED) and evidence

Supported properties:

• always terminates -- no infinite loops
• no deadlock -- no role waits forever
• no deletion -- no destructive operations
• X before Y -- message ordering constraint
• role cannot send message -- exclusion
• role exclusive message -- only this role may send this message
• confidence >= level -- minimum confidence threshold
• trust >= tier -- minimum trust tier
• all roles participate -- every role sends at least one message

lu_list_templates

Browse the 20 protocols in the Lingua Universale standard library.

Input: category (optional string) -- filter by: communication, data, business, ai_ml, security Output: JSON with template names, categories, property highlights, and usage instructions

Example workflow

# 1. Choose a protocol template
lu_list_templates(category="ai_ml")
# -> rag_pipeline, agent_delegation, tool_calling, human_in_loop, consensus

# 2. Load and inspect the protocol
lu_load_protocol("""
    protocol AgentDelegation:
        roles: supervisor, worker, validator
        supervisor asks worker to execute task
        worker returns result to supervisor
        supervisor asks validator to audit result
        validator returns verdict to supervisor
        properties:
            always terminates
            no deadlock
            all roles participate
            trust >= standard
""")

# 3. Verify messages as they flow
lu_verify_message(
    protocol_text=<above>,
    messages=[
        {"sender": "supervisor", "receiver": "worker", "action": "ask"}
    ],
    next_message={"sender": "worker", "receiver": "supervisor", "action": "return"}
)
# -> {"valid": true, "next_expected": "supervisor -> validator : audit_request"}

# 4. Check formal properties
lu_check_properties(<protocol_text>)
# -> all_passed: true, PROVED: always terminates, no deadlock, all roles participate

Protocol syntax (.lu)

protocol Name:
    roles: role1, role2, role3

    role1 asks role2 to do something
    role2 returns result to role1

    when role1 decides:
        approve:
            role1 tells role3 about decision
        reject:
            role1 sends error to role3

    properties:
        always terminates
        no deadlock
        all roles participate

Valid actions: asks, returns, sends, proposes, tells

Installation

# As a Claude Code MCP server
uvx openclaw-skill-lingua-universale

# Or install directly
pip install openclaw-skill-lingua-universale
lu-mcp  # starts stdio MCP server

Academic references

• Honda, Yoshida, Carbone - "Multiparty Asynchronous Session Types" (POPL 2008)
• Scribble Project - session type toolchain (Imperial College London)
• Dezani-Ciancaglini et al. - "Session Types for Access and Information Flow Control" (2021)