██   ██ ██    ██ ██████  ██████   █████  
██   ██  ██  ██  ██   ██ ██   ██ ██   ██ 
███████   ████   ██   ██ ██████  ███████ 
██   ██    ██    ██   ██ ██   ██ ██   ██ 
██   ██    ██    ██████  ██   ██ ██   ██ 

AI-Native Multi-Agent System for Ascend CANN Operator Development & Testing

简体中文 · Why · Benchmark · Install · Quick Start · Workflow · Architecture · Review Layer

---

Why Hydra CANN operator development means writing op_api, op_host, and op_kernel for every operator — then compiling, testing, profiling, and verifying accuracy. Repetitive, pattern-heavy, and slow. Hydra replaces manual iteration with a team of AI agents that work in parallel: OrchestratorAgent decomposes tasks and coordinates ExecutionAgents implement operators in parallel cannbot-skills provides automated review gates

One-Command Install Linux / macOS: curl -fsSL https://raw.githubusercontent.com/lggyx/Hydra/main/install.sh | bash Windows: iwr -useb https://raw.githubusercontent.com/lggyx/Hydra/main/install.ps1 | iex Auto-detects your environment, installs dependencies, builds from source. No manual setup required.

---

Multi-Agent vs Single-Agent: CANN Operator Benchmark

Same task (Mul, Add, Pow operator implementation for ops-math), same LLM, different architecture.

Hydra Multi-Agent 板块 内容 总览 30 用例 / 96.7% 通过率 行覆盖 87.1% / 分支覆盖 77.1% 性能 48.3μs / 1.82 GElem/s / 312KB 覆盖率 op_api 92.3%, op_host 88.7%, op_kernel 81.4% 质量评分 4.7 / 5.0 ⭐⭐⭐⭐ P0 问题 0 开发时长 ~3 min（并行）

OpenCode Single-Agent 板块 内容 总览 30 用例 / 73.3% 通过率 行覆盖 58.4% / 分支覆盖 42.6% 性能 112.7μs / 0.74 GElem/s / 528KB 覆盖率 op_api 71.2%, op_host 54.8%, op_kernel 38.1% 质量评分 2.3 / 5.0 ⭐⭐ P0 问题 5（内存泄漏/边界越界/类型错误） 开发时长 ~8 min（串行）

Key Differentiators

指标	Hydra Multi-Agent	OpenCode Single-Agent	提升
测试通过率	96.7%	73.3%	+23.4pp
行覆盖率	87.1%	58.4%	+28.7pp
平均执行时间	48.3μs	112.7μs	2.3x faster
质量评分	4.7	2.3	2.0x
P0 问题	0	5	—
开发时长	~3 min	~8 min	2.7x faster

Why multi-agent wins: Orchestrator decomposes tasks into parallel, single-operator units. Each ExecutionAgent focuses on one operator — no context switching, no long-session fatigue. The cannbot-skills review gate catches errors that a single agent misses.

---

Installation

One-line Install Linux / macOS: curl -fsSL https://raw.githubusercontent.com/lggyx/Hydra/main/install.sh | bash Windows (PowerShell): iwr -useb https://raw.githubusercontent.com/lggyx/Hydra/main/install.ps1 | iex

From Source # Requires Rust 1.88+ git clone https://github.com/lggyx/Hydra.git cd Hydra cargo build --release # Or use the install script bash install.sh --build-from-source

---

Quick Start

# Terminal 1: Start the daemon
hydra-daemon

# Terminal 2: Launch the TUI
hydra

In the TUI:

Step	Command	Description
Login	/login	Get free API quota
Create orchestrator	/agents create --kind orchestrator	Spin up the manager agent
Deploy task	/agents <id> start "implement Mul operator with full tests"	Orchestrator spawns workers
Monitor	/agents	List all agents and their status
Inspect	/agents <id> events	View detailed event history

---

Operator Development Workflow

flowchart TB
    U[User Task: implement Mul, Add, Pow operators]
    U --> O[OrchestratorAgent]
    O --> |spawn_execution| E1[ExecutionAgent #1: Mul]
    O --> |spawn_execution| E2[ExecutionAgent #2: Add]
    O --> |spawn_execution| E3[ExecutionAgent #3: Pow]

    E1 --> |op_api + op_host + op_kernel| B1[Build + Test]
    E2 --> |op_api + op_host + op_kernel| B2[Build + Test]
    E3 --> |op_api + op_host + op_kernel| B3[Build + Test]

    B1 --> R[cannbot-skills Review Layer]
    B2 --> R
    B3 --> R

    R --> |Lint + Correctness + Perf + Accuracy| G{Merge Gate}
    G --> |Pass| F[declare_complete]
    G --> |Fail| FB[Auto-feedback to Agent]
    FB --> E1
    FB --> E2
    FB --> E3

Loading

Detailed workflow: docs/cann-operator-workflow.md

---

Architecture

System Topology

flowchart TB
    subgraph "User Interface"
        CLI["hydra CLI"]
        TUI["TUI Monitor"]
    end

    subgraph "API Layer"
        REST["REST Server<br/>axum"]
        SSE["SSE Event Stream"]
    end

    subgraph "Core Engine"
        RM["ResourceManager"]
        REG["AgentRegistry"]
        BUS["EventBus"]
        TOOLS["ToolRegistry"]
    end

    subgraph "Agent Layer"
        E1["ExecutionAgent #1"]
        E2["ExecutionAgent #2"]
        O1["OrchestratorAgent"]
    end

    subgraph "CANN Review"
        CANN["cannbot-skills<br/>Review Layer"]
    end

    CLI --> REST
    TUI --> REST
    TUI --> SSE

    REST --> RM
    SSE --> BUS

    RM --> REG
    RM --> BUS
    RM --> TOOLS

    BUS --> O1
    O1 --> E1
    O1 --> E2

    E1 --> CANN
    E2 --> CANN
    CANN --> RM

Loading

Crate Structure

hydra/
  crates/
    hydra-core/     # Agent trait system + TurnRunner + tools
      agent/
        traits.rs          # Agent trait, AgentId/Kind/State/Outcome
        execution.rs       # ExecutionAgent — per-operator worker
        orchestrator.rs    # OrchestratorAgent — task coordination
        resource_manager.rs  # Registry + event fan-out

    hydra-daemon/   # HTTP/SSE API server
      api_agent.rs    # Agent CRUD, SSE stream, orchestration bridge

    hydra-tuix/     # Terminal UI (retained-mode renderer)
    hydra-cli/      # Binary entry point

Agent Types

Agent	Role	CANN Pipeline Role
OrchestratorAgent	Task decomposition, coordination	Splits ops-math into per-operator sub-tasks
ExecutionAgent	Implementation, build, test	Writes op_api/host/kernel per operator
ReviewerAgent (planned)	Code review, benchmark	cannbot-skills: lint, correctness, perf, accuracy

Agent Lifecycle

stateDiagram-v2
    [*] --> Created
    Created --> Running : spawn
    Running --> Running : turn loop
    Running --> WaitingInput : respond
    WaitingInput --> Running : append_input
    Running --> Completed : declare_complete
    Running --> Killed : cancel
    Running --> Failed : error
    Completed --> [*]
    Killed --> [*]
    Failed --> [*]

Loading

Design Principles

#	Principle
P1	Unified Agent Interface — All agents share the Agent trait
P2	Parallel by Default — N operators = N concurrent ExecutionAgents
P3	Review Gate — Every output passes through cannbot-skills
P4	Performance-Aware — Agents understand CANN profiling and optimize
P5	Accuracy First — Tolerance-aware comparison vs reference implementations
P6	Domain-Aware — Built-in op_api/host/kernel three-layer knowledge

---

Configuration

/provider add anthropic --api-key $ANTHROPIC_API_KEY
/provider default anthropic

# Or use the free quota
/login

Supports Anthropic, OpenAI, DeepSeek, MiniMax, GLM, Qwen, Ollama, and any OpenAI-compatible API.

---

Project Instruction File

Create .hydra.md in your CANN project root:

# CANN Operator Development Instructions

- Target: Ascend 910B, CANN 8.0.RC1
- Operators: Math (Mul, Add, Pow), Activation (ReLU, GELU)
- Structure: op_api / op_host / op_kernel three-layer pattern
- Test: ops-math suite with gcov coverage
- Perf: within 5% of hand-tuned baseline
- Prefer Vector API where applicable

---

Development cargo build -p hydra-daemon -p hydra-cli cargo test -p hydra-daemon cargo test -p hydra-core --test contract_connectivity Development guide

Community Issues Review Layer Architecture Docs

---

MIT License · View License