端到端组合范例。每个配方都可独立运行(mock provider 或真实 LLM 二选一)。


1. 流式 UI(SSE 推送)

让前端浏览器实时看到每个 token:

defmodule MyApp.AgentSSE do
  use Plug.Router

  plug :match
  plug :dispatch

  get "/agents/:id/stream" do
    conn = send_chunked(conn, 200)

    {:ok, session} = CMDC.create_agent(
      session_id: id,
      model: "anthropic:claude-sonnet-4-5"
    )

    CMDC.subscribe(session)
    CMDC.prompt(session, conn.params["q"])

    stream_loop(conn, session)
  end

  defp stream_loop(conn, session) do
    receive do
      {:cmdc_event, _sid, {:message_delta, %{delta: text}}} ->
        {:ok, conn} = chunk(conn, "data: #{Jason.encode!(%{delta: text})}\n\n")
        stream_loop(conn, session)

      {:cmdc_event, _sid, {:agent_end, _msgs, usage}} ->
        chunk(conn, "data: #{Jason.encode!(%{done: true, usage: usage})}\n\n")
        CMDC.stop(session)
        conn

      {:cmdc_event, _sid, {:tool_execution_start, name, _, _}} ->
        chunk(conn, "data: #{Jason.encode!(%{tool: name})}\n\n")
        stream_loop(conn, session)
    after
      30_000 ->
        CMDC.abort(session, reason: "timeout")
        conn
    end
  end
end

生产环境推荐直接用 cmdc_gateway 的 SSE / WebSocket 端点,不用自己撸。


2. HITL 审批(CLI 终端确认)

{:ok, session} = CMDC.create_agent(
  model: "anthropic:claude-sonnet-4-5",
  tools: [CMDC.Tool.Shell, CMDC.Tool.WriteFile],
  plugins: [
    {CMDC.Plugin.Builtin.HumanApproval, [
      tools: ["shell", "write_file"],
      permission_options: [:approve_once, :approve_always, :reject_once]
    ]}
  ]
)

CMDC.subscribe(session)
CMDC.prompt(session, "在 /tmp 创建一个 Elixir Hello World 项目")

handle_loop = fn loop ->
  receive do
    {:cmdc_event, _sid, {:approval_required, %{id: id, tool: tool, args: args}}} ->
      IO.puts("\n[Agent 想跑] #{tool} #{inspect(args)}")
      IO.write("(o)nce / (a)lways / (r)eject ? ")

      case IO.gets("") |> String.trim() do
        "o" -> CMDC.approve(session, id)
        "a" -> CMDC.approve(session, id, kind: :approve_always)
        _   -> CMDC.reject(session, id)
      end

      loop.(loop)

    {:cmdc_event, _sid, {:agent_end, _, _}} ->
      :done
  end
end

handle_loop.(handle_loop)

approve_always{tool, command_family} 加入 session 白名单,下次同类 工具调用直接放行。


3. 长会话不失忆(MemoryLoader + MemoryFlush 闭环)

{:ok, session} = CMDC.create_agent(
  model: "anthropic:claude-sonnet-4-5",
  working_dir: "/tmp/long-chat",
  plugins: [
    # 启动时加载 AGENTS.md / MEMORY.md 到 system prompt
    {CMDC.Plugin.Builtin.MemoryLoader, [
      files: ["AGENTS.md", "MEMORY.md"]
    ]},

    # 压缩前把待丢消息里的关键事实持久化
    {CMDC.Plugin.Builtin.MemoryFlush, [
      file: "MEMORY.md",
      max_facts_per_flush: 10,
      dedupe: true
    ]}
  ],
  compactor: [
    trigger: {:tokens, 50_000},
    keep: {:messages, 10}
  ]
)

# 第 1 次会话:记下用户偏好
CMDC.prompt(session, "我喜欢用 Phoenix LiveView 而不是 React 写前端,记住。")
{:ok, _} = CMDC.collect_reply(session)

# 跑很多轮触发 compact,期间 MemoryFlush 把"用户偏好 LiveView"写进 MEMORY.md
# ...

# 同 session_id 重新打开(或 BEAM 重启后)
{:ok, session2} = CMDC.create_agent(
  session_id: session.id,
  ...
)

# MemoryLoader 自动加载 MEMORY.md,新会话依然记得偏好
CMDC.prompt(session2, "帮我写个登录页")

详细示例见 CMDC.Plugin.Builtin.MemoryFlush


4. 多 Agent 协作(Task 工具派发)

父 Agent 通过 Tool.Task 派发子任务:

{:ok, parent} = CMDC.create_agent(
  model: "anthropic:claude-sonnet-4-5",
  tools: [CMDC.Tool.Task],
  subagents: [
    %CMDC.SubAgent{
      name: "researcher",
      description: "搜集资料的专家",
      system_prompt: "你只负责搜集和总结资料,不做评论",
      tools: [CMDC.Tool.Grep, CMDC.Tool.ReadFile]
    },
    %CMDC.SubAgent{
      name: "writer",
      description: "把资料整理成文章的作家",
      system_prompt: "把输入的资料写成结构化的中文报告",
      prompt_mode: :task  # 子代理用精简 system prompt 省 token
    }
  ]
)

CMDC.prompt(parent, """
帮我做一份"Elixir 在 AI Agent 领域应用"的调研报告:
1. 让 researcher 搜集本仓库 lib/ 下的相关代码
2. 让 writer 整合成 1500 字中文文章
""")

{:ok, report} = CMDC.collect_reply(parent)

子代理跑在独立 OTP 进程里,crash 不传染父进程。需要 DAG 编排(debate / hierarchy / router-llm)见 cmdc_orchestrator 子库。


5. 中段干预(Steering)

发现方向不对中途换路:

{:ok, session} = CMDC.create_agent(
  model: "anthropic:claude-sonnet-4-5",
  tools: [CMDC.Tool.Shell, CMDC.Tool.WriteFile]
)

CMDC.subscribe(session)
CMDC.prompt(session, "用 Python 实现一个简单的 Web 爬虫,目标 example.com")

# 等了几秒发现想换语言
Process.sleep(3_000)

ref = make_ref()
CMDC.steer(session, ref, """
打住——改用 Elixir 的 Req + Floki 实现,别用 Python。
""")

# 等待新方向的回复
{:ok, reply} = CMDC.collect_reply(session)

正在运行的 killable 工具被 brutal_kill,下次 LLM 调用看到合并后的 steering prompt。Plugin 可在 :before_steering hook 拦截恶意 steering。


6. 跨进程恢复(Checkpoint Facade,v0.5 推荐)

# 第一段:正常对话
{:ok, session} = CMDC.create_agent(
  session_id: "s-001",
  model: "anthropic:claude-sonnet-4-5"
)

CMDC.prompt(session, "帮我设计一个 Elixir Agent 框架,先列大纲。")
{:ok, _} = CMDC.collect_reply(session)

# v0.5 新增 facade:抓快照
{:ok, snapshot} = CMDC.checkpoint!(session,
  backend: CMDC.Checkpoint.Backend.DETS,
  label: "after-outline"
)

CMDC.stop(session)


# 第二段:BEAM 重启后 / 换台机器,从 snapshot 恢复
{:ok, snapshot} = CMDC.Checkpoint.load("s-001",
  backend: CMDC.Checkpoint.Backend.DETS
)

# v0.5 新增 facade:一行恢复 + 续 prompt
{:ok, restored} = CMDC.resume_session!(snapshot,
  session_id: "s-001-resumed"   # 避免 SessionRegistry 冲突,可选
)

CMDC.prompt(restored, "继续,第 2 章详细展开。")
{:ok, reply} = CMDC.collect_reply(restored)

序列化策略:保留 messages / tools / plugins(模块名) / user_data / prompt_mode 等可重建核心状态; plugin_states 不进 snapshot,resume 后 plugin 通过 :session_start hook 自动重建。

接 Cloak 加密

# write path
{:ok, snap} = CMDC.checkpoint!(session)
encrypted = CMDC.Checkpoint.Snapshot.redact(snap, &MyApp.Vault.encrypt/1)
CMDC.Checkpoint.Backend.DETS.save("s-001", encrypted, [])

# read path
{:ok, snap} = CMDC.Checkpoint.load("s-001", backend: CMDC.Checkpoint.Backend.DETS)
decrypted = CMDC.Checkpoint.Snapshot.redact(snap, &MyApp.Vault.decrypt/1)
{:ok, restored} = CMDC.resume_session!(decrypted, session_id: "s-001-resumed")

DETS backend 持久化到本地文件;生产推荐 PG backend(在 cmdc_memory_pg 子库提供)。


6.1 自动定时存档(AutoCheckpoint Plugin,v0.5 推荐)

避免每个集成方都自己写"每 N turn 自动 save"+"老 snapshot GC"逻辑:

{:ok, session} = CMDC.create_agent(
  session_id: "s-prod",
  model: "anthropic:claude-sonnet-4-5",
  plugins: [
    {CMDC.Plugin.Builtin.AutoCheckpoint,
     backend: CMDC.Checkpoint.Backend.DETS,    # 生产推 PG
     every_n_turns: 10,
     on_tools: ["shell", "edit_file"],          # 关键工具后强制存档
     on_events: [:approval_required, :session_end],
     max_checkpoints: 50,                       # 同 session 上限
     ttl_seconds: 30 * 86_400}                  # 30 天自动清理
  ]
)

# Plugin 自动按 OR 触发:任一条件命中即异步 save
# save + GC 走 CMDC.AsyncTaskSupervisor 异步,不阻塞 gen_statem callback
# snapshot.metadata 自动写 %{trigger: :auto, plugin: __MODULE__, turn: N}

# 订阅自动 save 事件做审计
CMDC.subscribe(session)
receive do
  {:cmdc_event, _, {:auto_checkpoint_saved, %{checkpoint_id: id, trigger: t}}} ->
    Logger.info("auto checkpoint saved: #{id} (trigger: #{t})")
end

6.2 长会话多租户 Hibernate(v0.5 新功能)

如果你的 Agent 是 1500+ idle 常驻多租户场景,单进程 heap ~8KB 累积内存爆炸。 新增 :hibernate_after_ms 配置后 idle 超时自动 hibernate,heap 8KB → 1.5KB(~80% 节省):

{:ok, session} = CMDC.create_agent(
  model: "anthropic:claude-sonnet-4-5",
  hibernate_after_ms: 60_000   # 1 分钟无事件自动 hibernate
)

推荐取值:

场景取值
短会话 / 单 Agentnil(默认,不启用)
长会话 / 主 Agent 多租户常驻60_000
极端多租户(>1000 idle)30_000
SubAgent 短任务nil(很快结束 hibernate 无收益)

:gen_statem 原生 :hibernate_after 选项,所有状态(idle / running / streaming / executing_tools) 下都会触发。配套 telemetry [:cmdc, :agent, :hibernate, :configured] 在 Agent init 时 emit 一次, 可观测有多少 session 启用了 hibernate。


6.3 多租户 Provider Profile(Provider Registry,v0.5.1+)

多租户 SaaS 场景下,每个租户有自己的 LLM provider api_key / base_url。 v0.5.1 新增 CMDC.Provider.Registry 替代 per-Agent provider_opts 拼装:

# ─────────────────────────────────────────────────────────────
# 启动时灌入:业务持久层 → cmdc Registry
# ─────────────────────────────────────────────────────────────

defmodule MyApp.ProviderProfileStore do
  @moduledoc "持久层(DB)维护所有租户 profile,启动时全量灌入 cmdc Registry。"

  use GenServer

  def start_link(_), do: GenServer.start_link(__MODULE__, [], name: __MODULE__)

  @impl true
  def init(_) do
    # 从 DB / 配置文件加载所有 profile(业务层持久化)
    profiles = MyApp.Repo.all(MyApp.AI.Profile)

    Enum.each(profiles, fn p ->
      :ok =
        CMDC.Provider.Registry.register(p.name,
          provider: p.provider,
          opts: [api_key: MyApp.Vault.decrypt(p.api_key), base_url: p.base_url]
        )
    end)

    {:ok, %{}}
  end
end

# ─────────────────────────────────────────────────────────────
# Blueprint 用字符串协议,一行替代 200µs/Agent opts 拼装
# ─────────────────────────────────────────────────────────────

defmodule MyApp.AI.TenantAgent do
  @behaviour CMDC.Blueprint

  def build(opts) do
    tenant_id = Keyword.fetch!(opts, :tenant_id)
    model_id = Keyword.fetch!(opts, :model_id)

    CMDC.Options.new!(
      # registry: 字符串协议 — Agent.init 自动解析 + 透写 provider_opts
      model: "registry:#{tenant_id}-anthropic:#{model_id}",
      tools: [CMDC.Tool.ReadFile, CMDC.Tool.Shell],
      working_dir: opts[:working_dir] || File.cwd!()
    )
  end
end

# ─────────────────────────────────────────────────────────────
# 跨节点同步:自实现 Broadcaster 接 Phoenix.PubSub(生产推荐)
# ─────────────────────────────────────────────────────────────

defmodule MyApp.ProviderRegistryBroadcaster do
  @behaviour CMDC.Provider.Registry.Broadcaster

  @topic "cmdc:provider_registry"

  # 可选:Registry init 时调用一次
  def setup(_opts) do
    Phoenix.PubSub.subscribe(MyApp.PubSub, @topic)
    :ok
  end

  @impl true
  def broadcast(event) do
    # 派发给所有订阅节点,远端节点 PubSub handler 调
    # send(GenServer.whereis(CMDC.Provider.Registry), {:cmdc_registry_remote, event})
    Phoenix.PubSub.broadcast(MyApp.PubSub, @topic, {:cmdc_registry_remote, event})
  end
end

# config/config.exs
config :cmdc, CMDC.Provider.Registry,
  broadcaster: MyApp.ProviderRegistryBroadcaster

# ─────────────────────────────────────────────────────────────
# Checkpoint resume 容错:profile 缺失自动重试
# ─────────────────────────────────────────────────────────────

defp resume_with_retry(snap, retries \\ 1) do
  case CMDC.resume_session!(snap) do
    {:ok, session} ->
      {:ok, session}

    {:error, {:registry_profile_missing, name}} when retries > 0 ->
      # 重新从 DB 加载 profile(可能其它节点已删/迁移)
      :ok = reload_profile_from_db(name)
      resume_with_retry(snap, retries - 1)

    {:error, _} = err ->
      err
  end
end

Profile name 约束:不能含 : 字符(解析时被吃掉一段)。 Studio 现有 "tenant-#{tid}-#{provider}" 格式继续可用;UUID 含 : 时替换为 -

性能Registry.lookup/1 ~0.5-1 µs ETS 读,比 per-Agent provider_opts 拼装快 200 倍。 完整设计见 docs/dev/rfc/2026-05-provider-registry.md(含 4 设计问题答疑 + 6 处评审订正照搬清单)。


7. 接 Langfuse / LangSmith / Tempo(Telemetry,v0.5.1 含 18 事件)

CMDC 只发 :telemetry 标准事件,sink 由你挂:

defmodule MyApp.LangfuseSink do
  def attach do
    :telemetry.attach_many(
      "cmdc-langfuse",
      CMDC.Telemetry.all_events(),       # v0.5.1 = 18 events (6 v0.4 + 10 v0.5 + 2 v0.5.1)
      &__MODULE__.handle_event/4,
      %{api_key: System.get_env("LANGFUSE_API_KEY")}
    )
  end

  # --- v0.4 core 6 events ---

  def handle_event([:cmdc, :llm, :request, :stop], measurements, metadata, config) do
    Langfuse.create_generation(%{
      session_id: metadata.session_id,
      model: metadata.model,
      input_tokens: metadata.tokens_in,
      output_tokens: metadata.tokens_out,
      latency_ms: measurements.duration_ms,
      api_key: config.api_key
    })
  end

  def handle_event([:cmdc, :tool, :exec, :stop], measurements, metadata, config) do
    Langfuse.create_span(%{
      session_id: metadata.session_id,
      name: "tool:#{metadata.tool}",
      duration_ms: measurements.duration_ms,
      error: metadata.error?,
      api_key: config.api_key
    })
  end

  # --- v0.5 new 10 events ---

  # Plugin Pipeline 整体跑完一次(hook 粒度,非 per-plugin),看 plugin 拖累
  def handle_event([:cmdc, :plugin, :pipeline, :stop], m, meta, _) do
    Metrics.histogram("cmdc.plugin.pipeline.duration_ms", m.duration_ms,
      tags: [hook: meta.hook, halted_by: meta.halted_by])
  end

  # Plugin 异常被 Pipeline rescue(含 plugin name / hook / reason)
  def handle_event([:cmdc, :plugin, :crash], _m, meta, _) do
    Sentry.capture_message("Plugin crash: #{inspect(meta.plugin)}",
      level: "warning",
      extra: %{hook: meta.hook, reason: meta.reason, session: meta.session_id})
  end

  # Compactor 压缩前后(看上下文压缩频率 + 压缩率)
  def handle_event([:cmdc, :compactor, :run, :stop], m, meta, _) do
    Metrics.gauge("cmdc.compactor.ratio", m.ratio,
      tags: [session: meta.session_id, strategy: meta.strategy])
  end

  # Checkpoint save/load 完成(snapshot_bytes / hit?)
  def handle_event([:cmdc, :checkpoint, :save, :stop], m, meta, _) do
    Metrics.histogram("cmdc.checkpoint.save.bytes", m.snapshot_bytes,
      tags: [backend: meta.backend, label: meta.label])
  end

  def handle_event([:cmdc, :checkpoint, :load, :stop], m, meta, _) do
    Metrics.counter("cmdc.checkpoint.load",
      tags: [backend: meta.backend, hit: m.hit?, outcome: meta.outcome])
  end

  # SubAgent 启停(看子代理生命周期)
  def handle_event([:cmdc, :subagent, :spawn, :start], _, meta, _) do
    Logger.info("subagent spawned: #{meta.name}",
      parent: meta.parent_session_id, sub: meta.sub_session_id)
  end

  # Hibernate 配置(init 时一次,看启用率)
  def handle_event([:cmdc, :agent, :hibernate, :configured], m, meta, _) do
    Metrics.gauge("cmdc.hibernate.enabled",
      if(meta.enabled?, do: 1, else: 0),
      tags: [session: meta.session_id, ms: m.ms])
  end

  def handle_event(_, _, _, _), do: :ok
end

# 在 application.ex 或合适入口处
MyApp.LangfuseSink.attach()

16 事件完整清单见 CMDC.Telemetry

设计说明:v0.5 新 10 事件:telemetry.execute 直接埋点,不经 EventBus 桥接, 避免 per-session 业务订阅者噪音(它们是 system-wide 可观测性事件)。 EventBus :tool_execution_metrics 业务事件保留给 Plugin / UI 用。


8. 大工具结果 0 token 占用

让一个 shell 工具返回 200KB 的 SQL 结果不再炸 LLM 上下文:

backend = CMDC.Backend.Filesystem.new(
  root_dir: "/tmp/cmdc-results",
  virtual_mode: true
)

{:ok, session} = CMDC.create_agent(
  model: "anthropic:claude-sonnet-4-5",
  tools: [CMDC.Tool.Shell, CMDC.Tool.ReadFile],
  plugins: [
    {CMDC.Plugin.Builtin.LargeResultOffload, [
      backend: backend,
      tool_token_limit_before_evict: 20_000  # ≈ 80KB chars
    ]}
  ]
)

CMDC.prompt(session, """
跑 `cat /var/log/nginx/access.log`,从输出里找 5xx 错误的 path 分布。
""")

工具返回 200KB 时 plugin 自动写到 backend /large_tool_results/<call_id>, LLM 看到的是 head + tail preview + 引导:"如需完整内容用 read_file(path, offset, limit) 分页"。


9. 内容安全双层防御

denylist 拦显式 + LLM-judge 拦语义:

{:ok, session} = CMDC.create_agent(
  model: "anthropic:claude-sonnet-4-5",
  plugins: [
    # L1 显式词拦截(快)
    {MyApp.SensitiveContentGuard, words: ["身份证号", "护照号", "信用卡"]},

    # L2 LLM-as-Judge(深度)
    {CMDC.Plugin.Builtin.ContentPolicy, [
      judge_model: "openai:gpt-4o-mini",
      judge_provider_opts: [temperature: 0.0],
      brand_keywords: ["MyProduct"]
    ]}
  ]
)

CMDC.prompt(session, "...")

10. 端到端模板:客服 Agent

把上面单点配方组合成一个完整 demo(多租户 + 工具 + 审批 + checkpoint + telemetry):

defmodule MyApp.SupportAgent do
  @behaviour CMDC.Blueprint

  @impl true
  def build(opts) do
    tenant_id = Keyword.fetch!(opts, :tenant_id)

    %CMDC.Options{
      model: "anthropic:claude-sonnet-4-5",
      system_prompt: "你是 Acme 公司的客服 agent,只回答与产品相关的问题。",
      working_dir: "/tmp/support/#{tenant_id}",
      user_data: %{tenant_id: tenant_id, user_tier: opts[:user_tier]},

      tools: [
        MyApp.QueryDB,
        MyApp.SearchKnowledgeBase,
        MyApp.RefundOrder,
        CMDC.Tool.AskUser
      ],

      plugins: [
        # 安全
        {CMDC.Plugin.Builtin.HumanApproval, [tools: ["refund_order"]]},
        {CMDC.Plugin.Builtin.ContentPolicy, []},

        # 优化
        {CMDC.Plugin.Builtin.MemoryLoader, [files: ["KB.md"]]},
        {CMDC.Plugin.Builtin.ModelRouter, [
          rules: [
            %{condition: {:user_tier, :free}, model: "openai:gpt-4o-mini"},
            %{condition: {:cost_gt, 0.10}, model: "openai:gpt-4o-mini"}
          ]
        ]},

        # 监控
        {CMDC.Plugin.Builtin.CostGuard, [max_usd: 0.50]},
        CMDC.Plugin.Builtin.EventLogger
      ],

      compactor: [trigger: {:tokens, 50_000}, keep: {:messages, 10}],
      event_buffer_size: 200  # 支持前端断网补帧
    }
  end
end

# 启动一个 session
{:ok, session} = CMDC.create_agent(MyApp.SupportAgent,
  tenant_id: "acme",
  user_tier: :pro
)

下一步

  • 升级指南 — v0.2 → v0.3 → v0.4 用户感知到的兼容边界