agent: plumb image ContentBlocks from tool results through to backend messages

toolExecutor now returns (string, []ContentBlock, error).
extractToolResult parses image blocks from structured tool output.
Session.update populates ContentBlocks on tool messages.
file_read detects PNG/JPEG/GIF/WEBP and returns base64-encoded image blocks.

pkg/agent/core.go

@@ -98,21 +98,21 @@ func BuildRuntime(cfg *config.Config, d tools.Dispatcher, cwd string, env []stri
 		}
 	}
 
-	exec := func(ctx context.Context, name string, args json.RawMessage) (string, error) {
-		server, ok := serverOf[name]
-		if !ok {
-			return "", fmt.Errorf("unknown tool: %s", name)
+		exec := func(ctx context.Context, name string, args json.RawMessage) (string, []backend.ContentBlock, error) {
+			server, ok := serverOf[name]
+			if !ok {
+				return "", nil, fmt.Errorf("unknown tool: %s", name)
+			}
+			raw, err := d.Dispatch(ctx, server, name, args)
+			if err != nil {
+				return "", nil, err
+			}
+			text, blocks, isErr := extractToolResult(raw)
+			if isErr {
+				return "", nil, fmt.Errorf("%s", text)
+			}
+			return text, blocks, nil
 		}
-		raw, err := d.Dispatch(ctx, server, name, args)
-		if err != nil {
-			return "", err
-		}
-		text, isErr := extractToolResult(raw)
-		if isErr {
-			return "", fmt.Errorf("%s", text)
-		}
-		return text, nil
-	}
 
 	var classify toolClassifier
 	if srv, ok := d.GetServer("execute"); ok {
@@ -1209,22 +1209,34 @@ func toolInfosToBackend(infos []tools.ToolInfo) []backend.Tool {
 	return out
 }
 
-func extractToolResult(raw json.RawMessage) (text string, isError bool) {
+func extractToolResult(raw json.RawMessage) (text string, contentBlocks []backend.ContentBlock, isError bool) {
 	var result struct {
 		IsError bool `json:"isError"`
 		Content []struct {
-			Type string `json:"type"`
-			Text string `json:"text"`
+			Type        string `json:"type"`
+			Text        string `json:"text"`
+			MediaType   string `json:"media_type"`
+			Data        string `json:"data"`
 		} `json:"content"`
 	}
 	if err := json.Unmarshal(raw, &result); err != nil {
-		return string(raw), false
+		return string(raw), nil, false
 	}
 	var parts []string
 	for _, c := range result.Content {
-		if c.Type == "text" {
+		switch c.Type {
+		case "text":
 			parts = append(parts, c.Text)
+		case "image":
+			contentBlocks = append(contentBlocks, backend.ContentBlock{
+				Type: "image",
+				ImageSource: &backend.ImageSource{
+					Type:      "base64",
+					MediaType: c.MediaType,
+					Data:      c.Data,
+				},
+			})
 		}
 	}
-	return strings.Join(parts, "\n"), result.IsError
+	return strings.Join(parts, "\n"), contentBlocks, result.IsError
 }

pkg/agent/core_test.go

@@ -184,7 +184,8 @@ func TestSubmit_StateTransitions(t *testing.T) {
 	unblock := make(chan struct{})
 	be := defaultBE()
 	be.respond = func(ctx context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return blockedStream(ctx, unblock), nil
+return blockedStream(ctx, unblock), nil
+
 	}
 	c := newCore(t, be, nil)
 
@@ -221,17 +222,19 @@ func TestSubmit_ToolCallStateTransitions(t *testing.T) {
 				StopReason: "tool_calls",
 			}
 			close(ch)
-			return ch, nil
+return ch, nil
+
 		}
-		return textStream("done"), nil
+return textStream("done"), nil
+
 	}
 
 	c := newCore(t, be, nil)
 
 	var stateAtExec string
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		stateAtExec = c.State()
-		return `{}`, nil
+return `{}`, nil, nil
 	}
 
 	collectEvents(context.Background(), c, "run tool")
@@ -252,7 +255,8 @@ func TestSubmit_PreTurnHookBlocks(t *testing.T) {
 	be := defaultBE()
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		callCount++
-		return textStream("should not be called"), nil
+return textStream("should not be called"), nil
+
 	}
 	c := newCore(t, be, Hooks{HookPreTurn: []string{"exit 2"}})
 
@@ -281,7 +285,7 @@ func TestSubmit_PreTurnHookContext(t *testing.T) {
 				lastUserMsg = m.Content
 			}
 		}
-		return textStream("ok"), nil
+	return textStream("ok"), nil
 	}
 	c := newCore(t, be, Hooks{HookPreTurn: []string{`echo "extra context"`}})
 
@@ -304,7 +308,8 @@ func TestSubmit_PostTurnHookContinue(t *testing.T) {
 	be := defaultBE()
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		callCount++
-		return textStream(fmt.Sprintf("response %d", callCount)), nil
+return textStream(fmt.Sprintf("response %d", callCount)), nil
+
 	}
 	c := newCore(t, be, Hooks{HookPostTurn: []string{hookScript}})
 
@@ -322,7 +327,7 @@ func TestSubmit_FIFODrain(t *testing.T) {
 	be := defaultBE()
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		callCount.Add(1)
-		return textStream("ok"), nil
+	return textStream("ok"), nil
 	}
 	c := newCore(t, be, nil)
 
@@ -350,7 +355,7 @@ func TestSubmit_PendingInjectAsNextTurn(t *testing.T) {
 	be := defaultBE()
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		callCount++
-		return textStream("ok"), nil
+	return textStream("ok"), nil
 	}
 	c := newCore(t, be, nil)
 
@@ -370,7 +375,8 @@ func TestInterrupt_Running(t *testing.T) {
 	unblock := make(chan struct{})
 	be := defaultBE()
 	be.respond = func(ctx context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return blockedStream(ctx, unblock), nil
+return blockedStream(ctx, unblock), nil
+
 	}
 	c := newCore(t, be, nil)
 
@@ -412,7 +418,8 @@ func TestSubmit_WhileRunning(t *testing.T) {
 	unblock := make(chan struct{})
 	be := defaultBE()
 	be.respond = func(ctx context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return blockedStream(ctx, unblock), nil
+return blockedStream(ctx, unblock), nil
+
 	}
 	c := newCore(t, be, nil)
 
@@ -449,7 +456,8 @@ func TestManualCompact(t *testing.T) {
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		callCount++
 		stateAtCompact = c.State()
-		return textStream("summary or answer"), nil
+return textStream("summary or answer"), nil
+
 	}
 
 	// Seed a session with more than hotTailSize+warmIndexSize messages so compact() doesn't short-circuit.
@@ -500,9 +508,11 @@ func TestAutoCompact(t *testing.T) {
 		callCount++
 		if callCount == 1 {
 			stateAtCompact = c.State()
-			return textStream("summary text for compaction"), nil
+return textStream("summary text for compaction"), nil
+
 		}
-		return textStream("answer"), nil
+return textStream("answer"), nil
+
 	}
 
 	c.session = newSession("goal")
@@ -664,7 +674,7 @@ func TestCommand_I_SetsInject(t *testing.T) {
 	done := make(chan struct{})
 	be.respond = func(_ context.Context, msgs []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		<-done
-		return textStream("ok"), nil
+	return textStream("ok"), nil
 	}
 	c := newCore(t, be, nil)
 	go c.Submit(context.Background(), "hello")
@@ -684,7 +694,7 @@ func TestCommand_I_SubmitsWhenIdle(t *testing.T) {
 	be := defaultBE()
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		callCount.Add(1)
-		return textStream("ok"), nil
+	return textStream("ok"), nil
 	}
 	c := newCore(t, be, nil)
 	c.Submit(context.Background(), "/i my prompt")
@@ -728,7 +738,8 @@ func TestCommand_Compact_WhileRunning(t *testing.T) {
 	unblock := make(chan struct{})
 	be := defaultBE()
 	be.respond = func(ctx context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return blockedStream(ctx, unblock), nil
+return blockedStream(ctx, unblock), nil
+
 	}
 	c := newCore(t, be, nil)
 
@@ -796,7 +807,8 @@ func TestCommand_Clear_WhileRunning(t *testing.T) {
 	unblock := make(chan struct{})
 	be := defaultBE()
 	be.respond = func(ctx context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return blockedStream(ctx, unblock), nil
+return blockedStream(ctx, unblock), nil
+
 	}
 	c := newCore(t, be, nil)
 
@@ -877,7 +889,8 @@ func TestRun_StreamInterrupted(t *testing.T) {
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		ch := make(chan backend.StreamEvent)
 		close(ch) // close without sending Done=true
-		return ch, nil
+return ch, nil
+
 	}
 	c := newCore(t, be, nil)
 	evs := collectEvents(context.Background(), c, "hello")
@@ -900,7 +913,8 @@ func TestRun_UnknownStopReason(t *testing.T) {
 		ch := make(chan backend.StreamEvent, 1)
 		ch <- backend.StreamEvent{Done: true, StopReason: "max_completion_tokens", Content: "partial"}
 		close(ch)
-		return ch, nil
+return ch, nil
+
 	}
 	c := newCore(t, be, nil)
 	evs := collectEvents(context.Background(), c, "hello")
@@ -930,9 +944,11 @@ func TestRun_ToolEmptyName(t *testing.T) {
 				StopReason: "tool_calls",
 			}
 			close(ch)
-			return ch, nil
+return ch, nil
+
 		}
-		return textStream("done"), nil
+return textStream("done"), nil
+
 	}
 	c := newCore(t, be, nil)
 	collectEvents(context.Background(), c, "run empty tool")
@@ -959,9 +975,11 @@ func TestRun_NoExec(t *testing.T) {
 				StopReason: "tool_calls",
 			}
 			close(ch)
-			return ch, nil
+return ch, nil
+
 		}
-		return textStream("done"), nil
+return textStream("done"), nil
+
 	}
 	c := newCore(t, be, nil)
 	c.runtime.Exec = nil // no executor
@@ -978,7 +996,7 @@ func TestHook_NonZeroExitCode(t *testing.T) {
 	be := defaultBE()
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		callCount++
-		return textStream("ok"), nil
+	return textStream("ok"), nil
 	}
 	// exit 1 is a non-blocking warning: turn should still proceed.
 	c := newCore(t, be, Hooks{HookPreTurn: []string{"exit 1"}})
@@ -1018,7 +1036,8 @@ func TestCommand_Backend_WhileRunning(t *testing.T) {
 	unblock := make(chan struct{})
 	be := defaultBE()
 	be.respond = func(ctx context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return blockedStream(ctx, unblock), nil
+return blockedStream(ctx, unblock), nil
+
 	}
 	c := newCore(t, be, nil)
 
@@ -1074,7 +1093,8 @@ func TestCommand_Model_WhileRunning(t *testing.T) {
 	unblock := make(chan struct{})
 	be := defaultBE()
 	be.respond = func(ctx context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return blockedStream(ctx, unblock), nil
+return blockedStream(ctx, unblock), nil
+
 	}
 	c := newCore(t, be, nil)
 
@@ -1156,7 +1176,8 @@ func TestCommand_Agent_WhileRunning(t *testing.T) {
 	unblock := make(chan struct{})
 	be := defaultBE()
 	be.respond = func(ctx context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return blockedStream(ctx, unblock), nil
+return blockedStream(ctx, unblock), nil
+
 	}
 	c := newCore(t, be, nil)
 
@@ -1448,7 +1469,7 @@ func TestRun_RateLimitRetry(t *testing.T) {
 		if callCount == 1 {
 			return nil, &backend.RateLimitError{RetryAfter: time.Millisecond}
 		}
-		return textStream("ok"), nil
+	return textStream("ok"), nil
 	}
 	c := newCore(t, be, nil)
 	var retryEvents []Event
@@ -1484,7 +1505,8 @@ func TestRun_ToolCancelledBeforeExec(t *testing.T) {
 			StopReason: "tool_calls",
 		}
 		close(ch)
-		return ch, nil
+return ch, nil
+
 	}
 	c := newCore(t, be, nil)
 	c.Submit(ctx, "hello")
@@ -1529,14 +1551,15 @@ func TestRun_ExecCancelledWithInject(t *testing.T) {
 			StopReason: "tool_calls",
 		}
 		close(ch)
-		return ch, nil
+return ch, nil
+
 	}
 	c := newCore(t, be, nil)
 	inject := "user interrupt"
 	c.pendingInject.Store(&inject)
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		cancel() // ctx cancelled during exec
-		return "", fmt.Errorf("exec cancelled")
+		return "", nil, fmt.Errorf("exec cancelled")
 	}
 	c.Submit(ctx, "hello")
 	if callCount != 1 {
@@ -1559,9 +1582,11 @@ func TestAutoCompact_WithHookContext(t *testing.T) {
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		callCount++
 		if callCount == 1 {
-			return textStream("summary text for compaction"), nil
+return textStream("summary text for compaction"), nil
+
 		}
-		return textStream("answer"), nil
+return textStream("answer"), nil
+
 	}
 	c.session = newSession("goal")
 	for range 15 {
@@ -1628,7 +1653,8 @@ func TestCore_SetGenerationParams_WhileRunning(t *testing.T) {
 	unblock := make(chan struct{})
 	be := defaultBE()
 	be.respond = func(ctx context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return blockedStream(ctx, unblock), nil
+return blockedStream(ctx, unblock), nil
+
 	}
 	c := newCore(t, be, nil)
 
@@ -1855,7 +1881,8 @@ func TestManualCompact_EmptySummary(t *testing.T) {
 		ch := make(chan backend.StreamEvent, 1)
 		ch <- backend.StreamEvent{Done: true, StopReason: "stop", Content: "   "}
 		close(ch)
-		return ch, nil
+return ch, nil
+
 	}
 	c.session = newSession("goal")
 	for i := range 15 {
@@ -1882,7 +1909,8 @@ func TestManualCompact_WithToolMessages(t *testing.T) {
 	be := defaultBE()
 	c := newCore(t, be, nil)
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return textStream("summary"), nil
+return textStream("summary"), nil
+
 	}
 	c.session = newSession("goal")
 	// Seed enough messages to pass the compaction threshold.
@@ -1992,7 +2020,7 @@ func TestCommand_Backend_Error(t *testing.T) {
 
 func TestExtractToolResult_Success(t *testing.T) {
 	raw := json.RawMessage(`{"isError":false,"content":[{"type":"text","text":"hello"}]}`)
-	text, isErr := extractToolResult(raw)
+	text, _, isErr := extractToolResult(raw)
 	if text != "hello" {
 		t.Errorf("text = %q; want %q", text, "hello")
 	}
@@ -2003,7 +2031,7 @@ func TestExtractToolResult_Success(t *testing.T) {
 
 func TestExtractToolResult_IsError(t *testing.T) {
 	raw := json.RawMessage(`{"isError":true,"content":[{"type":"text","text":"something failed"}]}`)
-	text, isErr := extractToolResult(raw)
+	text, _, isErr := extractToolResult(raw)
 	if text != "something failed" {
 		t.Errorf("text = %q; want %q", text, "something failed")
 	}
@@ -2014,7 +2042,7 @@ func TestExtractToolResult_IsError(t *testing.T) {
 
 func TestExtractToolResult_MultipleContentItems(t *testing.T) {
 	raw := json.RawMessage(`{"isError":false,"content":[{"type":"text","text":"a"},{"type":"text","text":"b"}]}`)
-	text, _ := extractToolResult(raw)
+	text, _, _ := extractToolResult(raw)
 	if text != "a\nb" {
 		t.Errorf("text = %q; want %q", text, "a\nb")
 	}
@@ -2022,15 +2050,33 @@ func TestExtractToolResult_MultipleContentItems(t *testing.T) {
 
 func TestExtractToolResult_NonTextItemsSkipped(t *testing.T) {
 	raw := json.RawMessage(`{"isError":false,"content":[{"type":"image","text":"ignored"},{"type":"text","text":"kept"}]}`)
-	text, _ := extractToolResult(raw)
+	text, _, _ := extractToolResult(raw)
 	if text != "kept" {
 		t.Errorf("text = %q; want %q", text, "kept")
 	}
 }
 
+func TestExtractToolResult_ImageBlock(t *testing.T) {
+	raw := json.RawMessage(`{"isError":false,"content":[{"type":"image","media_type":"image/png","data":"iVBOR"},{"type":"text","text":"screenshot above"}]}`)
+	text, blocks, isErr := extractToolResult(raw)
+	if text != "screenshot above" {
+		t.Errorf("text = %q; want %q", text, "screenshot above")
+	}
+	if isErr {
+		t.Error("isError should be false")
+	}
+	if len(blocks) != 1 {
+		t.Fatalf("blocks len = %d; want 1", len(blocks))
+	}
+	b := blocks[0]
+	if b.Type != "image" || b.ImageSource == nil || b.ImageSource.MediaType != "image/png" || b.ImageSource.Data != "iVBOR" {
+		t.Errorf("block = %+v", b)
+	}
+}
+
 func TestExtractToolResult_InvalidJSON(t *testing.T) {
 	raw := json.RawMessage(`not json`)
-	text, isErr := extractToolResult(raw)
+	text, _, isErr := extractToolResult(raw)
 	if text != "not json" {
 		t.Errorf("text = %q; want raw input on parse failure", text)
 	}
@@ -2297,7 +2343,7 @@ func TestBuildRuntime_ExecUnknownTool(t *testing.T) {
 	d := tools.NewDispatcher()
 	env := BuildRuntime(nil, d, t.TempDir(), nil)
 
-	_, err := env.Exec(context.Background(), "no_such_tool", json.RawMessage(`{}`))
+	_, _, err := env.Exec(context.Background(), "no_such_tool", json.RawMessage(`{}`))
 	if err == nil || !strings.Contains(err.Error(), "unknown tool") {
 		t.Errorf("expected unknown tool error; got %v", err)
 	}
@@ -2394,11 +2440,11 @@ func TestBuildRuntime_ExecDispatchSuccess(t *testing.T) {
 	d := &mockDispatcher{
 		tools: []tools.ToolInfo{{Server: "s1", Name: "mytool"}},
 		dispatch: func(ctx context.Context, server, tool string, args json.RawMessage) (json.RawMessage, error) {
-			return json.RawMessage(`{"content":[{"type":"text","text":"ok"}]}`), nil
+return json.RawMessage(`{"content":[{"type":"text","text":"ok"}]}`), nil
 		},
 	}
 	env := BuildRuntime(nil, d, t.TempDir(), nil)
-	result, err := env.Exec(context.Background(), "mytool", json.RawMessage(`{}`))
+	result, _, err := env.Exec(context.Background(), "mytool", json.RawMessage(`{}`))
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -2416,7 +2462,7 @@ func TestBuildRuntime_ExecDispatchError(t *testing.T) {
 		},
 	}
 	env := BuildRuntime(nil, d, t.TempDir(), nil)
-	_, err := env.Exec(context.Background(), "mytool", json.RawMessage(`{}`))
+	_, _, err := env.Exec(context.Background(), "mytool", json.RawMessage(`{}`))
 	if err == nil || !strings.Contains(err.Error(), "dispatch failed") {
 		t.Errorf("err = %v", err)
 	}
@@ -2427,11 +2473,11 @@ func TestBuildRuntime_ExecToolResultIsError(t *testing.T) {
 	d := &mockDispatcher{
 		tools: []tools.ToolInfo{{Server: "s1", Name: "mytool"}},
 		dispatch: func(ctx context.Context, server, tool string, args json.RawMessage) (json.RawMessage, error) {
-			return json.RawMessage(`{"isError":true,"content":[{"type":"text","text":"bad thing"}]}`), nil
+return json.RawMessage(`{"isError":true,"content":[{"type":"text","text":"bad thing"}]}`), nil
 		},
 	}
 	env := BuildRuntime(nil, d, t.TempDir(), nil)
-	_, err := env.Exec(context.Background(), "mytool", json.RawMessage(`{}`))
+	_, _, err := env.Exec(context.Background(), "mytool", json.RawMessage(`{}`))
 	if err == nil || !strings.Contains(err.Error(), "bad thing") {
 		t.Errorf("err = %v", err)
 	}

pkg/agent/loop.go

@@ -40,7 +40,7 @@ const stallThreshold = 5
 // task state re-injection into the conversation.
 const planReinjectInterval = 10
 
-type toolExecutor func(ctx context.Context, name string, args json.RawMessage) (string, error)
+type toolExecutor func(ctx context.Context, name string, args json.RawMessage) (string, []backend.ContentBlock, error)
 
 type agentConfig struct {
 	Backend              backend.Backend
@@ -287,9 +287,10 @@ func run(ctx context.Context, cfg agentConfig, state state) error {
 				}
 			}
 			var result string
+			var resultBlocks []backend.ContentBlock
 			var isErr bool
 			if cfg.Exec != nil {
-				out, err := cfg.Exec(ctx, tc.Name, tc.Arguments)
+				out, blocks, err := cfg.Exec(ctx, tc.Name, tc.Arguments)
 				if err != nil {
 					isErr = true
 					if ctx.Err() != nil {
@@ -305,6 +306,7 @@ func run(ctx context.Context, cfg agentConfig, state state) error {
 					result = fmt.Sprintf("error: %v", err)
 				} else {
 					result = out
+					resultBlocks = blocks
 				}
 			} else {
 				result = "error: no tool executor configured"
@@ -337,7 +339,7 @@ func run(ctx context.Context, cfg agentConfig, state state) error {
 			if !isErr && cfg.ClassifyTier != nil {
 				tier = cfg.ClassifyTier(tc.Name)
 			}
-			return toolResult{ToolCallID: tc.ID, Name: tc.Name, Content: result, IsError: isErr, Tier: tier}, false
+			return toolResult{ToolCallID: tc.ID, Name: tc.Name, Content: result, ContentBlocks: resultBlocks, IsError: isErr, Tier: tier}, false
 		}
 
 		isParallelSafe := func(name string) bool {

pkg/agent/loop_cache_test.go

@@ -21,9 +21,9 @@ func TestResultCache_HitSkipsExec(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ClassifyTool = func(string) bool { return true }
-	c.runtime.Exec = func(_ context.Context, name string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, name string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		atomic.AddInt32(&execCount, 1)
-		return "contents of a.txt", nil
+return "contents of a.txt", nil, nil
 	}
 
 	evs := collectEvents(context.Background(), c, "dup read")
@@ -54,9 +54,9 @@ func TestResultCache_DifferentArgsMiss(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ClassifyTool = func(string) bool { return true }
-	c.runtime.Exec = func(_ context.Context, name string, args json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, name string, args json.RawMessage) (string, []backend.ContentBlock, error) {
 		atomic.AddInt32(&execCount, 1)
-		return "result for " + string(args), nil
+return "result for " + string(args), nil, nil
 	}
 
 	collectEvents(context.Background(), c, "different args")
@@ -78,9 +78,9 @@ func TestResultCache_SerialToolNotCached(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ClassifyTool = func(name string) bool { return false } // all serial
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		atomic.AddInt32(&execCount, 1)
-		return "ok", nil
+return "ok", nil, nil
 	}
 
 	collectEvents(context.Background(), c, "serial tool")
@@ -101,9 +101,11 @@ func multiTurnToolsStream(turns [][]backend.ToolCall) func(context.Context, []ba
 			ch := make(chan backend.StreamEvent, 1)
 			ch <- backend.StreamEvent{ToolCalls: turns[i], Done: true, StopReason: "tool_calls"}
 			close(ch)
-			return ch, nil
+return ch, nil
+
 		}
-		return textStream("done"), nil
+return textStream("done"), nil
+
 	}
 }
 
@@ -121,12 +123,12 @@ func TestResultCache_ErrorNotCached(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ClassifyTool = func(string) bool { return true }
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		n := atomic.AddInt32(&execCount, 1)
 		if n == 1 {
-			return "", &mockErr{"transient failure"}
+			return "", nil, &mockErr{"transient failure"}
 		}
-		return "ok now", nil
+return "ok now", nil, nil
 	}
 
 	evs := collectEvents(context.Background(), c, "error then ok")
@@ -156,9 +158,9 @@ func TestResultCache_IdenticalParallelReads(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ClassifyTool = func(string) bool { return true } // both parallel-safe → same batch
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		atomic.AddInt32(&execCount, 1)
-		return "file contents", nil
+return "file contents", nil, nil
 	}
 
 	evs := collectEvents(context.Background(), c, "identical parallel")

pkg/agent/loop_errlimit_test.go

@@ -23,7 +23,8 @@ func alwaysFailStream() func(context.Context, []backend.Message, []backend.Tool,
 			Usage:      backend.Usage{InputTokens: 10, OutputTokens: 5},
 		}
 		close(ch)
-		return ch, nil
+return ch, nil
+
 	}
 }
 
@@ -32,8 +33,8 @@ func TestConsecutiveErrors_HardLimit(t *testing.T) {
 	be.respond = alwaysFailStream()
 
 	c := newCore(t, be, nil)
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
-		return "", fmt.Errorf("always fails")
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
+		return "", nil, fmt.Errorf("always fails")
 	}
 
 	evs := collectEvents(context.Background(), c, "do something")
@@ -62,7 +63,8 @@ func TestConsecutiveErrors_SoftLimitNudge(t *testing.T) {
 		if n > consecutiveErrorSoftLimit {
 			for _, m := range msgs {
 				if m.Role == "user" && strings.Contains(m.Content, "your last several tool calls all failed") {
-					return textStream("giving up"), nil
+return textStream("giving up"), nil
+
 				}
 			}
 			// Nudge not found — keep going (will hit hard limit if broken).
@@ -74,7 +76,8 @@ func TestConsecutiveErrors_SoftLimitNudge(t *testing.T) {
 				Usage:      backend.Usage{InputTokens: 10, OutputTokens: 5},
 			}
 			close(ch)
-			return ch, nil
+return ch, nil
+
 		}
 		ch := make(chan backend.StreamEvent, 1)
 		ch <- backend.StreamEvent{
@@ -84,12 +87,13 @@ func TestConsecutiveErrors_SoftLimitNudge(t *testing.T) {
 			Usage:      backend.Usage{InputTokens: 10, OutputTokens: 5},
 		}
 		close(ch)
-		return ch, nil
+return ch, nil
+
 	}
 
 	c := newCore(t, be, nil)
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
-		return "", fmt.Errorf("always fails")
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
+		return "", nil, fmt.Errorf("always fails")
 	}
 
 	evs := collectEvents(context.Background(), c, "do something")
@@ -124,10 +128,12 @@ func TestConsecutiveErrors_ResetOnSuccess(t *testing.T) {
 				Usage:      backend.Usage{InputTokens: 10, OutputTokens: 5},
 			}
 			close(ch)
-			return ch, nil
+return ch, nil
+
 		}
 		if n >= 11 {
-			return textStream("done"), nil
+return textStream("done"), nil
+
 		}
 		ch := make(chan backend.StreamEvent, 1)
 		ch <- backend.StreamEvent{
@@ -137,15 +143,16 @@ func TestConsecutiveErrors_ResetOnSuccess(t *testing.T) {
 			Usage:      backend.Usage{InputTokens: 10, OutputTokens: 5},
 		}
 		close(ch)
-		return ch, nil
+return ch, nil
+
 	}
 
 	c := newCore(t, be, nil)
-	c.runtime.Exec = func(_ context.Context, name string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, name string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		if name == "good_tool" {
-			return "ok", nil
+return "ok", nil, nil
 		}
-		return "", fmt.Errorf("fails")
+		return "", nil, fmt.Errorf("fails")
 	}
 
 	evs := collectEvents(context.Background(), c, "do something")

pkg/agent/loop_maxsteps_helpers_test.go

@@ -21,6 +21,7 @@ func sequentialStream(responses []mockResponse) func(context.Context, []backend.
 		i := int(atomic.AddInt32(&n, 1)) - 1
 		if i >= len(responses) {
 			return textStream("done"), nil
+
 		}
 		r := responses[i]
 		ch := make(chan backend.StreamEvent, 1)
@@ -32,6 +33,7 @@ func sequentialStream(responses []mockResponse) func(context.Context, []backend.
 		}
 		close(ch)
 		return ch, nil
+
 	}
 }
 

pkg/agent/loop_maxsteps_test.go

@@ -25,9 +25,9 @@ func TestMaxStepsZeroUnlimited(t *testing.T) {
 		Backend:  mb,
 		Tools:    []backend.Tool{{Name: "tool"}},
 		MaxSteps: 0,
-		Exec: func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
+		Exec: func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 			steps++
-			return "ok", nil
+			return "ok", nil, nil
 		},
 	}
 	if err := run(context.Background(), cfg, newState()); err != nil {
@@ -60,8 +60,8 @@ func TestMaxStepsSoftNudge(t *testing.T) {
 		Backend:  mb,
 		Tools:    []backend.Tool{{Name: "tool"}},
 		MaxSteps: 1,
-		Exec: func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
-			return "ok", nil
+		Exec: func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
+			return "ok", nil, nil
 		},
 		Output: func(ev Event) {
 			if ev.Role == "maxsteps" {
@@ -117,9 +117,9 @@ func TestMaxStepsExactBoundary(t *testing.T) {
 		Backend:  mb,
 		Tools:    []backend.Tool{{Name: "tool"}},
 		MaxSteps: limit,
-		Exec: func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
+		Exec: func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 			toolRounds++
-			return "ok", nil
+			return "ok", nil, nil
 		},
 	}
 	if err := run(context.Background(), cfg, newState()); err != nil {
@@ -149,8 +149,8 @@ func TestTaskStateReinjection(t *testing.T) {
 	cfg := agentConfig{
 		Backend: mb,
 		Tools:   []backend.Tool{{Name: "tool"}},
-		Exec: func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
-			return "ok", nil
+		Exec: func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
+			return "ok", nil, nil
 		},
 	}
 

pkg/agent/loop_parallel_test.go

@@ -21,9 +21,11 @@ func toolsStream(calls []backend.ToolCall) func(context.Context, []backend.Messa
 			ch := make(chan backend.StreamEvent, 1)
 			ch <- backend.StreamEvent{ToolCalls: calls, Done: true, StopReason: "tool_calls"}
 			close(ch)
-			return ch, nil
+return ch, nil
+
 		}
-		return textStream("done"), nil
+return textStream("done"), nil
+
 	}
 }
 
@@ -51,13 +53,13 @@ func TestParallel_ConcurrentExecution(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ClassifyTool = func(string) bool { return true }
-	c.runtime.Exec = func(ctx context.Context, name string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(ctx context.Context, name string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		started <- struct{}{}
 		select {
 		case <-gate:
-			return name + "-result", nil
+return name + "-result", nil, nil
 		case <-ctx.Done():
-			return "", ctx.Err()
+			return "", nil, ctx.Err()
 		}
 	}
 
@@ -93,11 +95,11 @@ func TestParallel_SerialToolBreaksBatch(t *testing.T) {
 	c.runtime.ClassifyTool = func(name string) bool {
 		return name == "read_a" || name == "read_c"
 	}
-	c.runtime.Exec = func(_ context.Context, name string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, name string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		mu.Lock()
 		order = append(order, name)
 		mu.Unlock()
-		return name + "-result", nil
+return name + "-result", nil, nil
 	}
 
 	collectEvents(context.Background(), c, "mixed tools")
@@ -134,10 +136,10 @@ func TestParallel_CancellationFillsRemaining(t *testing.T) {
 	c.runtime.ClassifyTool = func(name string) bool {
 		return name == "tool_a" || name == "tool_b"
 	}
-	c.runtime.Exec = func(execCtx context.Context, name string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(execCtx context.Context, name string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		cancel() // cancel on first execution; propagates to all
 		<-execCtx.Done()
-		return "", execCtx.Err()
+		return "", nil, execCtx.Err()
 	}
 
 	evs := collectEvents(ctx, c, "cancel mid-batch")
@@ -175,11 +177,11 @@ func TestParallel_NilClassifyToolIsSerial(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ClassifyTool = nil // no classifier → all serial
-	c.runtime.Exec = func(_ context.Context, name string, _ json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, name string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
 		mu.Lock()
 		order = append(order, name)
 		mu.Unlock()
-		return name + "-result", nil
+return name + "-result", nil, nil
 	}
 
 	evs := collectEvents(context.Background(), c, "serial fallback")

pkg/agent/loop_texttool_test.go

@@ -18,10 +18,12 @@ func TestTextToolCall_ParsedAndExecuted(t *testing.T) {
 		for _, m := range msgs {
 			if m.Role == "tool" {
 				// Second call after tool result: respond normally
-				return textStream("done"), nil
+return textStream("done"), nil
+
 			}
 		}
-		return textStream(`call_tool: {"calls":[{"tool":"file_read","args":["/tmp/test.go"]}]}`), nil
+return textStream(`call_tool: {"calls":[{"tool":"file_read","args":["/tmp/test.go"]}]}`), nil
+
 	}
 
 	var execCalled bool
@@ -30,12 +32,12 @@ func TestTextToolCall_ParsedAndExecuted(t *testing.T) {
 		{Name: "call_tool", Description: "run tools"},
 		{Name: "execute_code", Description: "run code"},
 	}
-	c.runtime.Exec = func(_ context.Context, name string, args json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, name string, args json.RawMessage) (string, []backend.ContentBlock, error) {
 		execCalled = true
 		if name != "call_tool" {
 			t.Errorf("expected tool name 'call_tool', got %q", name)
 		}
-		return "file contents here", nil
+return "file contents here", nil, nil
 	}
 
 	evs := collectEvents(context.Background(), c, "read a file")
@@ -55,10 +57,12 @@ func TestTextToolCall_RelaxedJSON(t *testing.T) {
 	be.respond = func(_ context.Context, msgs []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
 		for _, m := range msgs {
 			if m.Role == "tool" {
-				return textStream("done"), nil
+return textStream("done"), nil
+
 			}
 		}
-		return textStream(`execute_code: steps=[{code: "date"}]`), nil
+return textStream(`execute_code: steps=[{code: "date"}]`), nil
+
 	}
 
 	var execCalled bool
@@ -67,12 +71,12 @@ func TestTextToolCall_RelaxedJSON(t *testing.T) {
 		{Name: "execute_code", Description: "run code"},
 		{Name: "call_tool", Description: "run tools"},
 	}
-	c.runtime.Exec = func(_ context.Context, name string, args json.RawMessage) (string, error) {
+	c.runtime.Exec = func(_ context.Context, name string, args json.RawMessage) (string, []backend.ContentBlock, error) {
 		execCalled = true
 		if name != "execute_code" {
 			t.Errorf("expected tool name 'execute_code', got %q", name)
 		}
-		return "Sat May 23 11:00:00 UTC 2026", nil
+return "Sat May 23 11:00:00 UTC 2026", nil, nil
 	}
 
 	collectEvents(context.Background(), c, "what time is it")
@@ -87,7 +91,8 @@ func TestTextToolCall_RelaxedJSON(t *testing.T) {
 func TestTextToolCall_NoFalsePositive(t *testing.T) {
 	be := defaultBE()
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return textStream("Here is how execute_code works in general."), nil
+return textStream("Here is how execute_code works in general."), nil
+
 	}
 
 	c := newCore(t, be, nil)
@@ -110,7 +115,8 @@ func TestTextToolCall_NoFalsePositive(t *testing.T) {
 func TestTextToolCall_NoToolsConfigured(t *testing.T) {
 	be := defaultBE()
 	be.respond = func(_ context.Context, _ []backend.Message, _ []backend.Tool, _ backend.GenerationParams) (<-chan backend.StreamEvent, error) {
-		return textStream(`execute_code: steps=[{code: "print('hi')"}]`), nil
+return textStream(`execute_code: steps=[{code: "print('hi')"}]`), nil
+
 	}
 
 	c := newCore(t, be, nil)

pkg/agent/loop_truncate_test.go

@@ -23,8 +23,8 @@ func TestTruncation_LargeResultTruncated(t *testing.T) {
 	c := newCore(t, be, nil)
 	c.runtime.ClassifyTool = nil
 	c.runtime.ToolResultMaxBytes = maxBytes
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
-		return large, nil
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
+return large, nil, nil
 	}
 
 	evs := collectEvents(context.Background(), c, "big result")
@@ -55,8 +55,8 @@ func TestTruncation_SmallResultNotTruncated(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ToolResultMaxBytes = 8192
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
-		return "short result", nil
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
+return "short result", nil, nil
 	}
 
 	evs := collectEvents(context.Background(), c, "small result")
@@ -82,8 +82,8 @@ func TestTruncation_ZeroDisabled(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ToolResultMaxBytes = 0 // disabled
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
-		return large, nil
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
+return large, nil, nil
 	}
 
 	evs := collectEvents(context.Background(), c, "no limit")
@@ -110,8 +110,8 @@ func TestTruncation_ErrorNotTruncated(t *testing.T) {
 
 	c := newCore(t, be, nil)
 	c.runtime.ToolResultMaxBytes = maxBytes
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
-		return "", &mockErr{longErr}
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
+		return "", nil, &mockErr{longErr}
 	}
 
 	evs := collectEvents(context.Background(), c, "error result")
@@ -140,8 +140,8 @@ func TestTruncation_CachedResultAlreadyTruncated(t *testing.T) {
 	c := newCore(t, be, nil)
 	c.runtime.ClassifyTool = func(string) bool { return true } // read-safe → cacheable
 	c.runtime.ToolResultMaxBytes = maxBytes
-	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
-		return large, nil
+	c.runtime.Exec = func(_ context.Context, _ string, _ json.RawMessage) (string, []backend.ContentBlock, error) {
+return large, nil, nil
 	}
 
 	evs := collectEvents(context.Background(), c, "cached truncated")

pkg/agent/loop_turnerror_test.go

@@ -24,6 +24,7 @@ func errThenOKStream(err error) func(context.Context, []backend.Message, []backe
 			return nil, err
 		}
 		return textStream("recovered"), nil
+
 	}
 }
 

pkg/agent/session.go

@@ -218,9 +218,10 @@ func (s *Session) update(assistant backend.Message, results []toolResult) {
 			content = summarizeToolResult(r.Name, content)
 		}
 		s.messages = append(s.messages, backend.Message{
-			Role:       "tool",
-			Content:    content,
-			ToolCallID: r.ToolCallID,
+			Role:          "tool",
+			Content:       content,
+			ContentBlocks: r.ContentBlocks,
+			ToolCallID:    r.ToolCallID,
 		})
 	}
 }

pkg/agent/state.go

@@ -13,11 +13,12 @@ const (
 
 // toolResult holds the output of a single tool call.
 type toolResult struct {
-	ToolCallID string // may be empty (Ollama does not always set this)
-	Name       string
-	Content    string
-	IsError    bool
-	Tier       ResultTier // default TierHot; tools may self-classify
+	ToolCallID    string // may be empty (Ollama does not always set this)
+	Name          string
+	Content       string
+	ContentBlocks []backend.ContentBlock // image blocks extracted from tool output
+	IsError       bool
+	Tier          ResultTier // default TierHot; tools may self-classify
 }
 
 // state is the interface both ephemeral and bead-backed state must satisfy.