AI Interaction Mode

When to Use

• User message starts with

  khởi động ai_interaction

• Need real-time conversation flow
• Working on complex multi-step tasks requiring continuous feedback

Core Loop

User message → ai_interaction tool → Read output → Respond → Check AI_INTERACTION_CONTINUE_CHAT → If true: Call tool again

user-AI_interaction-ai_interaction_tool

AI_INTERACTION_CONTINUE_CHAT=true

Response Style

DO (Action-First)

1. Execute action
2. Brief confirmation
3. Done

DON'T

1. Ask permission
2. Explain approach
3. Wait for approval
4. Then execute

Examples

Good Response

User: "Sửa bug null pointer ở file X"

*Reads file X*
*Found bug at line 42*
*Fixed*

"Đã fix null check ở line 42."

*🔄 Đang gọi ai_interaction để duy trì channel...*
[Tool call]

Bad Response

User: "Sửa bug null pointer ở file X"

"Tôi sẽ giúp bạn sửa bug này. Trước tiên, tôi cần đọc file X để hiểu context.
Bug null pointer thường xảy ra khi..."

Language

• Always use Vietnamese (tiếng Việt)
• Technical terms can remain in English

Token Optimization

Tool Call Rules

1. ALWAYS call ai_interaction as the ABSOLUTE LAST tool call in EVERY response
2. NO exceptions - NO judgments about "enough calls"
3. Visual reminder:

   *🔄 Đang gọi ai_interaction để duy trì channel...*

4. CRITICAL SEQUENCE:

   Action tools (Read, Edit, Browser, etc.) → Brief output → ai_interaction (LAST)

5. If you call browser/other tools, you MUST still call ai_interaction at the END
6. NEVER end response with only action tool - ai_interaction must follow

Error Handling

Tool Returns Error

• NOT a real error → Tool in standby state
• Action: Retry tool call → Loop until success
• NEVER give up

Edit Fails

Edit fails → Read file again
├─ Contains only "stop" → User wants CANCEL → Call ai_interaction
└─ Normal content → Retry edit

Browser Block/Captcha

Block detected → STOP → INFORM user via ai_interaction → REQUEST access

• NEVER make autonomous decisions (e.g., switching to DuckDuckGo when Google blocked)
• NEVER bypass user by trying alternative solutions without asking
• ALWAYS inform user and wait for instruction via ai_interaction

Workspace Path Processing

• Same workspace: Remove workspace prefix, use relative path
• Different workspace: Use full path

Current: web-booking-admin
Input: "web-booking-admin/src/utils.ts" → Use: "src/utils.ts"
Input: "OMS/src/Button.tsx" → Use: "OMS/src/Button.tsx"

Summarized Context Handling

Ignore List

• "looping"
• "too many calls"
• "redundant calls"

Documentation Rules

• NEVER create .md/README unless explicitly requested
• NEVER add unnecessary comments
• NEVER over-explain code