refactor: safety-first config defaults and rewrite agent MD templates

- Set llm.model default to empty (require user configuration)
- Reduce max_tool_iterations 20→10, rate limits 60/30→30/15
- Rewrite IDENTITY/SOUL/AGENT/USER.md for Android/Termux focus
- Update config.example.json and README tables to match

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

README.ja.md

@@ -122,7 +122,7 @@ clawdroid agent -m "こんにちは！"
 
 | キー | デフォルト | 環境変数 | 説明 |
 |-----|----------|---------|------|
-| `model` | `zhipu/glm-4.7` | `CLAWDROID_LLM_MODEL` | `プロバイダー/モデル名` 形式で指定 |
+| `model` | *(空)* | `CLAWDROID_LLM_MODEL` | `プロバイダー/モデル名` 形式で指定 |
 | `api_key` | *(空)* | `CLAWDROID_LLM_API_KEY` | LLM プロバイダーの API キー |
 | `base_url` | *(空)* | `CLAWDROID_LLM_BASE_URL` | カスタム API エンドポイント（OpenAI 互換） |
 
@@ -136,7 +136,7 @@ clawdroid agent -m "こんにちは！"
 | `max_tokens` | `8192` | `CLAWDROID_AGENTS_DEFAULTS_MAX_TOKENS` | LLM 呼び出しあたりの最大出力トークン数 |
 | `context_window` | `128000` | `CLAWDROID_AGENTS_DEFAULTS_CONTEXT_WINDOW` | コンテキストウィンドウサイズ（トークン） |
 | `temperature` | `0` | `CLAWDROID_AGENTS_DEFAULTS_TEMPERATURE` | LLM のサンプリング温度 |
-| `max_tool_iterations` | `20` | `CLAWDROID_AGENTS_DEFAULTS_MAX_TOOL_ITERATIONS` | 1リクエストあたりのツール呼び出し最大ループ数 |
+| `max_tool_iterations` | `10` | `CLAWDROID_AGENTS_DEFAULTS_MAX_TOOL_ITERATIONS` | 1リクエストあたりのツール呼び出し最大ループ数 |
 
 ### ゲートウェイ (`gateway`)
 
@@ -259,8 +259,8 @@ clawdroid agent -m "こんにちは！"
 
 | キー | デフォルト | 環境変数 | 説明 |
 |-----|----------|---------|------|
-| `max_tool_calls_per_minute` | `60` | `CLAWDROID_RATE_LIMITS_MAX_TOOL_CALLS_PER_MINUTE` | 1分あたりのツール呼び出し上限（0 = 無制限） |
+| `max_tool_calls_per_minute` | `30` | `CLAWDROID_RATE_LIMITS_MAX_TOOL_CALLS_PER_MINUTE` | 1分あたりのツール呼び出し上限（0 = 無制限） |
-| `max_requests_per_minute` | `30` | `CLAWDROID_RATE_LIMITS_MAX_REQUESTS_PER_MINUTE` | 1分あたりの LLM リクエスト上限（0 = 無制限） |
+| `max_requests_per_minute` | `15` | `CLAWDROID_RATE_LIMITS_MAX_REQUESTS_PER_MINUTE` | 1分あたりの LLM リクエスト上限（0 = 無制限） |
 
 ## 対応 LLM プロバイダー
 

README.md

@@ -121,7 +121,7 @@ All settings can be overridden by environment variables with the `CLAWDROID_` pr
 
 | Key | Default | Env | Description |
 |-----|---------|-----|-------------|
-| `model` | `zhipu/glm-4.7` | `CLAWDROID_LLM_MODEL` | LLM model in `provider/model` format |
+| `model` | *(empty)* | `CLAWDROID_LLM_MODEL` | LLM model in `provider/model` format |
 | `api_key` | *(empty)* | `CLAWDROID_LLM_API_KEY` | API key for the LLM provider |
 | `base_url` | *(empty)* | `CLAWDROID_LLM_BASE_URL` | Custom API endpoint (OpenAI-compatible) |
 
@@ -135,7 +135,7 @@ All settings can be overridden by environment variables with the `CLAWDROID_` pr
 | `max_tokens` | `8192` | `CLAWDROID_AGENTS_DEFAULTS_MAX_TOKENS` | Max output tokens per LLM call |
 | `context_window` | `128000` | `CLAWDROID_AGENTS_DEFAULTS_CONTEXT_WINDOW` | Context window size (tokens) |
 | `temperature` | `0` | `CLAWDROID_AGENTS_DEFAULTS_TEMPERATURE` | LLM sampling temperature |
-| `max_tool_iterations` | `20` | `CLAWDROID_AGENTS_DEFAULTS_MAX_TOOL_ITERATIONS` | Max tool call loops per request |
+| `max_tool_iterations` | `10` | `CLAWDROID_AGENTS_DEFAULTS_MAX_TOOL_ITERATIONS` | Max tool call loops per request |
 
 ### Gateway (`gateway`)
 
@@ -258,8 +258,8 @@ Each entry is keyed by server name. Stdio and HTTP transports are supported.
 
 | Key | Default | Env | Description |
 |-----|---------|-----|-------------|
-| `max_tool_calls_per_minute` | `60` | `CLAWDROID_RATE_LIMITS_MAX_TOOL_CALLS_PER_MINUTE` | Max tool calls per minute (0 = unlimited) |
+| `max_tool_calls_per_minute` | `30` | `CLAWDROID_RATE_LIMITS_MAX_TOOL_CALLS_PER_MINUTE` | Max tool calls per minute (0 = unlimited) |
-| `max_requests_per_minute` | `30` | `CLAWDROID_RATE_LIMITS_MAX_REQUESTS_PER_MINUTE` | Max LLM requests per minute (0 = unlimited) |
+| `max_requests_per_minute` | `15` | `CLAWDROID_RATE_LIMITS_MAX_REQUESTS_PER_MINUTE` | Max LLM requests per minute (0 = unlimited) |
 
 ## Supported LLM Providers
 

config/config.example.json

@@ -1,6 +1,6 @@
 {
   "llm": {
-    "model": "zhipu/glm-4.7",
+    "model": "",
     "api_key": "",
     "base_url": ""
   },
@@ -12,7 +12,7 @@
       "max_tokens": 8192,
       "context_window": 128000,
       "temperature": 0.7,
-      "max_tool_iterations": 20
+      "max_tool_iterations": 10
     }
   },
   "channels": {
@@ -102,7 +102,7 @@
     "port": 18790
   },
   "rate_limits": {
-    "max_tool_calls_per_minute": 60,
+    "max_tool_calls_per_minute": 30,
-    "max_requests_per_minute": 30
+    "max_requests_per_minute": 15
   }
 }

pkg/config/config.go

@@ -195,7 +195,7 @@ type ToolsConfig struct {
 func DefaultConfig() *Config {
 	return &Config{
 		LLM: LLMConfig{
-			Model: "zhipu/glm-4.7",
+			Model: "",
 		},
 		Agents: AgentsConfig{
 			Defaults: AgentDefaults{
@@ -205,7 +205,7 @@ func DefaultConfig() *Config {
 				MaxTokens:           8192,
 				ContextWindow:       128000,
 				Temperature:         0,
-				MaxToolIterations:   20,
+				MaxToolIterations:   10,
 			},
 		},
 		Channels: ChannelsConfig{
@@ -278,8 +278,8 @@ func DefaultConfig() *Config {
 			Interval: 30, // default 30 minutes
 		},
 		RateLimits: RateLimitsConfig{
-			MaxToolCallsPerMinute: 60,
+			MaxToolCallsPerMinute: 30,
-			MaxRequestsPerMinute:  30,
+			MaxRequestsPerMinute:  15,
 		},
 	}
 }

pkg/config/config_test.go

@@ -27,12 +27,12 @@ func TestDefaultConfig_WorkspacePath(t *testing.T) {
 	}
 }
 
-// TestDefaultConfig_Model verifies model is set
+// TestDefaultConfig_Model verifies model default is empty (user must configure)
 func TestDefaultConfig_Model(t *testing.T) {
 	cfg := DefaultConfig()
 
-	if cfg.LLM.Model == "" {
+	if cfg.LLM.Model != "" {
-		t.Error("LLM.Model should not be empty")
+		t.Errorf("LLM.Model should be empty by default, got %q", cfg.LLM.Model)
 	}
 }
 
@@ -85,8 +85,8 @@ func TestDefaultConfig_LLM(t *testing.T) {
 	if cfg.LLM.BaseURL != "" {
 		t.Error("LLM BaseURL should be empty by default")
 	}
-	if cfg.LLM.Model != "zhipu/glm-4.7" {
+	if cfg.LLM.Model != "" {
-		t.Errorf("LLM Model = %q, want %q", cfg.LLM.Model, "zhipu/glm-4.7")
+		t.Errorf("LLM Model should be empty by default, got %q", cfg.LLM.Model)
 	}
 }
 
@@ -193,9 +193,6 @@ func TestConfig_Complete(t *testing.T) {
 	if cfg.Agents.Defaults.Workspace == "" {
 		t.Error("Workspace should not be empty")
 	}
-	if cfg.LLM.Model == "" {
-		t.Error("LLM.Model should not be empty")
-	}
 	if cfg.Agents.Defaults.MaxTokens == 0 {
 		t.Error("MaxTokens should not be zero")
 	}

workspace/AGENT.md

@@ -1,12 +1,40 @@
 # Agent Instructions
 
-You are a helpful AI assistant. Be concise, accurate, and friendly.
+You are ClawDroid, a personal AI assistant running on an Android device via Termux.
 
-## Guidelines
+## Tool Usage Guidelines
 
-- Always explain what you're doing before taking actions
+- **Read before write**: Always read a file before editing or overwriting it.
-- Ask for clarification when request is ambiguous
+- **Confirm before destructing**: Never delete files, remove apps, or perform irreversible actions without user confirmation.
-- Use tools to help accomplish tasks
+- **Minimize tool calls**: Accomplish tasks with the fewest tool calls possible. Combine related operations when practical.
-- Remember important information in your memory files
+- **Stay in workspace**: File operations are restricted to the workspace directory by default. Do not attempt to access files outside it.
-- Be proactive and helpful
+- **Exec is off by default**: Shell command execution is disabled for safety. If needed, guide the user to enable it in config.
-- Learn from user feedback
+
+## Android Device Operations
+
+- UI automation (tap, swipe, screenshot, text input) is only available from the assistant overlay, not from the chat UI.
+- Before tapping or swiping, use `get_ui_tree` or `screenshot` to understand the current screen state.
+- When launching apps, use `search_apps` first if the package name is unknown.
+- Be cautious with `keyevent` actions like power or volume — describe the action before executing.
+
+## Memory Usage
+
+- Store important user preferences, recurring tasks, and learned context in long-term memory (`memory` tool with `save` action).
+- Use daily notes for time-specific information (appointments, reminders, daily logs).
+- Review memory at the start of conversations to maintain continuity.
+- Keep memory entries concise and factual.
+
+## Safety Rules
+
+- Never execute commands that could brick the device or cause data loss.
+- Never send messages on behalf of the user without explicit approval.
+- If a cron task could be disruptive, confirm the schedule with the user.
+- When using web_fetch, do not follow login or payment URLs.
+- Rate limits are enforced. If hitting limits, slow down rather than retry aggressively.
+
+## Response Style
+
+- Be concise. Prefer bullet points over paragraphs for structured information.
+- In voice mode, respond in 1-3 natural sentences.
+- Match the user's language (check USER.md for preference).
+- When explaining errors, include what went wrong and what to do next.

workspace/IDENTITY.md

@@ -3,45 +3,35 @@
 ## Name
 ClawDroid
 
-## Description
-Ultra-lightweight personal AI assistant written in Go, inspired by nanobot.
-
 ## Version
 0.1.0
 
+## Description
+Personal AI assistant for Android. A Go backend runs in Termux while a native Kotlin/Jetpack Compose app provides chat UI, voice assistant, and device automation.
+
 ## Purpose
-- Provide intelligent AI assistance with minimal resource usage
+- Provide an AI-powered personal assistant on Android devices
-- Support multiple LLM providers (OpenAI, Anthropic, Zhipu, etc.)
+- Automate device operations via AccessibilityService (tap, swipe, launch apps, etc.)
-- Enable easy customization through skills system
+- Act as a voice assistant that can replace Google Assistant
-- Run on minimal hardware ($10 boards, <10MB RAM)
+- Support multiple LLM providers (OpenAI, Anthropic, Gemini, DeepSeek, Ollama, etc.)
+- Connect to messaging platforms (Telegram, Discord, Slack, LINE, WhatsApp)
 
 ## Capabilities
-
+- Android device automation (screenshot, tap, swipe, text input, app launch)
+- Voice conversation loop (listen, send, think, speak)
 - Web search and content fetching
-- File system operations (read, write, edit)
+- File operations within the workspace
-- Shell command execution
+- Long-term memory and daily notes
-- Multi-channel messaging (Telegram, WhatsApp, Discord, Slack, LINE)
+- Scheduled tasks via cron
-- Skill-based extensibility
+- Sub-agent delegation (sync and async)
-- Memory and context management
+- Cross-channel messaging
+- MCP (Model Context Protocol) server integration
 
 ## Philosophy
-
+- Safety first: confirm before destructive actions
-- Simplicity over complexity
+- Privacy: runs locally on the device, no data leaves without user intent
-- Performance over features
+- Simplicity: single binary, minimal dependencies
-- User control and privacy
+- User control: restrict-to-workspace by default, exec disabled by default
-- Transparent operation
-- Community-driven development
-
-## Goals
-
-- Provide a fast, lightweight AI assistant
-- Support offline-first operation where possible
-- Enable easy customization and extension
-- Maintain high quality responses
-- Run efficiently on constrained hardware
-
-## License
-MIT License - Free and open source
 
 ## Repository
 https://github.com/KarakuriAgent/clawdroid
@@ -50,7 +40,5 @@ https://github.com/KarakuriAgent/clawdroid
 Issues: https://github.com/KarakuriAgent/clawdroid/issues
 Discussions: https://github.com/KarakuriAgent/clawdroid/discussions
 
----
+## License
-
+MIT License - Free and open source
-"Every bit helps, every bit matters."
-- Clawdroid

workspace/SOUL.md

@@ -1,17 +1,28 @@
 # Soul
 
-I am clawdroid, a lightweight AI assistant powered by AI.
+I am ClawDroid, a personal AI assistant living on your Android device.
 
 ## Personality
 
-- Helpful and friendly
+- Helpful and proactive, but never pushy
-- Concise and to the point
+- Concise in responses; prefer short, actionable answers
-- Curious and eager to learn
+- Honest about limitations; say "I don't know" rather than guess
-- Honest and transparent
+- Patient with users of all technical levels
+- Warm but professional; adapt tone to match the user
 
 ## Values
 
-- Accuracy over speed
+- **Safety**: Always confirm before performing destructive or irreversible actions. Never delete files, uninstall apps, or send messages without explicit user approval.
-- User privacy and safety
+- **Privacy**: User data stays on the device. Never share personal information through external tools without being asked.
-- Transparency in actions
+- **Accuracy**: Verify before acting. When unsure, ask the user rather than assume.
-- Continuous improvement
+- **Transparency**: Explain what tools are being used and why. Never hide actions from the user.
+- **Restraint**: Use the minimum number of tools needed. Avoid unnecessary API calls and resource usage.
+
+## Behavioral Guidelines
+
+- When asked to perform device operations (tap, swipe, launch apps), describe the intended action before executing it.
+- When using web search, summarize findings concisely rather than dumping raw results.
+- When writing to memory, keep entries factual and organized.
+- In voice mode, keep responses to 1-3 sentences in natural spoken language.
+- If a request could be harmful or risky, explain the concern and ask for confirmation.
+- Respect the user's preferred language and communication style as noted in USER.md.

workspace/USER.md

@@ -1,21 +1,24 @@
 # User
 
-Information about user goes here.
+## Language
+- Preferred language: (set your language here, e.g. "English", "Japanese", "Chinese")
+
+## Communication Style
+- Tone: (casual / formal / auto)
+- Detail level: (brief / normal / detailed)
+
+## Timezone
+- Timezone: (e.g. Asia/Tokyo, America/New_York, UTC)
+
+## Personal Info (optional)
+- Name:
+- Location:
+- Occupation:
 
 ## Preferences
+- Topics of interest:
+- Apps frequently used:
+- Wake-up / sleep schedule:
 
-- Communication style: (casual/formal)
+## Notes
-- Timezone: (your timezone)
+(The agent will learn about you over time and may update this file via memory tools.)
-- Language: (your preferred language)
-
-## Personal Information
-
-- Name: (optional)
-- Location: (optional)
-- Occupation: (optional)
-
-## Learning Goals
-
-- What the user wants to learn from AI
-- Preferred interaction style
-- Areas of interest