Files
go-llm/v2/agent/agent_test.go
Steve Dudenhoeffer 87ec56a2be
All checks were successful
CI / Lint (push) Successful in 9m46s
CI / V2 Module (push) Successful in 12m5s
CI / Root Module (push) Successful in 12m6s
Add agent sub-package for composable LLM agents
Introduces v2/agent with a minimal API: Agent, New(), Run(), and AsTool().
Agents wrap a model + system prompt + tools. AsTool() turns an agent into
a llm.Tool, enabling parent agents to delegate to sub-agents through the
normal tool-call loop — no channels, pools, or orchestration needed.

Also exports NewClient(provider.Provider) for custom provider integration.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-07 23:17:19 -05:00

245 lines
7.1 KiB
Go

package agent
import (
"context"
"errors"
"sync"
"testing"
llm "gitea.stevedudenhoeffer.com/steve/go-llm/v2"
"gitea.stevedudenhoeffer.com/steve/go-llm/v2/provider"
)
// mockProvider is a test helper that implements provider.Provider.
type mockProvider struct {
mu sync.Mutex
completeFunc func(ctx context.Context, req provider.Request) (provider.Response, error)
requests []provider.Request
}
func (m *mockProvider) Complete(ctx context.Context, req provider.Request) (provider.Response, error) {
m.mu.Lock()
m.requests = append(m.requests, req)
m.mu.Unlock()
return m.completeFunc(ctx, req)
}
func (m *mockProvider) Stream(ctx context.Context, req provider.Request, events chan<- provider.StreamEvent) error {
close(events)
return nil
}
func (m *mockProvider) lastRequest() provider.Request {
m.mu.Lock()
defer m.mu.Unlock()
if len(m.requests) == 0 {
return provider.Request{}
}
return m.requests[len(m.requests)-1]
}
func newMockModel(fn func(ctx context.Context, req provider.Request) (provider.Response, error)) *llm.Model {
mp := &mockProvider{completeFunc: fn}
return llm.NewClient(mp).Model("mock-model")
}
func newSimpleMockModel(text string) *llm.Model {
return newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
return provider.Response{Text: text}, nil
})
}
func TestAgent_Run(t *testing.T) {
model := newSimpleMockModel("Hello from agent!")
a := New(model, "You are a helpful assistant.")
result, err := a.Run(context.Background(), "Say hello")
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
if result != "Hello from agent!" {
t.Errorf("expected 'Hello from agent!', got %q", result)
}
}
func TestAgent_Run_WithTools(t *testing.T) {
callCount := 0
model := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
callCount++
if callCount == 1 {
// First call: model requests a tool call
return provider.Response{
ToolCalls: []provider.ToolCall{
{ID: "tc1", Name: "greet", Arguments: `{}`},
},
}, nil
}
// Second call: model returns text after seeing tool result
return provider.Response{Text: "Tool said: hello!"}, nil
})
tool := llm.DefineSimple("greet", "Says hello", func(ctx context.Context) (string, error) {
return "hello!", nil
})
a := New(model, "You are helpful.", WithTools(llm.NewToolBox(tool)))
result, err := a.Run(context.Background(), "Use the greet tool")
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
if result != "Tool said: hello!" {
t.Errorf("expected 'Tool said: hello!', got %q", result)
}
if callCount != 2 {
t.Errorf("expected 2 calls (tool loop), got %d", callCount)
}
}
func TestAgent_AsTool(t *testing.T) {
// Create a child agent
childModel := newSimpleMockModel("child result: 42")
child := New(childModel, "You compute things.")
// Create the tool from the child agent
childTool := child.AsTool("compute", "Delegate computation to child agent")
// Verify tool metadata
if childTool.Name != "compute" {
t.Errorf("expected tool name 'compute', got %q", childTool.Name)
}
if childTool.Description != "Delegate computation to child agent" {
t.Errorf("expected correct description, got %q", childTool.Description)
}
// Execute the tool directly (simulating what the parent's Chat.Send loop does)
result, err := childTool.Execute(context.Background(), `{"input":"what is 6*7?"}`)
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
if result != "child result: 42" {
t.Errorf("expected 'child result: 42', got %q", result)
}
}
func TestAgent_AsTool_ParentChild(t *testing.T) {
// Child agent that always returns a fixed result
childModel := newSimpleMockModel("researched: Go generics are great")
child := New(childModel, "You are a researcher.")
// Parent agent: first call returns tool call, second returns text
parentCallCount := 0
parentModel := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
parentCallCount++
if parentCallCount == 1 {
return provider.Response{
ToolCalls: []provider.ToolCall{
{ID: "tc1", Name: "research", Arguments: `{"input":"Tell me about Go generics"}`},
},
}, nil
}
// After getting tool result, parent synthesizes final answer
return provider.Response{Text: "Based on research: Go generics are great"}, nil
})
parent := New(parentModel, "You coordinate tasks.",
WithTools(llm.NewToolBox(
child.AsTool("research", "Research a topic"),
)),
)
result, err := parent.Run(context.Background(), "Tell me about Go generics")
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
if result != "Based on research: Go generics are great" {
t.Errorf("expected synthesized result, got %q", result)
}
if parentCallCount != 2 {
t.Errorf("expected 2 parent calls (tool loop), got %d", parentCallCount)
}
}
func TestAgent_RunMessages(t *testing.T) {
model := newSimpleMockModel("I see the system and user messages")
a := New(model, "You are helpful.")
messages := []llm.Message{
llm.UserMessage("First question"),
llm.AssistantMessage("First answer"),
llm.UserMessage("Follow up"),
}
result, err := a.RunMessages(context.Background(), messages)
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
if result != "I see the system and user messages" {
t.Errorf("unexpected result: %q", result)
}
}
func TestAgent_ContextCancellation(t *testing.T) {
model := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
return provider.Response{}, ctx.Err()
})
a := New(model, "You are helpful.")
ctx, cancel := context.WithCancel(context.Background())
cancel() // Cancel immediately
_, err := a.Run(ctx, "This should fail")
if err == nil {
t.Fatal("expected error from cancelled context")
}
}
func TestAgent_WithRequestOptions(t *testing.T) {
var capturedReq provider.Request
model := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
capturedReq = req
return provider.Response{Text: "ok"}, nil
})
a := New(model, "You are helpful.",
WithRequestOptions(llm.WithTemperature(0.3), llm.WithMaxTokens(100)),
)
_, err := a.Run(context.Background(), "test")
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
if capturedReq.Temperature == nil || *capturedReq.Temperature != 0.3 {
t.Errorf("expected temperature 0.3, got %v", capturedReq.Temperature)
}
if capturedReq.MaxTokens == nil || *capturedReq.MaxTokens != 100 {
t.Errorf("expected maxTokens 100, got %v", capturedReq.MaxTokens)
}
}
func TestAgent_Run_Error(t *testing.T) {
wantErr := errors.New("model failed")
model := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
return provider.Response{}, wantErr
})
a := New(model, "You are helpful.")
_, err := a.Run(context.Background(), "test")
if err == nil {
t.Fatal("expected error, got nil")
}
}
func TestAgent_EmptySystem(t *testing.T) {
model := newSimpleMockModel("no system prompt")
a := New(model, "") // Empty system prompt
result, err := a.Run(context.Background(), "test")
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
if result != "no system prompt" {
t.Errorf("unexpected result: %q", result)
}
}