better tool prompt, fixes

chat.go

@@ -1,6 +1,7 @@
 package main
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
@@ -13,13 +14,13 @@ import (
 )
 
 type ToolCall struct {
-	ID      string  `json:"id"`
+	ID      string  `msgpack:"id"`
-	Name    string  `json:"name"`
+	Name    string  `msgpack:"name"`
-	Args    string  `json:"args"`
+	Args    string  `msgpack:"args"`
-	Result  string  `json:"result,omitempty"`
+	Result  string  `msgpack:"result,omitempty"`
-	Done    bool    `json:"done,omitempty"`
+	Done    bool    `msgpack:"done,omitempty"`
-	Invalid bool    `json:"invalid,omitempty"`
+	Invalid bool    `msgpack:"invalid,omitempty"`
-	Cost    float64 `json:"cost,omitempty"`
+	Cost    float64 `msgpack:"cost,omitempty"`
 }
 
 type TextFile struct {
@@ -94,12 +95,15 @@ func (t *ToolCall) AsToolMessage() openrouter.ChatCompletionMessage {
 	}
 }
 
-func (r *Request) Parse() (*openrouter.ChatCompletionRequest, error) {
+func (r *Request) Parse() (*openrouter.ChatCompletionRequest, int, error) {
-	var request openrouter.ChatCompletionRequest
+	var (
+		request   openrouter.ChatCompletionRequest
+		toolIndex int
+	)
 
 	model, ok := ModelMap[r.Model]
 	if !ok {
-		return nil, fmt.Errorf("unknown model: %q", r.Model)
+		return nil, 0, fmt.Errorf("unknown model: %q", r.Model)
 	}
 
 	request.Model = r.Model
@@ -113,11 +117,11 @@ func (r *Request) Parse() (*openrouter.ChatCompletionRequest, error) {
 	}
 
 	if r.Iterations < 1 || r.Iterations > 50 {
-		return nil, fmt.Errorf("invalid iterations (1-50): %d", r.Iterations)
+		return nil, 0, fmt.Errorf("invalid iterations (1-50): %d", r.Iterations)
 	}
 
 	if r.Temperature < 0 || r.Temperature > 2 {
-		return nil, fmt.Errorf("invalid temperature (0-2): %f", r.Temperature)
+		return nil, 0, fmt.Errorf("invalid temperature (0-2): %f", r.Temperature)
 	}
 
 	request.Temperature = float32(r.Temperature)
@@ -130,7 +134,7 @@ func (r *Request) Parse() (*openrouter.ChatCompletionRequest, error) {
 			request.Reasoning.Effort = &r.Reasoning.Effort
 		default:
 			if r.Reasoning.Tokens <= 0 || r.Reasoning.Tokens > 1024*1024 {
-				return nil, fmt.Errorf("invalid reasoning tokens (1-1048576): %d", r.Reasoning.Tokens)
+				return nil, 0, fmt.Errorf("invalid reasoning tokens (1-1048576): %d", r.Reasoning.Tokens)
 			}
 
 			request.Reasoning.MaxTokens = &r.Reasoning.Tokens
@@ -145,7 +149,7 @@ func (r *Request) Parse() (*openrouter.ChatCompletionRequest, error) {
 
 	prompt, err := BuildPrompt(r.Prompt, r.Metadata, model)
 	if err != nil {
-		return nil, err
+		return nil, 0, err
 	}
 
 	if prompt != "" {
@@ -156,9 +160,11 @@ func (r *Request) Parse() (*openrouter.ChatCompletionRequest, error) {
 		request.Tools = GetSearchTools()
 		request.ToolChoice = "auto"
 
+		toolIndex = len(request.Messages)
+
 		request.Messages = append(
 			request.Messages,
-			openrouter.SystemMessage(fmt.Sprintf(InternalToolsPrompt, r.Iterations-1)),
+			openrouter.SystemMessage(""),
 		)
 	}
 
@@ -194,9 +200,9 @@ func (r *Request) Parse() (*openrouter.ChatCompletionRequest, error) {
 
 				for i, file := range message.Files {
 					if len(file.Name) > 512 {
-						return nil, fmt.Errorf("file %d is invalid (name too long, max 512 characters)", i)
+						return nil, 0, fmt.Errorf("file %d is invalid (name too long, max 512 characters)", i)
 					} else if len(file.Content) > 4*1024*1024 {
-						return nil, fmt.Errorf("file %d is invalid (too big, max 4MB)", i)
+						return nil, 0, fmt.Errorf("file %d is invalid (too big, max 4MB)", i)
 					}
 
 					lines := strings.Count(file.Content, "\n") + 1
@@ -238,7 +244,7 @@ func (r *Request) Parse() (*openrouter.ChatCompletionRequest, error) {
 		}
 	}
 
-	return &request, nil
+	return &request, toolIndex, nil
 }
 
 func HandleChat(w http.ResponseWriter, r *http.Request) {
@@ -254,7 +260,7 @@ func HandleChat(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	request, err := raw.Parse()
+	request, toolIndex, err := raw.Parse()
 	if err != nil {
 		RespondJson(w, http.StatusBadRequest, map[string]any{
 			"error": err.Error(),
@@ -285,13 +291,25 @@ func HandleChat(w http.ResponseWriter, r *http.Request) {
 
 		response.WriteChunk(NewChunk(ChunkStart, nil))
 
-		if len(request.Tools) > 0 && iteration == raw.Iterations-1 {
+		if len(request.Tools) > 0 {
-			debug("no more tool calls")
+			if iteration == raw.Iterations-1 {
+				debug("no more tool calls")
 
-			request.Tools = nil
+				request.Tools = nil
-			request.ToolChoice = ""
+				request.ToolChoice = ""
+			}
 
-			request.Messages = append(request.Messages, openrouter.SystemMessage("You have reached the maximum number of tool calls for this conversation. Provide your final response based on the information you have gathered."))
+			// iterations - 1
+			total := raw.Iterations - (iteration + 1)
+
+			var tools bytes.Buffer
+
+			InternalToolsTmpl.Execute(&tools, map[string]any{
+				"total":     total,
+				"remaining": total - 1,
+			})
+
+			request.Messages[toolIndex].Content.Text = tools.String()
 		}
 
 		dump("chat.json", request)

internal/tools.txt

@@ -1,5 +1,13 @@
 # Tool Use
-Use at most 1 tool call per turn. You have %d turns with tool calls total. Choose your tool strategically - you may only get one chance.
+{{if eq .total 0}}
+No more tool calls available. Provide your final response now based on the information you have gathered.
+{{else if eq .total 1}}
+You have exactly 1 tool call available. After using it, you must provide your final response in the next turn based on the information you have gathered.
+{{else}}
+You have {{.total}} tool calls remaining. If you use a tool now, you will have {{.remaining}} more tool call(s) available in subsequent turns.
+{{end}}
+{{- if gt .total 0}}
+Use at most 1 tool call per turn (not multiple tools in the same turn).{{- if gt .total 1 }} After calling a tool, you will receive its results and can then decide whether to call another tool or provide your final response.{{- end }}
 
 **search_web({query, num_results?, intent?, recency?, domains?})**
 - Search the live web via Exa AI. Craft concise, specific queries (2-8 words typically work best).
@@ -28,4 +36,5 @@ Use at most 1 tool call per turn. You have %d turns with tool calls total. Choos
 - Get comprehensive repo overview via GitHub API: description, README content, file structure
 - Returns top-level files/directories with raw content links for direct access
 - Use when you need to understand project structure, setup instructions, or codebase overview
-- More efficient than searching for "repo_name GitHub" when you know the exact owner/repo
+- More efficient than searching for "repo_name GitHub" when you know the exact owner/repo
+{{end}}

prompts.go

@@ -32,6 +32,8 @@ var (
 	//go:embed internal/tools.txt
 	InternalToolsPrompt string
 
+	InternalToolsTmpl *template.Template
+
 	//go:embed internal/title.txt
 	InternalTitlePrompt string
 
@@ -42,6 +44,7 @@ var (
 )
 
 func init() {
+	InternalToolsTmpl = NewTemplate("internal-tools", InternalToolsPrompt)
 	InternalTitleTmpl = NewTemplate("internal-title", InternalTitlePrompt)
 
 	var err error

static/js/chat.js

@@ -1191,6 +1191,8 @@
 			chunk => {
 				stopLoadingTimeout();
 
+				console.log("chunk", chunk);
+
 				if (chunk === "aborted") {
 					chatController = null;