项目文件夹

文件
T
Asim Aslam ca87efef2f
goreleaser / goreleaser (push) Has been cancelled
feat(agent): expose run metadata + structured guardrail reasons to tool wrappers (#2981)
Closes the remaining ask in #2980 without adding a parallel callback API.
ToolResult.Refused tags a guardrail block with a reason (ai.RefusedLoop /
RefusedMaxSteps / RefusedApproval) so a wrapper can switch on it instead of
parsing the message. ai.RunInfo (RunID, ParentID, Agent) rides on the
context passed to the tool handler, giving wrappers run correlation and
delegation lineage. Before/after/retry/failure were already covered by
AgentWrapTool; this adds the metadata. Docs + tests included.

Co-authored-by: Claude <noreply@anthropic.com>
2026-06-19 08:09:22 +01:00

341 行
12 KiB
Go

package agent
import (
"context"
"encoding/json"
"fmt"
"strings"
"go-micro.dev/v6/ai"
codecBytes "go-micro.dev/v6/codec/bytes"
"go-micro.dev/v6/gateway/a2a"
"go-micro.dev/v6/store"
)
// Built-in agent tools. These are not service endpoints — they are
// capabilities the agent has over itself: maintaining a plan in its
// memory, and delegating a subtask to another agent.
//
// They are plain tools, wired into the agent's tool handler alongside
// the discovered service tools. There is no separate harness or graph:
// the LLM calls them like any other tool.
const (
toolPlan = "plan"
toolDelegate = "delegate"
)
// builtinTools returns the tool definitions exposed to the model in
// addition to the agent's scoped service tools.
func builtinTools() []ai.Tool {
return []ai.Tool{
{
Name: toolPlan,
OriginalName: toolPlan,
Description: "Record or update your plan as an ordered list of steps before doing multi-step work. " +
"Call this whenever the plan changes. The plan is saved to your memory and shown back to you on later turns.",
Properties: map[string]any{
"steps": map[string]any{
"type": "array",
"description": "Ordered plan steps. Each step has a 'task' (string) and a " +
"'status' (one of: pending, in_progress, done).",
},
},
},
{
Name: toolDelegate,
OriginalName: toolDelegate,
Description: "Delegate a self-contained subtask to another agent. If 'to' names an agent that already " +
"manages the relevant services, that agent handles it; otherwise a focused sub-agent is created for the " +
"subtask. The sub-agent works in an isolated context and returns only its result. Use this to keep your " +
"own context focused and to let domain experts handle their own services.",
Properties: map[string]any{
"task": map[string]any{
"type": "string",
"description": "The subtask to delegate, described completely and self-contained.",
},
"to": map[string]any{
"type": "string",
"description": "Optional. The agent or service name best suited to the subtask, or the URL of an external agent that speaks the A2A protocol.",
},
},
},
}
}
// Builtins returns the built-in agent tools (plan, delegate) together
// with a handler for them, so the same capabilities can be wired into a
// tool loop that isn't a running Agent — for example the `micro chat`
// fallback. The handler's third return value is false when the name is
// not a built-in, so callers can fall through to their own tools.
//
// Configure it with the same options as an Agent (Name, Provider,
// WithStore, WithRegistry, WithClient, ...); these back plan's memory
// and delegate's RPC/sub-agent behaviour.
func Builtins(opts ...Option) (tools []ai.Tool, handle func(name string, input map[string]any) (result any, content string, ok bool)) {
a := &agentImpl{opts: newOptions(opts...)}
handle = func(name string, input map[string]any) (any, string, bool) {
switch name {
case toolPlan:
r := a.handlePlan(ai.ToolCall{Name: name, Input: input})
return r.Value, r.Content, true
case toolDelegate:
r := a.handleDelegate(ai.ToolCall{Name: name, Input: input})
return r.Value, r.Content, true
}
return nil, "", false
}
return builtinTools(), handle
}
// toolHandler returns the agent's tool-call handler, composed as a stack
// of wrappers around a base handler — the same middleware shape as
// client/server wrappers. The base executes the call (custom tools,
// delegate, or RPC); the built-in guardrails wrap it; developer wrappers
// (WrapTool) wrap those, outermost, so they observe every call and its
// result including guardrail refusals. Ephemeral sub-agents get the bare
// service handler so they can neither plan nor re-delegate (which
// prevents runaway recursion).
func (a *agentImpl) toolHandler() ai.ToolHandler {
if a.ephemeral {
return a.tools.Handler()
}
// Innermost first: base, then guardrails (approve → loop → step →
// plan), then developer wrappers outermost. Wrapping reverses order,
// so the result runs plan → step → loop → approve → base.
h := a.baseHandler()
h = a.approveWrap(h)
h = a.loopWrap(h)
h = a.stepWrap(h)
h = a.planWrap(h)
for i := len(a.opts.wrappers) - 1; i >= 0; i-- {
h = a.opts.wrappers[i](h)
}
return h
}
// baseHandler executes a tool call: a developer custom tool, the built-in
// delegate, or an RPC to the service. It is the innermost handler.
func (a *agentImpl) baseHandler() ai.ToolHandler {
rpc := a.tools.Handler()
return func(ctx context.Context, call ai.ToolCall) ai.ToolResult {
for i := range a.opts.tools {
if a.opts.tools[i].def.Name == call.Name {
out, err := a.opts.tools[i].handler(ctx, call.Input)
if err != nil {
return errResult(call.ID, err.Error())
}
return ai.ToolResult{ID: call.ID, Value: out, Content: out}
}
}
if call.Name == toolDelegate {
return a.handleDelegate(call)
}
return rpc(ctx, call)
}
}
// planWrap handles the plan tool inline. plan is internal bookkeeping,
// not an action — it is never counted, loop-checked, or gated.
func (a *agentImpl) planWrap(next ai.ToolHandler) ai.ToolHandler {
return func(ctx context.Context, call ai.ToolCall) ai.ToolResult {
if call.Name == toolPlan {
return a.handlePlan(call)
}
return next(ctx, call)
}
}
// stepWrap bounds the number of actions per Ask (MaxSteps).
func (a *agentImpl) stepWrap(next ai.ToolHandler) ai.ToolHandler {
return func(ctx context.Context, call ai.ToolCall) ai.ToolResult {
if a.opts.MaxSteps > 0 {
a.steps++
if a.steps > a.opts.MaxSteps {
return refused(call.ID, ai.RefusedMaxSteps, fmt.Sprintf(
"step limit reached (%d). Do not call any more tools; stop and summarize what you have so far.",
a.opts.MaxSteps))
}
}
return next(ctx, call)
}
}
// loopWrap stops the agent repeating an identical action that makes no
// progress (which the step count alone won't catch).
func (a *agentImpl) loopWrap(next ai.ToolHandler) ai.ToolHandler {
return func(ctx context.Context, call ai.ToolCall) ai.ToolResult {
if a.opts.LoopLimit > 0 {
if a.calls == nil {
a.calls = map[string]int{}
}
args, _ := json.Marshal(call.Input)
fp := call.Name + ":" + string(args)
a.calls[fp]++
if a.calls[fp] > a.opts.LoopLimit {
return refused(call.ID, ai.RefusedLoop, fmt.Sprintf(
"loop detected: you have already called %q with the same arguments %d times and the result will not change. Stop repeating it — try a different approach, or finish with what you have.",
call.Name, a.opts.LoopLimit))
}
}
return next(ctx, call)
}
}
// approveWrap gates each action before it runs (ApproveTool).
func (a *agentImpl) approveWrap(next ai.ToolHandler) ai.ToolHandler {
return func(ctx context.Context, call ai.ToolCall) ai.ToolResult {
if a.opts.Approve != nil {
if ok, reason := a.opts.Approve(call.Name, call.Input); !ok {
msg := "tool call was not approved"
if reason != "" {
msg += ": " + reason
}
return refused(call.ID, ai.RefusedApproval, msg)
}
}
return next(ctx, call)
}
}
// handlePlan persists the supplied plan to the agent's memory and
// echoes it back so the model can see the stored state.
func (a *agentImpl) handlePlan(call ai.ToolCall) ai.ToolResult {
data, err := json.Marshal(call.Input)
if err != nil {
return errResult(call.ID, "invalid plan: "+err.Error())
}
a.stateStore().Write(&store.Record{Key: planKey, Value: data})
return ai.ToolResult{ID: call.ID, Value: call.Input, Content: string(data)}
}
// handleDelegate hands a subtask to another agent. Delegate-first:
// if 'to' names a registered agent, it is called via RPC. Otherwise an
// ephemeral sub-agent is created with a fresh, isolated context, asked
// the subtask, and its reply returned.
func (a *agentImpl) handleDelegate(call ai.ToolCall) ai.ToolResult {
input := call.Input
task, _ := input["task"].(string)
if task == "" {
return errResult(call.ID, "task is required")
}
to, _ := input["to"].(string)
// An external agent on another framework, addressed by A2A URL.
if strings.HasPrefix(to, "http://") || strings.HasPrefix(to, "https://") {
reply, err := a2a.NewClient(to).Send(context.Background(), task)
if err != nil {
return errResult(call.ID, "delegate to A2A agent "+to+": "+err.Error())
}
out := map[string]any{"agent": to, "reply": reply}
b, _ := json.Marshal(out)
return ai.ToolResult{ID: call.ID, Value: out, Content: string(b)}
}
// Delegate-first: an existing agent that owns the domain handles it.
if to != "" && a.isAgent(to) {
reply, err := a.callAgentRPC(context.Background(), to, task)
if err != nil {
return errResult(call.ID, "delegate to agent "+to+": "+err.Error())
}
out := map[string]any{"agent": to, "reply": reply}
b, _ := json.Marshal(out)
return ai.ToolResult{ID: call.ID, Value: out, Content: string(b)}
}
// Otherwise create a focused, ephemeral sub-agent. Fresh context:
// it loads no history and persists none.
var svcs []string
if to != "" {
svcs = []string{to}
}
sub := newEphemeral(
Name(a.opts.Name+".sub"),
Services(svcs...),
Prompt("You are a sub-agent handling a single delegated subtask. "+
"Complete it using the available tools and report the result concisely."),
Provider(a.opts.Provider),
Model(a.opts.Model),
APIKey(a.opts.APIKey),
WithRegistry(a.opts.Registry),
WithClient(a.opts.Client),
WithStore(a.opts.Store),
)
// Record lineage so the sub-agent's tool calls carry this run as parent.
sub.parentRunID = a.runID
resp, err := sub.Ask(context.Background(), task)
if err != nil {
return errResult(call.ID, "sub-agent: "+err.Error())
}
out := map[string]any{"reply": resp.Reply}
b, _ := json.Marshal(out)
return ai.ToolResult{ID: call.ID, Value: out, Content: string(b)}
}
// isAgent reports whether name resolves to a registered agent (a
// service advertising type=agent in its metadata).
func (a *agentImpl) isAgent(name string) bool {
if a.opts.Registry == nil {
return false
}
recs, err := a.opts.Registry.GetService(name)
if err != nil || len(recs) == 0 {
return false
}
if recs[0].Metadata != nil && recs[0].Metadata["type"] == "agent" {
return true
}
for _, n := range recs[0].Nodes {
if n.Metadata != nil && n.Metadata["type"] == "agent" {
return true
}
}
return false
}
// callAgentRPC calls another agent's Agent.Chat endpoint and returns
// its reply.
func (a *agentImpl) callAgentRPC(ctx context.Context, name, msg string) (string, error) {
body, _ := json.Marshal(map[string]string{"message": msg})
req := a.opts.Client.NewRequest(name, "Agent.Chat", &codecBytes.Frame{Data: body})
var rsp codecBytes.Frame
if err := a.opts.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
}
// planKey is the record key for an agent's plan within its scoped store.
const planKey = "plan"
// loadPlan returns the stored plan as a JSON string, or "" if none.
func (a *agentImpl) loadPlan() string {
recs, err := a.stateStore().Read(planKey)
if err != nil || len(recs) == 0 {
return ""
}
return string(recs[0].Value)
}
func errResult(id, msg string) ai.ToolResult {
m := map[string]string{"error": msg}
b, _ := json.Marshal(m)
return ai.ToolResult{ID: id, Value: m, Content: string(b)}
}
// refused is an error result a guardrail returns, tagged with a structured
// reason (ai.Refused*) so a tool wrapper can react to it without parsing
// the message.
func refused(id, reason, msg string) ai.ToolResult {
r := errResult(id, msg)
r.Refused = reason
return r
}