A Tidebreak turn is a durable loop: the selected model proposes the next step, Tidebreak applies the chat's permission rules, a tool runs when allowed, and the result goes back to the model. The loop continues until the model answers, the turn is stopped, or it needs you.
A turn, step by step
- Your message and current chat settings are saved.
- A worker claims the turn and calls the selected model with the conversation, available tools, attached sources, and relevant instructions.
- If the model calls a tool, Tidebreak validates its arguments and checks the chat's permission mode and standing grants.
- The tool result is recorded and the model receives it for the next step.
- Text, reasoning status, tool activity, and waiting states stream into the transcript as they are recorded.
- The final answer and any source references become part of the durable chat history.
The journal is the source of truth. Closing the window does not turn an accepted approval, question, or running background job into an unrecorded side effect; the app reconstructs the state when you return.
Plans
In Plan mode, mutating tools are refused rather than queued for approval. The agent can inspect available information and propose a plan. A plan review card lets you comment, request a revision, or execute it. Executing the plan switches the chat out of Plan so the work can proceed under the selected permission boundary.
During ordinary work, the agent may also maintain a visible task plan. It is a progress aid, not a source of authority: listing a step does not approve the tools needed to perform it.
Questions and approvals
When the agent needs a small choice, it can ask up to three structured questions and pause. The question remains in the transcript and the Inbox until you answer it. Approval cards work the same way: the turn waits on a durable decision rather than treating a closed window as consent or refusal.
Steering or queueing a running turn
If you send another message while a turn is active, Tidebreak can either queue it as the next turn or use it to steer the current one. Choose the default under Settings → Agents → While an agent is responding. Steering is applied at a safe model boundary; it does not interrupt a tool halfway through and pretend the tool never ran.
Use steering for corrections that matter to the current result. Queue a message when it is a separate follow-up you want handled after the current turn settles. The queue is durable and can be paused, reordered, edited, or advanced with Send now.
Background work
The foreground agent can delegate independent jobs to background agents. Each job has its own isolated workspace and visible status, cannot create more agents, and returns its result to the parent turn. Closing and reopening the app does not erase an admitted job.
Background work does not bypass consent. The foreground turn must be allowed to delegate, and any host resource passed to a child is named and checked explicitly.