Dep: make-fetch onnx embed gemma

.gitignore

@@ -15,3 +15,4 @@ gflt
 chat_exports/*.json
 ragimport
 .env
+onnx/

Makefile

@@ -30,6 +30,9 @@ lint: ## Run linters. Use make install-linters first.
 lintall: lint
 	noblanks ./...
 
+fetch-onnx:
+	mkdir -p onnx/embedgemma && curl -o onnx/embedgemma/config.json -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/config.json && curl -o onnx/embedgemma/tokenizer.json -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/tokenizer.json && curl -o onnx/embedgemma/model_q4.onnx -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/onnx/model_q4.onnx && curl -o onnx/embedgemma/model_q4.onnx_data -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/onnx/model_q4.onnx_data?download=true
+
 # Whisper STT Setup (in batteries directory)
 setup-whisper: build-whisper download-whisper-model
 

bot.go

@@ -3,6 +3,7 @@ package main
 import (
 	"bufio"
 	"bytes"
+	"compress/gzip"
 	"context"
 	"encoding/json"
 	"fmt"
@@ -63,7 +64,9 @@ var (
 		"google/gemma-3-27b-it:free",
 		"meta-llama/llama-3.3-70b-instruct:free",
 	}
-	LocalModels = []string{}
+	LocalModels     = []string{}
+	localModelsData *models.LCPModels
+	orModelsData    *models.ORModels
 )
 
 var thinkBlockRE = regexp.MustCompile(`(?s)<think>.*?</think>`)
@@ -265,9 +268,7 @@ func warmUpModel() {
 		// Continue with warmup attempt anyway
 	}
 	if loaded {
-		if err := notifyUser("model already loaded", "Model "+chatBody.Model+" is already loaded."); err != nil {
+		showToast("model already loaded", "Model "+chatBody.Model+" is already loaded.")
-			logger.Debug("failed to notify user", "error", err)
-		}
 		return
 	}
 	go func() {
@@ -355,6 +356,7 @@ func fetchORModels(free bool) ([]string, error) {
 	if err := json.NewDecoder(resp.Body).Decode(data); err != nil {
 		return nil, err
 	}
+	orModelsData = data
 	freeModels := data.ListModels(free)
 	return freeModels, nil
 }
@@ -416,6 +418,7 @@ func fetchLCPModelsWithStatus() (*models.LCPModels, error) {
 	if err := json.NewDecoder(resp.Body).Decode(data); err != nil {
 		return nil, err
 	}
+	localModelsData = data
 	return data, nil
 }
 
@@ -433,6 +436,33 @@ func isModelLoaded(modelID string) (bool, error) {
 	return false, nil
 }
 
+func ModelHasVision(api, modelID string) bool {
+	switch {
+	case strings.Contains(api, "deepseek"):
+		return false
+	case strings.Contains(api, "openrouter"):
+		resp, err := http.Get("https://openrouter.ai/api/v1/models")
+		if err != nil {
+			logger.Warn("failed to fetch OR models for vision check", "error", err)
+			return false
+		}
+		defer resp.Body.Close()
+		orm := &models.ORModels{}
+		if err := json.NewDecoder(resp.Body).Decode(orm); err != nil {
+			logger.Warn("failed to decode OR models for vision check", "error", err)
+			return false
+		}
+		return orm.HasVision(modelID)
+	default:
+		models, err := fetchLCPModelsWithStatus()
+		if err != nil {
+			logger.Warn("failed to fetch LCP models for vision check", "error", err)
+			return false
+		}
+		return models.HasVision(modelID)
+	}
+}
+
 // monitorModelLoad starts a goroutine that periodically checks if the specified model is loaded.
 func monitorModelLoad(modelID string) {
 	go func() {
@@ -451,9 +481,7 @@ func monitorModelLoad(modelID string) {
 					continue
 				}
 				if loaded {
-					if err := notifyUser("model loaded", "Model "+modelID+" is now loaded and ready."); err != nil {
+					showToast("model loaded", "Model "+modelID+" is now loaded and ready.")
-						logger.Debug("failed to notify user", "error", err)
-					}
 					refreshChatDisplay()
 					return
 				}
@@ -464,6 +492,17 @@ func monitorModelLoad(modelID string) {
 
 // extractDetailedErrorFromBytes extracts detailed error information from response body bytes
 func extractDetailedErrorFromBytes(body []byte, statusCode int) string {
+	// Try to decompress gzip if the response is compressed
+	if len(body) >= 2 && body[0] == 0x1f && body[1] == 0x8b {
+		reader, err := gzip.NewReader(bytes.NewReader(body))
+		if err == nil {
+			decompressed, err := io.ReadAll(reader)
+			reader.Close()
+			if err == nil {
+				body = decompressed
+			}
+		}
+	}
 	// Try to parse as JSON to extract detailed error information
 	var errorResponse map[string]any
 	if err := json.Unmarshal(body, &errorResponse); err == nil {
@@ -529,9 +568,7 @@ func sendMsgToLLM(body io.Reader) {
 	req, err := http.NewRequest("POST", cfg.CurrentAPI, body)
 	if err != nil {
 		logger.Error("newreq error", "error", err)
-		if err := notifyUser("error", "apicall failed:"+err.Error()); err != nil {
+		showToast("error", "apicall failed:"+err.Error())
-			logger.Error("failed to notify", "error", err)
-		}
 		streamDone <- true
 		return
 	}
@@ -543,9 +580,7 @@ func sendMsgToLLM(body io.Reader) {
 	resp, err := httpClient.Do(req)
 	if err != nil {
 		logger.Error("llamacpp api", "error", err)
-		if err := notifyUser("error", "apicall failed:"+err.Error()); err != nil {
+		showToast("error", "apicall failed:"+err.Error())
-			logger.Error("failed to notify", "error", err)
-		}
 		streamDone <- true
 		return
 	}
@@ -556,9 +591,7 @@ func sendMsgToLLM(body io.Reader) {
 		if err != nil {
 			logger.Error("failed to read error response body", "error", err, "status_code", resp.StatusCode)
 			detailedError := fmt.Sprintf("HTTP Status: %d, Failed to read response body: %v", resp.StatusCode, err)
-			if err := notifyUser("API Error", detailedError); err != nil {
+			showToast("API Error", detailedError)
-				logger.Error("failed to notify", "error", err)
-			}
 			resp.Body.Close()
 			streamDone <- true
 			return
@@ -566,9 +599,7 @@ func sendMsgToLLM(body io.Reader) {
 		// Parse the error response for detailed information
 		detailedError := extractDetailedErrorFromBytes(bodyBytes, resp.StatusCode)
 		logger.Error("API returned error status", "status_code", resp.StatusCode, "detailed_error", detailedError)
-		if err := notifyUser("API Error", detailedError); err != nil {
+		showToast("API Error", detailedError)
-			logger.Error("failed to notify", "error", err)
-		}
 		resp.Body.Close()
 		streamDone <- true
 		return
@@ -605,16 +636,12 @@ func sendMsgToLLM(body io.Reader) {
 				detailedError := fmt.Sprintf("Streaming connection closed unexpectedly (Status: %d). This may indicate an API error. Check your API provider and model settings.", resp.StatusCode)
 				logger.Error("error reading response body", "error", err, "detailed_error", detailedError,
 					"status_code", resp.StatusCode, "user_role", cfg.UserRole, "parser", chunkParser, "link", cfg.CurrentAPI)
-				if err := notifyUser("API Error", detailedError); err != nil {
+				showToast("API Error", detailedError)
-					logger.Error("failed to notify", "error", err)
-				}
 			} else {
 				logger.Error("error reading response body", "error", err, "line", string(line),
 					"user_role", cfg.UserRole, "parser", chunkParser, "link", cfg.CurrentAPI)
 				// if err.Error() != "EOF" {
-				if err := notifyUser("API error", err.Error()); err != nil {
+				showToast("API error", err.Error())
-					logger.Error("failed to notify", "error", err)
-				}
 			}
 			streamDone <- true
 			break
@@ -641,9 +668,7 @@ func sendMsgToLLM(body io.Reader) {
 		if err != nil {
 			logger.Error("error parsing response body", "error", err,
 				"line", string(line), "url", cfg.CurrentAPI)
-			if err := notifyUser("LLM Response Error", "Failed to parse LLM response: "+err.Error()); err != nil {
+			showToast("LLM Response Error", "Failed to parse LLM response: "+err.Error())
-				logger.Error("failed to notify user", "error", err)
-			}
 			streamDone <- true
 			break
 		}
@@ -718,7 +743,7 @@ func sendMsgToLLM(body io.Reader) {
 		}
 	interrupt:
 		if interruptResp { // read bytes, so it would not get into beginning of the next req
-			interruptResp = false
+			// interruptResp = false
 			logger.Info("interrupted bot response", "chunk_counter", counter)
 			streamDone <- true
 			break
@@ -772,6 +797,7 @@ func showSpinner() {
 }
 
 func chatRound(r *models.ChatRoundReq) error {
+	interruptResp = false
 	botRespMode = true
 	go showSpinner()
 	updateStatusLine()
@@ -937,6 +963,9 @@ out:
 	}
 	// Strip think blocks before parsing for tool calls
 	respTextNoThink := thinkBlockRE.ReplaceAllString(respText.String(), "")
+	if interruptResp {
+		return nil
+	}
 	if findCall(respTextNoThink, toolResp.String()) {
 		return nil
 	}
@@ -1174,17 +1203,59 @@ func findCall(msg, toolCall string) bool {
 	toolRunningMode = false
 	toolMsg := string(resp)
 	logger.Info("llm used a tool call", "tool_name", fc.Name, "too_args", fc.Args, "id", fc.ID, "tool_resp", toolMsg)
-	fmt.Fprintf(textView, "%s[-:-:b](%d) <%s>: [-:-:-]\n%s\n",
-		"\n\n", len(chatBody.Messages), cfg.ToolRole, toolMsg)
 	// Create tool response message with the proper tool_call_id
 	// Mark shell commands as always visible
 	isShellCommand := fc.Name == "execute_command"
-	toolResponseMsg := models.RoleMsg{
+	// Check if response is multimodal content (image)
-		Role:           cfg.ToolRole,
+	var toolResponseMsg models.RoleMsg
-		Content:        toolMsg,
+	if strings.HasPrefix(strings.TrimSpace(toolMsg), `{"type":"multimodal_content"`) {
-		ToolCallID:     lastToolCall.ID,
+		// Parse multimodal content response
-		IsShellCommand: isShellCommand,
+		multimodalResp := models.MultimodalToolResp{}
+		if err := json.Unmarshal([]byte(toolMsg), &multimodalResp); err == nil && multimodalResp.Type == "multimodal_content" {
+			// Create RoleMsg with ContentParts
+			var contentParts []any
+			for _, part := range multimodalResp.Parts {
+				partType := part["type"]
+				switch partType {
+				case "text":
+					contentParts = append(contentParts, models.TextContentPart{Type: "text", Text: part["text"]})
+				case "image_url":
+					contentParts = append(contentParts, models.ImageContentPart{
+						Type: "image_url",
+						ImageURL: struct {
+							URL string `json:"url"`
+						}{URL: part["url"]},
+					})
+				default:
+					continue
+				}
+			}
+			toolResponseMsg = models.RoleMsg{
+				Role:            cfg.ToolRole,
+				ContentParts:    contentParts,
+				HasContentParts: true,
+				ToolCallID:      lastToolCall.ID,
+				IsShellCommand:  isShellCommand,
+			}
+		} else {
+			// Fallback to regular content
+			toolResponseMsg = models.RoleMsg{
+				Role:           cfg.ToolRole,
+				Content:        toolMsg,
+				ToolCallID:     lastToolCall.ID,
+				IsShellCommand: isShellCommand,
+			}
+		}
+	} else {
+		toolResponseMsg = models.RoleMsg{
+			Role:           cfg.ToolRole,
+			Content:        toolMsg,
+			ToolCallID:     lastToolCall.ID,
+			IsShellCommand: isShellCommand,
+		}
 	}
+	fmt.Fprintf(textView, "%s[-:-:b](%d) <%s>: [-:-:-]\n%s\n",
+		"\n\n", len(chatBody.Messages), cfg.ToolRole, toolResponseMsg.GetText())
 	chatBody.Messages = append(chatBody.Messages, toolResponseMsg)
 	logger.Debug("findCall: added actual tool response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "tool_call_id", toolResponseMsg.ToolCallID, "message_count_after_add", len(chatBody.Messages))
 	// Clear the stored tool call ID after using it
@@ -1305,8 +1376,8 @@ func applyCharCard(cc *models.CharCard, loadHistory bool) {
 }
 
 func charToStart(agentName string, keepSysP bool) bool {
-	cc, ok := sysMap[agentName]
+	cc := GetCardByRole(agentName)
-	if !ok {
+	if cc == nil {
 		return false
 	}
 	applyCharCard(cc, keepSysP)
@@ -1339,6 +1410,7 @@ func updateModelLists() {
 				chatBody.Model = m
 				cachedModelColor = "green"
 				updateStatusLine()
+				updateToolCapabilities()
 				app.Draw()
 				return
 			}
@@ -1366,15 +1438,15 @@ func refreshLocalModelsIfEmpty() {
 
 func summarizeAndStartNewChat() {
 	if len(chatBody.Messages) == 0 {
-		_ = notifyUser("info", "No chat history to summarize")
+		showToast("info", "No chat history to summarize")
 		return
 	}
-	_ = notifyUser("info", "Summarizing chat history...")
+	showToast("info", "Summarizing chat history...")
 	// Call the summarize_chat tool via agent
 	summaryBytes := callToolWithAgent("summarize_chat", map[string]string{})
 	summary := string(summaryBytes)
 	if summary == "" {
-		_ = notifyUser("error", "Failed to generate summary")
+		showToast("error", "Failed to generate summary")
 		return
 	}
 	// Start a new chat
@@ -1393,7 +1465,7 @@ func summarizeAndStartNewChat() {
 	if err := updateStorageChat(activeChatName, chatBody.Messages); err != nil {
 		logger.Warn("failed to update storage after injecting summary", "error", err)
 	}
-	_ = notifyUser("info", "Chat summarized and new chat started with summary as tool response")
+	showToast("info", "Chat summarized and new chat started with summary as tool response")
 }
 
 func init() {
@@ -1452,6 +1524,23 @@ func init() {
 	if cfg.STT_ENABLED {
 		asr = NewSTT(logger, cfg)
 	}
+	if cfg.PlaywrightEnabled {
+		if err := checkPlaywright(); err != nil {
+			// slow, need a faster check if playwright install
+			if err := installPW(); err != nil {
+				logger.Error("failed to install playwright", "error", err)
+				cancel()
+				os.Exit(1)
+				return
+			}
+			if err := checkPlaywright(); err != nil {
+				logger.Error("failed to run playwright", "error", err)
+				cancel()
+				os.Exit(1)
+				return
+			}
+		}
+	}
 	// Initialize scrollToEndEnabled based on config
 	scrollToEndEnabled = cfg.AutoScrollEnabled
 	go updateModelLists()

config.example.toml

@@ -56,3 +56,6 @@ StripThinkingFromAPI = true  # Strip <think> blocks from messages before sending
 # Valid values: xhigh, high, medium, low, minimal, none (empty or none = disabled)
 # Models that support reasoning will include thinking content wrapped in <think> tags
 ReasoningEffort = "medium"
+# playwright tools
+PlaywrightEnabled = false
+PlaywrightDebug = false

config/config.go

@@ -70,6 +70,9 @@ type Config struct {
 	CharSpecificContextEnabled bool   `toml:"CharSpecificContextEnabled"`
 	CharSpecificContextTag     string `toml:"CharSpecificContextTag"`
 	AutoTurn                   bool   `toml:"AutoTurn"`
+	// playwright browser
+	PlaywrightEnabled bool `toml:"PlaywrightEnabled"`
+	PlaywrightDebug   bool `toml:"PlaywrightDebug"` // !headless
 }
 
 func LoadConfig(fn string) (*Config, error) {

docs/config.md

@@ -162,6 +162,15 @@ Those could be switched in program, but also bould be setup in config.
 #### ToolUse
 - Enable or disable explanation of tools to llm, so it could use them.
 
+#### Playwright Browser Automation
+These settings enable browser automation tools available to the LLM.
+
+- **PlaywrightEnabled** (`false`)
+  - Enable or disable Playwright browser automation tools for the LLM. When enabled, the LLM can use tools like `pw_browser`, `pw_close`, and `pw_status` to automate browser interactions.
+
+- **PlaywrightDebug** (`false`)
+  - Enable debug mode for Playwright browser. When set to `true`, the browser runs in visible (non-headless) mode, displaying the GUI for debugging purposes. When `false`, the browser runs in headless mode by default.
+
 ### StripThinkingFromAPI (`true`)
 - Strip thinking blocks from messages before sending to LLM. Keeps them in chat history for local viewing but reduces token usage in API calls.
 

go.mod

@@ -14,25 +14,36 @@ require (
 	github.com/jmoiron/sqlx v1.4.0
 	github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728
 	github.com/neurosnap/sentences v1.1.2
+	github.com/playwright-community/playwright-go v0.5700.1
 	github.com/rivo/tview v0.42.0
+	github.com/sugarme/tokenizer v0.3.0
+	github.com/yalue/onnxruntime_go v1.27.0
 	github.com/yuin/goldmark v1.4.13
 )
 
 require (
 	github.com/andybalholm/cascadia v1.3.3 // indirect
+	github.com/deckarep/golang-set/v2 v2.8.0 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/ebitengine/oto/v3 v3.4.0 // indirect
 	github.com/ebitengine/purego v0.9.1 // indirect
+	github.com/emirpasic/gods v1.18.1 // indirect
 	github.com/gdamore/encoding v1.0.1 // indirect
+	github.com/go-jose/go-jose/v3 v3.0.4 // indirect
+	github.com/go-stack/stack v1.8.1 // indirect
 	github.com/google/uuid v1.6.0 // indirect
 	github.com/hajimehoshi/go-mp3 v0.3.4 // indirect
 	github.com/hajimehoshi/oto/v2 v2.3.1 // indirect
 	github.com/lucasb-eyer/go-colorful v1.3.0 // indirect
 	github.com/mattn/go-isatty v0.0.20 // indirect
+	github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db // indirect
 	github.com/ncruces/go-strftime v1.0.0 // indirect
+	github.com/patrickmn/go-cache v2.1.0+incompatible // indirect
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect
 	github.com/rivo/uniseg v0.4.7 // indirect
+	github.com/schollz/progressbar/v2 v2.15.0 // indirect
+	github.com/sugarme/regexpset v0.0.0-20200920021344-4d4ec8eaf93c // indirect
 	golang.org/x/exp v0.0.0-20251209150349-8475f28825e9 // indirect
 	golang.org/x/net v0.48.0 // indirect
 	golang.org/x/sys v0.39.0 // indirect

go.sum

@@ -10,22 +10,32 @@ github.com/PuerkitoBio/goquery v1.11.0 h1:jZ7pwMQXIITcUXNH83LLk+txlaEy6NVOfTuP43
 github.com/PuerkitoBio/goquery v1.11.0/go.mod h1:wQHgxUOU3JGuj3oD/QFfxUdlzW6xPHfqyHre6VMY4DQ=
 github.com/andybalholm/cascadia v1.3.3 h1:AG2YHrzJIm4BZ19iwJ/DAua6Btl3IwJX+VI4kktS1LM=
 github.com/andybalholm/cascadia v1.3.3/go.mod h1:xNd9bqTn98Ln4DwST8/nG+H0yuB8Hmgu1YHNnWw0GeA=
+github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/deckarep/golang-set/v2 v2.8.0 h1:swm0rlPCmdWn9mESxKOjWk8hXSqoxOp+ZlfuyaAdFlQ=
+github.com/deckarep/golang-set/v2 v2.8.0/go.mod h1:VAky9rY/yGXJOLEDv3OMci+7wtDpOF4IN+y82NBOac4=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
 github.com/ebitengine/oto/v3 v3.4.0 h1:br0PgASsEWaoWn38b2Goe7m1GKFYfNgnsjSd5Gg+/bQ=
 github.com/ebitengine/oto/v3 v3.4.0/go.mod h1:IOleLVD0m+CMak3mRVwsYY8vTctQgOM0iiL6S7Ar7eI=
 github.com/ebitengine/purego v0.9.1 h1:a/k2f2HQU3Pi399RPW1MOaZyhKJL9w/xFpKAg4q1s0A=
 github.com/ebitengine/purego v0.9.1/go.mod h1:iIjxzd6CiRiOG0UyXP+V1+jWqUXVjPKLAI0mRfJZTmQ=
+github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc=
+github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
 github.com/gdamore/encoding v1.0.1 h1:YzKZckdBL6jVt2Gc+5p82qhrGiqMdG/eNs6Wy0u3Uhw=
 github.com/gdamore/encoding v1.0.1/go.mod h1:0Z0cMFinngz9kS1QfMjCP8TY7em3bZYeeklsSDPivEo=
 github.com/gdamore/tcell/v2 v2.13.2 h1:5j4srfF8ow3HICOv/61/sOhQtA25qxEB2XR3Q/Bhx2g=
 github.com/gdamore/tcell/v2 v2.13.2/go.mod h1:+Wfe208WDdB7INEtCsNrAN6O2m+wsTPk1RAovjaILlo=
 github.com/glebarez/go-sqlite v1.22.0 h1:uAcMJhaA6r3LHMTFgP0SifzgXg46yJkgxqyuyec+ruQ=
 github.com/glebarez/go-sqlite v1.22.0/go.mod h1:PlBIdHe0+aUEFn+r2/uthrWq4FxbzugL0L8Li6yQJbc=
+github.com/go-jose/go-jose/v3 v3.0.4 h1:Wp5HA7bLQcKnf6YYao/4kpRpVMp/yf6+pJKV8WFSaNY=
+github.com/go-jose/go-jose/v3 v3.0.4/go.mod h1:5b+7YgP7ZICgJDBdfjZaIt+H/9L9T/YQrVfLAMboGkQ=
 github.com/go-sql-driver/mysql v1.8.1 h1:LedoTUt/eveggdHS9qUFC1EFSa8bU2+1pZjSRpvNJ1Y=
 github.com/go-sql-driver/mysql v1.8.1/go.mod h1:wEBSXgmK//2ZFJyE+qWnIsVGmvmEKlqwuVSjsCm7DZg=
+github.com/go-stack/stack v1.8.1 h1:ntEHSVwIt7PNXNpgPmVfMrNhLtgjlmnZha2kOpuRiDw=
+github.com/go-stack/stack v1.8.1/go.mod h1:dcoOX6HbPZSZptuspn9bctJ+N/CnF5gGygcUP3XYfe4=
+github.com/google/go-cmp v0.5.9/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e h1:ijClszYn+mADRFY17kjQEVQ1XRhq2/JR1M3sGqeJoxs=
 github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e/go.mod h1:boTsfXsheKC2y+lKOCMpSfarhxDeIzfZG1jqGcPl3cA=
@@ -53,12 +63,18 @@ github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWE
 github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
 github.com/mattn/go-sqlite3 v1.14.22 h1:2gZY6PC6kBnID23Tichd1K+Z0oS6nE/XwU+Vz/5o4kU=
 github.com/mattn/go-sqlite3 v1.14.22/go.mod h1:Uh1q+B4BYcTPb+yiD3kU8Ct7aC0hY9fxUwlHK0RXw+Y=
+github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db h1:62I3jR2EmQ4l5rM/4FEfDWcRD+abF5XlKShorW5LRoQ=
+github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db/go.mod h1:l0dey0ia/Uv7NcFFVbCLtqEBQbrT4OCwCSKTEv6enCw=
 github.com/ncruces/go-strftime v1.0.0 h1:HMFp8mLCTPp341M/ZnA4qaf7ZlsbTc+miZjCLOFAw7w=
 github.com/ncruces/go-strftime v1.0.0/go.mod h1:Fwc5htZGVVkseilnfgOVb9mKy6w1naJmn9CehxcKcls=
 github.com/neurosnap/sentences v1.1.2 h1:iphYOzx/XckXeBiLIUBkPu2EKMJ+6jDbz/sLJZ7ZoUw=
 github.com/neurosnap/sentences v1.1.2/go.mod h1:/pwU4E9XNL21ygMIkOIllv/SMy2ujHwpf8GQPu1YPbQ=
+github.com/patrickmn/go-cache v2.1.0+incompatible h1:HRMgzkcYKYpi3C8ajMPV8OFXaaRUnok+kx1WdO15EQc=
+github.com/patrickmn/go-cache v2.1.0+incompatible/go.mod h1:3Qf8kWWT7OJRJbdiICTKqZju1ZixQ/KpMGzzAfe6+WQ=
 github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
 github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
+github.com/playwright-community/playwright-go v0.5700.1 h1:PNFb1byWqrTT720rEO0JL88C6Ju0EmUnR5deFLvtP/U=
+github.com/playwright-community/playwright-go v0.5700.1/go.mod h1:MlSn1dZrx8rszbCxY6x3qK89ZesJUYVx21B2JnkoNF0=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec h1:W09IVJc94icq4NjY3clb7Lk8O1qJ8BdBEF8z0ibU0rE=
@@ -67,8 +83,19 @@ github.com/rivo/tview v0.42.0 h1:b/ftp+RxtDsHSaynXTbJb+/n/BxDEi+W3UfF5jILK6c=
 github.com/rivo/tview v0.42.0/go.mod h1:cSfIYfhpSGCjp3r/ECJb+GKS7cGJnqV8vfjQPwoXyfY=
 github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=
 github.com/rivo/uniseg v0.4.7/go.mod h1:FN3SvrM+Zdj16jyLfmOkMNblXMcoc8DfTHruCPUcx88=
+github.com/schollz/progressbar/v2 v2.15.0 h1:dVzHQ8fHRmtPjD3K10jT3Qgn/+H+92jhPrhmxIJfDz8=
+github.com/schollz/progressbar/v2 v2.15.0/go.mod h1:UdPq3prGkfQ7MOzZKlDRpYKcFqEMczbD7YmbPgpzKMI=
+github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
+github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
 github.com/stretchr/testify v1.10.0 h1:Xv5erBjTwe/5IxqUQTdXv5kgmIvbHo3QQyRwhJsOfJA=
 github.com/stretchr/testify v1.10.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
+github.com/sugarme/regexpset v0.0.0-20200920021344-4d4ec8eaf93c h1:pwb4kNSHb4K89ymCaN+5lPH/MwnfSVg4rzGDh4d+iy4=
+github.com/sugarme/regexpset v0.0.0-20200920021344-4d4ec8eaf93c/go.mod h1:2gwkXLWbDGUQWeL3RtpCmcY4mzCtU13kb9UsAg9xMaw=
+github.com/sugarme/tokenizer v0.3.0 h1:FE8DYbNSz/kSbgEo9l/RjgYHkIJYEdskumitFQBE9FE=
+github.com/sugarme/tokenizer v0.3.0/go.mod h1:VJ+DLK5ZEZwzvODOWwY0cw+B1dabTd3nCB5HuFCItCc=
+github.com/yalue/onnxruntime_go v1.27.0 h1:c1YSgDNtpf0WGtxj3YeRIb8VC5LmM1J+Ve3uHdteC1U=
+github.com/yalue/onnxruntime_go v1.27.0/go.mod h1:b4X26A8pekNb1ACJ58wAXgNKeUCGEAQ9dmACut9Sm/4=
 github.com/yuin/goldmark v1.4.13 h1:fVcFKWvrslecOb/tg+Cc05dkeYx540o0FuFt3nUVDoE=
 github.com/yuin/goldmark v1.4.13/go.mod h1:6yULJ656Px+3vBD8DxQVa3kxgyrAnzto9xy5taEt/CY=
 golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=
@@ -152,6 +179,8 @@ golang.org/x/tools v0.21.1-0.20240508182429-e35e4ccd0d2d/go.mod h1:aiJjzUbINMkxb
 golang.org/x/tools v0.40.0 h1:yLkxfA+Qnul4cs9QA3KnlFu0lVmd8JJfoq+E41uSutA=
 golang.org/x/tools v0.40.0/go.mod h1:Ik/tzLRlbscWpqqMRjyWYDisX8bG13FrdXp3o4Sr9lc=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
 gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 modernc.org/cc/v4 v4.27.1 h1:9W30zRlYrefrDV2JE2O8VDtJ1yPGownxciz5rrbQZis=

helpfuncs.go

@@ -11,6 +11,7 @@ import (
 	"path"
 	"path/filepath"
 	"slices"
+	"strconv"
 	"strings"
 	"time"
 	"unicode"
@@ -197,7 +198,11 @@ func initSysCards() ([]string, error) {
 			logger.Warn("empty role", "file", cc.FilePath)
 			continue
 		}
-		sysMap[cc.Role] = cc
+		if cc.ID == "" {
+			cc.ID = models.ComputeCardID(cc.Role, cc.FilePath)
+		}
+		sysMap[cc.ID] = cc
+		roleToID[cc.Role] = cc.ID
 		labels = append(labels, cc.Role)
 	}
 	return labels, nil
@@ -286,24 +291,25 @@ func listRolesWithUser() []string {
 	return result
 }
 
-func loadImage() {
+func loadImage() error {
 	filepath := defaultImage
-	cc, ok := sysMap[cfg.AssistantRole]
+	cc := GetCardByRole(cfg.AssistantRole)
-	if ok {
+	if cc != nil {
 		if strings.HasSuffix(cc.FilePath, ".png") {
 			filepath = cc.FilePath
 		}
 	}
 	file, err := os.Open(filepath)
 	if err != nil {
-		panic(err)
+		return fmt.Errorf("failed to open image: %w", err)
 	}
 	defer file.Close()
 	img, _, err := image.Decode(file)
 	if err != nil {
-		panic(err)
+		return fmt.Errorf("failed to decode image: %w", err)
 	}
 	imgView.SetImage(img)
+	return nil
 }
 
 func strInSlice(s string, sl []string) bool {
@@ -376,9 +382,90 @@ func makeStatusLine() string {
 		roleInject := fmt.Sprintf(" | [%s:-:b]role injection[-:-:-] (alt+7)", boolColors[injectRole])
 		statusLine += roleInject
 	}
+	// context tokens
+	contextTokens := getContextTokens()
+	maxCtx := getMaxContextTokens()
+	if maxCtx == 0 {
+		maxCtx = 16384
+	}
+	if contextTokens > 0 {
+		contextInfo := fmt.Sprintf(" | context-estim: [orange:-:b]%d/%d[-:-:-]", contextTokens, maxCtx)
+		statusLine += contextInfo
+	}
 	return statusLine + imageInfo + shellModeInfo
 }
 
+func getContextTokens() int {
+	if chatBody == nil || chatBody.Messages == nil {
+		return 0
+	}
+	total := 0
+	messages := chatBody.Messages
+	for i := range messages {
+		msg := &messages[i]
+		if msg.Stats != nil && msg.Stats.Tokens > 0 {
+			total += msg.Stats.Tokens
+		} else if msg.GetText() != "" {
+			total += len(msg.GetText()) / 4
+		}
+	}
+	return total
+}
+
+const deepseekContext = 128000
+
+func getMaxContextTokens() int {
+	if chatBody == nil || chatBody.Model == "" {
+		return 0
+	}
+	modelName := chatBody.Model
+	switch {
+	case strings.Contains(cfg.CurrentAPI, "openrouter"):
+		if orModelsData != nil {
+			for i := range orModelsData.Data {
+				m := &orModelsData.Data[i]
+				if m.ID == modelName {
+					return m.ContextLength
+				}
+			}
+		}
+	case strings.Contains(cfg.CurrentAPI, "deepseek"):
+		return deepseekContext
+	default:
+		if localModelsData != nil {
+			for i := range localModelsData.Data {
+				m := &localModelsData.Data[i]
+				if m.ID == modelName {
+					for _, arg := range m.Status.Args {
+						if strings.HasPrefix(arg, "--ctx-size") {
+							if strings.Contains(arg, "=") {
+								val := strings.Split(arg, "=")[1]
+								if n, err := strconv.Atoi(val); err == nil {
+									return n
+								}
+							} else {
+								idx := -1
+								for j, a := range m.Status.Args {
+									if a == "--ctx-size" && j+1 < len(m.Status.Args) {
+										idx = j + 1
+										break
+									}
+								}
+								if idx != -1 {
+									if n, err := strconv.Atoi(m.Status.Args[idx]); err == nil {
+										return n
+									}
+								}
+							}
+						}
+					}
+				}
+			}
+		}
+	}
+	return 0
+}
+
 // set of roles within card definition and mention in chat history
 func listChatRoles() []string {
 	currentChat, ok := chatMap[activeChatName]
@@ -386,13 +473,9 @@ func listChatRoles() []string {
 	if !ok {
 		return cbc
 	}
-	currentCard, ok := sysMap[currentChat.Agent]
+	currentCard := GetCardByRole(currentChat.Agent)
-	if !ok {
+	if currentCard == nil {
-		// case which won't let to switch roles:
+		logger.Warn("failed to find current card", "agent", currentChat.Agent)
-		// started new chat (basic_sys or any other), at the start it yet be saved or have chatbody
-		// if it does not have a card or chars, it'll return an empty slice
-		// log error
-		logger.Warn("failed to find current card in sysMap", "agent", currentChat.Agent, "sysMap", sysMap)
 		return cbc
 	}
 	charset := []string{}
@@ -408,10 +491,7 @@ func listChatRoles() []string {
 func deepseekModelValidator() error {
 	if cfg.CurrentAPI == cfg.DeepSeekChatAPI || cfg.CurrentAPI == cfg.DeepSeekCompletionAPI {
 		if chatBody.Model != "deepseek-chat" && chatBody.Model != "deepseek-reasoner" {
-			if err := notifyUser("bad request", "wrong deepseek model name"); err != nil {
+			showToast("bad request", "wrong deepseek model name")
-				logger.Warn("failed ot notify user", "error", err)
-				return err
-			}
 			return nil
 		}
 	}
@@ -441,7 +521,7 @@ func updateFlexLayout() {
 	if shellMode {
 		flex.AddItem(shellInput, 0, 10, false)
 	} else {
-		flex.AddItem(textArea, 0, 10, false)
+		flex.AddItem(bottomFlex, 0, 10, true)
 	}
 	if positionVisible {
 		flex.AddItem(statusLineWidget, 0, 2, false)
@@ -611,9 +691,7 @@ func performSearch(term string) {
 		searchResults = nil
 		searchResultLengths = nil
 		notification := "Pattern not found: " + term
-		if err := notifyUser("search", notification); err != nil {
+		showToast("search", notification)
-			logger.Error("failed to send notification", "error", err)
-		}
 		return
 	}
 	// Store the formatted text positions and lengths for accurate highlighting
@@ -646,9 +724,7 @@ func highlightCurrentMatch() {
 	textView.Highlight(currentRegion).ScrollToHighlight()
 	// Send notification about which match we're at
 	notification := fmt.Sprintf("Match %d of %d", searchIndex+1, len(searchResults))
-	if err := notifyUser("search", notification); err != nil {
+	showToast("search", notification)
-		logger.Error("failed to send notification", "error", err)
-	}
 }
 
 // showSearchBar shows the search input field as an overlay
@@ -738,9 +814,7 @@ func addRegionTags(text string, positions []int, lengths []int, currentIdx int,
 // searchNext finds the next occurrence of the search term
 func searchNext() {
 	if len(searchResults) == 0 {
-		if err := notifyUser("search", "No search results to navigate"); err != nil {
+		showToast("search", "No search results to navigate")
-			logger.Error("failed to send notification", "error", err)
-		}
 		return
 	}
 	searchIndex = (searchIndex + 1) % len(searchResults)
@@ -750,9 +824,7 @@ func searchNext() {
 // searchPrev finds the previous occurrence of the search term
 func searchPrev() {
 	if len(searchResults) == 0 {
-		if err := notifyUser("search", "No search results to navigate"); err != nil {
+		showToast("search", "No search results to navigate")
-			logger.Error("failed to send notification", "error", err)
-		}
 		return
 	}
 	if searchIndex == 0 {

llm.go

@@ -3,7 +3,6 @@ package main
 import (
 	"bytes"
 	"encoding/json"
-	"errors"
 	"gf-lt/models"
 	"io"
 	"strings"
@@ -119,25 +118,22 @@ func (lcp LCPCompletion) FormMsg(msg, role string, resume bool) (io.Reader, erro
 	logger.Debug("formmsg lcpcompletion", "link", cfg.CurrentAPI)
 	localImageAttachmentPath := imageAttachmentPath
 	var multimodalData []string
-	if localImageAttachmentPath != "" {
-		imageURL, err := models.CreateImageURLFromPath(localImageAttachmentPath)
-		if err != nil {
-			logger.Error("failed to create image URL from path for completion",
-				"error", err, "path", localImageAttachmentPath)
-			return nil, err
-		}
-		// Extract base64 part from data URL (e.g., "data:image/jpeg;base64,...")
-		parts := strings.SplitN(imageURL, ",", 2)
-		if len(parts) == 2 {
-			multimodalData = append(multimodalData, parts[1])
-		} else {
-			logger.Error("invalid image data URL format", "url", imageURL)
-			return nil, errors.New("invalid image data URL format")
-		}
-		imageAttachmentPath = "" // Clear the attachment after use
-	}
 	if msg != "" { // otherwise let the bot to continue
-		newMsg := models.RoleMsg{Role: role, Content: msg}
+		var newMsg models.RoleMsg
+		if localImageAttachmentPath != "" {
+			newMsg = models.NewMultimodalMsg(role, []any{})
+			newMsg.AddTextPart(msg)
+			imageURL, err := models.CreateImageURLFromPath(localImageAttachmentPath)
+			if err != nil {
+				logger.Error("failed to create image URL from path for completion",
+					"error", err, "path", localImageAttachmentPath)
+				return nil, err
+			}
+			newMsg.AddImagePart(imageURL, localImageAttachmentPath)
+			imageAttachmentPath = "" // Clear the attachment after use
+		} else { // not a multimodal msg or image passed in tool call
+			newMsg = models.RoleMsg{Role: role, Content: msg}
+		}
 		newMsg = *processMessageTag(&newMsg)
 		chatBody.Messages = append(chatBody.Messages, newMsg)
 	}
@@ -146,22 +142,40 @@ func (lcp LCPCompletion) FormMsg(msg, role string, resume bool) (io.Reader, erro
 		chatBody.Messages = append(chatBody.Messages, models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
 	}
 	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
+	// Build prompt and extract images inline as we process each message
 	messages := make([]string, len(filteredMessages))
 	for i := range filteredMessages {
-		messages[i] = stripThinkingFromMsg(&filteredMessages[i]).ToPrompt()
+		m := stripThinkingFromMsg(&filteredMessages[i])
+		messages[i] = m.ToPrompt()
+		// Extract images from this message and add marker inline
+		if len(m.ContentParts) > 0 {
+			for _, part := range m.ContentParts {
+				var imgURL string
+				// Check for struct type
+				if imgPart, ok := part.(models.ImageContentPart); ok {
+					imgURL = imgPart.ImageURL.URL
+				} else if partMap, ok := part.(map[string]any); ok {
+					// Check for map type (from JSON unmarshaling)
+					if partType, exists := partMap["type"]; exists && partType == "image_url" {
+						if imgURLMap, ok := partMap["image_url"].(map[string]any); ok {
+							if url, ok := imgURLMap["url"].(string); ok {
+								imgURL = url
+							}
+						}
+					}
+				}
+				if imgURL != "" {
+					// Extract base64 part from data URL (e.g., "data:image/jpeg;base64,...")
+					parts := strings.SplitN(imgURL, ",", 2)
+					if len(parts) == 2 {
+						multimodalData = append(multimodalData, parts[1])
+						messages[i] += " <__media__>"
+					}
+				}
+			}
+		}
 	}
 	prompt := strings.Join(messages, "\n")
-	// Add multimodal media markers to the prompt text when multimodal data is present
-	// This is required by llama.cpp multimodal models so they know where to insert media
-	if len(multimodalData) > 0 {
-		// Add a media marker for each item in the multimodal data
-		var sb strings.Builder
-		sb.WriteString(prompt)
-		for range multimodalData {
-			sb.WriteString(" <__media__>") // llama.cpp default multimodal marker
-		}
-		prompt = sb.String()
-	}
 	// needs to be after <__media__> if there are images
 	if !resume {
 		botMsgStart := "\n" + botPersona + ":\n"
@@ -210,11 +224,9 @@ func (op LCPChat) ParseChunk(data []byte) (*models.TextChunk, error) {
 		logger.Error("failed to decode", "error", err, "line", string(data))
 		return nil, err
 	}
-
-	// Handle multiple choices safely
 	if len(llmchunk.Choices) == 0 {
-		logger.Warn("LCPChat ParseChunk: no choices in response", "data", string(data))
+		logger.Warn("LCPChat empty chunk choices", "raw_data", string(data), "chunk", llmchunk)
-		return &models.TextChunk{Finished: true}, nil
+		return &models.TextChunk{}, nil
 	}
 	lastChoice := llmchunk.Choices[len(llmchunk.Choices)-1]
 	resp := &models.TextChunk{
@@ -335,6 +347,10 @@ func (ds DeepSeekerCompletion) ParseChunk(data []byte) (*models.TextChunk, error
 		logger.Error("failed to decode", "error", err, "line", string(data))
 		return nil, err
 	}
+	if len(llmchunk.Choices) == 0 {
+		logger.Warn("empty chunk choices", "raw_data", string(data), "chunk", llmchunk)
+		return &models.TextChunk{}, nil
+	}
 	resp := &models.TextChunk{
 		Chunk: llmchunk.Choices[0].Text,
 	}
@@ -400,6 +416,10 @@ func (ds DeepSeekerChat) ParseChunk(data []byte) (*models.TextChunk, error) {
 		return nil, err
 	}
 	resp := &models.TextChunk{}
+	if len(llmchunk.Choices) == 0 {
+		logger.Warn("empty chunk choices", "raw_data", string(data), "chunk", llmchunk)
+		return resp, nil
+	}
 	if llmchunk.Choices[0].FinishReason != "" {
 		if llmchunk.Choices[0].Delta.Content != "" {
 			logger.Error("text inside of finish llmchunk", "chunk", llmchunk)
@@ -482,6 +502,10 @@ func (or OpenRouterCompletion) ParseChunk(data []byte) (*models.TextChunk, error
 		logger.Error("failed to decode", "error", err, "line", string(data))
 		return nil, err
 	}
+	if len(llmchunk.Choices) == 0 {
+		logger.Warn("empty chunk choices", "raw_data", string(data), "chunk", llmchunk)
+		return &models.TextChunk{}, nil
+	}
 	resp := &models.TextChunk{
 		Chunk: llmchunk.Choices[len(llmchunk.Choices)-1].Text,
 	}
@@ -544,6 +568,10 @@ func (or OpenRouterChat) ParseChunk(data []byte) (*models.TextChunk, error) {
 		logger.Error("failed to decode", "error", err, "line", string(data))
 		return nil, err
 	}
+	if len(llmchunk.Choices) == 0 {
+		logger.Warn("empty chunk choices", "raw_data", string(data), "chunk", llmchunk)
+		return &models.TextChunk{}, nil
+	}
 	lastChoice := llmchunk.Choices[len(llmchunk.Choices)-1]
 	resp := &models.TextChunk{
 		Chunk:     lastChoice.Delta.Content,

main.go

@@ -17,8 +17,9 @@ var (
 	shellHistoryPos   int = -1
 	thinkingCollapsed     = false
 	toolCollapsed         = true
-	statusLineTempl       = "help (F12) | chat: [orange:-:b]%s[-:-:-] (F1) | [%s:-:b]tool use[-:-:-] (ctrl+k) | model: [%s:-:b]%s[-:-:-] (ctrl+l) | [%s:-:b]skip LLM resp[-:-:-] (F10)\nAPI: [orange:-:b]%s[-:-:-] (ctrl+v) | writing as: [orange:-:b]%s[-:-:-] (ctrl+q) | bot will write as [orange:-:b]%s[-:-:-] (ctrl+x)"
+	statusLineTempl       = "help (F12) | chat: [orange:-:b]%s[-:-:-] (F1) | [%s:-:b]tool use[-:-:-] (ctrl+k) | model: [%s:-:b]%s[-:-:-] (ctrl+l) | [%s:-:b]skip LLM resp[-:-:-] (F10) | API: [orange:-:b]%s[-:-:-] (ctrl+v)\nwriting as: [orange:-:b]%s[-:-:-] (ctrl+q) | bot will write as [orange:-:b]%s[-:-:-] (ctrl+x)"
 	focusSwitcher         = map[tview.Primitive]tview.Primitive{}
+	app               *tview.Application
 )
 
 func main() {

models/card.go

@@ -1,6 +1,10 @@
 package models
 
-import "strings"
+import (
+	"crypto/md5"
+	"fmt"
+	"strings"
+)
 
 // https://github.com/malfoyslastname/character-card-spec-v2/blob/main/spec_v2.md
 // what a bloat; trim to Role->Msg pair and first msg
@@ -31,6 +35,7 @@ func (c *CharCardSpec) Simplify(userName, fpath string) *CharCard {
 	fm := strings.ReplaceAll(strings.ReplaceAll(c.FirstMes, "{{char}}", c.Name), "{{user}}", userName)
 	sysPr := strings.ReplaceAll(strings.ReplaceAll(c.Description, "{{char}}", c.Name), "{{user}}", userName)
 	return &CharCard{
+		ID:         ComputeCardID(c.Name, fpath),
 		SysPrompt:  sysPr,
 		FirstMsg:   fm,
 		Role:       c.Name,
@@ -39,7 +44,12 @@ func (c *CharCardSpec) Simplify(userName, fpath string) *CharCard {
 	}
 }
 
+func ComputeCardID(role, filePath string) string {
+	return fmt.Sprintf("%x", md5.Sum([]byte(role+filePath)))
+}
+
 type CharCard struct {
+	ID         string   `json:"id"`
 	SysPrompt  string   `json:"sys_prompt"`
 	FirstMsg   string   `json:"first_msg"`
 	Role       string   `json:"role"`

models/consts.go

@@ -1,7 +1,8 @@
 package models
 
 const (
-	LoadedMark = "(loaded) "
+	LoadedMark        = "(loaded) "
+	ToolRespMultyType = "multimodel_content"
 )
 
 type APIType int

models/models.go

@@ -391,7 +391,6 @@ func CreateImageURLFromPath(imagePath string) (string, error) {
 	if err != nil {
 		return "", err
 	}
-
 	// Determine the image format based on file extension
 	var mimeType string
 	switch {
@@ -408,10 +407,8 @@ func CreateImageURLFromPath(imagePath string) (string, error) {
 	default:
 		mimeType = "image/jpeg" // default
 	}
-
 	// Encode to base64
 	encoded := base64.StdEncoding.EncodeToString(data)
-
 	// Create data URL
 	return fmt.Sprintf("data:%s;base64,%s", mimeType, encoded), nil
 }
@@ -611,6 +608,20 @@ func (lcp *LCPModels) ListModels() []string {
 	return resp
 }
 
+func (lcp *LCPModels) HasVision(modelID string) bool {
+	for _, m := range lcp.Data {
+		if m.ID == modelID {
+			args := m.Status.Args
+			for i := 0; i < len(args)-1; i++ {
+				if args[i] == "--mmproj" {
+					return true
+				}
+			}
+		}
+	}
+	return false
+}
+
 type ResponseStats struct {
 	Tokens       int
 	Duration     float64
@@ -623,3 +634,8 @@ type ChatRoundReq struct {
 	Regen   bool
 	Resume  bool
 }
+
+type MultimodalToolResp struct {
+	Type  string              `json:"type"`
+	Parts []map[string]string `json:"parts"`
+}

models/openrouter.go

@@ -172,3 +172,16 @@ func (orm *ORModels) ListModels(free bool) []string {
 	}
 	return resp
 }
+
+func (orm *ORModels) HasVision(modelID string) bool {
+	for i := range orm.Data {
+		if orm.Data[i].ID == modelID {
+			for _, mod := range orm.Data[i].Architecture.InputModalities {
+				if mod == "image" {
+					return true
+				}
+			}
+		}
+	}
+	return false
+}

pngmeta/metareader.go

@@ -109,6 +109,12 @@ func ReadCardJson(fname string) (*models.CharCard, error) {
 	if err := json.Unmarshal(data, &card); err != nil {
 		return nil, err
 	}
+	if card.FilePath == "" {
+		card.FilePath = fname
+	}
+	if card.ID == "" {
+		card.ID = models.ComputeCardID(card.Role, card.FilePath)
+	}
 	return &card, nil
 }
 

popups.go

@@ -40,9 +40,7 @@ func showModelSelectionPopup() {
 		default:
 			message = "No llama.cpp models loaded. Ensure llama.cpp server is running with models."
 		}
-		if err := notifyUser("Empty list", message); err != nil {
+		showToast("Empty list", message)
-			logger.Error("failed to send notification", "error", err)
-		}
 		return
 	}
 	// Create a list primitive
@@ -119,9 +117,7 @@ func showAPILinkSelectionPopup() {
 	if len(apiLinks) == 0 {
 		logger.Warn("no API links available for selection")
 		message := "No API links available. Please configure API links in your config file."
-		if err := notifyUser("Empty list", message); err != nil {
+		showToast("Empty list", message)
-			logger.Error("failed to send notification", "error", err)
-		}
 		return
 	}
 	// Create a list primitive
@@ -143,6 +139,7 @@ func showAPILinkSelectionPopup() {
 	apiListWidget.SetSelectedFunc(func(index int, mainText string, secondaryText string, shortcut rune) {
 		// Update the API in config
 		cfg.CurrentAPI = mainText
+		// updateToolCapabilities()
 		// Update model list based on new API
 		// Helper function to get model list for a given API (same as in props_table.go)
 		getModelListForAPI := func(api string) []string {
@@ -160,8 +157,9 @@ func showAPILinkSelectionPopup() {
 		newModelList := getModelListForAPI(cfg.CurrentAPI)
 		// Ensure chatBody.Model is in the new list; if not, set to first available model
 		if len(newModelList) > 0 && !slices.Contains(newModelList, chatBody.Model) {
-			chatBody.Model = newModelList[0]
+			chatBody.Model = strings.TrimPrefix(newModelList[0], models.LoadedMark)
 			cfg.CurrentModel = chatBody.Model
+			updateToolCapabilities()
 		}
 		pages.RemovePage("apiLinkSelectionPopup")
 		app.SetFocus(textArea)
@@ -204,9 +202,7 @@ func showUserRoleSelectionPopup() {
 	if len(roles) == 0 {
 		logger.Warn("no roles available for selection")
 		message := "No roles available for selection."
-		if err := notifyUser("Empty list", message); err != nil {
+		showToast("Empty list", message)
-			logger.Error("failed to send notification", "error", err)
-		}
 		return
 	}
 	// Create a list primitive
@@ -283,9 +279,7 @@ func showBotRoleSelectionPopup() {
 	if len(roles) == 0 {
 		logger.Warn("no roles available for selection")
 		message := "No roles available for selection."
-		if err := notifyUser("Empty list", message); err != nil {
+		showToast("Empty list", message)
-			logger.Error("failed to send notification", "error", err)
-		}
 		return
 	}
 	// Create a list primitive
@@ -404,6 +398,66 @@ func showShellFileCompletionPopup(filter string) {
 	app.SetFocus(widget)
 }
 
+func showTextAreaFileCompletionPopup(filter string) {
+	baseDir := cfg.FilePickerDir
+	if baseDir == "" {
+		baseDir = "."
+	}
+	complMatches := scanFiles(baseDir, filter)
+	if len(complMatches) == 0 {
+		return
+	}
+	if len(complMatches) == 1 {
+		currentText := textArea.GetText()
+		atIdx := strings.LastIndex(currentText, "@")
+		if atIdx >= 0 {
+			before := currentText[:atIdx]
+			textArea.SetText(before+complMatches[0], true)
+		}
+		return
+	}
+	widget := tview.NewList().ShowSecondaryText(false).
+		SetSelectedBackgroundColor(tcell.ColorGray)
+	widget.SetTitle("file completion").SetBorder(true)
+	for _, m := range complMatches {
+		widget.AddItem(m, "", 0, nil)
+	}
+	widget.SetSelectedFunc(func(index int, mainText string, secondaryText string, shortcut rune) {
+		currentText := textArea.GetText()
+		atIdx := strings.LastIndex(currentText, "@")
+		if atIdx >= 0 {
+			before := currentText[:atIdx]
+			textArea.SetText(before+mainText, true)
+		}
+		pages.RemovePage("textAreaFileCompletionPopup")
+		app.SetFocus(textArea)
+	})
+	widget.SetInputCapture(func(event *tcell.EventKey) *tcell.EventKey {
+		if event.Key() == tcell.KeyEscape {
+			pages.RemovePage("textAreaFileCompletionPopup")
+			app.SetFocus(textArea)
+			return nil
+		}
+		if event.Key() == tcell.KeyRune && event.Rune() == 'x' {
+			pages.RemovePage("textAreaFileCompletionPopup")
+			app.SetFocus(textArea)
+			return nil
+		}
+		return event
+	})
+	modal := func(p tview.Primitive, width, height int) tview.Primitive {
+		return tview.NewFlex().
+			AddItem(nil, 0, 1, false).
+			AddItem(tview.NewFlex().SetDirection(tview.FlexRow).
+				AddItem(nil, 0, 1, false).
+				AddItem(p, height, 1, true).
+				AddItem(nil, 0, 1, false), width, 1, true).
+			AddItem(nil, 0, 1, false)
+	}
+	pages.AddPage("textAreaFileCompletionPopup", modal(widget, 80, 20), true, true)
+	app.SetFocus(widget)
+}
+
 func updateWidgetColors(theme *tview.Theme) {
 	bgColor := theme.PrimitiveBackgroundColor
 	fgColor := theme.PrimaryTextColor
@@ -450,9 +504,7 @@ func showColorschemeSelectionPopup() {
 	if len(schemeNames) == 0 {
 		logger.Warn("no colorschemes available for selection")
 		message := "No colorschemes available."
-		if err := notifyUser("Empty list", message); err != nil {
+		showToast("Empty list", message)
-			logger.Error("failed to send notification", "error", err)
-		}
 		return
 	}
 	// Create a list primitive

props_table.go

@@ -259,9 +259,7 @@ func makePropsTable(props map[string]float32) *tview.Table {
 				// Handle nil options
 				if data.Options == nil {
 					logger.Error("options list is nil for", "label", label)
-					if err := notifyUser("Configuration error", "Options list is nil for "+label); err != nil {
+					showToast("Configuration error", "Options list is nil for "+label)
-						logger.Error("failed to send notification", "error", err)
-					}
 					return
 				}
 
@@ -279,9 +277,7 @@ func makePropsTable(props map[string]float32) *tview.Table {
 							message = "No llama.cpp models loaded. Ensure llama.cpp server is running with models."
 						}
 					}
-					if err := notifyUser("Empty list", message); err != nil {
+					showToast("Empty list", message)
-						logger.Error("failed to send notification", "error", err)
-					}
 					return
 				}
 				// Create a list primitive

rag/embedder.go

@@ -9,6 +9,10 @@ import (
 	"gf-lt/models"
 	"log/slog"
 	"net/http"
+
+	"github.com/sugarme/tokenizer"
+	"github.com/sugarme/tokenizer/pretrained"
+	"github.com/yalue/onnxruntime_go"
 )
 
 // Embedder defines the interface for embedding text
@@ -134,11 +138,150 @@ func (a *APIEmbedder) EmbedSlice(lines []string) ([][]float32, error) {
 	return embeddings, nil
 }
 
-// TODO: ONNXEmbedder implementation would go here
-// This would require:
 // 1. Loading ONNX models locally
 // 2. Using a Go ONNX runtime (like gorgonia/onnx or similar)
 // 3. Converting text to embeddings without external API calls
-//
+type ONNXEmbedder struct {
-// For now, we'll focus on the API implementation which is already working in the current system,
+	session   *onnxruntime_go.DynamicAdvancedSession
-// and can be extended later when we have ONNX runtime integration
+	tokenizer *tokenizer.Tokenizer
+	dims      int // embedding dimension (e.g., 768)
+	logger    *slog.Logger
+}
+
+func NewONNXEmbedder(modelPath, tokenizerPath string, dims int, logger *slog.Logger) (*ONNXEmbedder, error) {
+	// Load tokenizer using sugarme/tokenizer
+	tok, err := pretrained.FromFile(tokenizerPath)
+	if err != nil {
+		return nil, fmt.Errorf("failed to load tokenizer: %w", err)
+	}
+	// Create ONNX session
+	session, err := onnxruntime_go.NewDynamicAdvancedSession(
+		modelPath, // onnx/embedgemma/model_q4.onnx
+		[]string{"input_ids", "attention_mask"},
+		[]string{"sentence_embedding"},
+		nil, // optional options
+	)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create ONNX session: %w", err)
+	}
+	return &ONNXEmbedder{
+		session:   session,
+		tokenizer: tok,
+		dims:      dims,
+		logger:    logger,
+	}, nil
+}
+
+func (e *ONNXEmbedder) Embed(text string) ([]float32, error) {
+	// 1. Tokenize
+	encoding, err := e.tokenizer.EncodeSingle(text)
+	if err != nil {
+		return nil, fmt.Errorf("tokenization failed: %w", err)
+	}
+	// 2. Convert to int64 and create attention mask
+	ids := encoding.Ids
+	inputIDs := make([]int64, len(ids))
+	attentionMask := make([]int64, len(ids))
+	for i, id := range ids {
+		inputIDs[i] = int64(id)
+		attentionMask[i] = 1
+	}
+	// 3. Create input tensors (shape: [1, seq_len])
+	seqLen := int64(len(inputIDs))
+	inputIDsTensor, err := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(1, seqLen),
+		inputIDs,
+	)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create input_ids tensor: %w", err)
+	}
+	defer inputIDsTensor.Destroy()
+	maskTensor, err := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(1, seqLen),
+		attentionMask,
+	)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create attention_mask tensor: %w", err)
+	}
+	defer maskTensor.Destroy()
+	// 4. Create output tensor
+	outputTensor, err := onnxruntime_go.NewEmptyTensor[float32](
+		onnxruntime_go.NewShape(1, int64(e.dims)),
+	)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create output tensor: %w", err)
+	}
+	defer outputTensor.Destroy()
+	// 5. Run inference
+	err = e.session.Run(
+		[]onnxruntime_go.Value{inputIDsTensor, maskTensor},
+		[]onnxruntime_go.Value{outputTensor},
+	)
+	if err != nil {
+		return nil, fmt.Errorf("inference failed: %w", err)
+	}
+	// 6. Copy output data
+	outputData := outputTensor.GetData()
+	embedding := make([]float32, len(outputData))
+	copy(embedding, outputData)
+	return embedding, nil
+}
+
+func (e *ONNXEmbedder) EmbedSlice(texts []string) ([][]float32, error) {
+	encodings := make([]*tokenizer.Encoding, len(texts))
+	maxLen := 0
+	for i, txt := range texts {
+		enc, err := e.tokenizer.EncodeSingle(txt)
+		if err != nil {
+			return nil, err
+		}
+		encodings[i] = enc
+		if l := len(enc.Ids); l > maxLen {
+			maxLen = l
+		}
+	}
+	batchSize := len(texts)
+	inputIDs := make([]int64, batchSize*maxLen)
+	attentionMask := make([]int64, batchSize*maxLen)
+	for i, enc := range encodings {
+		ids := enc.Ids
+		offset := i * maxLen
+		for j, id := range ids {
+			inputIDs[offset+j] = int64(id)
+			attentionMask[offset+j] = 1
+		}
+		// Remaining positions are already zero (padding)
+	}
+	// Create tensors with shape [batchSize, maxLen]
+	inputTensor, _ := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(int64(batchSize), int64(maxLen)),
+		inputIDs,
+	)
+	defer inputTensor.Destroy()
+	maskTensor, _ := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(int64(batchSize), int64(maxLen)),
+		attentionMask,
+	)
+	defer maskTensor.Destroy()
+	outputTensor, _ := onnxruntime_go.NewEmptyTensor[float32](
+		onnxruntime_go.NewShape(int64(batchSize), int64(e.dims)),
+	)
+	defer outputTensor.Destroy()
+	err := e.session.Run(
+		[]onnxruntime_go.Value{inputTensor, maskTensor},
+		[]onnxruntime_go.Value{outputTensor},
+	)
+	if err != nil {
+		return nil, err
+	}
+	// Extract embeddings per batch item
+	data := outputTensor.GetData()
+	embeddings := make([][]float32, batchSize)
+	for i := 0; i < batchSize; i++ {
+		start := i * e.dims
+		emb := make([]float32, e.dims)
+		copy(emb, data[start:start+e.dims])
+		embeddings[i] = emb
+	}
+	return embeddings, nil
+}

rag/rag.go

@@ -246,7 +246,7 @@ func (r *RAG) extractImportantPhrases(query string) string {
 				break
 			}
 		}
-		if isImportant || len(word) > 3 {
+		if isImportant || len(word) >= 3 {
 			important = append(important, word)
 		}
 	}

rag_issues.md

@@ -0,0 +1,104 @@
+# RAG Implementation Issues and Proposed Solutions
+
+## Overview
+The current RAG system fails to retrieve relevant information for specific queries, as demonstrated by the inability to find the "two she bears" reference in the KJV Bible (2 Kings 2:23-24). While the system retrieves documents containing the word "bear", it misses the actual verse, indicating fundamental flaws in chunking, query processing, retrieval, and answer synthesis. Below we dissect each problem and propose concrete solutions.
+
+---
+
+### Problem 1: Chunking Destroys Semantic Coherence
+- **Problem description**  
+  The current chunking splits text into sentences and groups them by a simple word count threshold (`RAGWordLimit`). This ignores document structure (chapters, headings) and can cut through narrative units, scattering related content across multiple chunks. For the Bible query, the story of Elisha and the bears likely spans multiple verses; splitting it prevents any single chunk from containing the full context, diluting the embedding signal and making retrieval difficult.
+
+- **Proposed solution**  
+  - **Structure-aware chunking**: Use the EPUB’s internal structure (chapters, sections) to create chunks that align with logical content units (e.g., by chapter or story).  
+  - **Overlap between chunks**: Add a configurable overlap (e.g., 10–20% of chunk size) to preserve continuity, ensuring key phrases like "two she bears" are not split across boundaries.  
+  - **Rich metadata**: Store book name, chapter, and verse numbers with each chunk to enable filtering and source attribution.  
+  - **Fallback to recursive splitting**: For documents without clear structure, use a recursive character text splitter with overlap (similar to LangChain’s `RecursiveCharacterTextSplitter`) to maintain semantic boundaries (paragraphs, sentences).
+
+---
+
+### Problem 2: Query Refinement Strips Critical Context
+- **Problem description**  
+  `RefineQuery` removes stop words and applies keyword-based filtering that discards semantically important modifiers. For "two she bears", the word "she" (a gender modifier) may be treated as a stop word, leaving "two bears". This loses the specificity of the query and causes the embedding to drift toward generic "bear" contexts. The rule-based approach cannot understand that "she bears" is a key phrase in the biblical story.
+
+- **Proposed solution**  
+  - **Entity-aware query preservation**: Use a lightweight NLP model (e.g., spaCy or a BERT-based NER tagger) to identify and retain key entities (quantities, animals, names) while only removing truly irrelevant stop words.  
+  - **Intelligent query rewriting**: Employ a small LLM (or a set of transformation rules) to generate query variations that reflect likely biblical phrasing, e.g., "two bears came out of the wood" or "Elisha and the bears".  
+  - **Contextual stop word removal**: Instead of a static list, use a POS tagger to keep adjectives, nouns, and verbs while removing only function words that don't carry meaning.  
+  - **Disable refinement for short queries**: If the query is already concise (like "two she bears"), skip aggressive filtering.
+
+---
+
+### Problem 3: Embedding Similarity Fails for Rare or Specific Phrases
+- **Problem description**  
+  Dense embeddings excel at capturing semantic similarity but can fail when the query contains rare phrases or when the relevant passage is embedded in a noisy chunk. The verse "there came forth two she bears out of the wood" shares only the word "bears" with the query, and its embedding may be pulled toward the average of surrounding verses. Consequently, the similarity score may be lower than that of other chunks containing the word "bear" in generic contexts.
+
+- **Proposed solution**  
+  - **Hybrid retrieval**: Combine dense embeddings with BM25 (keyword) search. BM25 excels at exact phrase matching and would likely retrieve the verse based on "two bears" even if the embedding is weak.  
+    - Use a library like [blevesearch](https://github.com/blevesearch/bleve) to index text alongside vectors.  
+    - Fuse results using Reciprocal Rank Fusion (RRF) or a weighted combination.  
+  - **Query expansion**: Add relevant terms to the query (e.g., "Elisha", "2 Kings") to improve embedding alignment.  
+  - **Fine-tuned embeddings**: Consider using an embedding model fine-tuned on domain-specific data (e.g., biblical texts) if this is a recurring use case.  
+
+---
+
+### Problem 4: Reranking Heuristics Are Insufficient
+- **Problem description**  
+  `RerankResults` boosts results based on simple keyword matching and file name heuristics. This coarse approach cannot reliably promote the correct verse over false positives. The adjustment `distance - score/100` is arbitrary and may not reflect true relevance.
+
+- **Proposed solution**  
+  - **Cross-encoder reranking**: After retrieving top candidates (e.g., top 20) with hybrid search, rerank them using a cross-encoder model that directly computes the relevance score between the query and each chunk.  
+    - Models like `cross-encoder/ms-marco-MiniLM-L-6-v2` are lightweight and can be run locally or via a microservice.  
+  - **Score normalization**: Use the cross-encoder scores to reorder results, discarding low-scoring ones.  
+  - **Contextual boosting**: If metadata (e.g., chapter/verse) is available, boost results that match the query’s expected location (if inferable).  
+
+---
+
+### Problem 5: Answer Synthesis Is Not Generative
+- **Problem description**  
+  `SynthesizeAnswer` embeds a prompt and attempts to retrieve a pre-stored answer, falling back to concatenating truncated chunks. This is fundamentally flawed: RAG requires an LLM to generate a coherent answer from retrieved context. In the Bible example, even if the correct verse were retrieved, the system would only output a snippet, not an answer explaining the reference.
+
+- **Proposed solution**  
+  - **Integrate an LLM for generation**: Use a local model (via Ollama, Llama.cpp) or a cloud API (OpenAI, etc.) to synthesize answers.  
+    - Construct a prompt that includes the retrieved chunks (with metadata) and the user query.  
+    - Instruct the model to answer based solely on the provided context and cite sources (e.g., "According to 2 Kings 2:24...").  
+  - **Implement a fallback**: If no relevant chunks are retrieved, return a message like "I couldn't find that information in your documents."  
+  - **Streaming support**: For better UX, stream the answer token-by-token.  
+
+---
+
+### Problem 6: Concurrency and Error Handling
+- **Problem description**  
+  The code uses a mutex only in `LoadRAG`, leaving other methods vulnerable to race conditions. The global status channel `LongJobStatusCh` may drop messages due to `select/default`, and errors are sometimes logged but not propagated. Ingestion is synchronous and slow.
+
+- **Proposed solution**  
+  - **Add context support**: Pass `context.Context` to all methods to allow cancellation and timeouts.  
+  - **Worker pools for embedding**: Parallelize batch embedding with a controlled number of workers to respect API rate limits and speed up ingestion.  
+  - **Retry logic**: Implement exponential backoff for transient API errors.  
+  - **Replace global channel**: Use a callback or an injectable status reporter to avoid dropping messages.  
+  - **Fine-grained locking**: Protect shared state (e.g., `storage`) with appropriate synchronization.  
+
+---
+
+### Problem 7: Lack of Monitoring and Evaluation
+- **Problem description**  
+  There are no metrics to track retrieval quality, latency, or user satisfaction. The failure case was discovered manually; without systematic evaluation, regressions will go unnoticed.
+
+- **Proposed solution**  
+  - **Log key metrics**: Record query, retrieved chunk IDs, scores, and latency for each search.  
+  - **User feedback**: Add a mechanism for users to rate answers (thumbs up/down) and use this data to improve retrieval.  
+  - **Offline evaluation**: Create a test set of queries and expected relevant chunks (e.g., the Bible example) to measure recall@k, MRR, etc., and run it after each change.  
+
+---
+
+## Summary
+Fixing the RAG pipeline requires a multi-pronged approach:
+1. **Structure-aware chunking** with metadata.
+2. **Hybrid retrieval** (dense + sparse).
+3. **Query understanding** via entity preservation and intelligent rewriting.
+4. **Cross-encoder reranking** for precision.
+5. **LLM-based answer generation**.
+6. **Robust concurrency and error handling**.
+7. **Monitoring and evaluation** to track improvements.
+
+Implementing these changes will transform the system from a brittle keyword matcher into a reliable knowledge assistant capable of handling nuanced queries like the "two she bears" reference.

session.go

@@ -168,8 +168,3 @@ func copyToClipboard(text string) error {
 	cmd.Stdin = strings.NewReader(text)
 	return cmd.Run()
 }
-
-func notifyUser(topic, message string) error {
-	cmd := exec.Command("notify-send", topic, message)
-	return cmd.Run()
-}

storage/migrate.go

@@ -10,16 +10,18 @@ import (
 //go:embed migrations/*
 var migrationsFS embed.FS
 
-func (p *ProviderSQL) Migrate() {
+func (p *ProviderSQL) Migrate() error {
 	// Get the embedded filesystem
 	migrationsDir, err := fs.Sub(migrationsFS, "migrations")
 	if err != nil {
 		p.logger.Error("Failed to get embedded migrations directory;", "error", err)
+		return fmt.Errorf("failed to get embedded migrations directory: %w", err)
 	}
 	// List all .up.sql files
 	files, err := migrationsFS.ReadDir("migrations")
 	if err != nil {
 		p.logger.Error("Failed to read migrations directory;", "error", err)
+		return fmt.Errorf("failed to read migrations directory: %w", err)
 	}
 	// Execute each .up.sql file
 	for _, file := range files {
@@ -27,11 +29,12 @@ func (p *ProviderSQL) Migrate() {
 			err := p.executeMigration(migrationsDir, file.Name())
 			if err != nil {
 				p.logger.Error("Failed to execute migration %s: %v", file.Name(), err)
-				panic(err)
+				return fmt.Errorf("failed to execute migration %s: %w", file.Name(), err)
 			}
 		}
 	}
 	p.logger.Debug("All migrations executed successfully!")
+	return nil
 }
 
 func (p *ProviderSQL) executeMigration(migrationsDir fs.FS, fileName string) error {

storage/storage.go

@@ -103,7 +103,10 @@ func NewProviderSQL(dbPath string, logger *slog.Logger) FullRepo {
 		return nil
 	}
 	p := ProviderSQL{db: db, logger: logger}
-	p.Migrate()
+	if err := p.Migrate(); err != nil {
+		logger.Error("migration failed, app cannot start", "error", err)
+		return nil
+	}
 	return p
 }
 

sysprompts/cluedo.json

@@ -1,7 +0,0 @@
-{
-  "sys_prompt": "A game of cluedo. Players are {{user}}, {{char}}, {{char2}};\n\nrooms: hall, lounge, dinning room kitchen, ballroom, conservatory, billiard room, library, study;\nweapons: candlestick, dagger, lead pipe, revolver, rope, spanner;\npeople: miss Scarlett, colonel Mustard, mrs. White, reverend Green, mrs. Peacock, professor Plum;\n\nA murder happened in a mansion with 9 rooms. Victim is dr. Black.\nPlayers goal is to find out who commited a murder, in what room and with what weapon.\nWeapons, people and rooms not involved in murder are distributed between players (as cards) by tool agent.\nThe objective of the game is to deduce the details of the murder. There are six characters, six murder weapons, and nine rooms, leaving the players with 324 possibilities. As soon as a player enters a room, they may make a suggestion as to the details, naming a suspect, the room they are in, and the weapon. For example: \"I suspect Professor Plum, in the Dining Room, with the candlestick\".\nOnce a player makes a suggestion, the others are called upon to disprove it.\nBefore the player's move, tool agent will remind that players their cards. There are two types of moves: making a suggestion (suggestion_move) and disproving other player suggestion (evidence_move);\nIn this version player wins when the correct details are named in the suggestion_move.\n\n<example_game>\n{{user}}:\nlet's start a game of cluedo!\ntool: cards of {{char}} are 'LEAD PIPE', 'BALLROOM', 'CONSERVATORY', 'STUDY', 'Mrs. White'; suggestion_move;\n{{char}}:\n(putting miss Scarlet into the Hall with the Revolver) \"I suspect miss Scarlett, in the Hall, with the revolver.\"\ntool: cards of {{char2}} are 'SPANNER', 'DAGGER', 'Professor Plum', 'LIBRARY', 'Mrs. Peacock'; evidence_move;\n{{char2}}:\n\"No objections.\" (no cards matching the suspicion of {{char}})\ntool: cards of {{user}} are 'Colonel Mustard', 'Miss Scarlett', 'DINNING ROOM', 'CANDLESTICK', 'HALL'; evidence_move;\n{{user}}:\n\"I object. Miss Scarlett is innocent.\" (shows card with 'Miss Scarlett')\ntool: cards of {{char2}} are 'SPANNER', 'DAGGER', 'Professor Plum', 'LIBRARY', 'Mrs. Peacock'; suggestion_move;\n{{char2}}:\n*So it was not Miss Scarlett, good to know.*\n(moves Mrs. White to the Billiard Room) \"It might have been Mrs. White, in the Billiard Room, with the Revolver.\"\ntool: cards of {{user}} are 'Colonel Mustard', 'Miss Scarlett', 'DINNING ROOM', 'CANDLESTICK', 'HALL'; evidence_move;\n{{user}}:\n(no matching cards for the assumption of {{char2}}) \"Sounds possible to me.\"\ntool: cards of {{char}} are 'LEAD PIPE', 'BALLROOM', 'CONSERVATORY', 'STUDY', 'Mrs. White'; evidence_move;\n{{char}}:\n(shows Mrs. White card) \"No. Was not Mrs. White\"\ntool: cards of {{user}} are 'Colonel Mustard', 'Miss Scarlett', 'DINNING ROOM', 'CANDLESTICK', 'HALL'; suggestion_move;\n{{user}}:\n*So not Mrs. White...* (moves Reverend Green into the Billiard Room) \"I suspect Reverend Green, in the Billiard Room, with the Revolver.\"\ntool: Correct. It was Reverend Green in the Billiard Room, with the revolver. {{user}} wins.\n</example_game>",
-  "role": "CluedoPlayer",
-  "role2": "CluedoEnjoyer",
-  "filepath": "sysprompts/cluedo.json",
-  "first_msg": "Hey guys! Want to play cluedo?"
-}

tables.go

@@ -147,9 +147,7 @@ func makeChatTable(chatMap map[string]models.Chat) *tview.Table {
 			if err := store.RemoveChat(sc.ID); err != nil {
 				logger.Error("failed to remove chat from db", "chat_id", sc.ID, "chat_name", sc.Name)
 			}
-			if err := notifyUser("chat deleted", selectedChat+" was deleted"); err != nil {
+			showToast("chat deleted", selectedChat+" was deleted")
-				logger.Error("failed to send notification", "error", err)
-			}
 			// load last chat
 			chatBody.Messages = loadOldChatOrGetNew()
 			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
@@ -159,27 +157,16 @@ func makeChatTable(chatMap map[string]models.Chat) *tview.Table {
 			// save updated card
 			fi := strings.Index(selectedChat, "_")
 			agentName := selectedChat[fi+1:]
-			cc, ok := sysMap[agentName]
+			cc := GetCardByRole(agentName)
-			if !ok {
+			if cc == nil {
 				logger.Warn("no such card", "agent", agentName)
-				//no:lint
+				showToast("error", "no such card: "+agentName)
-				if err := notifyUser("error", "no such card: "+agentName); err != nil {
-					logger.Warn("failed ot notify", "error", err)
-				}
 				return
 			}
-			// if chatBody.Messages[0].Role != "system" || chatBody.Messages[1].Role != agentName {
-			// 	if err := notifyUser("error", "unexpected chat structure; card: "+agentName); err != nil {
-			// 		logger.Warn("failed ot notify", "error", err)
-			// 	}
-			// 	return
-			// }
-			// change sys_prompt + first msg
 			cc.SysPrompt = chatBody.Messages[0].Content
 			cc.FirstMsg = chatBody.Messages[1].Content
 			if err := pngmeta.WriteToPng(cc.ToSpec(cfg.UserRole), cc.FilePath, cc.FilePath); err != nil {
-				logger.Error("failed to write charcard",
+				logger.Error("failed to write charcard", "error", err)
-					"error", err)
 			}
 			return
 		case "move sysprompt onto 1st msg":
@@ -190,33 +177,29 @@ func makeChatTable(chatMap map[string]models.Chat) *tview.Table {
 			pages.RemovePage(historyPage)
 			return
 		case "new_chat_from_card":
-			// Reread card from file and start fresh chat
 			fi := strings.Index(selectedChat, "_")
 			agentName := selectedChat[fi+1:]
-			cc, ok := sysMap[agentName]
+			cc := GetCardByRole(agentName)
-			if !ok {
+			if cc == nil {
 				logger.Warn("no such card", "agent", agentName)
-				if err := notifyUser("error", "no such card: "+agentName); err != nil {
+				showToast("error", "no such card: "+agentName)
-					logger.Warn("failed to notify", "error", err)
-				}
 				return
 			}
-			// Reload card from disk
 			newCard, err := pngmeta.ReadCard(cc.FilePath, cfg.UserRole)
 			if err != nil {
 				logger.Error("failed to reload charcard", "path", cc.FilePath, "error", err)
 				newCard, err = pngmeta.ReadCardJson(cc.FilePath)
 				if err != nil {
 					logger.Error("failed to reload charcard", "path", cc.FilePath, "error", err)
-					if err := notifyUser("error", "failed to reload card: "+cc.FilePath); err != nil {
+					showToast("error", "failed to reload card: "+cc.FilePath)
-						logger.Warn("failed to notify", "error", err)
-					}
 					return
 				}
 			}
-			// Update sysMap with fresh card data
+			if newCard.ID == "" {
-			sysMap[agentName] = newCard
+				newCard.ID = models.ComputeCardID(newCard.Role, newCard.FilePath)
-			// fetching sysprompt and first message anew from the card
+			}
+			sysMap[newCard.ID] = newCard
+			roleToID[newCard.Role] = newCard.ID
 			startNewChat(false)
 			pages.RemovePage(historyPage)
 			return
@@ -457,13 +440,13 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 			go func() {
 				if err := ragger.LoadRAG(fpath); err != nil {
 					logger.Error("failed to embed file", "chat", fpath, "error", err)
-					_ = notifyUser("RAG", "failed to embed file; error: "+err.Error())
+					showToast("RAG", "failed to embed file; error: "+err.Error())
 					app.QueueUpdate(func() {
 						pages.RemovePage(RAGPage)
 					})
 					return
 				}
-				_ = notifyUser("RAG", "file loaded successfully")
+				showToast("RAG", "file loaded successfully")
 				app.QueueUpdate(func() {
 					pages.RemovePage(RAGPage)
 				})
@@ -474,13 +457,13 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 			go func() {
 				if err := ragger.RemoveFile(f.name); err != nil {
 					logger.Error("failed to unload file from RAG", "filename", f.name, "error", err)
-					_ = notifyUser("RAG", "failed to unload file; error: "+err.Error())
+					showToast("RAG", "failed to unload file; error: "+err.Error())
 					app.QueueUpdate(func() {
 						pages.RemovePage(RAGPage)
 					})
 					return
 				}
-				_ = notifyUser("RAG", "file unloaded successfully")
+				showToast("RAG", "file unloaded successfully")
 				app.QueueUpdate(func() {
 					pages.RemovePage(RAGPage)
 				})
@@ -492,9 +475,7 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 				logger.Error("failed to delete file", "filename", fpath, "error", err)
 				return
 			}
-			if err := notifyUser("chat deleted", fpath+" was deleted"); err != nil {
+			showToast("chat deleted", fpath+" was deleted")
-				logger.Error("failed to send notification", "error", err)
-			}
 			return
 		default:
 			pages.RemovePage(RAGPage)
@@ -529,8 +510,8 @@ func makeAgentTable(agentList []string) *tview.Table {
 						SetSelectable(false))
 			case 1:
 				if actions[c-1] == "filepath" {
-					cc, ok := sysMap[agentList[r]]
+					cc := GetCardByRole(agentList[r])
-					if !ok {
+					if cc == nil {
 						continue
 					}
 					chatActTable.SetCell(r, c,
@@ -603,9 +584,7 @@ func makeAgentTable(agentList []string) *tview.Table {
 			if err := store.RemoveChat(sc.ID); err != nil {
 				logger.Error("failed to remove chat from db", "chat_id", sc.ID, "chat_name", sc.Name)
 			}
-			if err := notifyUser("chat deleted", selected+" was deleted"); err != nil {
+			showToast("chat deleted", selected+" was deleted")
-				logger.Error("failed to send notification", "error", err)
-			}
 			pages.RemovePage(agentPage)
 			return
 		default:
@@ -676,13 +655,9 @@ func makeCodeBlockTable(codeBlocks []string) *tview.Table {
 		switch tc.Text {
 		case "copy":
 			if err := copyToClipboard(selected); err != nil {
-				if err := notifyUser("error", err.Error()); err != nil {
+				showToast("error", err.Error())
-					logger.Error("failed to send notification", "error", err)
-				}
-			}
-			if err := notifyUser("copied", selected); err != nil {
-				logger.Error("failed to send notification", "error", err)
 			}
+			showToast("copied", selected)
 			pages.RemovePage(codeBlockPage)
 			app.SetFocus(textArea)
 			return
@@ -775,9 +750,7 @@ func makeImportChatTable(filenames []string) *tview.Table {
 			if err := store.RemoveChat(sc.ID); err != nil {
 				logger.Error("failed to remove chat from db", "chat_id", sc.ID, "chat_name", sc.Name)
 			}
-			if err := notifyUser("chat deleted", selected+" was deleted"); err != nil {
+			showToast("chat deleted", selected+" was deleted")
-				logger.Error("failed to send notification", "error", err)
-			}
 			pages.RemovePage(historyPage)
 			return
 		default:

tools.go

@@ -77,12 +77,17 @@ Your current tools:
 {
 "name":"file_create",
 "args": ["path", "content"],
-"when_to_use": "when asked to create a new file with optional content"
+"when_to_use": "when there is a need to create a new file with optional content"
 },
 {
 "name":"file_read",
 "args": ["path"],
-"when_to_use": "when asked to read the content of a file"
+"when_to_use": "when you need to read the content of a file"
+},
+{
+"name":"file_read_image",
+"args": ["path"],
+"when_to_use": "when you need to read or view an image file"
 },
 {
 "name":"file_write",
@@ -92,7 +97,7 @@ Your current tools:
 {
 "name":"file_write_append",
 "args": ["path", "content"],
-"when_to_use": "when asked to append content to a file; use sed to edit content"
+"when_to_use": "when you need append content to a file; use sed to edit content"
 },
 {
 "name":"file_edit",
@@ -107,22 +112,22 @@ Your current tools:
 {
 "name":"file_move",
 "args": ["src", "dst"],
-"when_to_use": "when asked to move a file from source to destination"
+"when_to_use": "when you need to move a file from source to destination"
 },
 {
 "name":"file_copy",
 "args": ["src", "dst"],
-"when_to_use": "when asked to copy a file from source to destination"
+"when_to_use": "copy a file from source to destination"
 },
 {
 "name":"file_list",
 "args": ["path"],
-"when_to_use": "when asked to list files in a directory; path is optional (default: current directory)"
+"when_to_use": "list files in a directory; path is optional (default: current directory)"
 },
 {
 "name":"execute_command",
 "args": ["command", "args"],
-"when_to_use": "when asked to execute a system command; args is optional; allowed commands: grep, sed, awk, find, cat, head, tail, sort, uniq, wc, ls, echo, cut, tr, cp, mv, rm, mkdir, rmdir, pwd, df, free, ps, top, du, whoami, date, uname, go"
+"when_to_use": "execute a system command; args is optional; allowed commands: grep, sed, awk, find, cat, head, tail, sort, uniq, wc, ls, echo, cut, tr, cp, mv, rm, mkdir, rmdir, pwd, df, free, ps, top, du, whoami, date, uname, go"
 }
 ]
 </tools>
@@ -157,40 +162,133 @@ After that you are free to respond to the user.
 	readURLSysPrompt   = `Extract and summarize the content from the webpage. Provide key information, main points, and any relevant details.`
 	summarySysPrompt   = `Please provide a concise summary of the following conversation. Focus on key points, decisions, and actions. Provide only the summary, no additional commentary.`
 	basicCard          = &models.CharCard{
+		ID:        models.ComputeCardID("assistant", "basic_sys"),
 		SysPrompt: basicSysMsg,
 		FirstMsg:  defaultFirstMsg,
-		Role:      "",
+		Role:      "assistant",
-		FilePath:  "",
+		FilePath:  "basic_sys",
 	}
-	sysMap    = map[string]*models.CharCard{"basic_sys": basicCard}
+	sysMap    = map[string]*models.CharCard{}
-	sysLabels = []string{"basic_sys"}
+	roleToID  = map[string]string{}
+	sysLabels = []string{"assistant"}
 
 	webAgentClient     *agent.AgentClient
 	webAgentClientOnce sync.Once
 	webAgentsOnce      sync.Once
 )
 
+var windowToolSysMsg = `
+Additional window tools (available only if xdotool and maim are installed):
+[
+{
+"name":"list_windows",
+"args": [],
+"when_to_use": "when asked to list visible windows; returns map of window ID to window name"
+},
+{
+"name":"capture_window",
+"args": ["window"],
+"when_to_use": "when asked to take a screenshot of a specific window; saves to /tmp; window can be ID or name substring; returns file path"
+},
+{
+"name":"capture_window_and_view",
+"args": ["window"],
+"when_to_use": "when asked to take a screenshot of a specific window and show it; saves to /tmp and returns image for viewing; window can be ID or name substring"
+}
+]
+`
+
 var WebSearcher searcher.WebSurfer
 
+var (
+	windowToolsAvailable bool
+	xdotoolPath          string
+	maimPath             string
+	modelHasVision       bool
+)
+
 func init() {
+	sysMap[basicCard.ID] = basicCard
+	roleToID["assistant"] = basicCard.ID
 	sa, err := searcher.NewWebSurfer(searcher.SearcherTypeScraper, "")
 	if err != nil {
-		panic("failed to init seachagent; error: " + err.Error())
+		if logger != nil {
+			logger.Warn("search agent unavailable; web_search tool disabled", "error", err)
+		}
+		WebSearcher = nil
+	} else {
+		WebSearcher = sa
 	}
-	WebSearcher = sa
 	if err := rag.Init(cfg, logger, store); err != nil {
 		logger.Warn("failed to init rag; rag_search tool will not be available", "error", err)
 	}
+	checkWindowTools()
+	registerWindowTools()
+}
+
+func GetCardByRole(role string) *models.CharCard {
+	cardID, ok := roleToID[role]
+	if !ok {
+		return nil
+	}
+	return sysMap[cardID]
+}
+
+func checkWindowTools() {
+	xdotoolPath, _ = exec.LookPath("xdotool")
+	maimPath, _ = exec.LookPath("maim")
+	windowToolsAvailable = xdotoolPath != "" && maimPath != ""
+	if windowToolsAvailable {
+		logger.Info("window tools available: xdotool and maim found")
+	} else {
+		if xdotoolPath == "" {
+			logger.Warn("xdotool not found, window listing tools will not be available")
+		}
+		if maimPath == "" {
+			logger.Warn("maim not found, window capture tools will not be available")
+		}
+	}
+}
+
+func updateToolCapabilities() {
+	if !cfg.ToolUse {
+		return
+	}
+	modelHasVision = false
+	if cfg == nil || cfg.CurrentAPI == "" {
+		logger.Warn("cannot determine model capabilities: cfg or CurrentAPI is nil")
+		registerWindowTools()
+		registerPlaywrightTools()
+		return
+	}
+	prevHasVision := modelHasVision
+	modelHasVision = ModelHasVision(cfg.CurrentAPI, cfg.CurrentModel)
+	if modelHasVision {
+		logger.Info("model has vision support", "model", cfg.CurrentModel, "api", cfg.CurrentAPI)
+	} else {
+		logger.Info("model does not have vision support", "model", cfg.CurrentModel, "api", cfg.CurrentAPI)
+		if windowToolsAvailable && !prevHasVision && !modelHasVision {
+			showToast("window tools", "Window capture-and-view unavailable: model lacks vision support")
+		}
+	}
+	registerWindowTools()
+	registerPlaywrightTools()
 }
 
 // getWebAgentClient returns a singleton AgentClient for web agents.
 func getWebAgentClient() *agent.AgentClient {
 	webAgentClientOnce.Do(func() {
 		if cfg == nil {
-			panic("cfg not initialized")
+			if logger != nil {
+				logger.Warn("web agent client unavailable: config not initialized")
+			}
+			return
 		}
 		if logger == nil {
-			panic("logger not initialized")
+			if logger != nil {
+				logger.Warn("web agent client unavailable: logger not initialized")
+			}
+			return
 		}
 		getToken := func() string {
 			if chunkParser == nil {
@@ -469,6 +567,43 @@ func fileRead(args map[string]string) []byte {
 	return jsonResult
 }
 
+func fileReadImage(args map[string]string) []byte {
+	path, ok := args["path"]
+	if !ok || path == "" {
+		msg := "path not provided to file_read_image tool"
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	path = resolvePath(path)
+	dataURL, err := models.CreateImageURLFromPath(path)
+	if err != nil {
+		msg := "failed to read image; error: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	// result := map[string]any{
+	// 	"type": "multimodal_content",
+	// 	"parts": []map[string]string{
+	// 		{"type": "text", "text": "Image at " + path},
+	// 		{"type": "image_url", "url": dataURL},
+	// 	},
+	// }
+	result := models.MultimodalToolResp{
+		Type: "multimodal_content",
+		Parts: []map[string]string{
+			{"type": "text", "text": "Image at " + path},
+			{"type": "image_url", "url": dataURL},
+		},
+	}
+	jsonResult, err := json.Marshal(result)
+	if err != nil {
+		msg := "failed to marshal result; error: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	return jsonResult
+}
+
 func fileWrite(args map[string]string) []byte {
 	path, ok := args["path"]
 	if !ok || path == "" {
@@ -1088,6 +1223,142 @@ func summarizeChat(args map[string]string) []byte {
 	return []byte(chatText)
 }
 
+func windowIDToHex(decimalID string) string {
+	id, err := strconv.ParseInt(decimalID, 10, 64)
+	if err != nil {
+		return decimalID
+	}
+	return fmt.Sprintf("0x%x", id)
+}
+
+func listWindows(args map[string]string) []byte {
+	if !windowToolsAvailable {
+		return []byte("window tools not available: xdotool or maim not found")
+	}
+	cmd := exec.Command(xdotoolPath, "search", "--name", ".")
+	output, err := cmd.Output()
+	if err != nil {
+		msg := "failed to list windows: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	windowIDs := strings.Fields(string(output))
+	windows := make(map[string]string)
+	for _, id := range windowIDs {
+		id = strings.TrimSpace(id)
+		if id == "" {
+			continue
+		}
+		nameCmd := exec.Command(xdotoolPath, "getwindowname", id)
+		nameOutput, err := nameCmd.Output()
+		if err != nil {
+			continue
+		}
+		name := strings.TrimSpace(string(nameOutput))
+		windows[id] = name
+	}
+	data, err := json.Marshal(windows)
+	if err != nil {
+		msg := "failed to marshal window list: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	return data
+}
+
+func captureWindow(args map[string]string) []byte {
+	if !windowToolsAvailable {
+		return []byte("window tools not available: xdotool or maim not found")
+	}
+	window, ok := args["window"]
+	if !ok || window == "" {
+		return []byte("window parameter required (window ID or name)")
+	}
+	var windowID string
+	if _, err := strconv.Atoi(window); err == nil {
+		windowID = window
+	} else {
+		cmd := exec.Command(xdotoolPath, "search", "--name", window)
+		output, err := cmd.Output()
+		if err != nil || len(strings.Fields(string(output))) == 0 {
+			return []byte("window not found: " + window)
+		}
+		windowID = strings.Fields(string(output))[0]
+	}
+	nameCmd := exec.Command(xdotoolPath, "getwindowname", windowID)
+	nameOutput, _ := nameCmd.Output()
+	windowName := strings.TrimSpace(string(nameOutput))
+	windowName = regexp.MustCompile(`[^a-zA-Z]+`).ReplaceAllString(windowName, "")
+	if windowName == "" {
+		windowName = "window"
+	}
+	timestamp := time.Now().Unix()
+	filename := fmt.Sprintf("/tmp/%s_%d.jpg", windowName, timestamp)
+	cmd := exec.Command(maimPath, "-i", windowIDToHex(windowID), filename)
+	if err := cmd.Run(); err != nil {
+		msg := "failed to capture window: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	return []byte("screenshot saved: " + filename)
+}
+
+func captureWindowAndView(args map[string]string) []byte {
+	if !windowToolsAvailable {
+		return []byte("window tools not available: xdotool or maim not found")
+	}
+	window, ok := args["window"]
+	if !ok || window == "" {
+		return []byte("window parameter required (window ID or name)")
+	}
+	var windowID string
+	if _, err := strconv.Atoi(window); err == nil {
+		windowID = window
+	} else {
+		cmd := exec.Command(xdotoolPath, "search", "--name", window)
+		output, err := cmd.Output()
+		if err != nil || len(strings.Fields(string(output))) == 0 {
+			return []byte("window not found: " + window)
+		}
+		windowID = strings.Fields(string(output))[0]
+	}
+	nameCmd := exec.Command(xdotoolPath, "getwindowname", windowID)
+	nameOutput, _ := nameCmd.Output()
+	windowName := strings.TrimSpace(string(nameOutput))
+	windowName = regexp.MustCompile(`[^a-zA-Z]+`).ReplaceAllString(windowName, "")
+	if windowName == "" {
+		windowName = "window"
+	}
+	timestamp := time.Now().Unix()
+	filename := fmt.Sprintf("/tmp/%s_%d.jpg", windowName, timestamp)
+	captureCmd := exec.Command(maimPath, "-i", windowIDToHex(windowID), filename)
+	if err := captureCmd.Run(); err != nil {
+		msg := "failed to capture window: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	dataURL, err := models.CreateImageURLFromPath(filename)
+	if err != nil {
+		msg := "failed to create image URL: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	result := models.MultimodalToolResp{
+		Type: "multimodal_content",
+		Parts: []map[string]string{
+			{"type": "text", "text": "Screenshot saved: " + filename},
+			{"type": "image_url", "url": dataURL},
+		},
+	}
+	jsonResult, err := json.Marshal(result)
+	if err != nil {
+		msg := "failed to marshal result: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	return jsonResult
+}
+
 type fnSig func(map[string]string) []byte
 
 var fnMap = map[string]fnSig{
@@ -1101,6 +1372,7 @@ var fnMap = map[string]fnSig{
 	"read_url_raw":      readURLRaw,
 	"file_create":       fileCreate,
 	"file_read":         fileRead,
+	"file_read_image":   fileReadImage,
 	"file_write":        fileWrite,
 	"file_write_append": fileWriteAppend,
 	"file_edit":         fileEdit,
@@ -1116,6 +1388,409 @@ var fnMap = map[string]fnSig{
 	"summarize_chat":    summarizeChat,
 }
 
+func removeWindowToolsFromBaseTools() {
+	windowToolNames := map[string]bool{
+		"list_windows":            true,
+		"capture_window":          true,
+		"capture_window_and_view": true,
+	}
+	var filtered []models.Tool
+	for _, tool := range baseTools {
+		if !windowToolNames[tool.Function.Name] {
+			filtered = append(filtered, tool)
+		}
+	}
+	baseTools = filtered
+	delete(fnMap, "list_windows")
+	delete(fnMap, "capture_window")
+	delete(fnMap, "capture_window_and_view")
+}
+
+func removePlaywrightToolsFromBaseTools() {
+	playwrightToolNames := map[string]bool{
+		"pw_start":               true,
+		"pw_stop":                true,
+		"pw_is_running":          true,
+		"pw_navigate":            true,
+		"pw_click":               true,
+		"pw_click_at":            true,
+		"pw_fill":                true,
+		"pw_extract_text":        true,
+		"pw_screenshot":          true,
+		"pw_screenshot_and_view": true,
+		"pw_wait_for_selector":   true,
+		"pw_drag":                true,
+	}
+	var filtered []models.Tool
+	for _, tool := range baseTools {
+		if !playwrightToolNames[tool.Function.Name] {
+			filtered = append(filtered, tool)
+		}
+	}
+	baseTools = filtered
+	delete(fnMap, "pw_start")
+	delete(fnMap, "pw_stop")
+	delete(fnMap, "pw_is_running")
+	delete(fnMap, "pw_navigate")
+	delete(fnMap, "pw_click")
+	delete(fnMap, "pw_click_at")
+	delete(fnMap, "pw_fill")
+	delete(fnMap, "pw_extract_text")
+	delete(fnMap, "pw_screenshot")
+	delete(fnMap, "pw_screenshot_and_view")
+	delete(fnMap, "pw_wait_for_selector")
+	delete(fnMap, "pw_drag")
+}
+
+func registerWindowTools() {
+	removeWindowToolsFromBaseTools()
+	if windowToolsAvailable {
+		fnMap["list_windows"] = listWindows
+		fnMap["capture_window"] = captureWindow
+		windowTools := []models.Tool{
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "list_windows",
+					Description: "List all visible windows with their IDs and names. Returns a map of window ID to window name.",
+					Parameters: models.ToolFuncParams{
+						Type:       "object",
+						Required:   []string{},
+						Properties: map[string]models.ToolArgProps{},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "capture_window",
+					Description: "Capture a screenshot of a specific window and save it to /tmp. Requires window parameter (window ID or name substring).",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"window"},
+						Properties: map[string]models.ToolArgProps{
+							"window": models.ToolArgProps{
+								Type:        "string",
+								Description: "window ID or window name (partial match)",
+							},
+						},
+					},
+				},
+			},
+		}
+		if modelHasVision {
+			fnMap["capture_window_and_view"] = captureWindowAndView
+			windowTools = append(windowTools, models.Tool{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "capture_window_and_view",
+					Description: "Capture a screenshot of a specific window, save it to /tmp, and return the image for viewing. Requires window parameter (window ID or name substring).",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"window"},
+						Properties: map[string]models.ToolArgProps{
+							"window": models.ToolArgProps{
+								Type:        "string",
+								Description: "window ID or window name (partial match)",
+							},
+						},
+					},
+				},
+			})
+		}
+		baseTools = append(baseTools, windowTools...)
+		toolSysMsg += windowToolSysMsg
+	}
+}
+
+func registerPlaywrightTools() {
+	removePlaywrightToolsFromBaseTools()
+	if cfg != nil && cfg.PlaywrightEnabled {
+		fnMap["pw_start"] = pwStart
+		fnMap["pw_stop"] = pwStop
+		fnMap["pw_is_running"] = pwIsRunning
+		fnMap["pw_navigate"] = pwNavigate
+		fnMap["pw_click"] = pwClick
+		fnMap["pw_click_at"] = pwClickAt
+		fnMap["pw_fill"] = pwFill
+		fnMap["pw_extract_text"] = pwExtractText
+		fnMap["pw_screenshot"] = pwScreenshot
+		fnMap["pw_screenshot_and_view"] = pwScreenshotAndView
+		fnMap["pw_wait_for_selector"] = pwWaitForSelector
+		fnMap["pw_drag"] = pwDrag
+		fnMap["pw_get_html"] = pwGetHTML
+		fnMap["pw_get_dom"] = pwGetDOM
+		fnMap["pw_search_elements"] = pwSearchElements
+		playwrightTools := []models.Tool{
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_start",
+					Description: "Start a Playwright browser instance. Call this first before using other pw_ tools. Uses headless mode by default (set PlaywrightHeadless=false in config for GUI).",
+					Parameters: models.ToolFuncParams{
+						Type:       "object",
+						Required:   []string{},
+						Properties: map[string]models.ToolArgProps{},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_stop",
+					Description: "Stop the Playwright browser instance. Call when done with browser automation.",
+					Parameters: models.ToolFuncParams{
+						Type:       "object",
+						Required:   []string{},
+						Properties: map[string]models.ToolArgProps{},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_is_running",
+					Description: "Check if Playwright browser is currently running.",
+					Parameters: models.ToolFuncParams{
+						Type:       "object",
+						Required:   []string{},
+						Properties: map[string]models.ToolArgProps{},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_navigate",
+					Description: "Navigate to a URL in the browser.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"url"},
+						Properties: map[string]models.ToolArgProps{
+							"url": models.ToolArgProps{
+								Type:        "string",
+								Description: "URL to navigate to",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_click",
+					Description: "Click on an element using CSS selector. Use 'index' for multiple matches (default 0).",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"selector"},
+						Properties: map[string]models.ToolArgProps{
+							"selector": models.ToolArgProps{
+								Type:        "string",
+								Description: "CSS selector for the element to click",
+							},
+							"index": models.ToolArgProps{
+								Type:        "string",
+								Description: "optional index for multiple matches (default 0)",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_fill",
+					Description: "Fill an input field with text using CSS selector.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"selector", "text"},
+						Properties: map[string]models.ToolArgProps{
+							"selector": models.ToolArgProps{
+								Type:        "string",
+								Description: "CSS selector for the input element",
+							},
+							"text": models.ToolArgProps{
+								Type:        "string",
+								Description: "text to fill into the input",
+							},
+							"index": models.ToolArgProps{
+								Type:        "string",
+								Description: "optional index for multiple matches (default 0)",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_extract_text",
+					Description: "Extract text content from the page or specific elements using CSS selector. Use 'body' for all page text.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"selector"},
+						Properties: map[string]models.ToolArgProps{
+							"selector": models.ToolArgProps{
+								Type:        "string",
+								Description: "CSS selector (use 'body' for all page text)",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_screenshot",
+					Description: "Take a screenshot of the page or a specific element. Returns file path to saved image.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{},
+						Properties: map[string]models.ToolArgProps{
+							"selector": models.ToolArgProps{
+								Type:        "string",
+								Description: "optional CSS selector for element to screenshot",
+							},
+							"full_page": models.ToolArgProps{
+								Type:        "string",
+								Description: "optional: 'true' to capture full page (default false)",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_screenshot_and_view",
+					Description: "Take a screenshot and return the image for viewing. Use when model needs to see the screenshot.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{},
+						Properties: map[string]models.ToolArgProps{
+							"selector": models.ToolArgProps{
+								Type:        "string",
+								Description: "optional CSS selector for element to screenshot",
+							},
+							"full_page": models.ToolArgProps{
+								Type:        "string",
+								Description: "optional: 'true' to capture full page (default false)",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_wait_for_selector",
+					Description: "Wait for an element to appear on the page.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"selector"},
+						Properties: map[string]models.ToolArgProps{
+							"selector": models.ToolArgProps{
+								Type:        "string",
+								Description: "CSS selector to wait for",
+							},
+							"timeout": models.ToolArgProps{
+								Type:        "string",
+								Description: "optional timeout in ms (default 30000)",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_drag",
+					Description: "Drag the mouse from one point to another.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"x1", "y1", "x2", "y2"},
+						Properties: map[string]models.ToolArgProps{
+							"x1": models.ToolArgProps{
+								Type:        "string",
+								Description: "starting X coordinate",
+							},
+							"y1": models.ToolArgProps{
+								Type:        "string",
+								Description: "starting Y coordinate",
+							},
+							"x2": models.ToolArgProps{
+								Type:        "string",
+								Description: "ending X coordinate",
+							},
+							"y2": models.ToolArgProps{
+								Type:        "string",
+								Description: "ending Y coordinate",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_get_html",
+					Description: "Get the HTML content of the page or a specific element.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{},
+						Properties: map[string]models.ToolArgProps{
+							"selector": models.ToolArgProps{
+								Type:        "string",
+								Description: "optional CSS selector (default: body)",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_get_dom",
+					Description: "Get a structured DOM representation of an element with tag, attributes, text, and children.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{},
+						Properties: map[string]models.ToolArgProps{
+							"selector": models.ToolArgProps{
+								Type:        "string",
+								Description: "optional CSS selector (default: body)",
+							},
+						},
+					},
+				},
+			},
+			{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "pw_search_elements",
+					Description: "Search for elements by text content or CSS selector. Returns matching elements with their tags, text, and HTML.",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{},
+						Properties: map[string]models.ToolArgProps{
+							"text": models.ToolArgProps{
+								Type:        "string",
+								Description: "text to search for in elements",
+							},
+							"selector": models.ToolArgProps{
+								Type:        "string",
+								Description: "CSS selector to search for",
+							},
+						},
+					},
+				},
+			},
+		}
+		baseTools = append(baseTools, playwrightTools...)
+		toolSysMsg += browserToolSysMsg
+	}
+}
+
 // callToolWithAgent calls the tool and applies any registered agent.
 func callToolWithAgent(name string, args map[string]string) []byte {
 	registerWebAgents()
@@ -1327,6 +2002,24 @@ var baseTools = []models.Tool{
 			},
 		},
 	},
+	// file_read_image
+	models.Tool{
+		Type: "function",
+		Function: models.ToolFunc{
+			Name:        "file_read_image",
+			Description: "Read an image file and return it for multimodal LLM viewing. Supports png, jpg, jpeg, gif, webp formats. Use when you need the LLM to see and analyze an image.",
+			Parameters: models.ToolFuncParams{
+				Type:     "object",
+				Required: []string{"path"},
+				Properties: map[string]models.ToolArgProps{
+					"path": models.ToolArgProps{
+						Type:        "string",
+						Description: "path of the image file to read",
+					},
+				},
+			},
+		},
+	},
 	// file_write
 	models.Tool{
 		Type: "function",
@@ -1580,3 +2273,56 @@ var baseTools = []models.Tool{
 		},
 	},
 }
+
+func init() {
+	if windowToolsAvailable {
+		baseTools = append(baseTools,
+			models.Tool{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "list_windows",
+					Description: "List all visible windows with their IDs and names. Returns a map of window ID to window name.",
+					Parameters: models.ToolFuncParams{
+						Type:       "object",
+						Required:   []string{},
+						Properties: map[string]models.ToolArgProps{},
+					},
+				},
+			},
+			models.Tool{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "capture_window",
+					Description: "Capture a screenshot of a specific window and save it to /tmp. Requires window parameter (window ID or name substring).",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"window"},
+						Properties: map[string]models.ToolArgProps{
+							"window": models.ToolArgProps{
+								Type:        "string",
+								Description: "window ID or window name (partial match)",
+							},
+						},
+					},
+				},
+			},
+			models.Tool{
+				Type: "function",
+				Function: models.ToolFunc{
+					Name:        "capture_window_and_view",
+					Description: "Capture a screenshot of a specific window, save it to /tmp, and return the image for viewing. Requires window parameter (window ID or name substring).",
+					Parameters: models.ToolFuncParams{
+						Type:     "object",
+						Required: []string{"window"},
+						Properties: map[string]models.ToolArgProps{
+							"window": models.ToolArgProps{
+								Type:        "string",
+								Description: "window ID or window name (partial match)",
+							},
+						},
+					},
+				},
+			},
+		)
+	}
+}

tools_playwright.go

@@ -0,0 +1,653 @@
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"gf-lt/models"
+	"os"
+	"strconv"
+	"strings"
+	"sync"
+
+	"github.com/playwright-community/playwright-go"
+)
+
+var browserToolSysMsg = `
+Additional browser automation tools (Playwright):
+[
+{
+    "name": "pw_start",
+    "args": [],
+    "when_to_use": "start a browser instance before doing any browser automation. Must be called first."
+},
+{
+    "name": "pw_stop",
+    "args": [],
+    "when_to_use": "stop the browser instance when done with automation."
+},
+{
+    "name": "pw_is_running",
+    "args": [],
+    "when_to_use": "check if browser is currently running."
+},
+{
+    "name": "pw_navigate",
+    "args": ["url"],
+    "when_to_use": "open a specific URL in the web browser."
+},
+{
+    "name": "pw_click",
+    "args": ["selector", "index"],
+    "when_to_use": "click on an element on the current webpage. Use 'index' for multiple matches (default 0)."
+},
+{
+    "name": "pw_fill",
+    "args": ["selector", "text", "index"],
+    "when_to_use": "type text into an input field. Use 'index' for multiple matches (default 0)."
+},
+{
+    "name": "pw_extract_text",
+    "args": ["selector"],
+    "when_to_use": "extract text content from the page or specific elements. Use selector 'body' for all page text."
+},
+{
+    "name": "pw_screenshot",
+    "args": ["selector", "full_page"],
+    "when_to_use": "take a screenshot of the page or a specific element. Returns a file path to the image. Use to verify actions or inspect visual state."
+},
+{
+    "name": "pw_screenshot_and_view",
+    "args": ["selector", "full_page"],
+    "when_to_use": "take a screenshot and return the image for viewing. Use to visually verify page state."
+},
+{
+    "name": "pw_wait_for_selector",
+    "args": ["selector", "timeout"],
+    "when_to_use": "wait for an element to appear on the page before proceeding with further actions."
+},
+{
+    "name": "pw_drag",
+    "args": ["x1", "y1", "x2", "y2"],
+    "when_to_use": "drag the mouse from point (x1,y1) to (x2,y2)."
+},
+{
+    "name": "pw_click_at",
+    "args": ["x", "y"],
+    "when_to_use": "click at specific X,Y coordinates on the page. Use when you know the exact position."
+},
+{
+    "name": "pw_get_html",
+    "args": ["selector"],
+    "when_to_use": "get the HTML content of the page or a specific element. Use to understand page structure or extract raw HTML."
+},
+{
+    "name": "pw_get_dom",
+    "args": ["selector"],
+    "when_to_use": "get a structured DOM representation with tag, attributes, text, and children. Use to inspect element hierarchy and properties."
+},
+{
+    "name": "pw_search_elements",
+    "args": ["text", "selector"],
+    "when_to_use": "search for elements by text content or CSS selector. Returns matching elements with their tags, text, and HTML."
+}
+]
+`
+
+var (
+	pw             *playwright.Playwright
+	browser        playwright.Browser
+	browserStarted bool
+	browserStartMu sync.Mutex
+	page           playwright.Page
+)
+
+func pwShutDown() error {
+	if pw == nil {
+		return nil
+	}
+	pwStop(nil)
+	return pw.Stop()
+}
+
+func installPW() error {
+	err := playwright.Install(&playwright.RunOptions{Verbose: false})
+	if err != nil {
+		logger.Warn("playwright not available", "error", err)
+		return err
+	}
+	return nil
+}
+
+func checkPlaywright() error {
+	var err error
+	pw, err = playwright.Run()
+	if err != nil {
+		logger.Warn("playwright not available", "error", err)
+		return err
+	}
+	return nil
+}
+
+func pwStart(args map[string]string) []byte {
+	browserStartMu.Lock()
+	defer browserStartMu.Unlock()
+	if browserStarted {
+		return []byte(`{"error": "Browser already started"}`)
+	}
+	var err error
+	browser, err = pw.Chromium.Launch(playwright.BrowserTypeLaunchOptions{
+		Headless: playwright.Bool(!cfg.PlaywrightDebug),
+	})
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to launch browser: %s"}`, err.Error()))
+	}
+	page, err = browser.NewPage()
+	if err != nil {
+		browser.Close()
+		return []byte(fmt.Sprintf(`{"error": "failed to create page: %s"}`, err.Error()))
+	}
+	browserStarted = true
+	return []byte(`{"success": true, "message": "Browser started"}`)
+}
+
+func pwStop(args map[string]string) []byte {
+	browserStartMu.Lock()
+	defer browserStartMu.Unlock()
+	if !browserStarted {
+		return []byte(`{"success": true, "message": "Browser was not running"}`)
+	}
+	if page != nil {
+		page.Close()
+		page = nil
+	}
+	if browser != nil {
+		browser.Close()
+		browser = nil
+	}
+	browserStarted = false
+	return []byte(`{"success": true, "message": "Browser stopped"}`)
+}
+
+func pwIsRunning(args map[string]string) []byte {
+	if browserStarted {
+		return []byte(`{"running": true, "message": "Browser is running"}`)
+	}
+	return []byte(`{"running": false, "message": "Browser is not running"}`)
+}
+
+func pwNavigate(args map[string]string) []byte {
+	url, ok := args["url"]
+	if !ok || url == "" {
+		return []byte(`{"error": "url not provided"}`)
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	_, err := page.Goto(url)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to navigate: %s"}`, err.Error()))
+	}
+	title, _ := page.Title()
+	pageURL := page.URL()
+	return []byte(fmt.Sprintf(`{"success": true, "title": "%s", "url": "%s"}`, title, pageURL))
+}
+
+func pwClick(args map[string]string) []byte {
+	selector, ok := args["selector"]
+	if !ok || selector == "" {
+		return []byte(`{"error": "selector not provided"}`)
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	index := 0
+	if args["index"] != "" {
+		if i, err := strconv.Atoi(args["index"]); err != nil {
+			logger.Warn("failed to parse index", "value", args["index"], "error", err)
+		} else {
+			index = i
+		}
+	}
+	locator := page.Locator(selector)
+	count, err := locator.Count()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to find elements: %s"}`, err.Error()))
+	}
+	if index >= count {
+		return []byte(fmt.Sprintf(`{"error": "Element not found at index %d (found %d elements)"}`, index, count))
+	}
+	err = locator.Nth(index).Click()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to click: %s"}`, err.Error()))
+	}
+	return []byte(`{"success": true, "message": "Clicked element"}`)
+}
+
+func pwFill(args map[string]string) []byte {
+	selector, ok := args["selector"]
+	if !ok || selector == "" {
+		return []byte(`{"error": "selector not provided"}`)
+	}
+	text := args["text"]
+	if text == "" {
+		text = ""
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	index := 0
+	if args["index"] != "" {
+		if i, err := strconv.Atoi(args["index"]); err != nil {
+			logger.Warn("failed to parse index", "value", args["index"], "error", err)
+		} else {
+			index = i
+		}
+	}
+	locator := page.Locator(selector)
+	count, err := locator.Count()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to find elements: %s"}`, err.Error()))
+	}
+	if index >= count {
+		return []byte(fmt.Sprintf(`{"error": "Element not found at index %d"}`, index))
+	}
+	err = locator.Nth(index).Fill(text)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to fill: %s"}`, err.Error()))
+	}
+	return []byte(`{"success": true, "message": "Filled input"}`)
+}
+
+func pwExtractText(args map[string]string) []byte {
+	selector := args["selector"]
+	if selector == "" {
+		selector = "body"
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	locator := page.Locator(selector)
+	count, err := locator.Count()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to find elements: %s"}`, err.Error()))
+	}
+	if count == 0 {
+		return []byte(`{"error": "No elements found"}`)
+	}
+	if selector == "body" {
+		text, err := page.Locator("body").TextContent()
+		if err != nil {
+			return []byte(fmt.Sprintf(`{"error": "failed to get text: %s"}`, err.Error()))
+		}
+		return []byte(fmt.Sprintf(`{"text": "%s"}`, text))
+	}
+	var texts []string
+	for i := 0; i < count; i++ {
+		text, err := locator.Nth(i).TextContent()
+		if err != nil {
+			continue
+		}
+		texts = append(texts, text)
+	}
+	return []byte(fmt.Sprintf(`{"text": "%s"}`, joinLines(texts)))
+}
+
+func joinLines(lines []string) string {
+	var sb strings.Builder
+	for i, line := range lines {
+		if i > 0 {
+			sb.WriteString("\n")
+		}
+		sb.WriteString(line)
+	}
+	return sb.String()
+}
+
+func pwScreenshot(args map[string]string) []byte {
+	selector := args["selector"]
+	fullPage := args["full_page"] == "true"
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	path := fmt.Sprintf("/tmp/pw_screenshot_%d.png", os.Getpid())
+	var err error
+	if selector != "" && selector != "body" {
+		locator := page.Locator(selector)
+		_, err = locator.Screenshot(playwright.LocatorScreenshotOptions{
+			Path: playwright.String(path),
+		})
+	} else {
+		_, err = page.Screenshot(playwright.PageScreenshotOptions{
+			Path:     playwright.String(path),
+			FullPage: playwright.Bool(fullPage),
+		})
+	}
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to take screenshot: %s"}`, err.Error()))
+	}
+	return []byte(fmt.Sprintf(`{"path": "%s"}`, path))
+}
+
+func pwScreenshotAndView(args map[string]string) []byte {
+	selector := args["selector"]
+	fullPage := args["full_page"] == "true"
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	path := fmt.Sprintf("/tmp/pw_screenshot_%d.png", os.Getpid())
+	var err error
+	if selector != "" && selector != "body" {
+		locator := page.Locator(selector)
+		_, err = locator.Screenshot(playwright.LocatorScreenshotOptions{
+			Path: playwright.String(path),
+		})
+	} else {
+		_, err = page.Screenshot(playwright.PageScreenshotOptions{
+			Path:     playwright.String(path),
+			FullPage: playwright.Bool(fullPage),
+		})
+	}
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to take screenshot: %s"}`, err.Error()))
+	}
+	dataURL, err := models.CreateImageURLFromPath(path)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to create image URL: %s"}`, err.Error()))
+	}
+	resp := models.MultimodalToolResp{
+		Type: "multimodal_content",
+		Parts: []map[string]string{
+			{"type": "text", "text": "Screenshot saved: " + path},
+			{"type": "image_url", "url": dataURL},
+		},
+	}
+	jsonResult, err := json.Marshal(resp)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to marshal result: %s"}`, err.Error()))
+	}
+	return jsonResult
+}
+
+func pwWaitForSelector(args map[string]string) []byte {
+	selector, ok := args["selector"]
+	if !ok || selector == "" {
+		return []byte(`{"error": "selector not provided"}`)
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	timeout := 30000
+	if args["timeout"] != "" {
+		if t, err := strconv.Atoi(args["timeout"]); err != nil {
+			logger.Warn("failed to parse timeout", "value", args["timeout"], "error", err)
+		} else {
+			timeout = t
+		}
+	}
+	locator := page.Locator(selector)
+	err := locator.WaitFor(playwright.LocatorWaitForOptions{
+		Timeout: playwright.Float(float64(timeout)),
+	})
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "element not found: %s"}`, err.Error()))
+	}
+	return []byte(`{"success": true, "message": "Element found"}`)
+}
+
+func pwDrag(args map[string]string) []byte {
+	x1, ok := args["x1"]
+	if !ok {
+		return []byte(`{"error": "x1 not provided"}`)
+	}
+	y1, ok := args["y1"]
+	if !ok {
+		return []byte(`{"error": "y1 not provided"}`)
+	}
+	x2, ok := args["x2"]
+	if !ok {
+		return []byte(`{"error": "x2 not provided"}`)
+	}
+	y2, ok := args["y2"]
+	if !ok {
+		return []byte(`{"error": "y2 not provided"}`)
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	var fx1, fy1, fx2, fy2 float64
+	if parsedX1, err := strconv.ParseFloat(x1, 64); err != nil {
+		logger.Warn("failed to parse x1", "value", x1, "error", err)
+	} else {
+		fx1 = parsedX1
+	}
+	if parsedY1, err := strconv.ParseFloat(y1, 64); err != nil {
+		logger.Warn("failed to parse y1", "value", y1, "error", err)
+	} else {
+		fy1 = parsedY1
+	}
+	if parsedX2, err := strconv.ParseFloat(x2, 64); err != nil {
+		logger.Warn("failed to parse x2", "value", x2, "error", err)
+	} else {
+		fx2 = parsedX2
+	}
+	if parsedY2, err := strconv.ParseFloat(y2, 64); err != nil {
+		logger.Warn("failed to parse y2", "value", y2, "error", err)
+	} else {
+		fy2 = parsedY2
+	}
+	mouse := page.Mouse()
+	err := mouse.Move(fx1, fy1)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to move mouse: %s"}`, err.Error()))
+	}
+	err = mouse.Down()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to mouse down: %s"}`, err.Error()))
+	}
+	err = mouse.Move(fx2, fy2)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to move mouse: %s"}`, err.Error()))
+	}
+	err = mouse.Up()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to mouse up: %s"}`, err.Error()))
+	}
+	return []byte(fmt.Sprintf(`{"success": true, "message": "Dragged from (%s,%s) to (%s,%s)"}`, x1, y1, x2, y2))
+}
+
+func pwClickAt(args map[string]string) []byte {
+	x, ok := args["x"]
+	if !ok {
+		return []byte(`{"error": "x not provided"}`)
+	}
+	y, ok := args["y"]
+	if !ok {
+		return []byte(`{"error": "y not provided"}`)
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	fx, err := strconv.ParseFloat(x, 64)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to parse x: %s"}`, err.Error()))
+	}
+	fy, err := strconv.ParseFloat(y, 64)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to parse y: %s"}`, err.Error()))
+	}
+	mouse := page.Mouse()
+	err = mouse.Click(fx, fy)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to click: %s"}`, err.Error()))
+	}
+	return []byte(fmt.Sprintf(`{"success": true, "message": "Clicked at (%s,%s)"}`, x, y))
+}
+
+func pwGetHTML(args map[string]string) []byte {
+	selector := args["selector"]
+	if selector == "" {
+		selector = "body"
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	locator := page.Locator(selector)
+	count, err := locator.Count()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to find elements: %s"}`, err.Error()))
+	}
+	if count == 0 {
+		return []byte(`{"error": "No elements found"}`)
+	}
+	html, err := locator.First().InnerHTML()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to get HTML: %s"}`, err.Error()))
+	}
+	return []byte(fmt.Sprintf(`{"html": %s}`, jsonString(html)))
+}
+
+type DOMElement struct {
+	Tag        string            `json:"tag,omitempty"`
+	Attributes map[string]string `json:"attributes,omitempty"`
+	Text       string            `json:"text,omitempty"`
+	Children   []DOMElement      `json:"children,omitempty"`
+	Selector   string            `json:"selector,omitempty"`
+	InnerHTML  string            `json:"innerHTML,omitempty"`
+}
+
+func buildDOMTree(locator playwright.Locator) ([]DOMElement, error) {
+	var results []DOMElement
+	count, err := locator.Count()
+	if err != nil {
+		return nil, err
+	}
+	for i := 0; i < count; i++ {
+		el := locator.Nth(i)
+		dom, err := elementToDOM(el)
+		if err != nil {
+			continue
+		}
+		results = append(results, dom)
+	}
+	return results, nil
+}
+
+func elementToDOM(el playwright.Locator) (DOMElement, error) {
+	dom := DOMElement{}
+	tag, err := el.Evaluate(`el => el.nodeName`, nil)
+	if err == nil {
+		dom.Tag = strings.ToLower(fmt.Sprintf("%v", tag))
+	}
+	attributes := make(map[string]string)
+	attrs, err := el.Evaluate(`el => {
+		let attrs = {};
+		for (let i = 0; i < el.attributes.length; i++) {
+			let attr = el.attributes[i];
+			attrs[attr.name] = attr.value;
+		}
+		return attrs;
+	}`, nil)
+	if err == nil {
+		if amap, ok := attrs.(map[string]any); ok {
+			for k, v := range amap {
+				if vs, ok := v.(string); ok {
+					attributes[k] = vs
+				}
+			}
+		}
+	}
+	if len(attributes) > 0 {
+		dom.Attributes = attributes
+	}
+	text, err := el.TextContent()
+	if err == nil && text != "" {
+		dom.Text = text
+	}
+	innerHTML, err := el.InnerHTML()
+	if err == nil && innerHTML != "" {
+		dom.InnerHTML = innerHTML
+	}
+	childCount, _ := el.Count()
+	if childCount > 0 {
+		childrenLocator := el.Locator("*")
+		children, err := buildDOMTree(childrenLocator)
+		if err == nil && len(children) > 0 {
+			dom.Children = children
+		}
+	}
+	return dom, nil
+}
+
+func pwGetDOM(args map[string]string) []byte {
+	selector := args["selector"]
+	if selector == "" {
+		selector = "body"
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	locator := page.Locator(selector)
+	count, err := locator.Count()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to find elements: %s"}`, err.Error()))
+	}
+	if count == 0 {
+		return []byte(`{"error": "No elements found"}`)
+	}
+	dom, err := elementToDOM(locator.First())
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to get DOM: %s"}`, err.Error()))
+	}
+	data, err := json.Marshal(dom)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to marshal DOM: %s"}`, err.Error()))
+	}
+	return []byte(fmt.Sprintf(`{"dom": %s}`, string(data)))
+}
+
+func pwSearchElements(args map[string]string) []byte {
+	text := args["text"]
+	selector := args["selector"]
+	if text == "" && selector == "" {
+		return []byte(`{"error": "text or selector not provided"}`)
+	}
+	if !browserStarted || page == nil {
+		return []byte(`{"error": "Browser not started. Call pw_start first."}`)
+	}
+	var locator playwright.Locator
+	if text != "" {
+		locator = page.GetByText(text)
+	} else {
+		locator = page.Locator(selector)
+	}
+	count, err := locator.Count()
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to search elements: %s"}`, err.Error()))
+	}
+	if count == 0 {
+		return []byte(`{"elements": []}`)
+	}
+	var results []map[string]string
+	for i := 0; i < count; i++ {
+		el := locator.Nth(i)
+		tag, _ := el.Evaluate(`el => el.nodeName`, nil)
+		text, _ := el.TextContent()
+		html, _ := el.InnerHTML()
+		results = append(results, map[string]string{
+			"index": strconv.Itoa(i),
+			"tag":   strings.ToLower(fmt.Sprintf("%v", tag)),
+			"text":  text,
+			"html":  html,
+		})
+	}
+	data, err := json.Marshal(results)
+	if err != nil {
+		return []byte(fmt.Sprintf(`{"error": "failed to marshal results: %s"}`, err.Error()))
+	}
+	return []byte(fmt.Sprintf(`{"elements": %s}`, string(data)))
+}
+
+func jsonString(s string) string {
+	b, _ := json.Marshal(s)
+	return string(b)
+}

tui.go

@@ -10,6 +10,7 @@ import (
 	"path"
 	"strconv"
 	"strings"
+	"time"
 
 	"github.com/gdamore/tcell/v2"
 	"github.com/rivo/tview"
@@ -21,7 +22,6 @@ func isFullScreenPageActive() bool {
 }
 
 var (
-	app                *tview.Application
 	pages              *tview.Pages
 	textArea           *tview.TextArea
 	editArea           *tview.TextArea
@@ -29,6 +29,8 @@ var (
 	statusLineWidget   *tview.TextView
 	helpView           *tview.TextView
 	flex               *tview.Flex
+	bottomFlex         *tview.Flex
+	notificationWidget *tview.TextView
 	imgView            *tview.Image
 	defaultImage       = "sysprompts/llama.png"
 	indexPickWindow    *tview.InputField
@@ -36,6 +38,7 @@ var (
 	roleEditWindow     *tview.InputField
 	shellInput         *tview.InputField
 	confirmModal       *tview.Modal
+	toastTimer         *time.Timer
 	confirmPageName    = "confirm"
 	fullscreenMode     bool
 	positionVisible    bool = true
@@ -137,6 +140,90 @@ func setShellMode(enabled bool) {
 	}()
 }
 
+// showToast displays a temporary notification in the bottom-right corner.
+// It auto-hides after 3 seconds.
+func showToast(title, message string) {
+	sanitize := func(s string, maxLen int) string {
+		sanitized := strings.Map(func(r rune) rune {
+			if r < 32 && r != '\t' {
+				return -1
+			}
+			return r
+		}, s)
+		if len(sanitized) > maxLen {
+			sanitized = sanitized[:maxLen-3] + "..."
+		}
+		return sanitized
+	}
+	title = sanitize(title, 50)
+	message = sanitize(message, 197)
+	if toastTimer != nil {
+		toastTimer.Stop()
+	}
+	// show blocking notification to not mess up flex
+	if fullscreenMode {
+		notification := tview.NewTextView().
+			SetTextAlign(tview.AlignCenter).
+			SetDynamicColors(true).
+			SetRegions(true).
+			SetText(fmt.Sprintf("[yellow]%s[-]\n", message)).
+			SetChangedFunc(func() {
+				app.Draw()
+			})
+		notification.SetTitleAlign(tview.AlignLeft).
+			SetBorder(true).
+			SetTitle(title)
+		// Wrap it in a full‑screen Flex to position it in the top‑right corner.
+		// Outer Flex (row) pushes content to the top; inner Flex (column) pushes to the right.
+		background := tview.NewFlex().SetDirection(tview.FlexRow).
+			AddItem(nil, 0, 1, false). // top spacer
+			AddItem(tview.NewFlex().SetDirection(tview.FlexColumn).
+				AddItem(nil, 0, 1, false).          // left spacer
+				AddItem(notification, 40, 1, true), // notification width 40
+				5, 1, false) // notification height 5
+		// Generate a unique page name (e.g., using timestamp) to allow multiple toasts.
+		pageName := fmt.Sprintf("toast-%d", time.Now().UnixNano())
+		pages.AddPage(pageName, background, true, true)
+		// Auto‑dismiss after 2 seconds, since blocking is more annoying
+		time.AfterFunc(2*time.Second, func() {
+			app.QueueUpdateDraw(func() {
+				if pages.HasPage(pageName) {
+					pages.RemovePage(pageName)
+				}
+			})
+		})
+		return
+	}
+	notificationWidget.SetTitle(title)
+	notificationWidget.SetText(fmt.Sprintf("[yellow]%s[-]", message))
+	go func() {
+		app.QueueUpdateDraw(func() {
+			flex.RemoveItem(bottomFlex)
+			flex.RemoveItem(statusLineWidget)
+			bottomFlex = tview.NewFlex().SetDirection(tview.FlexColumn).
+				AddItem(textArea, 0, 1, true).
+				AddItem(notificationWidget, 40, 1, false)
+			flex.AddItem(bottomFlex, 0, 10, true)
+			if positionVisible {
+				flex.AddItem(statusLineWidget, 0, 2, false)
+			}
+		})
+	}()
+	toastTimer = time.AfterFunc(3*time.Second, func() {
+		app.QueueUpdateDraw(func() {
+			flex.RemoveItem(bottomFlex)
+			flex.RemoveItem(statusLineWidget)
+			bottomFlex = tview.NewFlex().SetDirection(tview.FlexColumn).
+				AddItem(textArea, 0, 1, true).
+				AddItem(notificationWidget, 0, 0, false)
+			flex.AddItem(bottomFlex, 0, 10, true)
+			if positionVisible {
+				flex.AddItem(statusLineWidget, 0, 2, false)
+			}
+		})
+	})
+}
+
 func init() {
 	// Start background goroutine to update model color cache
 	startModelColorUpdater()
@@ -186,7 +273,7 @@ func init() {
 			shellHistoryPos = -1
 		}
 		// Handle Tab key for @ file completion
-		if event.Key() == tcell.KeyTab {
+		if event.Key() == tcell.KeyTab && shellMode {
 			currentText := shellInput.GetText()
 			atIndex := strings.LastIndex(currentText, "@")
 			if atIndex >= 0 {
@@ -213,8 +300,7 @@ func init() {
 			pages.SwitchToPage("main") // or whatever your main page is named
 		})
 	confirmModal.SetInputCapture(func(event *tcell.EventKey) *tcell.EventKey {
-		switch event.Key() {
+		if event.Key() == tcell.KeyRune {
-		case tcell.KeyRune:
 			switch event.Rune() {
 			case 'y', 'Y':
 				persona := cfg.UserRole
@@ -238,12 +324,26 @@ func init() {
 		SetDynamicColors(true).
 		SetRegions(true).
 		SetChangedFunc(func() {
+			// INFO:
+			// https://github.com/rivo/tview/wiki/Concurrency#event-handlers
+			// although already called by default per tview specs
+			// calling it explicitly makes text streaming to look more smooth
 			app.Draw()
 		})
+	notificationWidget = tview.NewTextView().
+		SetTextAlign(tview.AlignCenter).
+		SetDynamicColors(true).
+		SetRegions(true).
+		SetChangedFunc(func() {
+		})
+	notificationWidget.SetBorder(true).SetTitle("notification")
+	bottomFlex = tview.NewFlex().SetDirection(tview.FlexColumn).
+		AddItem(textArea, 0, 1, true).
+		AddItem(notificationWidget, 0, 0, false)
 	//
 	flex = tview.NewFlex().SetDirection(tview.FlexRow).
 		AddItem(textView, 0, 40, false).
-		AddItem(textArea, 0, 10, true) // Restore original height
+		AddItem(bottomFlex, 0, 10, true)
 	if positionVisible {
 		flex.AddItem(statusLineWidget, 0, 2, false)
 	}
@@ -312,10 +412,14 @@ func init() {
 	// 	y += h / 2
 	// 	return x, y, w, h
 	// })
+	notificationWidget.SetDrawFunc(func(screen tcell.Screen, x, y, w, h int) (int, int, int, int) {
+		y += h / 2
+		return x, y, w, h
+	})
 	// Initially set up flex without search bar
 	flex = tview.NewFlex().SetDirection(tview.FlexRow).
 		AddItem(textView, 0, 40, false).
-		AddItem(textArea, 0, 10, true) // Restore original height
+		AddItem(bottomFlex, 0, 10, true)
 	if positionVisible {
 		flex.AddItem(statusLineWidget, 0, 2, false)
 	}
@@ -328,9 +432,7 @@ func init() {
 			defer colorText()
 			editedMsg := editArea.GetText()
 			if editedMsg == "" {
-				if err := notifyUser("edit", "no edit provided"); err != nil {
+				showToast("edit", "no edit provided")
-					logger.Error("failed to send notification", "error", err)
-				}
 				pages.RemovePage(editMsgPage)
 				return nil
 			}
@@ -360,9 +462,7 @@ func init() {
 			case tcell.KeyEnter:
 				newRole := roleEditWindow.GetText()
 				if newRole == "" {
-					if err := notifyUser("edit", "no role provided"); err != nil {
+					showToast("edit", "no role provided")
-						logger.Error("failed to send notification", "error", err)
-					}
 					pages.RemovePage(roleEditPage)
 					return
 				}
@@ -389,9 +489,7 @@ func init() {
 			siInt, err := strconv.Atoi(si)
 			if err != nil {
 				logger.Error("failed to convert provided index", "error", err, "si", si)
-				if err := notifyUser("cancel", "no index provided, copying user input"); err != nil {
+				showToast("cancel", "no index provided, copying user input")
-					logger.Error("failed to send notification", "error", err)
-				}
 				if err := copyToClipboard(textArea.GetText()); err != nil {
 					logger.Error("failed to copy to clipboard", "error", err)
 				}
@@ -402,9 +500,7 @@ func init() {
 			if len(chatBody.Messages)-1 < selectedIndex || selectedIndex < 0 {
 				msg := "chosen index is out of bounds, will copy user input"
 				logger.Warn(msg, "index", selectedIndex)
-				if err := notifyUser("error", msg); err != nil {
+				showToast("error", msg)
-					logger.Error("failed to send notification", "error", err)
-				}
 				if err := copyToClipboard(textArea.GetText()); err != nil {
 					logger.Error("failed to copy to clipboard", "error", err)
 				}
@@ -430,9 +526,7 @@ func init() {
 				}
 				previewLen := min(30, len(msgText))
 				notification := fmt.Sprintf("msg '%s' was copied to the clipboard", msgText[:previewLen])
-				if err := notifyUser("copied", notification); err != nil {
+				showToast("copied", notification)
-					logger.Error("failed to send notification", "error", err)
-				}
 				hideIndexBar() // Hide overlay after copying
 			}
 			return nil
@@ -464,9 +558,7 @@ func init() {
 				logger.Error("failed to upsert chat", "error", err, "chat", currentChat)
 			}
 			notification := fmt.Sprintf("renamed chat to '%s'", activeChatName)
-			if err := notifyUser("renamed", notification); err != nil {
+			showToast("renamed", notification)
-				logger.Error("failed to send notification", "error", err)
-			}
 		}
 		return event
 	})
@@ -576,9 +668,7 @@ func init() {
 			if scrollToEndEnabled {
 				status = "enabled"
 			}
-			if err := notifyUser("autoscroll", "Auto-scrolling "+status); err != nil {
+			showToast("autoscroll", "Auto-scrolling "+status)
-				logger.Error("failed to send notification", "error", err)
-			}
 			updateStatusLine()
 		}
 		// Handle Alt+7 to toggle injectRole
@@ -595,9 +685,7 @@ func init() {
 			if thinkingCollapsed {
 				status = "collapsed"
 			}
-			if err := notifyUser("thinking", "Thinking blocks "+status); err != nil {
+			showToast("thinking", "Thinking blocks "+status)
-				logger.Error("failed to send notification", "error", err)
-			}
 			return nil
 		}
 		// Handle Ctrl+T to toggle tool call/response visibility
@@ -609,9 +697,7 @@ func init() {
 			if toolCollapsed {
 				status = "collapsed"
 			}
-			if err := notifyUser("tools", "Tool calls/responses "+status); err != nil {
+			showToast("tools", "Tool calls/responses "+status)
-				logger.Error("failed to send notification", "error", err)
-			}
 			return nil
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == 'i' && event.Modifiers()&tcell.ModAlt != 0 {
@@ -631,9 +717,7 @@ func init() {
 			// Check if there are no chats for this agent
 			if len(chatList) == 0 {
 				notification := "no chats found for agent: " + cfg.AssistantRole
-				if err := notifyUser("info", notification); err != nil {
+				showToast("info", notification)
-					logger.Error("failed to send notification", "error", err)
-				}
 				return nil
 			}
 			chatMap := make(map[string]models.Chat)
@@ -651,9 +735,7 @@ func init() {
 		if event.Key() == tcell.KeyF2 && !botRespMode {
 			// regen last msg
 			if len(chatBody.Messages) == 0 {
-				if err := notifyUser("info", "no messages to regenerate"); err != nil {
+				showToast("info", "no messages to regenerate")
-					logger.Error("failed to send notification", "error", err)
-				}
 				return nil
 			}
 			chatBody.Messages = chatBody.Messages[:len(chatBody.Messages)-1]
@@ -679,9 +761,7 @@ func init() {
 				return nil
 			}
 			if len(chatBody.Messages) == 0 {
-				if err := notifyUser("info", "no messages to delete"); err != nil {
+				showToast("info", "no messages to delete")
-					logger.Error("failed to send notification", "error", err)
-				}
 				return nil
 			}
 			chatBody.Messages = chatBody.Messages[:len(chatBody.Messages)-1]
@@ -727,6 +807,7 @@ func init() {
 		if event.Key() == tcell.KeyF6 {
 			interruptResp = true
 			botRespMode = false
+			toolRunningMode = false
 			return nil
 		}
 		if event.Key() == tcell.KeyF7 {
@@ -739,9 +820,7 @@ func init() {
 			}
 			previewLen := min(30, len(msgText))
 			notification := fmt.Sprintf("msg '%s' was copied to the clipboard", msgText[:previewLen])
-			if err := notifyUser("copied", notification); err != nil {
+			showToast("copied", notification)
-				logger.Error("failed to send notification", "error", err)
-			}
 			return nil
 		}
 		if event.Key() == tcell.KeyF8 {
@@ -755,9 +834,7 @@ func init() {
 			text := textView.GetText(false)
 			cb := codeBlockRE.FindAllString(text, -1)
 			if len(cb) == 0 {
-				if err := notifyUser("notify", "no code blocks in chat"); err != nil {
+				showToast("notify", "no code blocks in chat")
-					logger.Error("failed to send notification", "error", err)
-				}
 				return nil
 			}
 			table := makeCodeBlockTable(cb)
@@ -772,9 +849,7 @@ func init() {
 			// read files in chat_exports
 			filelist, err := os.ReadDir(exportDir)
 			if err != nil {
-				if err := notifyUser("failed to load exports", err.Error()); err != nil {
+				showToast("failed to load exports", err.Error())
-					logger.Error("failed to send notification", "error", err)
-				}
 				return nil
 			}
 			fli := []string{}
@@ -804,9 +879,7 @@ func init() {
 				logger.Error("failed to export chat;", "error", err, "chat_name", activeChatName)
 				return nil
 			}
-			if err := notifyUser("exported chat", "chat: "+activeChatName+" was exported"); err != nil {
+			showToast("exported chat", "chat: "+activeChatName+" was exported")
-				logger.Error("failed to send notification", "error", err)
-			}
 			return nil
 		}
 		if event.Key() == tcell.KeyCtrlP {
@@ -845,9 +918,7 @@ func init() {
 			labels, err := initSysCards()
 			if err != nil {
 				logger.Error("failed to read sys dir", "error", err)
-				if err := notifyUser("error", "failed to read: "+cfg.SysDir); err != nil {
+				showToast("error", "failed to read: "+cfg.SysDir)
-					logger.Debug("failed to notify user", "error", err)
-				}
 				return nil
 			}
 			at := makeAgentTable(labels)
@@ -860,6 +931,7 @@ func init() {
 		if event.Key() == tcell.KeyCtrlK {
 			// add message from tools
 			cfg.ToolUse = !cfg.ToolUse
+			updateToolCapabilities()
 			updateStatusLine()
 			return nil
 		}
@@ -871,21 +943,27 @@ func init() {
 				if err != nil {
 					logger.Error("failed to open attached image", "path", lastImg, "error", err)
 					// Fall back to showing agent image
-					loadImage()
+					if err := loadImage(); err != nil {
+						logger.Warn("failed to load agent image", "error", err)
+					}
 				} else {
 					defer file.Close()
 					img, _, err := image.Decode(file)
 					if err != nil {
 						logger.Error("failed to decode attached image", "path", lastImg, "error", err)
 						// Fall back to showing agent image
-						loadImage()
+						if err := loadImage(); err != nil {
+							logger.Warn("failed to load agent image", "error", err)
+						}
 					} else {
 						imgView.SetImage(img)
 					}
 				}
 			} else {
 				// No attached image, show agent image as before
-				loadImage()
+				if err := loadImage(); err != nil {
+					logger.Warn("failed to load agent image", "error", err)
+				}
 			}
 			pages.AddPage(imgPage, imgView, true, true)
 			return nil
@@ -897,9 +975,7 @@ func init() {
 				if err != nil {
 					msg := "failed to inference user speech; error:" + err.Error()
 					logger.Error(msg)
-					if err := notifyUser("stt error", msg); err != nil {
+					showToast("stt error", msg)
-						logger.Error("failed to notify user", "error", err)
-					}
 					return nil
 				}
 				if userSpeech != "" {
@@ -957,6 +1033,17 @@ func init() {
 			showBotRoleSelectionPopup()
 			return nil
 		}
+		// INFO: shutdown
+		if event.Key() == tcell.KeyCtrlC {
+			logger.Info("caught Ctrl+C via tcell event")
+			go func() {
+				if err := pwShutDown(); err != nil {
+					logger.Error("shutdown failed", "err", err)
+				}
+				app.Stop()
+			}()
+			return nil // swallow the event
+		}
 		if event.Key() == tcell.KeyCtrlG {
 			// cfg.RAGDir is the directory with files to use with RAG
 			// rag load
@@ -968,26 +1055,20 @@ func init() {
 					// Create the RAG directory if it doesn't exist
 					if mkdirErr := os.MkdirAll(cfg.RAGDir, 0755); mkdirErr != nil {
 						logger.Error("failed to create RAG directory", "dir", cfg.RAGDir, "error", mkdirErr)
-						if notifyerr := notifyUser("failed to create RAG directory", mkdirErr.Error()); notifyerr != nil {
+						showToast("failed to create RAG directory", mkdirErr.Error())
-							logger.Error("failed to send notification", "error", notifyerr)
-						}
 						return nil
 					}
 					// Now try to read the directory again after creating it
 					files, err = os.ReadDir(cfg.RAGDir)
 					if err != nil {
 						logger.Error("failed to read dir after creating it", "dir", cfg.RAGDir, "error", err)
-						if notifyerr := notifyUser("failed to read RAG directory", err.Error()); notifyerr != nil {
+						showToast("failed to read RAG directory", err.Error())
-							logger.Error("failed to send notification", "error", notifyerr)
-						}
 						return nil
 					}
 				} else {
 					// Other error (permissions, etc.)
 					logger.Error("failed to read dir", "dir", cfg.RAGDir, "error", err)
-					if notifyerr := notifyUser("failed to open RAG files dir", err.Error()); notifyerr != nil {
+					showToast("failed to open RAG files dir", err.Error())
-						logger.Error("failed to send notification", "error", notifyerr)
-					}
 					return nil
 				}
 			}
@@ -1017,9 +1098,7 @@ func init() {
 		if event.Key() == tcell.KeyRune && event.Modifiers() == tcell.ModAlt && event.Rune() == '9' {
 			// Warm up (load) the currently selected model
 			go warmUpModel()
-			if err := notifyUser("model warmup", "loading model: "+chatBody.Model); err != nil {
+			showToast("model warmup", "loading model: "+chatBody.Model)
-				logger.Debug("failed to notify user", "error", err)
-			}
 			return nil
 		}
 		// cannot send msg in editMode or botRespMode
@@ -1072,6 +1151,18 @@ func init() {
 			chatRoundChan <- &models.ChatRoundReq{Role: persona, UserMsg: msgText}
 			return nil
 		}
+		if event.Key() == tcell.KeyTab && !shellMode {
+			currentF := app.GetFocus()
+			if currentF == textArea {
+				currentText := textArea.GetText()
+				atIndex := strings.LastIndex(currentText, "@")
+				if atIndex >= 0 {
+					filter := currentText[atIndex+1:]
+					showTextAreaFileCompletionPopup(filter)
+				}
+			}
+			return nil
+		}
 		if event.Key() == tcell.KeyPgUp || event.Key() == tcell.KeyPgDn {
 			currentF := app.GetFocus()
 			app.SetFocus(focusSwitcher[currentF])