feat(#3996): fall back to an explicit user-prompt filename for one unnamed image

pkg/chat/media.go

@@ -23,14 +23,17 @@ type MediaDelta struct {
 	// runtime accumulator synthesizes one when needed.
 	Name string `json:"name,omitempty"`
 
-	// RequestedPath is the prompt-directed target path the model asked for
-	// (e.g. echoed from an "as sunshine.jpg" instruction), when one exists.
-	// It is untrusted model input: the runtime routes it through
+	// RequestedPath is the prompt-directed target path for this blob, when
+	// one exists. It is untrusted input: the runtime routes it through
 	// workspacemedia.ClassifyRequestedPath, and a path escaping the workspace
-	// requires an explicit user confirmation before it is honored. Response
-	// marker extraction (the "[media-file: ...]" protocol) will populate it;
-	// until that lands, providers leave it empty and materialization falls
-	// back to Name.
+	// requires an explicit user confirmation before it is honored. The
+	// runtime's response marker filter (the "[media-file: ...]" protocol,
+	// pkg/runtime/generated_media_markers.go) populates it by pairing marker
+	// paths with blobs in response order; a single otherwise-unnamed blob may
+	// instead get it from deterministic explicit-filename extraction on the
+	// triggering user message (pkg/runtime/generated_media_prompt_filename.go).
+	// Blobs neither source names keep it empty and materialization falls back
+	// to Name, then a generic name.
 	RequestedPath string `json:"requested_path,omitempty"`
 
 	// Size is the byte length of Data, cached because Data itself is

pkg/model/provider/gemini/client.go

@@ -761,6 +761,7 @@ func (c *Client) CreateChatCompletionStream(
 
 	if c.wantsImageResponseModalities(imageOutputEnabled) {
 		config.ResponseModalities = []string{string(genai.ModalityText), string(genai.ModalityImage)}
+		applyImageOutputMediaFileInstruction(config)
 	}
 
 	// Start with Google built-in tools (search, maps, code execution) from provider_opts

pkg/model/provider/gemini/image_output_instruction.go

@@ -0,0 +1,31 @@
+package gemini
+
+import "google.golang.org/genai"
+
+// imageOutputMediaFileInstruction is appended as a system instruction on
+// declared image-output chat requests (see wantsImageResponseModalities) so
+// generated images arrive with a machine-readable filename: the runtime
+// strips these exact marker lines from the reply and uses the paths to name
+// the materialized workspace files (pkg/runtime/generated_media_markers.go).
+// The single-image steering keeps one request yielding one predictably named
+// file; every blob the model actually returns is still persisted.
+const imageOutputMediaFileInstruction = `When you generate images, name each one with a marker line, placed alone on its own line, in this exact format:
+[media-file: relative/path.ext]
+Rules:
+- Emit exactly one marker line per generated image, in the same order as the images.
+- If the user asked for a specific file name or path, echo it in the marker exactly as requested.
+- Otherwise choose a short, meaningful, kebab-case file name.
+- Generate a single image unless the user explicitly asks for multiple images or variations.
+- Never emit a marker line for an image you did not generate.`
+
+// applyImageOutputMediaFileInstruction appends the marker-protocol
+// instruction to the request's system instruction, preserving any parts
+// already present. Callers gate it on wantsImageResponseModalities so only
+// declared image-output chat requests carry it.
+func applyImageOutputMediaFileInstruction(config *genai.GenerateContentConfig) {
+	if config.SystemInstruction == nil {
+		config.SystemInstruction = &genai.Content{}
+	}
+	config.SystemInstruction.Parts = append(config.SystemInstruction.Parts,
+		genai.NewPartFromText(imageOutputMediaFileInstruction))
+}

pkg/model/provider/gemini/image_output_instruction_test.go

@@ -0,0 +1,188 @@
+package gemini
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/docker/docker-agent/pkg/chat"
+	"github.com/docker/docker-agent/pkg/config/latest"
+	"github.com/docker/docker-agent/pkg/environment"
+	"github.com/docker/docker-agent/pkg/model/provider/options"
+)
+
+// systemInstructionTextsInBody decodes body's systemInstruction part texts
+// (genai serializes GenerateContentConfig.SystemInstruction under the
+// top-level "systemInstruction" key for the Gemini Developer API), returning
+// nil when the key is absent.
+func systemInstructionTextsInBody(t *testing.T, body []byte) []string {
+	t.Helper()
+
+	var req map[string]any
+	require.NoError(t, json.Unmarshal(body, &req))
+
+	si, ok := req["systemInstruction"].(map[string]any)
+	if !ok {
+		return nil
+	}
+	parts, ok := si["parts"].([]any)
+	if !ok {
+		return nil
+	}
+	var out []string
+	for _, p := range parts {
+		partMap, ok := p.(map[string]any)
+		if !ok {
+			continue
+		}
+		if text, ok := partMap["text"].(string); ok {
+			out = append(out, text)
+		}
+	}
+	return out
+}
+
+// TestCreateChatCompletionStream_MediaFileInstruction_PositiveRoutes pins
+// that ordinary image-output chat requests carry the media-file marker
+// instruction exactly once on every supported Google surface.
+func TestCreateChatCompletionStream_MediaFileInstruction_PositiveRoutes(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name    string
+		cfg     func(serverURL string) *latest.ModelConfig
+		env     map[string]string
+		gateway bool
+	}{
+		{
+			name: "gateway",
+			cfg: func(string) *latest.ModelConfig {
+				return &latest.ModelConfig{Provider: "google", Model: "gemini-2.5-flash-image", OutputCapabilities: &latest.OutputCapabilitiesConfig{Image: new(true)}}
+			},
+			env:     map[string]string{environment.DockerDesktopTokenEnv: "test-dd-token"},
+			gateway: true,
+		},
+		{
+			name: "direct Gemini API",
+			cfg: func(serverURL string) *latest.ModelConfig {
+				return &latest.ModelConfig{Provider: "google", Model: "gemini-2.5-flash-image", BaseURL: serverURL, OutputCapabilities: &latest.OutputCapabilitiesConfig{Image: new(true)}}
+			},
+			env: map[string]string{"GOOGLE_API_KEY": "test-key"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			server, captured := newBodyCapturingGeminiServer(t, writeGeminiSSEResponse)
+			opts := []options.Opt(nil)
+			if tt.gateway {
+				opts = append(opts, options.WithGateway(server.URL))
+			}
+			client, err := NewClient(t.Context(), tt.cfg(server.URL), environment.NewMapEnvProvider(tt.env), opts...)
+			require.NoError(t, err)
+
+			stream, err := client.CreateChatCompletionStream(t.Context(), []chat.Message{{Role: chat.MessageRoleUser, Content: "generate an image of a red panda"}}, nil)
+			require.NoError(t, err)
+			drainStream(t, stream)
+
+			bodies := captured.all()
+			require.Len(t, bodies, 1)
+			texts := systemInstructionTextsInBody(t, bodies[0])
+			require.Len(t, texts, 1, "the instruction must be sent exactly once")
+			assert.Equal(t, imageOutputMediaFileInstruction, texts[0])
+			assert.Equal(t, 1, strings.Count(texts[0], "[media-file: "), "the instruction must show the marker format exactly once")
+		})
+	}
+}
+
+// TestCreateChatCompletionStream_MediaFileInstruction_AbsentOnOtherRoutes
+// pins that non-image-output and internal text-only requests send no marker
+// instruction: gateway calls without image output enabled, plus gateway
+// title-generation and compaction calls.
+func TestCreateChatCompletionStream_MediaFileInstruction_AbsentOnOtherRoutes(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name    string
+		cfg     func(serverURL string) *latest.ModelConfig
+		env     map[string]string
+		gateway bool
+		opts    []options.Opt
+	}{
+		{
+			name: "gateway, declared false: absent",
+			cfg: func(string) *latest.ModelConfig {
+				return &latest.ModelConfig{
+					Provider: "google", Model: "gemini-2.5-flash-image",
+					OutputCapabilities: &latest.OutputCapabilitiesConfig{Image: new(false)},
+				}
+			},
+			env:     map[string]string{environment.DockerDesktopTokenEnv: "test-dd-token"},
+			gateway: true,
+		},
+		{
+			name: "gateway, declaration missing: absent",
+			cfg: func(string) *latest.ModelConfig {
+				return &latest.ModelConfig{Provider: "google", Model: "gemini-2.5-flash-image"}
+			},
+			env:     map[string]string{environment.DockerDesktopTokenEnv: "test-dd-token"},
+			gateway: true,
+		},
+		{
+			name: "gateway, declared true, generating title: absent",
+			cfg: func(string) *latest.ModelConfig {
+				return &latest.ModelConfig{
+					Provider: "google", Model: "gemini-2.5-flash-image",
+					OutputCapabilities: &latest.OutputCapabilitiesConfig{Image: new(true)},
+				}
+			},
+			env:     map[string]string{environment.DockerDesktopTokenEnv: "test-dd-token"},
+			gateway: true,
+			opts:    []options.Opt{options.WithGeneratingTitle()},
+		},
+		{
+			name: "gateway, declared true, compacting: absent",
+			cfg: func(string) *latest.ModelConfig {
+				return &latest.ModelConfig{
+					Provider: "google", Model: "gemini-2.5-flash-image",
+					OutputCapabilities: &latest.OutputCapabilitiesConfig{Image: new(true)},
+				}
+			},
+			env:     map[string]string{environment.DockerDesktopTokenEnv: "test-dd-token"},
+			gateway: true,
+			opts:    []options.Opt{options.WithCompacting()},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			server, captured := newBodyCapturingGeminiServer(t, writeGeminiSSEResponse)
+
+			cfg := tt.cfg(server.URL)
+			env := environment.NewMapEnvProvider(tt.env)
+			opts := tt.opts
+			if tt.gateway {
+				opts = append(opts, options.WithGateway(server.URL))
+			}
+			client, err := NewClient(t.Context(), cfg, env, opts...)
+			require.NoError(t, err)
+
+			stream, err := client.CreateChatCompletionStream(t.Context(), []chat.Message{
+				{Role: chat.MessageRoleUser, Content: "hello"},
+			}, nil)
+			require.NoError(t, err)
+			drainStream(t, stream)
+
+			bodies := captured.all()
+			require.Len(t, bodies, 1)
+			assert.Nil(t, systemInstructionTextsInBody(t, bodies[0]), "the marker instruction must be absent on this route")
+		})
+	}
+}