AI Chat

Connect your AI account

• Connect Codex with your ChatGPT subscription account or use OpenAI·Anthropic API key. You can connect.
• API keys and ChatGPT login runtime are encrypted with AES-256-GCM for each user. Save it in Manager DB.
• Codex runtime first preserves the authentication status and stores data within 50MB from the most recent session. Save. Excluded old sessions are safely transferred to new sessions the next time they are used. Switch.
• If there is no follow-up request for 30 minutes after the AI ​​task is completed or forced to stop, the thread is stopped. Automatic initialization. When all threads in the same account expire, runtime It also organizes sessions and maintains login information, chat and action logs.
• When multiple AI connections are available, select the one used by each website. The connection window shows the connection method, account email or API key suffix, and the active state.
• When you disconnect an AI connection, the current conversation thread for that connection is also cleaned up.

Built-in agent that already knows your website

AI Chat starts a new conversation with the current website ID and name, package.json, and README.md. It provides requirements, common guidelines, and recent conversations. Afterwards, package.json Or, only when README.md changes, the content is re-delivered, and Directly check the current source, settings, log, and data with Cake20 MCP.

• File changes and testing are only done with the website-specific Cake20 MCP tool.
• The Codex working directory is read-only and network-blocked. Changes pass through website-scoped MCP path and activity-recording rules.
• Claude also empties the general tools and only allows the MCP tools for the current website.

Dialogue and control while working

• Simple greetings, confirmations, short explanations and read-only questions that can be answered immediately We respond immediately without a plan and do not create or open a plan card.
• Single screen capture and one or two-step correction/diagnosis are processed immediately without a plan. Shows the entire plan only when there are three or more independent complex steps. Each stage is in progress when it starts and when it ends. It is updated in order as completed, and error/interruption status and completed plans are also communicated. It remains on record.
• The current Plan is displayed as a floating card at the top of the chat scroll area. When collapsed, only one line is visible, giving priority to the steps in progress, and when all are completed, the last line is visible. It displays steps and can be closed separately.
• A separate completion badge is not attached to the AI ​​response, and only images or videos are left. and is accompanied by a short completion explanation.
• If the conversation content is shorter than the screen height, it is stacked starting from the bottom just above the input window. As new conversations are added, existing content is pushed up. When you pass the screen Scroll as usual.
• Request analysis summary, task sequence, execution tool, target file/search conditions and completion status Stacks and displays in real time on AI progress cards in chat. Explanation written by AI It displays immediately upon creation even before the task is completed.
• Even while working, you can queue up follow-up requests, modify, delete, or cancel the current task. You can send it immediately after stopping. Immediate transmission of pending messages is currently Stopping a task and sending the corresponding message are processed sequentially in one action.
• The transfer icon on the right side of the input window changes to a stop icon if the currently selected AI is working. will change and be highlighted in orange, and actions started in other tabs or sites will also be displayed. Stop. Even during an operation, you can send subsequent requests to the queue with Enter.
• Paste PNG, JPEG, GIF, or WebP images into the input window, or use image, PDF, Drag and drop text and source files to request up to 10 files and a total of 8MB. You can attach it.
• AI supports images, MP4/WebM videos, and YouTube players in answers. Up to 6 can be displayed directly. Click on the image to open it in its original form. The video plays within the chat. If YouTube playback is blocked by authentication verification You can continue viewing in a new browser using the Open External button.
• Only public·app/assets media and secure HTTPS addresses from your site source. display, and the YouTube address verifies the video ID and identifies the current site origin. Converts to the specified player. Does not execute arbitrary HTML or iframe code It doesn't.
• The HTTP/HTTPS address of the conversation is displayed while hovering over the mouse and pressing Ctrl or Cmd. It is activated as a link, and when clicked, it opens in a separate browser pop-up.
• Dragging the border between the script and Chat upwards minimizes the body of the code. You can expand the conversation area by leaving just 30px.
• The attached original is kept inside the website tmp only for the duration of the current AI request and immediately It is deleted, and only the file name, format, and size remain in the conversation history.
• The stop action on the progress card or input cancels the current response, MCP request, and Preview at any stage.
• If the same AI account is processing requests from other sites or tabs, the error message A button will stop that action and automatically resend the current message.
• Conversations maintain up to 100 messages per user, website, and AI.
• The Chat log shows every built-in AI request run on the current website, regardless of user or AI connection, and remains after a conversation reset.
• search_chat_history searches prior AI requests within the current user and website scope.
• Files modified by AI are tracked by the editor while working and are currently previewed when completed. Refreshes automatically.
• Chat displays conversations and progress output even if you switch to the Console tab. Maintain. MCP commands do not automatically switch chat screens when the AI ​​is responding.
• During MCP modification, the rotation dotted line is displayed on the AI ​​coding button, not in the log.

Handing over tasks from external AI to internal AI

An external AI using the Cake20 MCP address linked to your account will Context can be handed over directly to the editor’s built-in AI Chat. user Active AI on the same website without having to copy and retell conversations The handover content and internal AI response are saved in the conversation.

• External AI completes source modification, commit, and inspection and checks the modification status with finish_work. After completion, call handoff_to_ai.
• Handover includes user goals, completed items, key decisions, verification results, remaining issues and Recommended actions include:
• Internal AI ‘received all tasks from external AI. now with me Start with a confirmation, ‘You can proceed’ and summarize the handover status.
• During the handover call, the user's Wait for next instructions. Afterwards, when the user sends a new request in AI Chat, The internal AI checks the screen and source and continues the work.
• handoff_to_ai is not exposed to internal AI, so it does not make recursive calls between AIs. Can only be used from external MCP addresses that contain account information.

Preview check only used when requested

Cake20 AI Chat allows users to take screen captures or visual confirmation of the current request. Opens a dedicated Preview browser only when specified to view rendered screens and You can check for browser errors.

For general implementation, testing, and completion reporting, the Preview tool that uses image tokens is used. Verify with source and target tests, readiness checks, requests and logs without calls.

• open_preview and interact_preview are only used for explicit visual confirmation of the current request.
• Preview images captured through explicit requests are automatically displayed in the corresponding AI Chat response.
• We do not repeatedly capture responsive screens or operational states without user request.

Request scope and execution authority

• A request to implement, change, or recover a website authorizes the routine source edits, Git commit, and review needed to complete that result without a separate approval popup.
• Database, Redis, and Storage changes stay inside the targets and scope named in the current request. Unrequested destructive work such as a full deletion is not performed.
• Production deployment, ownership transfer, and website deletion run only when the current request explicitly includes that larger action.
• Secret values never appear in responses or activity records; settings expose only the names that are configured.

Connected to screen editing function

• Dynamic phrases and external images that cannot be saved directly in content management are currently stored. Send to AI Chat with screen context.
• Send one error/warning or all results of the distribution check via AI Chat request.
• Send unresolved screen comments from the inspection share as an edit request that includes the location and path.
• AI is used to select pages and files in the open editor, refresh preview, and start/stop. Control Console·Chat switching to display work results on the same screen.