feat: complete v0.3.1 agent workflow

This commit is contained in:
loveuer
2026-09-15 03:39:22 -07:00
parent 47671f3098
commit 065b30a84c
33 changed files with 4921 additions and 170 deletions
+286
View File
@@ -3,8 +3,13 @@ package agent
import (
"context"
"encoding/json"
"errors"
"fmt"
"os"
"path/filepath"
"reflect"
"strings"
"sync"
"testing"
"agentu/internal/tools"
@@ -177,6 +182,173 @@ func (p *contentProvider) ChatStream(ctx context.Context, req llm.ChatRequest, e
return emit(llm.StreamEvent{Content: p.text})
}
type btwProvider struct {
req llm.ChatRequest
err error
}
func (p *btwProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
p.req = req
if p.err != nil {
return p.err
}
return emit(llm.StreamEvent{Content: "side answer", Usage: &llm.Usage{TotalTokens: 999}})
}
func TestBTWUsesContextWithoutMutatingConversation(t *testing.T) {
provider := &btwProvider{}
a := New(Options{Provider: provider, Model: "test", Thinking: "high", ThinkingEnabled: true})
a.SetMessages([]llm.Message{
{Role: llm.RoleUser, Content: "main question"},
{Role: llm.RoleAssistant, Content: "main answer"},
})
a.SetLastUsage(&llm.Usage{TotalTokens: 7})
before := a.Messages()
var out strings.Builder
if err := a.BTW(context.Background(), "what does that mean?", &out); err != nil {
t.Fatal(err)
}
if out.String() != "side answer" {
t.Fatalf("output = %q", out.String())
}
if !reflect.DeepEqual(a.Messages(), before) {
t.Fatalf("conversation changed:\nbefore: %+v\nafter: %+v", before, a.Messages())
}
if usage := a.LastUsage(); usage == nil || usage.TotalTokens != 7 {
t.Fatalf("usage changed: %+v", usage)
}
if len(provider.req.Tools) != 0 || provider.req.ToolChoice != "" {
t.Fatalf("btw exposed tools: %+v / %q", provider.req.Tools, provider.req.ToolChoice)
}
last := provider.req.Messages[len(provider.req.Messages)-1]
if last.Role != llm.RoleUser || last.Content != "what does that mean?" {
t.Fatalf("last request message = %+v", last)
}
if provider.req.Extra["thinking"] != "high" {
t.Fatalf("thinking extra = %+v", provider.req.Extra)
}
}
func TestBTWRejectsEmptyQuestionAndReturnsProviderError(t *testing.T) {
provider := &btwProvider{err: errors.New("boom")}
a := New(Options{Provider: provider, Model: "test"})
if err := a.BTW(context.Background(), " ", nil); err == nil {
t.Fatal("empty question did not fail")
}
if err := a.BTW(context.Background(), "question", nil); err == nil || !strings.Contains(err.Error(), "boom") {
t.Fatalf("provider error = %v", err)
}
}
func TestDotContinuesCurrentTaskWithoutRewritingHistory(t *testing.T) {
provider := &btwProvider{}
a := New(Options{Provider: provider, Model: "test"})
if err := a.RunTurn(context.Background(), " . ", nil, nil); err != nil {
t.Fatal(err)
}
requestMessage := provider.req.Messages[len(provider.req.Messages)-1]
if requestMessage.Content != continuationPrompt {
t.Fatalf("request content = %q, want %q", requestMessage.Content, continuationPrompt)
}
history := a.Messages()
if history[0].Content != "." {
t.Fatalf("stored input = %q, want literal dot", history[0].Content)
}
}
func TestExpandContinuationMessagesOnlyExpandsSingleDot(t *testing.T) {
messages := []llm.Message{
{Role: llm.RoleUser, Content: "."},
{Role: llm.RoleUser, Content: " . "},
{Role: llm.RoleUser, Content: ".."},
{Role: llm.RoleUser, Content: ". continue"},
{Role: llm.RoleAssistant, Content: "."},
}
expanded := expandContinuationMessages(messages)
if expanded[0].Content != continuationPrompt || expanded[1].Content != continuationPrompt {
t.Fatalf("single dots were not expanded: %+v", expanded)
}
for _, index := range []int{2, 3, 4} {
if expanded[index].Content != messages[index].Content {
t.Fatalf("message %d changed from %q to %q", index, messages[index].Content, expanded[index].Content)
}
}
if messages[0].Content != "." || messages[1].Content != " . " {
t.Fatalf("source messages were mutated: %+v", messages)
}
}
func TestCompleteConversationPrefixDropsPendingToolRound(t *testing.T) {
messages := []llm.Message{
{Role: llm.RoleUser, Content: "first"},
{Role: llm.RoleAssistant, Content: "answer"},
{Role: llm.RoleUser, Content: "main"},
{Role: llm.RoleAssistant, ToolCalls: []llm.ToolCall{{ID: "call_1"}, {ID: "call_2"}}},
{Role: llm.RoleTool, ToolCallID: "call_1", Content: "one"},
}
got := completeConversationPrefix(messages)
if len(got) != 3 || got[len(got)-1].Content != "main" {
t.Fatalf("complete prefix = %+v", got)
}
}
type concurrentBTWProvider struct {
mainStarted chan struct{}
releaseMain chan struct{}
once sync.Once
mu sync.Mutex
btwRequest llm.ChatRequest
}
func (p *concurrentBTWProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
last := req.Messages[len(req.Messages)-1].Content
if last == "main task" {
p.once.Do(func() { close(p.mainStarted) })
select {
case <-p.releaseMain:
return emit(llm.StreamEvent{Content: "main answer"})
case <-ctx.Done():
return ctx.Err()
}
}
p.mu.Lock()
p.btwRequest = req
p.mu.Unlock()
return emit(llm.StreamEvent{Content: "side answer"})
}
func TestBTWCanRunDuringMainTurn(t *testing.T) {
provider := &concurrentBTWProvider{mainStarted: make(chan struct{}), releaseMain: make(chan struct{})}
a := New(Options{Provider: provider, Model: "test"})
mainDone := make(chan error, 1)
go func() { mainDone <- a.RunTurn(context.Background(), "main task", nil, nil) }()
<-provider.mainStarted
var side strings.Builder
if err := a.BTW(context.Background(), "side question", &side); err != nil {
t.Fatal(err)
}
if side.String() != "side answer" {
t.Fatalf("side output = %q", side.String())
}
provider.mu.Lock()
req := provider.btwRequest
provider.mu.Unlock()
if len(req.Messages) < 2 || req.Messages[len(req.Messages)-2].Content != "main task" {
t.Fatalf("btw request did not include active main turn: %+v", req.Messages)
}
close(provider.releaseMain)
if err := <-mainDone; err != nil {
t.Fatal(err)
}
if len(a.Messages()) != 2 {
t.Fatalf("main conversation contains btw messages: %+v", a.Messages())
}
}
type compactProvider struct {
summaryCalls int
answerCalls int
@@ -629,6 +801,120 @@ type parallelProvider struct {
calls int
}
type authorizationProvider struct {
call llm.ToolCallDelta
calls int
toolOutput string
}
func (p *authorizationProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
p.calls++
if p.calls == 1 {
return emit(llm.StreamEvent{ToolCalls: []llm.ToolCallDelta{p.call}})
}
p.toolOutput = req.Messages[len(req.Messages)-1].Content
return emit(llm.StreamEvent{Content: "done"})
}
type recordingAuthorizer struct {
decision ToolAuthorizationDecision
requests []ToolAuthorizationRequest
}
func (a *recordingAuthorizer) AuthorizeToolCall(ctx context.Context, req ToolAuthorizationRequest) (ToolAuthorizationDecision, error) {
a.requests = append(a.requests, req)
return a.decision, ctx.Err()
}
func TestAgentAuthorizesMutationAndCreatesRollbackSnapshot(t *testing.T) {
dir := t.TempDir()
provider := &authorizationProvider{call: llm.ToolCallDelta{
Index: 0, ID: "write_1", Type: "function", Name: "file_write",
Arguments: `{"path":"new.txt","content":"hello"}`,
}}
authorizer := &recordingAuthorizer{decision: ToolAuthorizationAllow}
rollback := tools.NewRollbackStore()
a := New(Options{
Provider: provider,
Model: "test",
ToolRegistry: tools.Builtins(dir),
Authorizer: authorizer,
RollbackStore: rollback,
})
if err := a.RunTurn(context.Background(), "write it", nil, nil); err != nil {
t.Fatal(err)
}
if len(authorizer.requests) != 1 || !authorizer.requests[0].RollbackAvailable {
t.Fatalf("authorization requests = %+v", authorizer.requests)
}
if !strings.Contains(authorizer.requests[0].Preview, "write 5 bytes to new.txt") {
t.Fatalf("preview = %q", authorizer.requests[0].Preview)
}
rollbackID := rollback.LastID()
if rollbackID == "" || !strings.Contains(provider.toolOutput, rollbackID) {
t.Fatalf("tool output %q missing rollback ID %q", provider.toolOutput, rollbackID)
}
if _, err := rollback.Restore(rollbackID); err != nil {
t.Fatal(err)
}
if _, err := os.Stat(filepath.Join(dir, "new.txt")); !os.IsNotExist(err) {
t.Fatalf("new file still exists after rollback: %v", err)
}
}
func TestAgentDeniedMutationDoesNotExecute(t *testing.T) {
dir := t.TempDir()
provider := &authorizationProvider{call: llm.ToolCallDelta{
Index: 0, ID: "write_1", Type: "function", Name: "file_write",
Arguments: `{"path":"denied.txt","content":"no"}`,
}}
authorizer := &recordingAuthorizer{decision: ToolAuthorizationDeny}
a := New(Options{Provider: provider, Model: "test", ToolRegistry: tools.Builtins(dir), Authorizer: authorizer})
if err := a.RunTurn(context.Background(), "write it", nil, nil); err != nil {
t.Fatal(err)
}
if _, err := os.Stat(filepath.Join(dir, "denied.txt")); !os.IsNotExist(err) {
t.Fatalf("denied file exists: %v", err)
}
if !strings.Contains(provider.toolOutput, "denied by user") {
t.Fatalf("tool output = %q", provider.toolOutput)
}
}
func TestAgentDoesNotAuthorizeApplyPatchDryRun(t *testing.T) {
dir := t.TempDir()
provider := &authorizationProvider{call: llm.ToolCallDelta{
Index: 0, ID: "patch_1", Type: "function", Name: "apply_patch",
Arguments: `{"patch":"*** Begin Patch\n*** Add File: dry.txt\n+dry\n*** End Patch","dry_run":true}`,
}}
authorizer := &recordingAuthorizer{decision: ToolAuthorizationDeny}
a := New(Options{Provider: provider, Model: "test", ToolRegistry: tools.Builtins(dir), Authorizer: authorizer})
if err := a.RunTurn(context.Background(), "preview", nil, nil); err != nil {
t.Fatal(err)
}
if len(authorizer.requests) != 0 {
t.Fatalf("dry run requested authorization: %+v", authorizer.requests)
}
if _, err := os.Stat(filepath.Join(dir, "dry.txt")); !os.IsNotExist(err) {
t.Fatalf("dry run created a file: %v", err)
}
}
func TestSetMessagesRefreshesCurrentSystemInstructions(t *testing.T) {
a := New(Options{SystemPrompt: "current instructions"})
a.SetMessages([]llm.Message{
{Role: llm.RoleSystem, Content: "old instructions"},
{Role: llm.RoleUser, Content: "hello"},
})
messages := a.Messages()
if messages[0].Content != "current instructions" || messages[1].Content != "hello" {
t.Fatalf("messages = %+v", messages)
}
}
func (p *parallelProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
p.calls++
if p.calls <= len(p.responses) {