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

318 行
8.8 KiB
Go

// Package flow provides event-driven workflows for go-micro services.
//
// A Flow is a workflow in the sense of Anthropic's "Building Effective
// Agents": LLMs and tools orchestrated through a predefined path. It
// subscribes to a broker topic and, for each event, runs one augmented
// LLM step — the registered services as tools, a fixed prompt — and
// lets the model decide which RPCs to call. Use a Flow when the task is
// well-defined and you want a deterministic trigger; use an Agent (see
// the agent package) when the work needs to direct itself dynamically.
//
// Usage:
//
// f := flow.New("onboard-user",
// flow.Trigger("events.user.created"),
// flow.Prompt("New user created: {{.Data}}. Send welcome email and create workspace."),
// flow.Provider("anthropic"),
// flow.APIKey(key),
// )
// f.Register(service)
// service.Run()
package flow
import (
"bytes"
"context"
"encoding/json"
"fmt"
"strconv"
"sync"
"text/template"
"time"
"github.com/google/uuid"
"go-micro.dev/v5/ai"
"go-micro.dev/v5/broker"
"go-micro.dev/v5/client"
codecbytes "go-micro.dev/v5/codec/bytes"
"go-micro.dev/v5/logger"
"go-micro.dev/v5/registry"
// Register default providers.
_ "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"
)
// Flow is an event-driven LLM orchestration unit. It subscribes to
// a broker topic, discovers services as tools, and feeds each event
// into an LLM that decides which RPCs to call.
type Flow struct {
name string
opts Options
model ai.Model
toolSet *ai.Tools
client client.Client
tmpl *template.Template
log logger.Logger
checkpoint Checkpoint
reg registry.Registry
sub broker.Subscriber
registration *registry.Service
mu sync.Mutex
results []Result
}
// Result records one flow execution.
type Result struct {
FlowName string `json:"flow"`
Trigger string `json:"trigger"`
Prompt string `json:"prompt"`
Reply string `json:"reply,omitempty"`
Answer string `json:"answer,omitempty"`
ToolCalls []string `json:"tool_calls,omitempty"`
Error string `json:"error,omitempty"`
Timestamp time.Time `json:"timestamp"`
Duration float64 `json:"duration_seconds"`
}
// New creates a Flow with the given name and options.
func New(name string, opts ...Option) *Flow {
o := Options{
Provider: "openai",
SystemPrompt: "You are a service orchestrator. Use the available tools to fulfill the request. Explain what you do.",
HistoryLimit: 20,
}
for _, opt := range opts {
opt(&o)
}
var tmpl *template.Template
if o.Prompt != "" {
var err error
tmpl, err = template.New(name).Parse(o.Prompt)
if err != nil {
tmpl = template.Must(template.New(name).Parse("{{.Data}}"))
}
}
return &Flow{
name: name,
opts: o,
tmpl: tmpl,
log: logger.DefaultLogger,
checkpoint: defaultCheckpoint(name, o),
}
}
// Register wires the flow into a running service. It sets up the
// model, discovers tools from the registry, and subscribes to the
// trigger topic on the broker. Call this before service.Run().
func (f *Flow) Register(reg registry.Registry, br broker.Broker, cl client.Client) error {
f.client = cl
f.reg = reg
f.toolSet = ai.NewTools(reg, ai.ToolClient(cl))
// A flow that dispatches to an agent doesn't run its own model — the
// agent is the engine. Otherwise, set up the augmented LLM.
if f.opts.Agent == "" {
var modelOpts []ai.Option
if f.opts.APIKey != "" {
modelOpts = append(modelOpts, ai.WithAPIKey(f.opts.APIKey))
}
if f.opts.Model != "" {
modelOpts = append(modelOpts, ai.WithModel(f.opts.Model))
}
if f.opts.BaseURL != "" {
modelOpts = append(modelOpts, ai.WithBaseURL(f.opts.BaseURL))
}
modelOpts = append(modelOpts, ai.WithTools(f.toolSet))
f.model = ai.New(f.opts.Provider, modelOpts...)
if f.model == nil {
return fmt.Errorf("unknown provider: %s", f.opts.Provider)
}
}
if f.opts.TriggerTopic != "" {
sub, err := br.Subscribe(f.opts.TriggerTopic, func(p broker.Event) error {
data := string(p.Message().Body)
if err := f.Execute(context.Background(), data); err != nil {
f.log.Logf(logger.ErrorLevel, "Flow %s failed: %v", f.name, err)
}
return nil
})
if err != nil {
return fmt.Errorf("subscribe to %s: %w", f.opts.TriggerTopic, err)
}
f.sub = sub
f.log.Logf(logger.InfoLevel, "Flow %s subscribed to %s", f.name, f.opts.TriggerTopic)
// Announce the flow in the registry so it's discoverable like a
// service or agent (e.g. `micro flow list`). This is liveness only:
// Stop deregisters it. Durable run history lives in the store.
f.registration = &registry.Service{
Name: f.name,
Version: "latest",
Metadata: map[string]string{
"type": "flow",
"trigger": f.opts.TriggerTopic,
"steps": strconv.Itoa(len(f.opts.Steps)),
},
Nodes: []*registry.Node{{
Id: f.name + "-" + uuid.New().String()[:8],
Address: "flow://" + f.name,
Metadata: map[string]string{"type": "flow"},
}},
}
if err := reg.Register(f.registration); err != nil {
f.log.Logf(logger.ErrorLevel, "Flow %s registry register: %v", f.name, err)
f.registration = nil
}
}
return nil
}
// Stop unsubscribes the flow from its trigger and deregisters it from the
// registry. In-flight and past runs remain in the store; Stop only ends
// the flow's liveness, mirroring how a service leaves the registry when
// it shuts down.
func (f *Flow) Stop() error {
if f.sub != nil {
_ = f.sub.Unsubscribe()
f.sub = nil
}
if f.registration != nil && f.reg != nil {
err := f.reg.Deregister(f.registration)
f.registration = nil
return err
}
return nil
}
// Execute runs the flow once with the given input data. This is
// called automatically on each broker event, but can also be
// invoked directly for testing or one-shot use.
func (f *Flow) Execute(ctx context.Context, data string) error {
// Stepped flows run the ordered, checkpointed step loop.
if len(f.opts.Steps) > 0 {
_, err := f.startRun(ctx, data)
return err
}
start := time.Now()
prompt := data
if f.tmpl != nil {
var buf bytes.Buffer
f.tmpl.Execute(&buf, map[string]string{"Data": data})
prompt = buf.String()
}
result := Result{
FlowName: f.name,
Trigger: f.opts.TriggerTopic,
Prompt: prompt,
Timestamp: start,
}
// Flow triggers, Agent reasons: hand the event to the named agent.
if f.opts.Agent != "" {
reply, err := f.callAgent(ctx, f.opts.Agent, prompt)
result.Duration = time.Since(start).Seconds()
if err != nil {
result.Error = err.Error()
f.record(result)
return err
}
result.Reply = reply
f.record(result)
f.log.Logf(logger.InfoLevel, "Flow %s dispatched to agent %s in %.1fs",
f.name, f.opts.Agent, result.Duration)
return nil
}
// Otherwise run a single augmented-LLM step with the services as tools.
discovered, err := f.toolSet.Discover()
if err != nil {
result.Duration = time.Since(start).Seconds()
result.Error = err.Error()
f.record(result)
return fmt.Errorf("discover tools: %w", err)
}
resp, err := f.model.Generate(ctx, &ai.Request{
Prompt: prompt,
SystemPrompt: f.opts.SystemPrompt,
Tools: discovered,
})
result.Duration = time.Since(start).Seconds()
if err != nil {
result.Error = err.Error()
f.record(result)
return err
}
result.Reply = resp.Reply
result.Answer = resp.Answer
for _, tc := range resp.ToolCalls {
args, _ := json.Marshal(tc.Input)
result.ToolCalls = append(result.ToolCalls, fmt.Sprintf("%s(%s)", tc.Name, args))
}
f.record(result)
f.log.Logf(logger.InfoLevel, "Flow %s completed in %.1fs: %d tool calls",
f.name, result.Duration, len(result.ToolCalls))
return nil
}
// callAgent hands the rendered prompt to a registered agent's Agent.Chat
// endpoint over RPC and returns its reply.
func (f *Flow) callAgent(ctx context.Context, name, message string) (string, error) {
body, _ := json.Marshal(map[string]string{"message": message})
req := f.client.NewRequest(name, "Agent.Chat", &codecbytes.Frame{Data: body})
var rsp codecbytes.Frame
if err := f.client.Call(ctx, req, &rsp); err != nil {
return "", err
}
var out struct {
Reply string `json:"reply"`
}
if err := json.Unmarshal(rsp.Data, &out); err != nil {
return "", err
}
return out.Reply, nil
}
// Results returns a copy of all recorded execution results.
func (f *Flow) Results() []Result {
f.mu.Lock()
defer f.mu.Unlock()
out := make([]Result, len(f.results))
copy(out, f.results)
return out
}
// Name returns the flow name.
func (f *Flow) Name() string {
return f.name
}
func (f *Flow) record(r Result) {
f.mu.Lock()
f.results = append(f.results, r)
f.mu.Unlock()
if f.opts.OnResult != nil {
f.opts.OnResult(r)
}
}