Conversations and context

Start, continue and organize chats while keeping their model, context and destination clear.

A Mellow chat combines your request, the agent's context, and the actions needed to produce a result. Before sending, check four things: the agent, model, project, and execution location. They answer different questions and can change independently.

Orient the conversation

A remote host selected in the conversation is not merely a visual filter. Check its name before sending a file-changing task. Likewise, a provider called Cloud does not mean the conversation is automatically a task assigned to a published Cloud workspace agent.

Start, continue, or branch the work

Use a new chat for a fresh objective. Continue an existing chat when its history is still relevant. If two tasks have unrelated assumptions or different folders, separate conversations keep their context clearer.

Give a task an observable result:

Compare these two proposals. Return a table of scope, cost assumptions,
and unresolved dependencies. Quote the source section for each difference.
Do not send the comparison to anyone.

For a follow-up, identify what should change: “Keep the table, but add a column for the owner of each unresolved dependency.” This is more precise than asking the model to redo everything.

Choosing a model

Open the model control and choose an available provider/model pair. Where supported, select its reasoning or effort option. Model-specific options are not universal; an option present for one provider may not apply to another.

Readiness has several stages: installed or listed, authenticated, loaded where necessary, and successfully responding. A visible model name is not proof of all four. Test with a short message after changing providers.

A long context or a second local model may require additional memory. If loading fails, inspect the model status rather than sending the same message repeatedly.

Working Folder

Select a folder in the composer for work on local files. Each conversation retains its own selection. A fresh project chat can inherit the project's folder; otherwise an agent's remembered folder can seed the choice.

Review the chip after restoring a chat or switching tasks. Clear or change the folder when it no longer belongs to the work. Choosing a folder affects the execution arrangement and can disable Sandbox mode for that agent.

Follow actions and permissions

Tool rows show operations such as reads, searches, and commands. A permission card describes a proposed action. Approving one action does not mean every later operation succeeded. Inspect the returned tool result and any generated file.

If an action is denied, revise the task or change the relevant setting deliberately. Do not interpret the agent's explanation as proof that the denied operation occurred.

Use Stop when the active work should end. A tool that already changed a file may need a separate review or undo; stopping generation is not a transaction rollback.

Reviewing file changes

The inspector's File Changes view exposes supported recorded edits. Review individual operations or the final file state. Revert controls act on retained history and may depend on whether the file has changed again.

Open artifacts to inspect deliverables. A link to a generated file is more useful than an unsupported claim that it was created. Keep normal backups and version control for important projects.

History and multiple conversations

Use sidebar search and project organization to find prior work. Pin items you return to often. Deleting a conversation is different from removing it from a project and can also remove associated history; inspect the action before confirming.

Separate windows or tabs can hold independent conversations. Their agents, folders, and models may differ. Keeping a window visible does not change the permissions of its task.

Voice input

Configure speech recognition in Voice, grant microphone access, and confirm the relevant speech model is ready. Dictation should place recognized text in the intended composer or target field. Voice playback uses a different path and may require its own model.

If recognition works in Setup but not in a test field, inspect the destination and typing permission rather than redownloading a working recognizer immediately.

Context compaction

Long conversations can exceed the selected model's useful context budget. Compaction condenses earlier material so work can continue. A condensed account can omit exact wording or incidental details; retain important requirements in project instructions or source documents and restate them when needed.

Troubleshooting

For shortcuts, use the app's menus and shortcut settings: custom assignments can differ from defaults. For execution details, continue to Tasks and files.