Merge pull request #1020 from shikihane/feat/agent-vision-pipeline-v2

feat(agent): add vision/image support with streaming base64 and filetype detection

go.mod

@@ -37,6 +37,8 @@ require (
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/elliotchance/orderedmap/v3 v3.1.0 // indirect
 	github.com/gdamore/encoding v1.0.1 // indirect
+	github.com/gdamore/tcell/v2 v2.13.8 // indirect
+	github.com/h2non/filetype v1.1.3 // indirect
 	github.com/inconshreveable/mousetrap v1.1.0 // indirect
 	github.com/lucasb-eyer/go-colorful v1.3.0 // indirect
 	github.com/mattn/go-colorable v0.1.14 // indirect

go.sum

@@ -98,6 +98,8 @@ github.com/gorilla/websocket v1.5.3 h1:saDtZ6Pbx/0u+bgYQ3q96pZgCzfhKXGPqt7kZ72aN
 github.com/gorilla/websocket v1.5.3/go.mod h1:YR8l580nyteQvAITg2hZ9XVh4b55+EU/adAjf1fMHhE=
 github.com/grbit/go-json v0.11.0 h1:bAbyMdYrYl/OjYsSqLH99N2DyQ291mHy726Mx+sYrnc=
 github.com/grbit/go-json v0.11.0/go.mod h1:IYpHsdybQ386+6g3VE6AXQ3uTGa5mquBme5/ZWmtzek=
+github.com/h2non/filetype v1.1.3 h1:FKkx9QbD7HR/zjK1Ia5XiBsq9zdLi5Kf3zGyFTAFkGg=
+github.com/h2non/filetype v1.1.3/go.mod h1:319b3zT68BvV+WRj7cwy856M2ehB3HqNOt6sy1HndBY=
 github.com/hashicorp/golang-lru/v2 v2.0.7 h1:a+bsQ5rvGLjzHuww6tVxozPZFVghXaHOwFs4luLUK2k=
 github.com/hashicorp/golang-lru/v2 v2.0.7/go.mod h1:QeFd9opnmA6QUJc5vARoKUSoFhyfM2/ZepoAG6RGpeM=
 github.com/hpcloud/tail v1.0.0/go.mod h1:ab1qPbhIpdTxEkNHXyeSf5vhxWSCs/tWer42PpOxQnU=

pkg/agent/context.go

@@ -527,10 +527,14 @@ func (cb *ContextBuilder) BuildMessages(
 
 	// Add current user message
 	if strings.TrimSpace(currentMessage) != "" {
-		messages = append(messages, providers.Message{
+		msg := providers.Message{
 			Role:    "user",
 			Content: currentMessage,
-		})
+		}
+		if len(media) > 0 {
+			msg.Media = media
+		}
+		messages = append(messages, msg)
 	}
 
 	return messages

pkg/agent/loop.go

@@ -47,14 +47,15 @@ type AgentLoop struct {
 
 // processOptions configures how a message is processed
 type processOptions struct {
-	SessionKey      string // Session identifier for history/context
-	Channel         string // Target channel for tool execution
-	ChatID          string // Target chat ID for tool execution
-	UserMessage     string // User message content (may include prefix)
-	DefaultResponse string // Response when LLM returns empty
-	EnableSummary   bool   // Whether to trigger summarization
-	SendResponse    bool   // Whether to send response via bus
-	NoHistory       bool   // If true, don't load session history (for heartbeat)
+	SessionKey      string   // Session identifier for history/context
+	Channel         string   // Target channel for tool execution
+	ChatID          string   // Target chat ID for tool execution
+	UserMessage     string   // User message content (may include prefix)
+	Media           []string // media:// refs from inbound message
+	DefaultResponse string   // Response when LLM returns empty
+	EnableSummary   bool     // Whether to trigger summarization
+	SendResponse    bool     // Whether to send response via bus
+	NoHistory       bool     // If true, don't load session history (for heartbeat)
 }
 
 const defaultResponse = "I've completed processing but have no response to give. Increase `max_tool_iterations` in config.json."
@@ -497,6 +498,7 @@ func (al *AgentLoop) processMessage(ctx context.Context, msg bus.InboundMessage)
 		Channel:         msg.Channel,
 		ChatID:          msg.ChatID,
 		UserMessage:     msg.Content,
+		Media:           msg.Media,
 		DefaultResponse: defaultResponse,
 		EnableSummary:   true,
 		SendResponse:    false,
@@ -603,11 +605,15 @@ func (al *AgentLoop) runAgentLoop(
 		history,
 		summary,
 		opts.UserMessage,
-		nil,
+		opts.Media,
 		opts.Channel,
 		opts.ChatID,
 	)
 
+	// Resolve media:// refs to base64 data URLs (streaming)
+	maxMediaSize := al.cfg.Agents.Defaults.GetMaxMediaSize()
+	messages = resolveMediaRefs(messages, al.mediaStore, maxMediaSize)
+
 	// 3. Save user message to session
 	agent.Sessions.AddMessage(opts.SessionKey, "user", opts.UserMessage)
 

pkg/agent/loop_media.go

@@ -0,0 +1,122 @@
+// PicoClaw - Ultra-lightweight personal AI agent
+// Inspired by and based on nanobot: https://github.com/HKUDS/nanobot
+// License: MIT
+//
+// Copyright (c) 2026 PicoClaw contributors
+
+package agent
+
+import (
+	"bytes"
+	"encoding/base64"
+	"io"
+	"os"
+	"strings"
+
+	"github.com/h2non/filetype"
+
+	"github.com/sipeed/picoclaw/pkg/logger"
+	"github.com/sipeed/picoclaw/pkg/media"
+	"github.com/sipeed/picoclaw/pkg/providers"
+)
+
+// resolveMediaRefs replaces media:// refs in message Media fields with base64 data URLs.
+// Uses streaming base64 encoding (file handle → encoder → buffer) to avoid holding
+// both raw bytes and encoded string in memory simultaneously.
+// Returns a new slice; original messages are not mutated.
+func resolveMediaRefs(messages []providers.Message, store media.MediaStore, maxSize int) []providers.Message {
+	if store == nil {
+		return messages
+	}
+
+	result := make([]providers.Message, len(messages))
+	copy(result, messages)
+
+	for i, m := range result {
+		if len(m.Media) == 0 {
+			continue
+		}
+
+		resolved := make([]string, 0, len(m.Media))
+		for _, ref := range m.Media {
+			if !strings.HasPrefix(ref, "media://") {
+				resolved = append(resolved, ref)
+				continue
+			}
+
+			localPath, meta, err := store.ResolveWithMeta(ref)
+			if err != nil {
+				logger.WarnCF("agent", "Failed to resolve media ref", map[string]any{
+					"ref":   ref,
+					"error": err.Error(),
+				})
+				continue
+			}
+
+			info, err := os.Stat(localPath)
+			if err != nil {
+				logger.WarnCF("agent", "Failed to stat media file", map[string]any{
+					"path":  localPath,
+					"error": err.Error(),
+				})
+				continue
+			}
+			if info.Size() > int64(maxSize) {
+				logger.WarnCF("agent", "Media file too large, skipping", map[string]any{
+					"path":     localPath,
+					"size":     info.Size(),
+					"max_size": maxSize,
+				})
+				continue
+			}
+
+			// Determine MIME type: prefer metadata, fallback to magic-bytes detection
+			mime := meta.ContentType
+			if mime == "" {
+				kind, ftErr := filetype.MatchFile(localPath)
+				if ftErr != nil || kind == filetype.Unknown {
+					logger.WarnCF("agent", "Unknown media type, skipping", map[string]any{
+						"path": localPath,
+					})
+					continue
+				}
+				mime = kind.MIME.Value
+			}
+
+			// Streaming base64: open file → base64 encoder → buffer
+			// Peak memory: ~1.33x file size (buffer only, no raw bytes copy)
+			f, err := os.Open(localPath)
+			if err != nil {
+				logger.WarnCF("agent", "Failed to open media file", map[string]any{
+					"path":  localPath,
+					"error": err.Error(),
+				})
+				continue
+			}
+
+			prefix := "data:" + mime + ";base64,"
+			encodedLen := base64.StdEncoding.EncodedLen(int(info.Size()))
+			var buf bytes.Buffer
+			buf.Grow(len(prefix) + encodedLen)
+			buf.WriteString(prefix)
+
+			encoder := base64.NewEncoder(base64.StdEncoding, &buf)
+			if _, err := io.Copy(encoder, f); err != nil {
+				f.Close()
+				logger.WarnCF("agent", "Failed to encode media file", map[string]any{
+					"path":  localPath,
+					"error": err.Error(),
+				})
+				continue
+			}
+			encoder.Close()
+			f.Close()
+
+			resolved = append(resolved, buf.String())
+		}
+
+		result[i].Media = resolved
+	}
+
+	return result
+}

pkg/agent/loop_test.go

@@ -6,12 +6,14 @@ import (
 	"os"
 	"path/filepath"
 	"slices"
+	"strings"
 	"testing"
 	"time"
 
 	"github.com/sipeed/picoclaw/pkg/bus"
 	"github.com/sipeed/picoclaw/pkg/channels"
 	"github.com/sipeed/picoclaw/pkg/config"
+	"github.com/sipeed/picoclaw/pkg/media"
 	"github.com/sipeed/picoclaw/pkg/providers"
 	"github.com/sipeed/picoclaw/pkg/tools"
 )
@@ -808,3 +810,142 @@ func TestHandleReasoning(t *testing.T) {
 		}
 	})
 }
+
+func TestResolveMediaRefs_ResolvesToBase64(t *testing.T) {
+	store := media.NewFileMediaStore()
+	dir := t.TempDir()
+
+	// Create a minimal valid PNG (8-byte header is enough for filetype detection)
+	pngPath := filepath.Join(dir, "test.png")
+	// PNG magic: 0x89 P N G \r \n 0x1A \n + minimal IHDR
+	pngHeader := []byte{
+		0x89, 0x50, 0x4E, 0x47, 0x0D, 0x0A, 0x1A, 0x0A, // PNG signature
+		0x00, 0x00, 0x00, 0x0D, // IHDR length
+		0x49, 0x48, 0x44, 0x52, // "IHDR"
+		0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01, 0x08, 0x02, // 1x1 RGB
+		0x00, 0x00, 0x00, // no interlace
+		0x90, 0x77, 0x53, 0xDE, // CRC
+	}
+	if err := os.WriteFile(pngPath, pngHeader, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	ref, err := store.Store(pngPath, media.MediaMeta{}, "test")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	messages := []providers.Message{
+		{Role: "user", Content: "describe this", Media: []string{ref}},
+	}
+	result := resolveMediaRefs(messages, store, config.DefaultMaxMediaSize)
+
+	if len(result[0].Media) != 1 {
+		t.Fatalf("expected 1 resolved media, got %d", len(result[0].Media))
+	}
+	if !strings.HasPrefix(result[0].Media[0], "data:image/png;base64,") {
+		t.Fatalf("expected data:image/png;base64, prefix, got %q", result[0].Media[0][:40])
+	}
+}
+
+func TestResolveMediaRefs_SkipsOversizedFile(t *testing.T) {
+	store := media.NewFileMediaStore()
+	dir := t.TempDir()
+
+	bigPath := filepath.Join(dir, "big.png")
+	// Write PNG header + padding to exceed limit
+	data := make([]byte, 1024+1) // 1KB + 1 byte
+	copy(data, []byte{0x89, 0x50, 0x4E, 0x47, 0x0D, 0x0A, 0x1A, 0x0A})
+	if err := os.WriteFile(bigPath, data, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	ref, _ := store.Store(bigPath, media.MediaMeta{}, "test")
+
+	messages := []providers.Message{
+		{Role: "user", Content: "hi", Media: []string{ref}},
+	}
+	// Use a tiny limit (1KB) so the file is oversized
+	result := resolveMediaRefs(messages, store, 1024)
+
+	if len(result[0].Media) != 0 {
+		t.Fatalf("expected 0 media (oversized), got %d", len(result[0].Media))
+	}
+}
+
+func TestResolveMediaRefs_SkipsUnknownType(t *testing.T) {
+	store := media.NewFileMediaStore()
+	dir := t.TempDir()
+
+	txtPath := filepath.Join(dir, "readme.txt")
+	if err := os.WriteFile(txtPath, []byte("hello world"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	ref, _ := store.Store(txtPath, media.MediaMeta{}, "test")
+
+	messages := []providers.Message{
+		{Role: "user", Content: "hi", Media: []string{ref}},
+	}
+	result := resolveMediaRefs(messages, store, config.DefaultMaxMediaSize)
+
+	if len(result[0].Media) != 0 {
+		t.Fatalf("expected 0 media (unknown type), got %d", len(result[0].Media))
+	}
+}
+
+func TestResolveMediaRefs_PassesThroughNonMediaRefs(t *testing.T) {
+	messages := []providers.Message{
+		{Role: "user", Content: "hi", Media: []string{"https://example.com/img.png"}},
+	}
+	result := resolveMediaRefs(messages, nil, config.DefaultMaxMediaSize)
+
+	if len(result[0].Media) != 1 || result[0].Media[0] != "https://example.com/img.png" {
+		t.Fatalf("expected passthrough of non-media:// URL, got %v", result[0].Media)
+	}
+}
+
+func TestResolveMediaRefs_DoesNotMutateOriginal(t *testing.T) {
+	store := media.NewFileMediaStore()
+	dir := t.TempDir()
+	pngPath := filepath.Join(dir, "test.png")
+	pngHeader := []byte{
+		0x89, 0x50, 0x4E, 0x47, 0x0D, 0x0A, 0x1A, 0x0A,
+		0x00, 0x00, 0x00, 0x0D, 0x49, 0x48, 0x44, 0x52,
+		0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01, 0x08, 0x02,
+		0x00, 0x00, 0x00, 0x90, 0x77, 0x53, 0xDE,
+	}
+	os.WriteFile(pngPath, pngHeader, 0o644)
+	ref, _ := store.Store(pngPath, media.MediaMeta{}, "test")
+
+	original := []providers.Message{
+		{Role: "user", Content: "hi", Media: []string{ref}},
+	}
+	originalRef := original[0].Media[0]
+
+	resolveMediaRefs(original, store, config.DefaultMaxMediaSize)
+
+	if original[0].Media[0] != originalRef {
+		t.Fatal("resolveMediaRefs mutated original message slice")
+	}
+}
+
+func TestResolveMediaRefs_UsesMetaContentType(t *testing.T) {
+	store := media.NewFileMediaStore()
+	dir := t.TempDir()
+
+	// File with JPEG content but stored with explicit content type
+	jpegPath := filepath.Join(dir, "photo")
+	jpegHeader := []byte{0xFF, 0xD8, 0xFF, 0xE0} // JPEG magic bytes
+	os.WriteFile(jpegPath, jpegHeader, 0o644)
+	ref, _ := store.Store(jpegPath, media.MediaMeta{ContentType: "image/jpeg"}, "test")
+
+	messages := []providers.Message{
+		{Role: "user", Content: "hi", Media: []string{ref}},
+	}
+	result := resolveMediaRefs(messages, store, config.DefaultMaxMediaSize)
+
+	if len(result[0].Media) != 1 {
+		t.Fatalf("expected 1 media, got %d", len(result[0].Media))
+	}
+	if !strings.HasPrefix(result[0].Media[0], "data:image/jpeg;base64,") {
+		t.Fatalf("expected jpeg prefix, got %q", result[0].Media[0][:30])
+	}
+}

pkg/config/config.go

@@ -180,6 +180,16 @@ type AgentDefaults struct {
 	MaxTokens                 int      `json:"max_tokens"                      env:"PICOCLAW_AGENTS_DEFAULTS_MAX_TOKENS"`
 	Temperature               *float64 `json:"temperature,omitempty"           env:"PICOCLAW_AGENTS_DEFAULTS_TEMPERATURE"`
 	MaxToolIterations         int      `json:"max_tool_iterations"             env:"PICOCLAW_AGENTS_DEFAULTS_MAX_TOOL_ITERATIONS"`
+	MaxMediaSize              int      `json:"max_media_size,omitempty"        env:"PICOCLAW_AGENTS_DEFAULTS_MAX_MEDIA_SIZE"`
+}
+
+const DefaultMaxMediaSize = 20 * 1024 * 1024 // 20 MB
+
+func (d *AgentDefaults) GetMaxMediaSize() int {
+	if d.MaxMediaSize > 0 {
+		return d.MaxMediaSize
+	}
+	return DefaultMaxMediaSize
 }
 
 // GetModelName returns the effective model name for the agent defaults.

pkg/providers/openai_compat/provider.go

@@ -116,7 +116,7 @@ func (p *Provider) Chat(
 
 	requestBody := map[string]any{
 		"model":    model,
-		"messages": stripSystemParts(messages),
+		"messages": serializeMessages(messages),
 	}
 
 	if len(tools) > 0 {
@@ -296,19 +296,55 @@ type openaiMessage struct {
 	ToolCallID       string     `json:"tool_call_id,omitempty"`
 }
 
-// stripSystemParts converts []Message to []openaiMessage, dropping the
-// SystemParts field so it doesn't leak into the JSON payload sent to
-// OpenAI-compatible APIs (some strict endpoints reject unknown fields).
-func stripSystemParts(messages []Message) []openaiMessage {
-	out := make([]openaiMessage, len(messages))
-	for i, m := range messages {
-		out[i] = openaiMessage{
-			Role:             m.Role,
-			Content:          m.Content,
-			ReasoningContent: m.ReasoningContent,
-			ToolCalls:        m.ToolCalls,
-			ToolCallID:       m.ToolCallID,
+// serializeMessages converts internal Message structs to the OpenAI wire format.
+// - Strips SystemParts (unknown to third-party endpoints)
+// - Converts messages with Media to multipart content format (text + image_url parts)
+// - Preserves ToolCallID, ToolCalls, and ReasoningContent for all messages
+func serializeMessages(messages []Message) []any {
+	out := make([]any, 0, len(messages))
+	for _, m := range messages {
+		if len(m.Media) == 0 {
+			out = append(out, openaiMessage{
+				Role:             m.Role,
+				Content:          m.Content,
+				ReasoningContent: m.ReasoningContent,
+				ToolCalls:        m.ToolCalls,
+				ToolCallID:       m.ToolCallID,
+			})
+			continue
 		}
+
+		// Multipart content format for messages with media
+		parts := make([]map[string]any, 0, 1+len(m.Media))
+		if m.Content != "" {
+			parts = append(parts, map[string]any{
+				"type": "text",
+				"text": m.Content,
+			})
+		}
+		for _, mediaURL := range m.Media {
+			parts = append(parts, map[string]any{
+				"type": "image_url",
+				"image_url": map[string]any{
+					"url": mediaURL,
+				},
+			})
+		}
+
+		msg := map[string]any{
+			"role":    m.Role,
+			"content": parts,
+		}
+		if m.ToolCallID != "" {
+			msg["tool_call_id"] = m.ToolCallID
+		}
+		if len(m.ToolCalls) > 0 {
+			msg["tool_calls"] = m.ToolCalls
+		}
+		if m.ReasoningContent != "" {
+			msg["reasoning_content"] = m.ReasoningContent
+		}
+		out = append(out, msg)
 	}
 	return out
 }

pkg/providers/openai_compat/provider_test.go

@@ -5,8 +5,11 @@ import (
 	"net/http"
 	"net/http/httptest"
 	"net/url"
+	"strings"
 	"testing"
 	"time"
+
+	"github.com/sipeed/picoclaw/pkg/providers/protocoltypes"
 )
 
 func TestProviderChat_UsesMaxCompletionTokensForGLM(t *testing.T) {
@@ -416,3 +419,97 @@ func TestProvider_FunctionalOptionRequestTimeoutNonPositive(t *testing.T) {
 		t.Fatalf("http timeout = %v, want %v", p.httpClient.Timeout, defaultRequestTimeout)
 	}
 }
+
+func TestSerializeMessages_PlainText(t *testing.T) {
+	messages := []protocoltypes.Message{
+		{Role: "user", Content: "hello"},
+		{Role: "assistant", Content: "hi", ReasoningContent: "thinking..."},
+	}
+	result := serializeMessages(messages)
+
+	data, err := json.Marshal(result)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	var msgs []map[string]any
+	json.Unmarshal(data, &msgs)
+
+	if msgs[0]["content"] != "hello" {
+		t.Fatalf("expected plain string content, got %v", msgs[0]["content"])
+	}
+	if msgs[1]["reasoning_content"] != "thinking..." {
+		t.Fatalf("reasoning_content not preserved, got %v", msgs[1]["reasoning_content"])
+	}
+}
+
+func TestSerializeMessages_WithMedia(t *testing.T) {
+	messages := []protocoltypes.Message{
+		{Role: "user", Content: "describe this", Media: []string{"data:image/png;base64,abc123"}},
+	}
+	result := serializeMessages(messages)
+
+	data, _ := json.Marshal(result)
+	var msgs []map[string]any
+	json.Unmarshal(data, &msgs)
+
+	content, ok := msgs[0]["content"].([]any)
+	if !ok {
+		t.Fatalf("expected array content for media message, got %T", msgs[0]["content"])
+	}
+	if len(content) != 2 {
+		t.Fatalf("expected 2 content parts, got %d", len(content))
+	}
+
+	textPart := content[0].(map[string]any)
+	if textPart["type"] != "text" || textPart["text"] != "describe this" {
+		t.Fatalf("text part mismatch: %v", textPart)
+	}
+
+	imgPart := content[1].(map[string]any)
+	if imgPart["type"] != "image_url" {
+		t.Fatalf("expected image_url type, got %v", imgPart["type"])
+	}
+	imgURL := imgPart["image_url"].(map[string]any)
+	if imgURL["url"] != "data:image/png;base64,abc123" {
+		t.Fatalf("image url mismatch: %v", imgURL["url"])
+	}
+}
+
+func TestSerializeMessages_MediaWithToolCallID(t *testing.T) {
+	messages := []protocoltypes.Message{
+		{Role: "tool", Content: "image result", Media: []string{"data:image/png;base64,xyz"}, ToolCallID: "call_1"},
+	}
+	result := serializeMessages(messages)
+
+	data, _ := json.Marshal(result)
+	var msgs []map[string]any
+	json.Unmarshal(data, &msgs)
+
+	if msgs[0]["tool_call_id"] != "call_1" {
+		t.Fatalf("tool_call_id not preserved with media, got %v", msgs[0]["tool_call_id"])
+	}
+	// Content should be multipart array
+	if _, ok := msgs[0]["content"].([]any); !ok {
+		t.Fatalf("expected array content, got %T", msgs[0]["content"])
+	}
+}
+
+func TestSerializeMessages_StripsSystemParts(t *testing.T) {
+	messages := []protocoltypes.Message{
+		{
+			Role:    "system",
+			Content: "you are helpful",
+			SystemParts: []protocoltypes.ContentBlock{
+				{Type: "text", Text: "you are helpful"},
+			},
+		},
+	}
+	result := serializeMessages(messages)
+
+	data, _ := json.Marshal(result)
+	raw := string(data)
+	if strings.Contains(raw, "system_parts") {
+		t.Fatal("system_parts should not appear in serialized output")
+	}
+}

pkg/providers/protocoltypes/types.go

@@ -65,6 +65,7 @@ type ContentBlock struct {
 type Message struct {
 	Role             string         `json:"role"`
 	Content          string         `json:"content"`
+	Media            []string       `json:"media,omitempty"`
 	ReasoningContent string         `json:"reasoning_content,omitempty"`
 	SystemParts      []ContentBlock `json:"system_parts,omitempty"` // structured system blocks for cache-aware adapters
 	ToolCalls        []ToolCall     `json:"tool_calls,omitempty"`