Skip to main content
The assistant uses the current workspace, selected library-entry context, authorized tools, recent conversation history and explicitly uploaded attachments. A filename or natural-language request helps it locate candidates; it does not grant permission to read arbitrary storage or invent a matching file.

Context versus file bytes

Selecting Current file supplies a library reference. The tools can inspect metadata and available indexed text under the actor’s permissions. They do not open every original automatically. To ask a model to inspect an image/PDF directly, use an explicit supported attachment and a compatible model. Ordinary library references and temporary chat attachments have different storage lifecycles. A natural request such as “Find the latest launch brief and open it” can use actual library search/list results without requiring the user to type its exact extension. Ambiguous candidates still need clarification. The assistant must receive an action receipt before claiming it opened a file.

Temporary attachments

Markdown is accepted when detected as supported plain text; an arbitrary MIME alias is not automatically accepted. AVIF, video and Office attachments are not in the shipped staging allowlist. The selected provider/model may impose lower limits. An accepted attachment format does not guarantee provider support. Temporary attachments are private and scoped to their workspace, owner and run. Staging uses assistant.attachment_directory, falling back to storage_path('app/private/media-studio-assistant'). Access/current file state is validated again for provider delivery. Expired/unavailable attachments must be uploaded again. Configure assistant maintenance and retention for their lifecycle; uploading a chat attachment does not add a managed library asset.

Authorized tool actions

Tools execute as the current principal inside the selected workspace and tenant. Underlying file/source permissions and revision checks still apply. A model’s request cannot grant extra access. Trashing and creating a share are real actions when the user requests them; the assistant is not an unrestricted administrator. The current assistant does not generate or edit image bytes; the separate image-editing integration publishes explicit user selections.

Turn and history bounds

The shipped agent permits at most 6 tool steps and 4096 response tokens. Provider streaming has a 120-second timeout; the queued run job has a 180-second timeout. These are implementation limits, not currently exposed model-profile settings. Stop requests cancellation; it is not a provider billing refund or rollback of a tool action already committed. Provider context uses at most 58 historical messages within a 64 KiB prompt/history budget. Saved chats can be much longer because their UI and database are paginated. This is not unlimited model memory. See history pagination and operations for concurrency, retention and failure handling.