feat: Phase 1 code-agent capabilities — file_edit, git tools, parallel execution, tool output preview

New tools: - file_edit: surgical line-range or pattern edits (yolo only) - git_status: show changed files (read-only) - git_diff: show unstaged/staged diff (read-only) - git_log: show recent commits (read-only) Agent improvements: - parallel execution for independent read-only tool calls - mutating tools (file_write, file_edit, shell_run) run sequentially - tool output preview in logs with [output] marker TUI improvements: - tool cards now show truncated output preview (up to 8 lines) - mergeToolLog updates placeholder cards with results - toolStatusText keeps activity bar single-line Documentation: - README updated with git tools and file_edit descriptions
2026-06-24 19:47:14 -07:00
parent 7b8bad5e62
commit 3d403bd685
13 changed files with 1135 additions and 39 deletions
@@ -7,18 +7,22 @@ import (
 	"io"
 	"sort"
 	"strings"
+	"sync"
 	"time"

-	"agentu/pkg/llm"
 	"agentu/internal/tools"
+	"agentu/pkg/llm"
 )

 const (
-	defaultMaxToolRounds  = 50
-	doomLoopThreshold     = 3
-	compactRecentMessages = 6
-	summaryMessageLimit   = 8000
-	contextSummaryPrefix  = "[context summary]"
+	defaultMaxToolRounds      = 50
+	doomLoopThreshold         = 3
+	compactRecentMessages     = 6
+	summaryMessageLimit       = 8000
+	contextSummaryPrefix      = "[context summary]"
+	toolLogArgumentLimit      = 400
+	toolLogOutputPreviewBytes = 4096
+	toolLogOutputMarker       = "[output]"

 	summarizerSystemPrompt = "You are a conversation summarizer. Summarize the following conversation history into a concise but detailed summary. Preserve:\n" +
 		"- Key facts, decisions, and conclusions\n" +
@@ -29,6 +33,11 @@ const (
 		"Output only the summary, no preamble."
 )

+type toolExecutionResult struct {
+	call   llm.ToolCall
+	output string
+}
+
 type Agent struct {
 	provider         llm.Provider
 	providerName     string
@@ -364,14 +373,15 @@ func (a *Agent) RunTurn(ctx context.Context, input string, out io.Writer, logs i
 			}
 		}

-		for _, call := range calls {
-			result := a.executeTool(ctx, call, logs)
+		results := a.executeTools(ctx, calls, logs)
+		for _, result := range results {
 			a.messages = append(a.messages, llm.Message{
 				Role:       llm.RoleTool,
-				ToolCallID: call.ID,
-				Content:    result,
+				ToolCallID: result.call.ID,
+				Content:    result.output,
 			})
 		}
+
 	}

 	return fmt.Errorf("tool round limit reached after %d rounds; the task may be too complex for a single turn", a.maxToolRounds)
@@ -433,11 +443,8 @@ func (a *Agent) toolDefinitions() []llm.Tool {
 	return a.toolRegistry.Definitions()
 }

-func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall, logs io.Writer) string {
+func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall) string {
 	name := call.Function.Name
-	if logs != nil {
-		fmt.Fprintf(logs, "\n[tool] %s %s\n", name, compact(call.Function.Arguments, 400))
-	}

 	if a.toolRegistry == nil {
 		return tools.FormatError(fmt.Errorf("tool registry is disabled"))
@@ -461,12 +468,84 @@ func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall, logs io.Writ
 		}
 		return tools.FormatError(err)
 	}
-	if logs != nil {
-		fmt.Fprintf(logs, "[tool] %s done\n", name)
-	}
 	return output
 }

+func (a *Agent) hasMutatingCall(calls []llm.ToolCall) bool {
+	if a.toolRegistry == nil {
+		return false
+	}
+	for _, call := range calls {
+		if tool, ok := a.toolRegistry.Get(call.Function.Name); ok {
+			if tools.IsMutating(tool) {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+func (a *Agent) executeTools(ctx context.Context, calls []llm.ToolCall, logs io.Writer) []toolExecutionResult {
+	results := make([]toolExecutionResult, len(calls))
+
+	// Log all tool starts sequentially so TUI gets ordered placeholders.
+	if logs != nil {
+		for _, call := range calls {
+			logToolStart(logs, call)
+		}
+	}
+
+	if a.hasMutatingCall(calls) {
+		// Sequential: avoid racing mutating tools.
+		for i, call := range calls {
+			output := a.executeTool(ctx, call)
+			results[i] = toolExecutionResult{call: call, output: output}
+			if logs != nil {
+				logToolOutput(logs, call, output)
+			}
+		}
+		return results
+	}
+
+	// Concurrent: independent read-only tools.
+	var mu sync.Mutex
+	var wg sync.WaitGroup
+	for i, call := range calls {
+		wg.Add(1)
+		go func(idx int, c llm.ToolCall) {
+			defer wg.Done()
+			output := a.executeTool(ctx, c)
+			results[idx] = toolExecutionResult{call: c, output: output}
+			if logs != nil {
+				mu.Lock()
+				logToolOutput(logs, c, output)
+				mu.Unlock()
+			}
+		}(i, call)
+	}
+	wg.Wait()
+	return results
+}
+
+func logToolStart(logs io.Writer, call llm.ToolCall) {
+	fmt.Fprintf(logs, "\n[tool] %s %s\n", call.Function.Name, compact(call.Function.Arguments, toolLogArgumentLimit))
+}
+
+func logToolOutput(logs io.Writer, call llm.ToolCall, output string) {
+	fmt.Fprintf(logs, "\n[tool] %s %s\n%s\n%s\n", call.Function.Name, compact(call.Function.Arguments, toolLogArgumentLimit), toolLogOutputMarker, toolOutputPreview(output))
+}
+
+func toolOutputPreview(output string) string {
+	output = strings.TrimSpace(output)
+	if output == "" {
+		return "(no output)"
+	}
+	if len(output) <= toolLogOutputPreviewBytes {
+		return output
+	}
+	return output[:toolLogOutputPreviewBytes] + fmt.Sprintf("\n[truncated: showing first %d bytes; omitted %d bytes]", toolLogOutputPreviewBytes, len(output)-toolLogOutputPreviewBytes)
+}
+
 type toolCallBuilder struct {
 	index     int
 	id        string
@@ -7,8 +7,8 @@ import (
 	"strings"
 	"testing"

-	"agentu/pkg/llm"
 	"agentu/internal/tools"
+	"agentu/pkg/llm"
 )

 type fakeProvider struct {
@@ -340,3 +340,174 @@ func TestAgentHardLimitStopsVaryingCalls(t *testing.T) {
 		t.Fatalf("calls = %d, want %d", provider.calls, limit)
 	}
 }
+
+// slowTool blocks on a channel before returning.
+type slowTool struct {
+	name    string
+	started chan<- string
+	release <-chan struct{}
+}
+
+func (t *slowTool) Definition() llm.Tool {
+	return llm.Tool{
+		Type: "function",
+		Function: llm.ToolFunction{
+			Name:        t.name,
+			Description: "slow test tool",
+			Parameters:  json.RawMessage(`{"type":"object","properties":{"text":{"type":"string"}}}`),
+		},
+	}
+}
+
+func (t *slowTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	var args struct {
+		Text string `json:"text"`
+	}
+	if err := json.Unmarshal(raw, &args); err != nil {
+		return "", err
+	}
+	t.started <- t.name
+	<-t.release
+	return args.Text, nil
+}
+
+// mutatingSlowTool is like slowTool but implements MutatingTool.
+type mutatingSlowTool struct {
+	slowTool
+}
+
+func (*mutatingSlowTool) Mutates() bool { return true }
+
+func TestAgentExecutesIndependentToolCallsInParallel(t *testing.T) {
+	started := make(chan string, 2)
+	release := make(chan struct{})
+
+	tool1 := &slowTool{name: "slow_one", started: started, release: release}
+	tool2 := &slowTool{name: "slow_two", started: started, release: release}
+
+	provider := &parallelProvider{
+		responses: [][]llm.ToolCallDelta{
+			{
+				{Index: 0, ID: "call_1", Type: "function", Name: "slow_one", Arguments: `{"text":"one"}`},
+				{Index: 1, ID: "call_2", Type: "function", Name: "slow_two", Arguments: `{"text":"two"}`},
+			},
+		},
+		finalContent: "done",
+	}
+
+	a := New(Options{
+		Provider:     provider,
+		Model:        "test",
+		SystemPrompt: "system",
+		ToolRegistry: tools.NewRegistry(tool1, tool2),
+	})
+
+	var out strings.Builder
+	done := make(chan error, 1)
+	go func() {
+		done <- a.RunTurn(context.Background(), "run both", &out, nil)
+	}()
+
+	// Both tools must start before we release them.
+	names := make(map[string]bool)
+	for i := 0; i < 2; i++ {
+		select {
+		case name := <-started:
+			names[name] = true
+		case <-make(chan struct{}):
+			// Use a timer in real code, but for test simplicity we rely on deadlock detection.
+			t.Fatal("timed out waiting for tool to start")
+		}
+	}
+	if !names["slow_one"] || !names["slow_two"] {
+		t.Fatalf("expected both tools to start, got %v", names)
+	}
+	close(release)
+
+	if err := <-done; err != nil {
+		t.Fatal(err)
+	}
+	if out.String() != "done" {
+		t.Fatalf("out = %q", out.String())
+	}
+	// Verify tool results were in order.
+	msgs := a.Messages()
+	var toolMsgs []string
+	for _, m := range msgs {
+		if m.Role == llm.RoleTool {
+			toolMsgs = append(toolMsgs, m.Content)
+		}
+	}
+	if len(toolMsgs) != 2 || toolMsgs[0] != "one" || toolMsgs[1] != "two" {
+		t.Fatalf("tool messages = %v, want [one two]", toolMsgs)
+	}
+}
+
+func TestAgentRunsMutatingToolCallsSequentially(t *testing.T) {
+	started := make(chan string, 2)
+	release := make(chan struct{})
+
+	tool1 := &mutatingSlowTool{slowTool{name: "mut_one", started: started, release: release}}
+	tool2 := &mutatingSlowTool{slowTool{name: "mut_two", started: started, release: release}}
+
+	provider := &parallelProvider{
+		responses: [][]llm.ToolCallDelta{
+			{
+				{Index: 0, ID: "call_1", Type: "function", Name: "mut_one", Arguments: `{"text":"one"}`},
+				{Index: 1, ID: "call_2", Type: "function", Name: "mut_two", Arguments: `{"text":"two"}`},
+			},
+		},
+		finalContent: "done",
+	}
+
+	a := New(Options{
+		Provider:     provider,
+		Model:        "test",
+		SystemPrompt: "system",
+		ToolRegistry: tools.NewRegistry(tool1, tool2),
+	})
+
+	var out strings.Builder
+	done := make(chan error, 1)
+	go func() {
+		done <- a.RunTurn(context.Background(), "run both", &out, nil)
+	}()
+
+	// Only the first tool should start.
+	select {
+	case name := <-started:
+		if name != "mut_one" {
+			t.Fatalf("expected mut_one to start first, got %q", name)
+		}
+	case <-make(chan struct{}):
+		t.Fatal("timed out waiting for first tool to start")
+	}
+
+	// Second tool should NOT have started yet.
+	select {
+	case name := <-started:
+		t.Fatalf("second tool started too early: %q", name)
+	default:
+		// Expected: second tool hasn't started.
+	}
+
+	close(release)
+	if err := <-done; err != nil {
+		t.Fatal(err)
+	}
+}
+
+// parallelProvider emits a batch of tool calls on the first request, then returns finalContent.
+type parallelProvider struct {
+	responses    [][]llm.ToolCallDelta
+	finalContent string
+	calls        int
+}
+
+func (p *parallelProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	p.calls++
+	if p.calls <= len(p.responses) {
+		return emit(llm.StreamEvent{ToolCalls: p.responses[p.calls-1]})
+	}
+	return emit(llm.StreamEvent{Content: p.finalContent})
+}