项目文件夹

文件
T
Asim Aslam 9fdcc24cce Implement durable execution and scoped state management for flows (#2972)
* docs: design note for flow steps + Checkpoint durable execution

* docs: fold in durable-execution decisions (State struct, single Step, run retention, retry)

* docs: rename State.Payload to State.Data

* feat(flow): ordered steps + Checkpoint durable execution

A flow can now be an ordered list of steps (a task with stages) instead
of a single LLM turn. State carries typed Data plus a Stage marker; each
step is checkpointed before and after via a pluggable Checkpoint
(store-backed by default), so a run survives a crash and resumes where it
stopped without re-running completed steps. Flow-level Retry with a
per-step override; runs retained for audit unless DeleteOnSuccess.

Step actions: Call (RPC), LLM (augmented turn), Dispatch (to an agent),
or any StepFunc. Single-step and agent-dispatch flows are unchanged.

* feat(flow): top-level re-exports + durable flow example

Expose the step/checkpoint API from the micro package (FlowSteps,
FlowStep, FlowState, FlowRetry, FlowWithCheckpoint, FlowCall/LLM/Dispatch,
Checkpoint, StoreCheckpoint) and add a runnable, key-free example
demonstrating crash + resume.

* docs: document durable flow steps (guide, README, CLI help)

* docs: blog post + changelog for durable workflows

* fix(flow): scope checkpoint keys by flow name (flow/{name}/runs/{id})

Run keys were flow/runs/{id} — a single global keyspace shared by every
flow on the default store. Namespace them by flow name so each flow's
state is kept apart. StoreCheckpoint now takes a scope argument (the flow
passes its name by default).

* feat(store): Scope handle; scope agent and flow state by name

Add store.Scope(s, database, table) — a store handle that confines every
operation to a database/table without mutating the shared store, so
co-located components don't clobber each other's table (the failure mode
of the global Init(Table(...)) approach).

Use it to keep each agent's memory and plan in its own table
(agent/{name}) and each flow's runs in its own (flow/{name}), instead of
one global table partitioned only by key prefix. Services already scope
by service name.

* feat: consistent state model — service store scoping, flow registry, list/history CLI

- service: scope store via store.Scope (database service / table name),
  retiring the Init(store.Table(name)) global-mutation hack; bridge the
  default store so handlers using store.DefaultStore stay isolated.
- flow: register in the registry as type=flow while running (with trigger
  and step count), deregister on Stop. Live discovery, like agents.
- cli: micro flow list (registry), micro flow runs <name> (durable store),
  micro agent history <name> (durable store). list = running, runs/history
  = durable, mirroring the service model.

* test: mini-universe end-to-end harness + scheduled GitHub Action

internal/harness/universe boots a small but real go-micro world — four
services, a durable checkout flow that crashes at payment and resumes,
and a guardrailed agent with a tool wrapper reached over RPC — drives the
scenario, asserts the end state (10 checks), and shuts down. Everything
is real except the LLM (mocked), so it's deterministic and needs no key;
-provider anthropic runs it live. Exits non-zero on failure, so it's an
end-to-end test, not just a demo.

Adds .github/workflows/universe.yml (push/PR/daily/dispatch) running the
universe + existing harnesses on the mock provider, plus an opt-in job
that runs live when ANTHROPIC_API_KEY is set. 'make harness' runs them
locally.

* ci: run the live universe job against AtlasCloud (ATLASCLOUD_API_KEY)

* ci: run the live universe job only on schedule or manual dispatch

The deterministic mock job still runs on push/PR/daily; the live
(AtlasCloud) job runs daily and on manual workflow_dispatch only, so
changes don't burn API credits on every PR but can still be checked
against a real model on demand.

---------

Co-authored-by: Claude <noreply@anthropic.com>
2026-06-17 12:02:47 +01:00

313 行
7.5 KiB
Go

// Package agent provides the Agent abstraction for Go Micro.
//
// An Agent is a service with an LLM inside it. It registers a Chat
// RPC endpoint, discovers its assigned services' tools, and
// orchestrates them intelligently.
//
// agent := micro.NewAgent("task-mgr",
// micro.AgentServices("task"),
// micro.AgentPrompt("You manage tasks."),
// micro.AgentProvider("anthropic"),
// )
// agent.Run()
package agent
import (
"context"
"encoding/json"
"fmt"
"strings"
"sync"
pb "go-micro.dev/v5/agent/proto"
"go-micro.dev/v5/ai"
"go-micro.dev/v5/server"
"go-micro.dev/v5/store"
_ "go-micro.dev/v5/ai/anthropic"
_ "go-micro.dev/v5/ai/atlascloud"
_ "go-micro.dev/v5/ai/gemini"
_ "go-micro.dev/v5/ai/groq"
_ "go-micro.dev/v5/ai/mistral"
_ "go-micro.dev/v5/ai/openai"
_ "go-micro.dev/v5/ai/together"
)
// Agent is the interface for an AI agent that manages services.
type Agent interface {
Name() string
Init(...Option)
Options() Options
Ask(ctx context.Context, message string) (*Response, error)
Run() error
Stop() error
String() string
}
// Response is what an agent returns from Chat.
type Response struct {
Reply string
ToolCalls []ai.ToolCall
Agent string
}
type agentImpl struct {
opts Options
model ai.Model
tools *ai.Tools
mem Memory
server server.Server
mu sync.Mutex
// ephemeral marks a short-lived sub-agent created by delegation.
// Ephemeral agents run with an isolated context: they load and
// persist no history, and have no built-in tools (so they cannot
// plan or re-delegate).
ephemeral bool
// steps counts tool executions in the current Ask, for MaxSteps.
steps int
// calls counts identical tool calls (name+args) in the current Ask,
// for LoopLimit.
calls map[string]int
}
// New creates a new Agent.
func New(opts ...Option) Agent {
return &agentImpl{
opts: newOptions(opts...),
}
}
// newEphemeral creates a short-lived sub-agent for a delegated subtask.
// It shares the parent's provider, model, and infrastructure but runs
// with an isolated context: it loads and persists no history and has no
// built-in tools (so it can neither plan nor re-delegate). Returns the
// concrete type because ephemeral is an internal construction detail,
// not a public option.
func newEphemeral(opts ...Option) *agentImpl {
return &agentImpl{
opts: newOptions(opts...),
ephemeral: true,
}
}
func (a *agentImpl) Name() string {
return a.opts.Name
}
func (a *agentImpl) Init(opts ...Option) {
for _, o := range opts {
o(&a.opts)
}
a.setup()
}
func (a *agentImpl) Options() Options {
return a.opts
}
func (a *agentImpl) String() string {
return "agent"
}
func (a *agentImpl) setup() {
var modelOpts []ai.Option
modelOpts = append(modelOpts, ai.WithAPIKey(a.opts.APIKey))
if a.opts.Model != "" {
modelOpts = append(modelOpts, ai.WithModel(a.opts.Model))
}
a.tools = ai.NewTools(a.opts.Registry, ai.ToolClient(a.opts.Client))
modelOpts = append(modelOpts, ai.WithToolHandler(a.toolHandler()))
a.model = ai.New(a.opts.Provider, modelOpts...)
// Memory is pluggable. Use the configured one, otherwise the default
// store-backed memory — except ephemeral sub-agents, which keep an
// isolated, non-persistent context.
switch {
case a.opts.Memory != nil:
a.mem = a.opts.Memory
case a.ephemeral:
a.mem = NewInMemory(a.opts.HistoryLimit)
default:
a.mem = NewMemory(a.stateStore(), "history", a.opts.HistoryLimit)
}
}
// stateStore returns the agent's own state store, scoped to its name so
// memory and plan live in their own table ("agent/{name}") rather than a
// shared global one. The scoped handle injects the database/table per
// operation without mutating the underlying store.
func (a *agentImpl) stateStore() store.Store {
s := a.opts.Store
if s == nil {
s = store.DefaultStore
}
return store.Scope(s, "agent", a.opts.Name)
}
// Ask sends a message and returns the agent's response.
// This is the programmatic API for direct use.
func (a *agentImpl) Ask(ctx context.Context, message string) (*Response, error) {
a.mu.Lock()
defer a.mu.Unlock()
if a.model == nil {
a.setup()
}
toolList, err := a.discoverTools()
if err != nil {
return nil, fmt.Errorf("discover tools: %w", err)
}
a.mem.Add("user", message)
a.steps = 0
a.calls = map[string]int{}
resp, err := a.model.Generate(ctx, &ai.Request{
Prompt: message,
SystemPrompt: a.buildPrompt(),
Tools: toolList,
Messages: a.mem.Messages(),
})
if err != nil {
return nil, err
}
if resp.Reply != "" {
a.mem.Add("assistant", resp.Reply)
}
if resp.Answer != "" {
a.mem.Add("assistant", resp.Answer)
}
reply := resp.Reply
if resp.Answer != "" {
if reply != "" {
reply += "\n\n"
}
reply += resp.Answer
}
return &Response{
Reply: reply,
ToolCalls: resp.ToolCalls,
Agent: a.opts.Name,
}, nil
}
// Chat implements the proto AgentHandler interface for RPC.
// @example {"message": "What tasks are overdue?"}
func (a *agentImpl) Chat(ctx context.Context, req *pb.ChatRequest, rsp *pb.ChatResponse) error {
resp, err := a.Ask(ctx, req.Message)
if err != nil {
return err
}
rsp.Reply = resp.Reply
rsp.Agent = resp.Agent
for _, tc := range resp.ToolCalls {
input, _ := json.Marshal(tc.Input)
rsp.ToolCalls = append(rsp.ToolCalls, &pb.ToolCall{
Id: tc.ID,
Name: tc.Name,
Input: string(input),
Result: tc.Result,
})
}
return nil
}
// Run starts the agent as a service with a Chat RPC endpoint.
func (a *agentImpl) Run() error {
if a.model == nil {
a.setup()
}
a.server = server.NewServer(
server.Name(a.opts.Name),
server.Registry(a.opts.Registry),
server.Metadata(map[string]string{
"type": "agent",
"services": strings.Join(a.opts.Services, ","),
}),
)
pb.RegisterAgentHandler(a.server, a)
if err := a.server.Start(); err != nil {
return fmt.Errorf("failed to start agent: %w", err)
}
fmt.Printf("Agent %s registered (manages: %s)\n", a.opts.Name, strings.Join(a.opts.Services, ", "))
ch := make(chan struct{})
<-ch
return nil
}
func (a *agentImpl) Stop() error {
if a.server != nil {
return a.server.Stop()
}
return nil
}
func (a *agentImpl) discoverTools() ([]ai.Tool, error) {
all, err := a.tools.Discover()
if err != nil {
return nil, err
}
var scoped []ai.Tool
for _, t := range all {
if strings.HasPrefix(t.OriginalName, a.opts.Name+".") {
continue
}
if len(a.opts.Services) == 0 {
scoped = append(scoped, t)
continue
}
for _, svc := range a.opts.Services {
if strings.HasPrefix(t.OriginalName, svc+".") {
scoped = append(scoped, t)
break
}
}
}
// Developer-registered custom tools (WithTool).
for i := range a.opts.tools {
scoped = append(scoped, a.opts.tools[i].def)
}
// Expose the agent's own capabilities (plan, delegate) as tools.
// Ephemeral sub-agents don't get them.
if !a.ephemeral {
scoped = append(scoped, builtinTools()...)
}
return scoped, nil
}
func (a *agentImpl) buildPrompt() string {
var base string
switch {
case a.opts.Prompt != "":
base = a.opts.Prompt
case len(a.opts.Services) > 0:
base = fmt.Sprintf("You are the %s agent. You manage these services: %s. Use the available tools to fulfill requests.",
a.opts.Name, strings.Join(a.opts.Services, ", "))
default:
base = fmt.Sprintf("You are the %s agent. Use the available tools to fulfill requests.", a.opts.Name)
}
// Keep the agent oriented: surface its saved plan, if any.
if !a.ephemeral {
if plan := a.loadPlan(); plan != "" {
base += "\n\nYour current plan (update it with the plan tool as you make progress):\n" + plan
}
}
return base
}