run/executor.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"strings"
 	"time"
 
 	"gitea.stevedudenhoeffer.com/steve/majordomo/agent"
@@ -320,7 +321,7 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 	// One WithSteer drains BOTH the session mailbox (a tool's AttachImages) and
 	// the critic's nudges before each step.
 	steer := func() []llm.Message { return append(mailbox.drain(), critic.drainSteer()...) }
-	runRes, runErr := ag.Run(runCtx, input, agent.WithSteer(steer))
+	runRes, runErr := runAgent(runCtx, ag, input, inv.Images, agent.WithSteer(steer))
 
 	status := statusFor(runCtx, runErr)
 	if runRes != nil {
@@ -440,3 +441,28 @@ func detach(ctx context.Context) context.Context {
 	_ = cancel // bounded by the timeout; nothing to cancel early
 	return c
 }
+
+// runAgent dispatches the majordomo agent loop. majordomo's Run takes a text-only
+// input arg, so when the invocation carries images they're folded into the first
+// user message (text + image parts) via WithHistory and Run is called with an
+// empty input — the model then sees a multimodal opening turn. The image-less path
+// passes the prompt straight through.
+//
+// The text part is omitted when input is blank (image-only run), matching
+// runSession.AttachImages so no empty TextPart is sent.
+func runAgent(ctx context.Context, ag *agent.Agent, input string, images []llm.ImagePart, opts ...agent.RunOption) (*agent.Result, error) {
+	if len(images) == 0 {
+		return ag.Run(ctx, input, opts...)
+	}
+	parts := make([]llm.Part, 0, len(images)+1)
+	if strings.TrimSpace(input) != "" {
+		parts = append(parts, llm.Text(input))
+	}
+	for _, img := range images {
+		parts = append(parts, img)
+	}
+	// Copy opts before appending so a caller-supplied backing array is never
+	// mutated/aliased (the variadic slice can have spare capacity).
+	opts = append(opts[:len(opts):len(opts)], agent.WithHistory([]llm.Message{llm.UserParts(parts...)}))
+	return ag.Run(ctx, "", opts...)
+}

run/images_test.go

@@ -0,0 +1,121 @@
+package run_test
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/fake"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/run"
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// TestExecutorFoldsInitialImages: when the invocation carries Images, they're
+// folded into the first user message (alongside the prompt text) instead of being
+// dropped — majordomo's Run input arg is text-only, so the executor seeds the
+// multimodal opening turn via history.
+func TestExecutorFoldsInitialImages(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("saw the image"))
+	m, _ := fp.Model("m")
+
+	img := llm.ImagePart{MIME: "image/png", Data: []byte("PNGDATA")}
+	inv := tool.Invocation{RunID: "r1", Images: []llm.ImagePart{img}}
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   func(ctx context.Context, _ string) (context.Context, llm.Model, error) { return ctx, m, nil },
+	})
+	res := ex.Run(context.Background(), run.RunnableAgent{ModelTier: "m"}, inv, "describe this")
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+
+	calls := fp.Calls()
+	if len(calls) == 0 {
+		t.Fatal("no model calls recorded")
+	}
+	// The text + image must be CO-LOCATED in a single user message (not split
+	// across two), so the model reads them as one multimodal turn.
+	coLocated := false
+	for _, msg := range calls[0].Request.Messages {
+		sawImage, sawText := false, false
+		for _, p := range msg.Parts {
+			switch pp := p.(type) {
+			case llm.ImagePart:
+				if string(pp.Data) == "PNGDATA" {
+					sawImage = true
+				}
+			case llm.TextPart:
+				if strings.Contains(pp.Text, "describe this") {
+					sawText = true
+				}
+			}
+		}
+		if sawImage && sawText {
+			coLocated = true
+		}
+	}
+	if !coLocated {
+		t.Error("image + prompt text were not folded into the SAME user message")
+	}
+}
+
+// TestExecutorImageOnlyNoBlankText: an image-only run (blank prompt) must NOT emit
+// an empty TextPart — the message carries just the image, matching
+// runSession.AttachImages's guard.
+func TestExecutorImageOnlyNoBlankText(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("saw it"))
+	m, _ := fp.Model("m")
+
+	inv := tool.Invocation{RunID: "r3", Images: []llm.ImagePart{{MIME: "image/png", Data: []byte("IMG")}}}
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   func(ctx context.Context, _ string) (context.Context, llm.Model, error) { return ctx, m, nil },
+	})
+	res := ex.Run(context.Background(), run.RunnableAgent{ModelTier: "m"}, inv, "   ")
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	for _, msg := range fp.Calls()[0].Request.Messages {
+		for _, p := range msg.Parts {
+			if tp, ok := p.(llm.TextPart); ok && strings.TrimSpace(tp.Text) == "" {
+				t.Error("image-only run emitted a blank TextPart")
+			}
+		}
+	}
+}
+
+// TestExecutorTextOnlyUnchanged: with no Images, the prompt flows through as the
+// text input (regression guard that the fold path didn't break the common case).
+func TestExecutorTextOnlyUnchanged(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("ok"))
+	m, _ := fp.Model("m")
+
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   func(ctx context.Context, _ string) (context.Context, llm.Model, error) { return ctx, m, nil },
+	})
+	res := ex.Run(context.Background(), run.RunnableAgent{ModelTier: "m"}, tool.Invocation{RunID: "r2"}, "plain prompt")
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	calls := fp.Calls()
+	if len(calls) == 0 {
+		t.Fatal("no model calls recorded")
+	}
+	sawText := false
+	for _, msg := range calls[0].Request.Messages {
+		for _, p := range msg.Parts {
+			if tp, ok := p.(llm.TextPart); ok && strings.Contains(tp.Text, "plain prompt") {
+				sawText = true
+			}
+		}
+	}
+	if !sawText {
+		t.Error("text-only prompt did not reach the model")
+	}
+}