Chat, context, and project memory

The chat composer controls both what Xorin is allowed to do and which project information is emphasized for a request.

Choose an interaction mode

Automatic project context

Each request includes a compact snapshot of the active scene, current selection, scene statistics, recent console errors, render pipeline, scripting backend, tags, and layers. It is useful orientation, not a complete dump of your project.

Use @mentions for precise context

Type @ and at least two characters to search. Mention chips can refer to:

• Project files and assets, including scripts and prefabs.
• GameObjects from the current scene scope.
• @selection for selected objects.
• @console for recent logs and errors.
• @scene for the current scene hierarchy.

Drag context into the composer

You can drag supported Unity objects and project content onto the complete composer surface. Accepted non-image items become removable context chips without adding synthetic text to your prompt.

• Images: PNG, JPEG, GIF, and WebP, including external image files.
• C# scripts, prefabs, and ScriptableObject assets from the project.
• Materials (.mat) and scenes (.unity).
• Shader sources and graphs: .shader, .hlsl, .cginc, .compute, and .shadergraph.
• Text, configuration, and Unity UI files: .json, .txt, .md, .xml, .yaml, .asmdef, .asmref, .uxml, and .uss.
• GameObjects or components from the Hierarchy; components resolve to their owning GameObject.
• Indexed folders under Assets or Packages.

Xorin accepts up to 20 pending non-image context items. A folder counts as one item and expands to supported indexed project files, including the additional formats above, when the request is sent. Non-image files and folders must be under the indexed Assets or Packages roots. Unsupported or over-limit items are skipped with a short composer message; accepted items in the same drop remain attached.

Attach visual context

Select the paperclip attachment button beside Send to upload an image or capture the current Scene View, Game View, or a four-angle Scene View composite. Xorin can also request Scene or Game View captures while diagnosing a task.

Attachments are visual evidence. They do not replace checking the hierarchy, Inspector, console, or runtime state.

Reuse a previous prompt

Select Reuse on one of your previous messages to copy its text back into the composer. Xorin also restores saved mentions and image attachments that are still available. Review the restored request before sending it again.

Reuse is unavailable while another request is running. If the composer already contains a draft, Xorin asks before replacing its text, mentions, and attachments. A notice identifies references that could not be restored so you can add them again manually.

Resume a task at its turn limit

A long-running request can pause when it reaches its agent turn limit. Use the inline card to add turns and resume the same request, or select Stop to end it. Open Adjust… to change how many turns to add and where the increased limit applies: This request, This chat, or Default on this device.

A chat limit also applies to future requests in that chat. A device default applies to chats without their own override. Continuing work can consume additional credits or provider usage. If the request has ended and cannot be resumed, start a follow-up request.

Continue an incomplete response

If a model reaches its output limit before finishing, Xorin preserves the partial response and displays Continue. Continuing starts a new request with an instruction to finish the incomplete task without repeating completed work. Xorin asks before replacing an existing composer draft.

Open project-file links

File links in Xorin responses can open an existing file under the Unity project's Assets directory or a registered Packages directory in your configured external editor. Links can also open a specific line when one is provided.

Project memory

Open Memory from the Xorin header or choose Xorin → Memory to save instructions that should accompany every request in the current project. Rules are stored in .xorin/rules.md. Good rules describe conventions such as namespaces, folder layout, input system, target platform, or art direction.

The Memory panel shows a word counter and allows up to approximately 1,600 words.

Keep rules short and stable. Task-specific instructions belong in the request rather than project memory.

Project instructions from AGENTS.md

Place shared project instructions in Assets/AGENTS.md. Xorin reads them alongside .xorin/rules.md before each Ask, Plan, or Agent request. Your explicit request takes precedence; when the project files disagree, .xorin/rules.md takes precedence over AGENTS.md and its referenced documents.

If the default file is absent, Xorin looks for other AGENTS.md files in the project. When it finds alternatives, a card lets you select Use file, Skip, or Choose another… before the request continues. Your choice is saved locally for the project. Select Change on the answered card to update future requests.

Link to other project Markdown documents from AGENTS.md to include coding conventions, architecture notes, or other guidance. Relative paths resolve from the referring document; paths starting with Assets/ or .xorin/ resolve from the project root. Xorin follows nested references and refreshes the content for each request. References must stay inside the project. Missing files, unreadable content, and loading limits produce notices; external URLs are not imported.

Chat history and context usage

• Chats persist locally and can be reopened, renamed, or deleted.
• History, Memory, and Settings open in integrated side panels from the Xorin header.
• Responses render Markdown, including tables, lists, and code blocks with Copy controls. Safe links to existing project and package files can open directly in your configured external editor.
• While a request runs, Xorin shows the current user-facing action. Successful execution history is retained but hidden after ordinary completion unless Show detailed work activity is enabled in Settings. Failures and interruptions remain visible.
• The context ring reports the most recent model call's estimated context usage and changes color as the conversation approaches its limit.

Select the context ring to see free space and usage attributed to conversation messages, the system prompt, project memory, tool definitions, provider or format overhead, and cache activity.