Hermes Kanban Deep Dive with a Real-World Public Account Matrix Management System

What is Hermes Kanban?

Hermes Kanban is the multi‑agent orchestration component of the Hermes Agent framework. It combines a Kanban‑style visual board with autonomous large‑language‑model execution, providing automatic task decomposition, parallel execution, dependency handling, state tracking, and optional human intervention.

Core Features

Task orchestration : automatically splits large projects into parallelizable subtasks.

Role division : predefined specialist roles (researcher, analyst, writer, reviewer, backend‑eng, frontend‑eng, ops, pm).

Dependency management : resolves task precedence automatically.

Status tracking : records the full lifecycle of each task.

Human intervention : tasks can be paused for manual decisions.

Core Concepts

Specialist Roster

researcher

: information gathering, fact verification, data organization (workspace scratch). analyst: synthesis, deduplication, recommendation generation (workspace scratch). writer: content creation (workspace scratch or dir:). reviewer: quality audit and approval (workspace scratch). backend‑eng: backend code development (workspace worktree). frontend‑eng: frontend UI development (workspace worktree). ops: deployment and script execution (workspace dir:). pm: requirement analysis and specification (workspace scratch).

Workspace Types

scratch

: temporary directory cleared after the task. dir:<path>: persistent shared directory for long‑term state. worktree: Git worktree for code development and version control.

Task State Flow

todo → ready → in_progress → done
               ↓
           blocked (awaiting human)

todo : initial state, waiting for dependencies.

ready : dependencies satisfied, ready for scheduling.

in_progress : currently executing.

done : task completed.

blocked : paused for manual decision.

Case Study: Public‑Account Matrix Management System

Project Background

The system must manage multiple public‑account accounts, article drafts, media assets, one‑click publishing, API‑key security, and large‑model cost control.

Task Decomposition

T0: Project planning (pm)
   ↓
T1: Account management (backend‑eng)
T2: Article management (backend‑eng)
T3: One‑click publishing (backend‑eng)
T4: API key management (backend‑eng)
T5: Large‑model management (backend‑eng)
   ↓
T6: Integration testing (ops)
   ↓
T7: Code review (reviewer)

Implementation (Python)

import os
from hermes_tools import kanban_create, kanban_complete

# Tenant isolation
tenant = os.environ.get("HERMES_TENANT")

# T0: Project planning
t0 = kanban_create(
    title="制定公众号矩阵管理系统技术规格",
    assignee="pm",
    body="""
    需求：
    1. 支持多公众号账号管理（至少 10 个账号）
    2. 文章草稿、素材库、定时发布
    3. 一键发布到多个平台
    4. API 密钥安全存储
    5. 大模型调用成本统计

    输出：技术规格说明书，包含 API 设计、数据模型、安全策略
    """,
    tenant=tenant,
)["task_id"]

# T1: Account management module
t1 = kanban_create(
    title="开发公众号管理模块",
    assignee="backend-eng",
    body="""
    功能：
    - 公众号账号 CRUD
    - OAuth2 认证流程
    - 账号权限配置
    - 账号健康检查

    技术栈：FastAPI + SQLAlchemy + Redis
    输出：/modules/account_manager.py + 单元测试
    """,
    parents=[t0],
)["task_id"]

# T2: Article management module
t2 = kanban_create(
    title="开发文章管理模块",
    assignee="backend-eng",
    body="""
    功能：
    - 文章草稿箱（Markdown 格式）
    - 素材库管理（图片、视频）
    - 定时任务调度
    - 版本历史

    输出：/modules/article_manager.py + 单元测试
    """,
    parents=[t0],
)["task_id"]

# T3: One‑click publishing module
t3 = kanban_create(
    title="开发一键发布模块",
    assignee="backend-eng",
    body="""
    功能：
    - 多平台适配器（微信、知乎、掘金）
    - 发布策略配置（立即/定时/灰度）
    - 发布状态追踪
    - 失败重试机制

    输出：/modules/publisher.py + 适配器实现
    """,
    parents=[t0],
)["task_id"]

# T4: API key management module
t4 = kanban_create(
    title="开发 API 密钥管理模块",
    assignee="backend-eng",
    body="""
    功能：
    - 密钥加密存储（AES-256）
    - 密钥轮换机制
    - 访问日志审计
    - 异常访问告警

    输出：/modules/key_manager.py + 加密工具类
    """,
    parents=[t0],
)["task_id"]

# T5: Large‑model management module
t5 = kanban_create(
    title="开发大模型管理模块",
    assignee="backend-eng",
    body="""
    功能：
    - 多模型接入（Qwen、Claude、GPT）
    - 调用成本统计
    - 智能路由（根据任务类型选择模型）
    - Token 使用监控

    输出：/modules/llm_manager.py + 成本计算器
    """,
    parents=[t0],
)["task_id"]

# T6: Integration testing
t6 = kanban_create(
    title="执行集成测试",
    assignee="ops",
    body="""
    测试范围：
    1. 公众号绑定流程端到端测试
    2. 文章发布完整链路测试
    3. 密钥轮换不影响服务
    4. 大模型故障降级

    输出：测试报告 + 覆盖率统计
    """,
    parents=[t1, t2, t3, t4, t5],
)["task_id"]

# T7: Code review
t7 = kanban_create(
    title="代码审核与安全审计",
    assignee="reviewer",
    body="""
    审核重点：
    1. 密钥是否硬编码
    2. SQL 注入风险
    3. 认证授权逻辑
    4. 敏感信息脱敏

    输出：审核报告 + 问题列表
    """,
    parents=[t6],
)["task_id"]

# Mark the workflow as complete
kanban_complete(
    summary="创建公众号矩阵管理系统开发任务树，共 8 个任务（T0‑T7）",
    metadata={
        "total_tasks": 8,
        "modules": ["account", "article", "publisher", "key_manager", "llm_manager"],
        "task_graph": {
            "T0": "pm: 技术规格",
            "T1‑T5": "backend‑eng: 5 个模块并行开发",
            "T6": "ops: 集成测试",
            "T7": "reviewer: 代码审核"
        }
    }
)

Execution Effects

Parallel execution of T1‑T5 : five backend tasks start simultaneously without blocking each other.

Dependency waiting : T6 remains in ready until all T1‑T5 finish.

Automatic state transition : completion of a task triggers its downstream task.

Full audit logging : execution logs and outputs are persisted for traceability.

Advanced Usage

Human Intervention (Block/Unblock)

When a task requires a manual decision, add a comment and block the task.

from hermes_tools import kanban_block, kanban_comment

# Add contextual comment
kanban_comment(
    task_id="t_xxx",
    body="""
    当前情况：
    - 微信公众号 API 需要企业认证才能申请发布权限
    - 测试账号只能使用草稿箱功能
    - 生产环境需要营业执照扫描件

    需要决策：是否先用测试环境继续开发？
    """
)

# Block the task awaiting human input
kanban_block(reason="需要企业认证才能继续，是否切换到测试环境开发？")

Dependency Graph (DAG)

Complex projects can be expressed as a directed acyclic graph.

T0
    / | \
   T1 T2 T3
    \ | /
     T4
     |
     T5

Metadata Propagation

Downstream tasks can read metadata produced by upstream tasks.

# Upstream task completion
kanban_complete(
    summary="完成用户认证模块",
    metadata={
        "endpoints": ["/login", "/logout", "/refresh"],
        "security_level": "high",
        "test_coverage": 0.95
    }
)

# Downstream task can access the above metadata for further checks

Common Issues

Q1: Task stuck in todo state

Cause : parent task not finished or dependency misconfiguration.

Solution :

# View task details
hermes kanban show <task_id>

# Check status of parent tasks
hermes kanban list --status in_progress

Q2: Agent execution failure recovery

Three recovery methods:

Reclaim : abort current execution and reset to ready. hermes kanban reclaim <task_id> Reassign : change the executing role.

hermes kanban reassign <task_id> <new-profile> --reclaim

Model change : switch the underlying large model and rerun.

hermes -p <profile> model set qwen-max
hermes kanban reclaim <task_id>

Q3: Debugging task execution

# Live logs
hermes kanban tail <task_id>

# Full execution history
hermes kanban history <task_id>

# Export audit log as JSON
hermes kanban export <task_id> --format json

Best Practices

Task Decomposition Principles

Atomicity : each task can be executed independently.

Verifiability : clear completion criteria.

Rollbackability : failure does not affect other tasks.

Traceability : outputs are consumable by downstream tasks.

Metadata Standards

metadata = {
    "changed_files": ["module.py", "test_module.py"],
    "tests_run": 15,
    "tests_passed": 15,
    "security_scan": "passed",
    "decisions": ["使用 Redis 缓存会话"]
}

Error Handling

# Block when confidence is low
if confidence < 0.8:
    kanban_block(reason="需要确认业务逻辑：XXX 场景下是否允许并发发布？")
    return

# Do not hard‑code secrets
API_KEY = os.environ["API_KEY"]

Conclusion

Hermes Kanban transforms traditional Kanban into an autonomous workflow with automatic task splitting, smart assignment, real‑time state transitions, and built‑in auditability. The public‑account matrix case demonstrates parallel acceleration, specialist division, quality assurance, and full traceability across a complex development project.