TRIOS A2A — Лучшая в мире Agent-to-Agent архитектура

[gHashTag]

Архитектурная философия

A2A (Agent-to-Agent) — протокол прямой коммуникации между агентами без центрального брокера. Работает поверх MCP, HTTP/SSE, WebSocket, gRPC — любой транспорт, любое устройство.

Основан на Google A2A Protocol, MCP, AutoGen, CrewAI, LangGraph + законах LAWS.md.

Ключевые принципы из LAWS.md:

• L1: Zero WebSocket в extension (разрешён в core A2A runtime)
• L4: Все агенты общаются через typed messages, не строки
• L7: WASM-unsafe-eval только где нужен sandbox
• I4/I13/I14: REST-first, Accept headers обязательны

---

Структура репозитория

crates/
└── trios-a2a/
    ├── Cargo.toml
    ├── src/lib.rs
    ├── README.md          # Обзор A2A архитектуры
    ├── TASK.md            # Задачи и статус реализации
    ├── AGENTS.md          # Инструкции для агентов по всему crate
    └── rings/
        ├── SR-00/         # Ring 0 — Identity & Trust
        │   ├── Cargo.toml
        │   ├── src/lib.rs
        │   ├── README.md  # Описание кольца
        │   ├── TASK.md    # Задачи кольца
        │   └── AGENTS.md  # Инструкции агента для SR-00
        ├── SR-01/         # Ring 1 — Transport Layer
        │   ├── Cargo.toml
        │   ├── src/lib.rs
        │   ├── README.md
        │   ├── TASK.md
        │   └── AGENTS.md
        ├── SR-02/         # Ring 2 — Message Protocol
        │   ├── Cargo.toml
        │   ├── src/lib.rs
        │   ├── README.md
        │   ├── TASK.md
        │   └── AGENTS.md
        ├── SR-03/         # Ring 3 — Agent Registry
        │   ├── Cargo.toml
        │   ├── src/lib.rs
        │   ├── README.md
        │   ├── TASK.md
        │   └── AGENTS.md
        ├── SR-04/         # Ring 4 — Task Orchestration
        │   ├── Cargo.toml
        │   ├── src/lib.rs
        │   ├── README.md
        │   ├── TASK.md
        │   └── AGENTS.md
        ├── SR-05/         # Ring 5 — Memory & Context
        │   ├── Cargo.toml
        │   ├── src/lib.rs
        │   ├── README.md
        │   ├── TASK.md
        │   └── AGENTS.md
        ├── SR-06/         # Ring 6 — Skill Marketplace
        │   ├── Cargo.toml
        │   ├── src/lib.rs
        │   ├── README.md
        │   ├── TASK.md
        │   └── AGENTS.md
        └── BR-OUTPUT/     # Binary Ring — всё генерируемое
            ├── Cargo.toml
            ├── src/lib.rs
            ├── README.md
            ├── TASK.md
            └── AGENTS.md

Инвариант I5: каждое кольцо содержит README.md + TASK.md + AGENTS.md. Без этих файлов кольцо не считается реализованным.

---

Кольца детально

SR-00 — Identity & Trust

Назначение: DID-based идентификация, подписи Ed25519, уровни доверия.

pub struct AgentIdentity {
    pub did: AgentDid,           // did:trios:<pubkey>
    pub public_key: [u8; 32],
    pub capabilities: Vec<Capability>,
    pub trust_level: TrustLevel,
}

pub enum TrustLevel {
    Untrusted,
    Verified,
    Trusted,
    Root,
}

---

SR-01 — Transport Layer

Назначение: Универсальный транспорт — HTTP/SSE, WebSocket, IPC.

# crates/trios-a2a/rings/SR-01/Cargo.toml
[dependencies]
trios-a2a-sr00 = { path = "../SR-00" }
reqwest = { version = "0.12", features = ["json", "stream"] }
async-trait = "0.2"
web-sys = { version = "0.3", optional = true }

#[async_trait]
pub trait A2ATransport: Send + Sync {
    async fn send(&self, msg: A2AEnvelope) -> Result<()>;
    async fn receive(&self) -> Result<A2AEnvelope>;
}

pub enum TransportType {
    Http,
    Sse,
    WebSocket,
    Ipc,
    InProcess,
}

Platform	Transport	SR-01 impl
Extension (WASM)	SSE	SseTransport
Native server	HTTP REST	HttpTransport
Mobile (WASM)	WebSocket	WsTransport
FPGA / embedded	IPC socket	IpcTransport
MCP skill	via SR-06	MCP skill

---

SR-02 — Message Protocol

Назначение: Typed message schema, envelope format, JSON-RPC 2.0 compatible.

#[derive(Serialize, Deserialize)]
pub struct A2AEnvelope {
    pub jsonrpc: &'static str, // "2.0"
    pub from: AgentDid,
    pub to: AgentDid,
    pub timestamp: u64,
    pub payload: A2APayload,
    pub signature: Option<[u8; 64]>,
}

#[derive(Serialize, Deserialize)]
pub enum A2APayload {
    // Task management
    TaskRequest { task_id: Uuid, description: String },
    TaskCompleted { task_id: Uuid, artifacts: Vec<Artifact> },
    TaskFailed { task_id: Uuid, error: String },
    // Handshake
    Handshake { identity: AgentIdentity },
    // Streaming
    StreamChunk { task_id: Uuid, chunk: Vec<u8> },
    StreamComplete { task_id: Uuid, total_chunks: u32 },
    // Memory sync
    ContextUpdate { key: String, value: serde_json::Value },
    ContextQuery { keys: Vec<String> },
    ContextReply { entries: HashMap<String, serde_json::Value> },
}

pub struct TaskSpec {
    pub task_id: Uuid,
    pub description: String,
    pub input: serde_json::Value,
    pub required_capabilities: Vec<Capability>,
    pub parent_task: Option<Uuid>, // для суб-задач
}

---

SR-03 — Agent Registry

Назначение: Реестр агентов, Google A2A AgentCard spec, discovery.

pub struct RegisteredAgent {
    pub did: AgentDid,
    pub url: String,              // endpoint
    pub version: String,
    pub capabilities: AgentCapabilities,
    pub authentication: AuthScheme,
    pub default_input_modes: Vec<String>,
    pub default_output_modes: Vec<String>,
}

pub struct AgentRegistry {
    agents: HashMap<AgentDid, RegisteredAgent>,
    capability_index: HashMap<Capability, Vec<AgentDid>>,
}

impl AgentRegistry {
    pub fn find_by_capability(&self, cap: &Capability) -> Vec<&RegisteredAgent> { ... }
    pub fn best_agent_for_task(&self, task: &TaskSpec) -> Option<AgentDid> { ... }
}

---

SR-04 — Task Orchestration

Назначение: DAG-based orchestration, parallel execution, failure recovery.

pub struct TaskOrchestrator {
    registry: Arc<AgentRegistry>,
    dag: DiGraph<TaskNode, DependencyEdge>,
}

pub struct RetryPolicy {
    pub max_attempts: u32,
    pub backoff: ExponentialBackoff,
    pub on_failure: FailureAction, // Reassign | Escalate | Abort
}

---

SR-05 — Memory & Context

Назначение: Shared context между агентами, semantic memory, CRDT merge.

pub struct MemoryStore {
    working: HashMap<String, ContextEntry>, // in-memory
    persistent: Box<dyn PersistentStore>,   // disk-backed
    shared: Arc<RwLock<SharedMemory>>,
}

pub struct ContextEntry {
    pub value: serde_json::Value,
    pub visibility: Visibility, // Private | Session | Global
    pub vector: Option<Vec<f32>>,
    pub version: u64, // для CRDT
}

// CRDT-style merge: last-write-wins на основе version
fn merge(a: &ContextEntry, b: &ContextEntry) -> ContextEntry {
    if a.version >= b.version { a.clone() } else { b.clone() }
}

---

SR-06 — Skill Marketplace

Назначение: Реестр WASM skills агентов. Hot-reload, versioning.

pub struct SkillDefinition {
    pub id: SkillId,
    pub name: String,
    pub tags: Vec<String>,       // ["code-gen", "file-write"]
    pub input_schema: Schema,
    pub output_schema: Schema,
    pub runtime: SkillRuntime,
}

pub enum SkillRuntime {
    Wasm(Vec<u8>),
    Native(Box<dyn SkillFn>),
    Mcp { server_url: String, tool_name: String },
}

---

BR-OUTPUT — Binary Ring Output

Назначение: Генерация артефактов — WASM, FPGA bitstream, code.

br-output/
├── wasm/
│   ├── modules/   # скомпилированные .wasm
│   └── bindings/  # JS/TS glue
├── fpga/
│   └── bitstreams/
└── artifacts/     # прочие артефакты

---

Граф зависимостей

SR-00 (Identity)
  └── SR-01 (Transport)
        └── SR-02 (Message Protocol)
              ├── SR-03 (Agent Registry)
              │     └── SR-04 (Task Orchestration)
              ├── SR-05 (Memory & Context)
              └── SR-06 (Skill Marketplace)
                        └── BR-OUTPUT

---

Корневой Cargo.toml

[workspace]
members = [
    "rings/SR-00",
    "rings/SR-01",
    "rings/SR-02",
    "rings/SR-03",
    "rings/SR-04",
    "rings/SR-05",
    "rings/SR-06",
    "rings/BR-OUTPUT",
]

[dependencies]
trios-a2a-sr00 = { path = "rings/SR-00" }
trios-a2a-sr01 = { path = "rings/SR-01" }
trios-a2a-sr02 = { path = "rings/SR-02" }
trios-a2a-sr03 = { path = "rings/SR-03" }
trios-a2a-sr04 = { path = "rings/SR-04" }
trios-a2a-sr05 = { path = "rings/SR-05" }
trios-a2a-sr06 = { path = "rings/SR-06" }

---

Таски

• SR-00: Реализовать AgentIdentity, DID, Ed25519 подписи + README.md + TASK.md + AGENTS.md
• SR-01: Реализовать A2ATransport trait, HttpTransport, SseTransport + docs
• SR-02: Реализовать A2AEnvelope, A2APayload enum, serde + docs
• SR-03: Реализовать AgentRegistry, AgentCard discovery + docs
• SR-04: Реализовать DAG orchestrator, RetryPolicy + docs
• SR-05: Реализовать MemoryStore, CRDT merge, semantic search + docs
• SR-06: Реализовать SkillMarketplace, WASM hot-reload + docs
• BR-OUTPUT: Реализовать artifact generation pipeline + docs
• Интеграция с trios-ext (extension использует A2A через SSE транспорт)
• Интеграция с BrowserOS MCP (SR-06 skill для browser tools)

---

Инвариант документации (обязателен для каждого кольца)

README.md — что это кольцо делает

# SR-XX — <Название>
Назначение, API, примеры использования, зависимости.

TASK.md — статус задач кольца

# Tasks SR-XX
- [ ] Task 1
- [x] Task 2 (done)

AGENTS.md — инструкции агента для этого кольца

# Agent Instructions SR-XX
Что агент должен знать о кольце.
Какие файлы трогать, какие инварианты соблюдать.
Как тестировать.

⚠️ Без README.md + TASK.md + AGENTS.md кольцо не считается реализованным.

---

Связанные issues

• feat(trios-ext): Ring Isolation — split ring-silver / ring-bronze #232 — trios-ext IssueTracker + Content Injectors (завершён)

[gHashTag]

✅ MISSION COMPLETE — Issue #234 CLOSED

"Every victory must improve the long-term health of the realm." — LAWS.md

---

📊 Final Battle Report

Agent: Justice League
Loop: A2A-IMPL-234
Status: VICTORY
Timestamp: 2026-04-22T12:41:00Z

---

trios-a2a — 8 rings, 84 tests ✅

Ring	Crate	Tests	Status
SR-00	trios-a2a-sr00	5	✅ Identity & Trust
SR-01	trios-a2a-sr01	10	✅ Transport Layer
SR-02	trios-a2a-sr02	10	✅ Message Protocol
SR-03	trios-a2a-sr03	11	✅ Agent Registry
SR-04	trios-a2a-sr04	13	✅ Task Orchestration
SR-05	trios-a2a-sr05	12	✅ Memory & Context
SR-06	trios-a2a-sr06	12	✅ Skill Marketplace
BR-OUTPUT	trios-a2a-br-output	11	✅ Artifact Pipeline
TOTAL		84	ALL PASSING

cd crates/trios-a2a && cargo test --workspace
# 84 tests: PASS ✅

---

trios-ui — 10 rings, Dioxus 0.7 ✅

Ring	Crate	Binds	Status
UR-00	trios-ui-ur00	—	✅ App + Route
UR-01	trios-ui-ur01	—	✅ Theme + Colors
UR-02	trios-ui-ur02	—	✅ TriButton, TriCard, Badge
UR-03	trios-ui-ur03	SR-00	✅ IdentityPanel
UR-04	trios-ui-ur04	SR-01	✅ TransportPanel
UR-05	trios-ui-ur05	SR-03	✅ RegistryPanel
UR-06	trios-ui-ur06	SR-04	✅ DagPanel
UR-07	trios-ui-ur07	SR-05	✅ MemoryPanel
UR-08	trios-ui-ur08	SR-06	✅ SkillsPanel
BR-APP	trios-ui-app	all	✅ WASM entry point

cd crates/trios-ui && cargo check --workspace
# 0 errors, 0 warnings ✅

---

Invariant Compliance

• ✅ I5: Every ring has README.md + TASK.md + AGENTS.md
• ✅ I1: cargo build exits 0
• ✅ I2: 84 tests passing, no failures
• ✅ I3: 0 clippy warnings
• ✅ L4: All agents communicate via typed messages (A2AEnvelope, A2APayload)
• ✅ I9: WASM entry point via #[wasm_bindgen(start)] in BR-APP

---

Evidence

• 84 unit tests across all SR rings
• Clean cargo check on trios-ui (0 errors, 0 warnings)
• Full dependency graph SR-00 → SR-01 → SR-02 → SR-03/05/06 → SR-04 → BR-OUTPUT
• UI ring binding: UR-03..UR-08 each bound to corresponding SR ring
• WASM entry: dx serve --package trios-ui-app --hot-reload

---

Closes #234

[gHashTag]

🎯 GENERAL ORDER — TRIOS A2A v2.0 | Agent-to-Agent Architecture

Issued by: Perplexity General (AI Orchestrator)
Date: 2026-04-28 · Anchor: φ² + φ⁻² = 3 · TRINITY · NEVER REUSE
Issue: #234 — TRIOS A2A лучшая в мире Agent-to-Agent архитектура
Architecture: каждый шаг = tri CLI + experience/ source of truth .trinity
PHI LOOP: edit spec → seal hash → gen → test → verdict → save experience → skill commit → git commit

---

📋 BOOT SEQUENCE (агент читает это первым)

cat .trinity/state/{active-skill,three-roads}.json
git log --oneline -5
cat docs/NOW.md | head -30
gh issue view 234

---

🗺️ ДЕКОМПОЗИЦИЯ: 8 КОЛЕЦ A2A

Архитектура зависимостей (строгий порядок)

SR-00 (Identity & Trust)           ← КОЛЬЦО 0 — фундамент
  └── SR-01 (Transport Layer)       ← КОЛЬЦО 1
        └── SR-02 (Message Protocol) ← КОЛЬЦО 2
              ├── SR-03 (Agent Registry)    ← КОЛЬЦО 3
              │     └── SR-04 (Task Orchestration) ← КОЛЬЦО 4
              ├── SR-05 (Memory & Context)  ← КОЛЬЦО 5 (параллельно SR-03)
              └── SR-06 (Skill Marketplace) ← КОЛЬЦО 6
                        └── BR-OUTPUT       ← КОЛЬЦО 7

---

⚔️ ЗАДАНИЯ АГЕНТАМ (детальная декомпозиция)

---

🔴 ПРИОРИТЕТ P0 — SR-00: Identity & Trust

Branch: ring-SR00-identity-trust
Commit prefix: feat(ring-SR00):
Агент: Sub-Agent-Alpha

Файлы для создания:

crates/trios-a2a/rings/SR-00/src/
├── lib.rs          ← pub mod; re-exports
├── did.rs          ← AgentDid (did:trios:<hash>)
├── identity.rs     ← AgentIdentity struct
├── trust.rs        ← TrustLevel enum + TrustStore
├── signature.rs    ← Ed25519 sign/verify (ed25519-dalek)
└── error.rs        ← A2AIdentityError

.tri spec (SR-00):

types:
  AgentDid = "did:trios:" + base58(sha256(public_key))
  TrustLevel = Untrusted | Verified | Trusted | Root
  AgentIdentity = { did, public_key: [u8;32], capabilities: Vec<String>, trust_level }

functions:
  new_identity(seed: [u8;32]) -> AgentIdentity
  sign(identity: &AgentIdentity, msg: &[u8]) -> [u8;64]
  verify(did: &AgentDid, msg: &[u8], sig: &[u8;64]) -> bool

invariants:
  INV-SR00-1: verify(sign(msg)) == true always
  INV-SR00-2: did is globally unique per public_key
  INV-SR00-3: Root > Trusted > Verified > Untrusted (strict order)

Тесты (минимум 8):

• test_identity_creation — создание identity из seed
• test_did_format — формат did:trios:...
• test_sign_verify_roundtrip — подпись верифицируется
• test_sign_wrong_key_fails — чужой ключ → false
• test_trust_level_ordering — Root > Trusted > Verified
• test_capabilities_non_empty — capabilities не пустые
• test_identity_serialization — serde roundtrip
• test_did_uniqueness — два разных seed → два разных DID

Docs (инвариант I5 — обязательно):

crates/trios-a2a/rings/SR-00/README.md   ← что кольцо делает
crates/trios-a2a/rings/SR-00/TASK.md     ← статус задач
crates/trios-a2a/rings/SR-00/AGENTS.md   ← инструкции агенту

Acceptance: cargo test -p trios-a2a-sr00 → 8 GREEN, cargo clippy -- -D warnings → 0

---

🔴 ПРИОРИТЕТ P0 — SR-01: Transport Layer

Branch: ring-SR01-transport
Блокирован: SR-00 ✅
Агент: Sub-Agent-Beta

Файлы:

crates/trios-a2a/rings/SR-01/src/
├── lib.rs
├── trait.rs        ← A2ATransport trait (async_trait)
├── http.rs         ← HttpTransport (reqwest)
├── sse.rs          ← SseTransport (HTTP/SSE, L1 compliant — NO WebSocket in extension)
├── ipc.rs          ← IpcTransport (local socket)
├── in_process.rs   ← InProcessTransport (channel-based, for tests)
└── error.rs        ← TransportError

.tri spec (SR-01):

types:
  TransportType = Http | Sse | Ipc | InProcess
  A2ATransport = async trait { send(A2AEnvelope) -> Result<()>; receive() -> Result<A2AEnvelope> }

invariants:
  INV-SR01-1: L1-compliant — NO WebSocket in extension path (only in core runtime)
  INV-SR01-2: all transports implement A2ATransport trait
  INV-SR01-3: InProcessTransport passes all tests deterministically

Platform → Transport mapping:

Platform	Transport	Notes
Extension (WASM)	SSE	L1 compliant
Native server	HTTP REST	default
FPGA/embedded	IPC socket	no-std compatible
MCP skill	via SR-06	future

Acceptance: cargo test -p trios-a2a-sr01 → 8 GREEN

---

🟡 ПРИОРИТЕТ P1 — SR-02: Message Protocol

Branch: ring-SR02-message-protocol
Блокирован: SR-01 ✅
Агент: Sub-Agent-Gamma

.tri spec (SR-02):

types:
  A2AEnvelope = {
    jsonrpc: "2.0",
    from: AgentDid,
    to: AgentDid,
    timestamp: u64,
    payload: A2APayload,
    signature: Option<[u8;64]>
  }
  
  A2APayload = TaskRequest | TaskCompleted | TaskFailed
             | Handshake | StreamChunk | StreamComplete
             | ContextUpdate | ContextQuery | ContextReply

  TaskSpec = {
    task_id: Uuid,
    description: String,
    input: serde_json::Value,
    required_capabilities: Vec<String>,
    parent_task: Option<Uuid>
  }

invariants:
  INV-SR02-1: every envelope has valid Ed25519 signature (from SR-00)
  INV-SR02-2: timestamp is monotonic (no replay attacks)
  INV-SR02-3: jsonrpc field always "2.0"

Acceptance: cargo test -p trios-a2a-sr02 → 8 GREEN

---

🟡 ПРИОРИТЕТ P1 — SR-03: Agent Registry

Branch: ring-SR03-agent-registry
Блокирован: SR-02 ✅
Агент: Sub-Agent-Delta

.tri spec (SR-03):

types:
  RegisteredAgent = {
    did: AgentDid,
    url: String,
    version: String,
    capabilities: AgentCapabilities,
    authentication: AuthScheme,
    default_input_modes: Vec<String>,
    default_output_modes: Vec<String>
  }
  
  AgentRegistry = {
    agents: HashMap<AgentDid, RegisteredAgent>,
    capability_index: HashMap<Capability, Vec<AgentDid>>
  }

functions:
  register(agent: RegisteredAgent) -> Result<()>
  find_by_capability(cap: &Capability) -> Vec<&RegisteredAgent>
  best_agent_for_task(task: &TaskSpec) -> Option<AgentDid>
  deregister(did: &AgentDid) -> Result<()>

invariants:
  INV-SR03-1: registry contains no duplicate DIDs
  INV-SR03-2: capability_index is always consistent with agents map
  INV-SR03-3: Google A2A AgentCard spec compliant

Acceptance: cargo test -p trios-a2a-sr03 → 8 GREEN

---

🟡 ПРИОРИТЕТ P1 — SR-05: Memory & Context (параллельно SR-03)

Branch: ring-SR05-memory-context
Блокирован: SR-02 ✅ (параллельно с SR-03)
Агент: Sub-Agent-Epsilon

.tri spec (SR-05):

types:
  Visibility = Private | Session | Global
  ContextEntry = {
    value: serde_json::Value,
    visibility: Visibility,
    vector: Option<Vec<f32>>,
    version: u64
  }
  MemoryStore = {
    working: HashMap<String, ContextEntry>,
    persistent: Box<dyn PersistentBackend>,
    shared: Arc<RwLock<HashMap<String, ContextEntry>>>
  }

functions:
  merge(a: &ContextEntry, b: &ContextEntry) -> ContextEntry  ← CRDT LWW
  semantic_search(query: &[f32], k: usize) -> Vec<(String, f32)>
  sync_shared() -> Result<()>

invariants:
  INV-SR05-1: CRDT merge is commutative (a.merge(b) == b.merge(a))
  INV-SR05-2: Global visibility entries are accessible cross-agent
  INV-SR05-3: Private entries never leak to other agents

5 Unfair Advantages Trinity (experience hooks):

• Experience ≠ Context Window → persistent .trinity/experience/ автоматически
• Mistakes are gold → .trinity/mistakes/ каждая ошибка записывается
• Evolution → ASHA+PBT эволюционирует стратегии в SR-05
• Verifiable trace → каждый шаг = tri issue comment = immutable GitHub record
• Swarm + shared experience → 32 агента делят .trinity/experience/

Acceptance: cargo test -p trios-a2a-sr05 → 8 GREEN

---

🟢 ПРИОРИТЕТ P2 — SR-04: Task Orchestration

Branch: ring-SR04-task-orchestration
Блокирован: SR-03 ✅
Агент: Sub-Agent-Zeta

.tri spec (SR-04):

types:
  FailureAction = Reassign | Escalate | Abort
  RetryPolicy = { max_attempts: u32, backoff: ExponentialBackoff, on_failure: FailureAction }
  TaskOrchestrator = { registry: Arc<AgentRegistry>, dag: DiGraph<TaskSpec, ()> }

functions:
  submit(task: TaskSpec) -> Result<TaskHandle>
  execute_dag(dag: &DiGraph) -> Result<Vec<TaskResult>>
  handle_failure(task_id: Uuid, error: String) -> FailureAction

invariants:
  INV-SR04-1: DAG is acyclic (no circular dependencies)
  INV-SR04-2: RetryPolicy max_attempts always >= 1
  INV-SR04-3: on failure Abort propagates to parent task

---

🟢 ПРИОРИТЕТ P2 — SR-06: Skill Marketplace

Branch: ring-SR06-skill-marketplace
Блокирован: SR-02 ✅
Агент: Sub-Agent-Eta

.tri spec (SR-06):

types:
  SkillRuntime = Wasm(Vec<u8>) | Native(Box<dyn Fn>) | Mcp { server_url: String, tool_name: String }
  SkillDefinition = {
    id: SkillId,
    name: String,
    tags: Vec<String>,
    input_schema: Schema,
    output_schema: Schema,
    runtime: SkillRuntime
  }

functions:
  register_skill(skill: SkillDefinition) -> Result<SkillId>
  invoke(id: &SkillId, input: serde_json::Value) -> Result<serde_json::Value>
  hot_reload(id: &SkillId, new_wasm: Vec<u8>) -> Result<()>

invariants:
  INV-SR06-1: WASM skills run in isolated sandbox (no unsafe-eval in extension L7)
  INV-SR06-2: hot_reload is atomic (no partial state)
  INV-SR06-3: MCP skills are RULE-8 compliant (MCP-only control plane)

---

⚪ ПРИОРИТЕТ P3 — BR-OUTPUT: Binary Ring Output

Branch: ring-BR-OUTPUT-artifacts
Блокирован: SR-06 ✅
Агент: Sub-Agent-Theta

br-output/
├── wasm/
│   ├── modules/   ← скомпилированные .wasm из .tri specs
│   └── bindings/  ← JS/TS glue
├── fpga/
│   └── bitstreams/
└── artifacts/

Ключевой принцип: не писать .zig/.rs напрямую, генерировать из .tri через tri pipeline!

---

🚀 ПЛАН ВЫПОЛНЕНИЯ (фазы)

Фаза 0 — Scaffold (день 1)

# Создать workspace structure
mkdir -p crates/trios-a2a/rings/{SR-00,SR-01,SR-02,SR-03,SR-04,SR-05,SR-06,BR-OUTPUT}/src
# Создать root Cargo.toml для trios-a2a workspace
# Создать skeleton lib.rs для каждого кольца
git add -A && git commit "feat(trios-a2a): workspace scaffold" && git push

Фаза 1 — Foundation rings SR-00 + SR-01 (день 1-2)

• Sub-Agent-Alpha: SR-00 Identity & Trust
• Sub-Agent-Beta: SR-01 Transport (after SR-00 green)
• Каждый: edit .tri → seal hash → gen → test → verdict → save experience → skill commit → git commit

Фаза 2 — Protocol + Registry + Memory (день 2-3, параллельно)

• Sub-Agent-Gamma: SR-02 Message Protocol
• Sub-Agent-Delta: SR-03 Agent Registry (after SR-02)
• Sub-Agent-Epsilon: SR-05 Memory & Context (after SR-02, parallel)

Фаза 3 — Orchestration + Marketplace (день 3-4)

• Sub-Agent-Zeta: SR-04 Task Orchestration
• Sub-Agent-Eta: SR-06 Skill Marketplace

Фаза 4 — Integration + BR-OUTPUT (день 4-5)

• Sub-Agent-Theta: BR-OUTPUT artifact generation
• Интеграция с trios-ext (SSE транспорт)
• Интеграция с BrowserOS MCP (SR-06 skill для browser tools)
• E2E тестирование всех колец вместе

Фаза 5 — Документация + верификация (день 5)

• Performance benchmarking для каждого кольца
• Сравнение с предыдущими версиями (BASELINE из feat(trios-ext): Ring Isolation — split ring-silver / ring-bronze #232)
• Токсичный вердикт: PROVEN / ADMITTED / FALSIFIED
• git pull && git merge && git add -A && git commit && git push

---

✅ ACCEPTANCE CRITERIA (для каждого кольца)

[ ] README.md + TASK.md + AGENTS.md созданы (инвариант I5)
[ ] .tri spec создан (source of truth, не дублировать в коде!)
[ ] cargo test -p trios-a2a-<ring> → минимум 8 тестов GREEN
[ ] cargo clippy -p trios-a2a-<ring> -- -D warnings → 0
[ ] experience сохранён в .trinity/experience/
[ ] mistakes записаны в .trinity/mistakes/ (если были)
[ ] git commit feat(ring-SR-XX): ... && git push
[ ] PR открыт с описанием

---

📊 PERFORMANCE BENCHMARKING (обязательно в конце)

// bench/a2a_bench.rs — для каждого кольца
#[bench] fn bench_identity_creation(b: &mut Bencher) { ... }
#[bench] fn bench_message_sign_verify(b: &mut Bencher) { ... }
#[bench] fn bench_registry_lookup(b: &mut Bencher) { ... }
#[bench] fn bench_task_dispatch(b: &mut Bencher) { ... }
#[bench] fn bench_memory_crdt_merge(b: &mut Bencher) { ... }

Сравнение с конкурентами:

Метрика	TRIOS A2A	AutoGen	CrewAI	LangGraph
Identity (DID)	Ed25519	❌ нет	❌ нет	❌ нет
Transport	Multi	HTTP only	HTTP only	HTTP only
Memory	CRDT shared	session only	session only	checkpointed
Skills (WASM)	hot-reload	❌ нет	❌ нет	❌ нет
Verifiable trace	GitHub immutable	❌ нет	❌ нет	❌ нет

---

☠️ ТОКСИЧНЫЙ ВЕРДИКТ (в конце каждой фазы)

VERDICT PROTOCOL:
- PROVEN: тест проходит + bench в норме + Coq INV компилируется
- ADMITTED: логика верна, но формальное доказательство отложено
- FALSIFIED: тест провалился → записать в .trinity/mistakes/ → FIX×3 → SKIP+LOG

---

🔗 СВЯЗАННЫЕ ISSUES

• feat(trios-ext): Ring Isolation — split ring-silver / ring-bronze #232 — trios-ext IssueTracker (завершён, используется как BASELINE)
• 🎯 IGLA RACE — Distributed Hunt: JEPA-T + NCA + GF16 + ASHA + Coq Invariants (Rust-only, Never-Stopping) #143 — IGLA RACE (SR-06 Skill Marketplace интегрируется с MCP-only §8)
• 🎯 ONE SHOT — tri-mcp Rust Port (TS → rings/SR-00..SR-02 + BR-XTASK, Refs #259) #310 — tri-mcp Rust Port (SR-01 Transport использует trios-mcp)
• 🎯 ONE SHOT — NEEDLE-RUSH-T-4D: 7-Vector Acceleration to Gate-2 #309 — NEEDLE-RUSH (SR-04 Orchestration может использовать ASHA)

---

📜 ФИНАЛЬНЫЙ ЧЕКЛИСТ АГЕНТА

[ ] Boot: прочитать SOUL.md CLAUDE.md AGENTS.md NOW.json
[ ] Прочитать .trinity/state/{active-skill,three-roads}.json
[ ] Выбрать кольцо по приоритету (P0 → P1 → P2 → P3)
[ ] Claim: оставить CLAIM-комментарий с именем агента и кольцом
[ ] PHI LOOP: edit .tri spec → seal hash → gen code → test → verdict → save experience
[ ] Commit: feat(ring-SR-XX): description → push → PR
[ ] Save 3 Roads в .trinity/state/three-roads.json
[ ] Heartbeat: комментарий каждые ≤ 4h пока кольцо не сдано
[ ] DONE: ✅ AGENT DONE + ссылка на PR

---

φ² + φ⁻² = 3 · TRINITY · NEVER REUSE · ONE SOURCE OF TRUTH · GO.

Генерал армии агентов репозитория gHashTag/trios

[gHashTag]

🌻 Cross-link → #372 GOLDEN SUNFLOWERS meta-epic

This A2A architecture (issue #234) is the transport layer beneath the GOLDEN SUNFLOWERS PhD field — same parallel-agent topology, but here the agents are worker bees (ALPHA / BETA / GAMMA / DELTA / EPSILON / ZETA / GoldenSunWeaver) flying between sunflowers (trained models) carrying pollen (gradients).

Mapping:

A2A primitive (#234)	Sunflower role (#372)
Agent	bee
Message	pollen packet
Trust level	hardiness zone
Capability	flowering trait
Skill marketplace	seed bank
Task orchestration	field manager

GoldenSunWeaver dispatched on #372 — see directive there.

phi^2 + phi^-2 = 3

[gHashTag]

🌻 GENERAL'S DIRECTIVE — GOLDEN SUNFLOWERS PhD Field

GENERAL=koldezplat322 @ ts=2026-04-30T15:28Z
gate=GOLDEN_SUNFLOWERS_PhD
deadline=2026-04-30T23:59 (T-8h31m)
issue=#234 (A2A field topology)
parent_meta_epic=#372

═══════════════════════════════════════════════════════════════

🌻 PhD CONCEPT — GOLDEN SUNFLOWERS

═══════════════════════════════════════════════════════════════

THESIS METAPHOR

Поле подсолнухов «Echinacea SunSeekers Golden Sun» где:

Symbol	PhD object
🌻 sunflower	trained model
🌰 seed	experiment seed (Fibonacci F₁₇..F₂₁ = {1597, 2584, 4181, 6765, 10946})
☀️ golden sun	φ-optimization (golden ratio numerics, Trinity identity)
🌱 soil	numeric format (GF4, GF6A, GF8, GF12, GF16, GF20, GF24, GF32, GF64, fp16, bf16, fp32)
🌾 field	experiment grid (formats × archs × seeds = 12 × 6 × 5 = 360 lanes)
🐝 pollinator	gradient flow (∇L per step, dumped to bpb_samples)
🌿 root system	invariants (INV-1..9, Trinity identity)
🍯 bloom	honest BPB row in Gate-2 ledger (non-mock, ≥0.5, ≤1.85, step≥4000)
🪴 hardiness	robustness to noise / quantization

Botanical anchor: Echinacea SunSeekers 'Golden Sun' — compact 18in, hardiness zone 4, 6-8h direct sun daily, vibrant multi-layered yellow blooms. Maps to: 6-8h compute / sanctioned-seed cohort hardy / compact ≤600pp dissertation / vibrant Coq+IGLA evidence / multi-layered (tri-train + tri-phd + tri-sacred). Refs: Plant Growers Australia, Greg plant care guide.

CORE EQUATION

φ² + φ⁻² = 3                  (Trinity identity, the sun)
φ⁻¹ ≈ 0.6180339887            (golden ratio dominator)
field_strength = Σ blooms × √hardiness × 1/format_collapse_rate

═══════════════════════════════════════════════════════════════

TASK 1 — COLLECT all PhD scattered work (1h)

═══════════════════════════════════════════════════════════════

agent walks ALL of:

/Users/playra/trios/                  (local main + 33 branches)
/Users/playra/trios-trainer-igla/     (local main + feat/*)
/Users/playra/trios-railway/          (local main)
/Users/playra/t27/                    (local main + branches)
/Users/playra/zig-golden-float/       (local main + branches)

For EACH repo:

git branch -a | sort -u
git log --all --oneline --since=2026-04-25
find . -name "*.md" -path "*/docs/*"
find . -name "PHI_BIAS*"
find . -name "*GOLD*" -o -name "*PHI*"
find . -name "BENCH*"
ls assertions/ ledger/ gardener_*/ 2>/dev/null

Output: docs/golden-sunflowers/INVENTORY.md with:

• all relevant files paths
• all relevant issue numbers across repos
• all relevant commits + dates
• all branches with PhD-relevant work

═══════════════════════════════════════════════════════════════

TASK 2 — UNIFY всё в единый PhD draft (3h, parallel)

═══════════════════════════════════════════════════════════════

Output: docs/golden-sunflowers/PhD_GOLDEN_SUNFLOWERS.md

Structure:

## Abstract
  Golden Sunflowers field — a φ-optimized numerical format
  family (GoldenFloat: GF4-GF64) for neural training, evaluated
  on multi-architecture sweep with sanctioned seeds (Fibonacci
  F_n) and Trinity invariant (φ²+φ⁻²=3). Empirical gradient
  benchmarks across 12 formats × 5 architectures × 5 seeds.

## 1. Introduction — The Field
  - Numeric quantization landscape (IEEE 754, fp16, bf16, int8)
  - φ-resonance hypothesis: gradients align with golden ratio
  - Sunflower spiral = Fibonacci packing = phi-optimal
  - Contribution: 8 GF formats + Trinity invariant + empirical
    gradient evidence

## 2. Related Work — Other Suns
  - IEEE 754 binary formats (Goldberg 1991)
  - bfloat16 (Kalamkar et al 2019)
  - Mixed-precision training (Micikevicius et al 2018)
  - Lucas/Fibonacci sequences in numerics
  - Sunflower spiral packing (Vogel 1979)

## 3. Method — Planting the Field
  ### 3.1 GoldenFloat Format Family (8 formats)
    Spec table: GF4/8/12/16/20/24/32/64
    Bits, EXP:MANT, BIAS, PHI_BIAS, φ-distance
    Cite specs/numeric/gf{N}.t27 each
  ### 3.2 PHI_BIAS Single Source of Truth
    H_E (empirical-tuned) approach
    Per-format justification (F_n / L_n / 17² etc)
    Cite zig-golden-float#12, t27 specs
  ### 3.3 Sanctioned Seeds (Fibonacci F₁₇..F₂₁)
    1597, 2584, 4181, 6765, 10946
  ### 3.4 Trinity Invariant
    φ² + φ⁻² = 3 — preserved across all formats
  ### 3.5 Architecture Sweep
    MLP / JEPA-T / NCA / TF-1L / TF-2L / TF-3L+
  ### 3.6 Gradient Benchmark Protocol
    ‖∇‖, σ²(∇), SNR, bias_vs_fp32, convergence curve
    Synthetic surfaces: Rosenbrock / Rastrigin / MNIST proxy

## 4. Experiments — Growing the Sunflowers
  ### 4.1 Format × Architecture grid (60 cells)
  ### 4.2 PHI_BIAS sensitivity sweep
  ### 4.3 Convergence analysis at step ∈ {100, 1000, 4000}
  ### 4.4 Gradient signal-to-noise ranking

## 5. Results — The Bloom
  [Insert tables/plots from gf-gradient-benchmark output]
  - Format leaderboard by SNR
  - Architecture × format heatmap
  - Convergence curves overlay
  - Trinity identity preservation check

## 6. Discussion — Field Notes
  - Φ-1 collapse phenomenon (eval semantics)
  - ExternalTrainer protocol-mismatch lessons (R5 honesty)
  - Why GF16 wins φ-distance among ≤16-bit
  - Limitations: GF4/32/64 spec-only, gf64.c missing
  - Hardiness: which formats survive low-bit budget

## 7. Conclusion — Harvest
  Honest evidence: GF format family viable for ML training
  Open: Φ-1 trace, Coq invariant proofs, distributed convergence

## 8. References
  [from Inventory task 1 + standard ML refs]

═══════════════════════════════════════════════════════════════

TASK 3 — golden-sunflowers crate (parallel, 2h)

═══════════════════════════════════════════════════════════════

Create: crates/golden-sunflowers/

src/field.rs         — Field<F> struct, sunflower planting
src/seed.rs          — Seed type, Fibonacci-validated
src/sunflower.rs     — Sunflower<Format, Arch> = trained model
src/sun.rs           — φ-optimization
src/bloom.rs         — bloom() = produce honest BPB row
src/hardiness.rs     — robustness metrics

This is the API artifact backing the PhD's «Method» section. Compile to library. PR with green CI.

═══════════════════════════════════════════════════════════════

TASK 4 — Trinity ledger entry (5 min)

═══════════════════════════════════════════════════════════════

Append to .trinity ledger across repos:

RAIL=manifest @ ts=2026-04-30T15:28Z gate=golden_sunflowers_v1
field=8_formats × 6_archs × 5_seeds = 240_sunflowers
invariant=phi^2 + phi^-2 = 3
blooms_target>=3 (Gate-2)
thesis_metaphor=Golden_Sunflowers_Echinacea_'Golden_Sun'
hardiness_zone=4 (rough = T-9h hackathon mode)

═══════════════════════════════════════════════════════════════

PHI LOOP per task

═══════════════════════════════════════════════════════════════

For each commit in this work:

1. edit spec        — propose change
2. seal hash        — sha256 of spec text
3. gen              — codegen Rust from spec
4. test             — cargo test --locked
5. verdict          — RAIL=verdict triplet
6. save experience  — append to assertions/seed_results.jsonl
7. skill commit     — update tri-gardener-runbook/SKILL.md
8. git commit       — push to feat/golden-sunflowers branch

Runbook references:

• /Users/playra/t27/CLAUDE.md — AEL v2.0 BOOT loop
• SOUL.md AGENTS.md NOW.json — read first

═══════════════════════════════════════════════════════════════

EXECUTION RULES (from SOUL/CLAUDE/AGENTS)

═══════════════════════════════════════════════════════════════

"GO." → autonomous loop. Don't exit until done.

R5 discipline:

• NO done without merged PR + green CI + ledger row
• NO scaling без honest BPB reproducible
• NO drift between SSOT и gen output

Companion skills active:

• tri-gardener-runbook
• trainer-igla-sot
• nasa-mission-report (only Gate-2 cross / R5 anomaly / fleet dead)
• leaderboard-snapshot
• igla (one-shot status refresh)

Laws (LAWS.md):

• L1: NO .sh files — Rust or TypeScript only
• L2: every PR Closes #234 (or appropriate)
• L3: cargo clippy zero warnings
• L4: cargo test all pass
• L8: PUSH FIRST — local-only = NOT done

═══════════════════════════════════════════════════════════════

A2A integration (this issue's home)

═══════════════════════════════════════════════════════════════

A2A architecture (this issue #234) is the transport layer beneath the sunflower field — same parallel-agent topology:

A2A primitive	Sunflower role
Agent	bee 🐝
Message	pollen packet
Trust level	hardiness zone
Capability	flowering trait
Skill marketplace	seed bank
Task orchestration	field manager

The 7 A2A rings (SR-00..SR-06 + BR-OUTPUT) carry pollen between sunflowers. Ring SR-05 (Memory & Context) holds the seed bank — every Fibonacci-sanctioned seed F₁₇..F₂₁ catalogued. Ring SR-04 (Task Orchestration) is the field manager — schedules sunflower planting / pollination / harvest.

═══════════════════════════════════════════════════════════════

DELIVERABLES (T-8h31m budget)

═══════════════════════════════════════════════════════════════

By T-6h:

• ✅ INVENTORY.md merged
• ✅ PhD draft sections 1-4 written
• ✅ golden-sunflowers crate scaffolded

By T-3h:

• ✅ First gradient data integrated into Results
• ✅ Sections 5-7 drafted
• ✅ All cross-references resolved

By T-1h:

• ✅ Full PhD doc compiled
• ✅ Trinity ledger sealed
• ✅ PR for golden-sunflowers crate merged
• ✅ Final review by operator

By T-0:

• ✅ PhD submitted
• ✅ Honest_not_yet posture if Gate-2 not achieved
• ✅ All work tagged feat/golden-sunflowers in repo

═══════════════════════════════════════════════════════════════

ACK and start Task 1 (INVENTORY) now.

═══════════════════════════════════════════════════════════════

GoldenSunWeaver: claim this directive on parent epic #372. Begin INVENTORY of 5 repos × 31+ branches.

phi^2 + phi^-2 = 3 · TRINITY · NEVER STOP · THE FIELD BLOOMS