본문으로 건너뛰기

Plan work, choose effort, and pursue a goal

Plan, effort and goals are available in Grida Desktop on macOS and Linux and in headless and ACP clients. This guide describes Desktop; clients use the client guide.

Decide what the agent may do​

Type /plan and select it, or choose + → Plan mode, to add a Plan badge. Describe what to investigate and send. The agent reads and proposes without changing files or running commands. The badge stays for later messages until you remove it (hover it and choose ×) or choose Plan mode again. Adding the badge starts no work and keeps the rest of your draft.

Messages you queue keep the mode they were sent with; a correction steered into a running turn follows that turn's mode. Plan does not change the agent's access or effort. Understand agent permissions

Choose how much effort to use​

The model button at the bottom right of the composer shows the model name followed by its effort. Select it and drag the Effort slider: Low, Medium, High, Extra high and Ultra. Medium is the default; the reset button returns to it.

Each chat remembers its model and effort together, including unsent changes. Opening another chat restores that chat's pair. A new chat uses the complete pair from your last explicit model or effort change; simply reopening a chat never updates that default. Accepted messages and Goals keep their captured effort when you change the composer for future work.

Low through Extra high set how much the model reasons before answering, for models whose provider supports a reasoning setting. Other models keep their own reasoning, and the panel says so. Any effort works with Plan and applies to your next message. Steering keeps the current turn's effort, and editing a queued or earlier message keeps its original effort.

Explicit delegation in ordinary Act work is available at every effort level on a supported native connection. Ask the agent to divide independent work; Ultra is not required. General workers have a separately bounded allowance and the parent's delegated permissions. Plan and persistent Goals retain their existing narrower delegation contracts.

Ultra gives an ordinary task more inference steps and lets the agent ask one independent reviewer to check its work while it continues. It is available on every supported native model connection, including ChatGPT subscriptions and compatible endpoints. On models with a reasoning setting, Ultra also selects the highest supported reasoning level. Other models keep their own reasoning and still receive the larger work allowance and review option.

The reviewer uses the same model connection and account as the main agent. It can read the permitted project but does not see your chat. Its work appears in the turn's activity. Choosing Ultra permits a review; the agent decides whether one is useful for the task. Work with subagents

The agent and reviewer work within bounded inference-step allowances. Provider token limits still apply when the connection supports them. Ultra does not set a subscription-quota or spending limit.

If the Desktop host cannot start Ultra, the message stays unsent with your selected effort. Restore the connection or update Desktop before retrying; Grida does not send that message with ordinary effort instead.

Keep an objective across turns​

  1. Type /goal and select it, or choose + → Goal. A Goal badge appears in the composer without starting work; hover it and choose × to remove it. A chat holds one goal at a time; clear or finish the current one first.
  2. Describe the objective in your own words and send. If the agent is still working or messages are waiting, the goal waits in the queue as Goal: … and starts when its turn comes; until then you can reorder or remove it like a queued message. The message is marked Sent as goal in the chat. The agent works from that objective and continues on its own within eight work turns and 64 inference steps. Choose Ultra before sending to let one independent review share the goal's allowance. Other effort levels do not change a goal's reasoning: goal turns use the model's default reasoning, as the effort panel notes while the Goal badge is on.
  3. While the goal exists, a strip above the composer shows its status and objective, for example Pursuing goal or Paused goal. Use its buttons to clear the goal, pause or resume it, and edit it. Editing opens the objective text; saving keeps the goal's used allowance and stops any current goal turn while keeping its recorded work.
  4. When the strip shows Review goal, compare the result with your objective and choose Confirm when it is met. The strip then shows Goal complete, with Dismiss goal in place of Clear goal. The chat keeps the proposal and confirmation so you can return to them later.

If the goal stops before it is done, the strip says why:

  • Goal needs input, Goal stalled or Goal can't continue: answer or fix what the latest reply asks for, then choose Resume goal.
  • Goal needs review: check the work against the objective before resuming or clearing the goal.
  • Goal interrupted: Grida could not confirm how the last goal turn ended. Review its work, then clear the goal and start a new one if more is needed.
  • Goal out of allowance: automatic work has ended. Clear the goal and start a new one to keep going.

If an Ultra review finishes after its goal turn, choose Refresh goal allowance on the strip before resuming.

Goal entry takes text and @ mentions; attachments are unavailable while the Goal badge is on. Add the Plan badge as well for a goal that investigates and proposes; later goal turns keep the instructions accepted when the goal started.

For example, set the goal “Find the open decisions in the project notes and the section where each appears.” Check the answer against the notes before confirming. Inspect and save results

To hold a goal, choose Pause goal on its strip, then Resume goal to continue with the allowance it already has. Pausing with the composer's main button pauses only the current turn. Guide and interrupt work

Desktop does not ask for separate success criteria, allowances, a spending limit, a named reviewer or a check-in when starting a goal. Clients can set those through the goal protocol. Use goals from a client · Set a goal's spending limit