feat(tools): stream shell_run output live into the TUI

README.md

@@ -167,7 +167,8 @@ Web tools are also available in read-only mode:
 
 - `file_write` for creating or replacing whole files
 - `file_edit` for surgical line-range or pattern edits to existing files
-- `shell_run` for non-interactive shell commands
+- `shell_run` for non-interactive shell commands; output streams live into the
+  TUI while the command runs instead of appearing only after completion
 
 Use `--yolo` for prompts that ask agentu to run or inspect local commands, for
 example:

internal/agent/agent.go

@@ -24,6 +24,7 @@ const (
 	toolLogArgumentLimit          = 400
 	toolLogOutputPreviewBytes     = 4096
 	toolLogOutputMarker           = "[output]"
+	toolStreamLogLimit            = tools.MaxToolOutputBytes
 
 	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" +
@@ -182,6 +183,7 @@ func (a *Agent) SetLastUsage(usage *llm.Usage) {
 	copy := *usage
 	a.lastUsage = &copy
 }
+
 // compactRetainedTokenBudget returns the number of estimated tokens to retain
 // based on compactRecentRetentionPercent, using integer ceiling division.
 func compactRetainedTokenBudget(totalTokens int) int {
@@ -237,7 +239,6 @@ func compactRecentStart(messages []llm.Message, prefixEnd int) int {
 	return compactTurnBoundary(messages, prefixEnd, start)
 }
 
-
 // Compact summarizes older conversation messages while preserving the system
 // prompt and recent messages intact.
 func (a *Agent) Compact(ctx context.Context) error {
@@ -495,32 +496,75 @@ func (a *Agent) toolDefinitions() []llm.Tool {
 	return a.toolRegistry.Definitions()
 }
 
-func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall) string {
+func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall, logs io.Writer) (string, bool) {
 	name := call.Function.Name
 
 	if a.toolRegistry == nil {
-		return tools.FormatError(fmt.Errorf("tool registry is disabled"))
+		return tools.FormatError(fmt.Errorf("tool registry is disabled")), false
 	}
 	tool, ok := a.toolRegistry.Get(name)
 	if !ok {
-		return tools.FormatError(fmt.Errorf("unknown tool: %s", name))
+		return tools.FormatError(fmt.Errorf("unknown tool: %s", name)), false
 	}
 
 	args := strings.TrimSpace(call.Function.Arguments)
 	if args == "" {
 		args = "{}"
 	}
+	raw := json.RawMessage(args)
 
 	toolCtx, cancel := context.WithTimeout(ctx, a.toolTimeout)
 	defer cancel()
-	output, err := tool.Execute(toolCtx, json.RawMessage(args))
+
+	if streaming, ok := tool.(tools.StreamingTool); ok {
+		return a.executeStreamingTool(streaming, toolCtx, raw, name, args, logs)
+	}
+
+	output, err := tool.Execute(toolCtx, raw)
 	if err != nil {
 		if output != "" {
-			return output + "\n" + tools.FormatError(err)
+			return output + "\n" + tools.FormatError(err), false
 		}
-		return tools.FormatError(err)
+		return tools.FormatError(err), false
 	}
-	return output
+	return output, false
+}
+
+func (a *Agent) executeStreamingTool(streaming tools.StreamingTool, ctx context.Context, raw json.RawMessage, name, args string, logs io.Writer) (string, bool) {
+	var streamed int
+	emit := func(chunk string) error {
+		if chunk == "" || logs == nil {
+			return nil
+		}
+		if streamed >= toolStreamLogLimit {
+			return nil
+		}
+		if len(chunk) > toolStreamLogLimit-streamed {
+			chunk = chunk[:toolStreamLogLimit-streamed]
+			streamed = toolStreamLogLimit
+			_, err := fmt.Fprintf(logs, "\n[tool] %s %s\n%s\n%s\n[streaming output truncated after %d bytes]", name, compact(args, toolLogArgumentLimit), toolLogOutputMarker, chunk, toolStreamLogLimit)
+			return err
+		}
+		streamed += len(chunk)
+		_, err := fmt.Fprintf(logs, "\n[tool] %s %s\n%s\n%s", name, compact(args, toolLogArgumentLimit), toolLogOutputMarker, chunk)
+		return err
+	}
+
+	output, err := streaming.ExecuteStream(ctx, raw, emit)
+	if logs != nil {
+		if err != nil {
+			_ = emit("\n" + tools.FormatError(err))
+		} else if streamed == 0 {
+			_ = emit("(no output)")
+		}
+	}
+	if err != nil {
+		if output != "" {
+			return output + "\n" + tools.FormatError(err), true
+		}
+		return tools.FormatError(err), true
+	}
+	return output, true
 }
 
 func (a *Agent) hasMutatingCall(calls []llm.ToolCall) bool {
@@ -550,9 +594,9 @@ func (a *Agent) executeTools(ctx context.Context, calls []llm.ToolCall, logs io.
 	if a.hasMutatingCall(calls) {
 		// Sequential: avoid racing mutating tools.
 		for i, call := range calls {
-			output := a.executeTool(ctx, call)
+			output, streamed := a.executeTool(ctx, call, logs)
 			results[i] = toolExecutionResult{call: call, output: output}
-			if logs != nil {
+			if logs != nil && !streamed {
 				logToolOutput(logs, call, output)
 			}
 		}
@@ -566,9 +610,9 @@ func (a *Agent) executeTools(ctx context.Context, calls []llm.ToolCall, logs io.
 		wg.Add(1)
 		go func(idx int, c llm.ToolCall) {
 			defer wg.Done()
-			output := a.executeTool(ctx, c)
+			output, streamed := a.executeTool(ctx, c, logs)
 			results[idx] = toolExecutionResult{call: c, output: output}
-			if logs != nil {
+			if logs != nil && !streamed {
 				mu.Lock()
 				logToolOutput(logs, c, output)
 				mu.Unlock()

internal/agent/agent_test.go

@@ -87,6 +87,86 @@ func TestAgentRunsToolLoop(t *testing.T) {
 	}
 }
 
+type streamToolProvider struct {
+	calls int
+	t     *testing.T
+}
+
+func (p *streamToolProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	p.calls++
+	switch p.calls {
+	case 1:
+		return emit(llm.StreamEvent{ToolCalls: []llm.ToolCallDelta{
+			{Index: 0, ID: "call_1", Type: "function", Name: "test_stream", Arguments: `{}`},
+		}})
+	case 2:
+		last := req.Messages[len(req.Messages)-1]
+		if last.Role != llm.RoleTool || last.Content != "full result" {
+			p.t.Fatalf("last message = %#v", last)
+		}
+		return emit(llm.StreamEvent{Content: "done"})
+	default:
+		p.t.Fatalf("unexpected call count %d", p.calls)
+		return nil
+	}
+}
+
+type streamEchoTool struct{}
+
+func (streamEchoTool) Definition() llm.Tool {
+	return llm.Tool{
+		Type: "function",
+		Function: llm.ToolFunction{
+			Name:        "test_stream",
+			Description: "streaming test",
+			Parameters:  json.RawMessage(`{"type":"object"}`),
+		},
+	}
+}
+
+func (streamEchoTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	return "full result", nil
+}
+
+func (streamEchoTool) ExecuteStream(ctx context.Context, raw json.RawMessage, emit func(string) error) (string, error) {
+	if err := emit("one\n"); err != nil {
+		return "", err
+	}
+	if err := emit("two\n"); err != nil {
+		return "", err
+	}
+	return "full result", nil
+}
+
+func TestAgentStreamsToolOutputToLogs(t *testing.T) {
+	provider := &streamToolProvider{t: t}
+	a := New(Options{
+		Provider:     provider,
+		Model:        "test",
+		SystemPrompt: "system",
+		ToolRegistry: tools.NewRegistry(streamEchoTool{}),
+	})
+
+	var out strings.Builder
+	var logs strings.Builder
+	if err := a.RunTurn(context.Background(), "stream", &out, &logs); err != nil {
+		t.Fatal(err)
+	}
+	if out.String() != "done" {
+		t.Fatalf("out = %q", out.String())
+	}
+	logText := logs.String()
+	if strings.Count(logText, "[output]") != 2 {
+		t.Fatalf("logs = %q", logText)
+	}
+	if !strings.Contains(logText, "one\n") || !strings.Contains(logText, "two\n") {
+		t.Fatalf("logs missing streamed chunks: %q", logText)
+	}
+	if strings.Contains(logText, "[output]\nfull result") {
+		t.Fatalf("logs contain duplicate full output: %q", logText)
+	}
+}
+
 type contentProvider struct {
 	called bool
 	text   string

internal/tools/shell.go

@@ -1,6 +1,7 @@
 package tools
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
@@ -8,6 +9,7 @@ import (
 	"os"
 	"os/exec"
 	"strings"
+	"sync"
 
 	"agentu/pkg/llm"
 )
@@ -46,6 +48,10 @@ type shellRunArgs struct {
 }
 
 func (t *ShellRunTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	return t.ExecuteStream(ctx, raw, nil)
+}
+
+func (t *ShellRunTool) ExecuteStream(ctx context.Context, raw json.RawMessage, emit func(string) error) (string, error) {
 	var args shellRunArgs
 	if err := json.Unmarshal(raw, &args); err != nil {
 		return "", fmt.Errorf("parse arguments: %w", err)
@@ -66,16 +72,45 @@ func (t *ShellRunTool) Execute(ctx context.Context, raw json.RawMessage) (string
 	cmd := exec.CommandContext(ctx, shell, "-lc", args.Command)
 	cmd.Dir = workdir
 
-	output, err := cmd.CombinedOutput()
-	text := FormatResult(string(output))
+	var mu sync.Mutex
+	var combined bytes.Buffer
+	writer := &shellStreamWriter{mu: &mu, buf: &combined, emit: emit}
+	cmd.Stdout = writer
+	cmd.Stderr = writer
+
+	runErr := cmd.Run()
+	text := FormatResult(combined.String())
 	if errors.Is(ctx.Err(), context.DeadlineExceeded) {
 		return text, fmt.Errorf("command timed out: %w", ctx.Err())
 	}
-	if err != nil {
+	if runErr != nil {
 		if text != "" {
-			return text, fmt.Errorf("command failed: %w", err)
+			return text, fmt.Errorf("command failed: %w", runErr)
 		}
-		return "", fmt.Errorf("command failed: %w", err)
+		return "", fmt.Errorf("command failed: %w", runErr)
 	}
 	return text, nil
 }
+
+// shellStreamWriter captures command output for the final result while
+// forwarding every write to the streaming callback as it happens.
+type shellStreamWriter struct {
+	mu   *sync.Mutex
+	buf  *bytes.Buffer
+	emit func(string) error
+}
+
+func (w *shellStreamWriter) Write(p []byte) (int, error) {
+	w.mu.Lock()
+	n, writeErr := w.buf.Write(p)
+	w.mu.Unlock()
+	if writeErr != nil {
+		return n, writeErr
+	}
+	if w.emit != nil {
+		if err := w.emit(string(p)); err != nil {
+			return 0, err
+		}
+	}
+	return len(p), nil
+}

internal/tools/shell_test.go

@@ -0,0 +1,71 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"sync"
+	"testing"
+)
+
+func TestShellRunStreamsOutputWhileRunning(t *testing.T) {
+	tool := NewShellRunTool(t.TempDir())
+	var mu sync.Mutex
+	var chunks []string
+	output, err := tool.ExecuteStream(context.Background(), json.RawMessage(`{"command":"printf 'a\\n'; sleep 0.05; printf 'b\\n'"}`), func(chunk string) error {
+		mu.Lock()
+		chunks = append(chunks, chunk)
+		mu.Unlock()
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(output, "a") || !strings.Contains(output, "b") {
+		t.Fatalf("output = %q", output)
+	}
+	mu.Lock()
+	joined := strings.Join(chunks, "")
+	mu.Unlock()
+	if len(chunks) == 0 {
+		t.Fatal("expected at least one streamed chunk")
+	}
+	if !strings.Contains(joined, "a") || !strings.Contains(joined, "b") {
+		t.Fatalf("streamed chunks = %q", joined)
+	}
+}
+
+func TestShellRunExecuteCollectsFullOutput(t *testing.T) {
+	tool := NewShellRunTool(t.TempDir())
+	output, err := tool.Execute(context.Background(), json.RawMessage(`{"command":"printf 'hello world'"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.TrimSpace(output) != "hello world" {
+		t.Fatalf("output = %q", output)
+	}
+}
+
+func TestShellRunStreamsStderrAndStdout(t *testing.T) {
+	tool := NewShellRunTool(t.TempDir())
+	var mu sync.Mutex
+	var chunks []string
+	output, err := tool.ExecuteStream(context.Background(), json.RawMessage(`{"command":"echo out; echo err >&2"}`), func(chunk string) error {
+		mu.Lock()
+		chunks = append(chunks, chunk)
+		mu.Unlock()
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(output, "out") || !strings.Contains(output, "err") {
+		t.Fatalf("output = %q", output)
+	}
+	mu.Lock()
+	joined := strings.Join(chunks, "")
+	mu.Unlock()
+	if !strings.Contains(joined, "out") || !strings.Contains(joined, "err") {
+		t.Fatalf("streamed chunks = %q", joined)
+	}
+}

internal/tools/tools.go

@@ -19,6 +19,15 @@ type Tool interface {
 	Execute(ctx context.Context, args json.RawMessage) (string, error)
 }
 
+// StreamingTool is optionally implemented by tools that can report partial
+// output while they execute. emit receives chunks of the tool's output as
+// they are produced; the returned string is still the complete result that
+// gets sent back to the model.
+type StreamingTool interface {
+	Tool
+	ExecuteStream(ctx context.Context, args json.RawMessage, emit func(string) error) (string, error)
+}
+
 // MutatingTool is optionally implemented by tools that modify external state.
 // It is used by the agent to determine whether tool calls can run in parallel.
 type MutatingTool interface {

internal/tui/app.go

@@ -683,15 +683,27 @@ func (m *model) mergeToolLog(text string) bool {
 	if incoming.output == "" {
 		return false
 	}
+	incomingHeader, _, ok := strings.Cut(text, "\n[output]\n")
+	if !ok {
+		return false
+	}
 	for i := len(m.messages) - 1; i >= 0; i-- {
 		if m.messages[i].role != roleTool {
 			continue
 		}
 		existing := parseToolLog(m.messages[i].content)
-		if existing.name == incoming.name && existing.args == incoming.args && existing.output == "" {
-			m.messages[i].content = text
-			return true
+		if existing.name != incoming.name || existing.args != incoming.args {
+			continue
 		}
+		if !strings.Contains(m.messages[i].content, "\n[output]\n") {
+			// Placeholder card from the tool-start log; first chunk fills it.
+			m.messages[i].content = text
+		} else {
+			// Streaming tools send incremental chunks; append the raw delta
+			// so line boundaries are preserved inside the same card.
+			m.messages[i].content += strings.TrimPrefix(text, incomingHeader+"\n[output]\n")
+		}
+		return true
 	}
 	return false
 }

internal/tui/app_test.go

@@ -395,6 +395,28 @@ func TestToolLogOutputUpdatesExistingToolCard(t *testing.T) {
 	}
 }
 
+func TestToolLogStreamingChunksAppendToSameCard(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	m.messages = append(m.messages, message{role: roleTool, content: `shell_run echo hi`})
+
+	if !m.mergeToolLog("shell_run echo hi\n[output]\none\n") {
+		t.Fatal("expected first mergeToolLog to return true")
+	}
+	if !m.mergeToolLog("shell_run echo hi\n[output]\ntwo\n") {
+		t.Fatal("expected second mergeToolLog to return true")
+	}
+	if len(m.messages) != 1 {
+		t.Fatalf("messages count = %d, want 1", len(m.messages))
+	}
+	plain := stripANSI(m.renderMessages())
+	if !strings.Contains(plain, "one") || !strings.Contains(plain, "two") {
+		t.Fatalf("rendered output missing streamed lines:\n%s", plain)
+	}
+	if strings.Contains(plain, "onetwo") {
+		t.Fatalf("streamed lines were concatenated:\n%s", plain)
+	}
+}
+
 type tuiCompactProvider struct{}
 
 func (tuiCompactProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {