OpenCode’s October 29, 2026 release gives developers and small software teams two ways to hand off ongoing work: let a coding session continue toward a goal, or arrange prompts for later. Version 2.2.1 adds goal controls in the terminal and Desktop, alongside experimental, opt-in scheduling.
The practical distinction between goals and schedules is the question each answers: whether an objective is finished, or when another prompt should arrive. Neither is a guarantee of correct code or lower costs.
A goal is checked by the model, not a person
The new /goal <objective> command keeps a session working across turns, with controls to pause, resume or clear the goal. The feature’s design uses the session’s own AI model for the completion review—not a human or a separately selected reviewer model.
The review prompt asks that model to assess conversation records and tool output, rather than accept an unsupported claim of success. That adds a checking step, not human sign-off. For an outcome-based task, define observable completion criteria and review the resulting changes before adopting them.
The released implementation pauses after three consecutive stalled rounds or its 50-round limit; resuming starts a fresh round allowance. These are continuation limits, not a maximum of 50 model calls or a spending ceiling. The feature also has no dedicated token budget to cap model usage for that goal.
Schedules send prompts, not proof of completion
Scheduling is disabled by default. It requires experimental.schedules in the configuration where the session runs. Once enabled, /schedule asks the agent to turn a natural-language request into a schedule.
The timer supports one-off prompts, recurring intervals and calendar-based timing, with optional time zones and endpoints based on time or number of firings. Those endpoints limit prompt delivery; they do not establish that the requested checks succeeded.
For a recurring project check, inspect the created prompt, time zone, next run and endpoint—the schedule tools expose those details. The timer runs in the OpenCode server, so establish where that server will remain available; scheduling is not a promise of work continuing while a local machine is off.
This release does not include the proposed conditional shell checks that would stop repetition when a command meets a condition. Do not treat “repeat until tests pass” as a built-in schedule condition; a goal’s model-based completion review is a different mechanism.
Set spending limits separately
According to OpenCode’s billing guide, Console usage draws on workspace credits, while usage through customers’ own providers is billed by those providers. Automatic credit replenishment can buy more credits as the balance falls, so an initial balance is not necessarily an overall spending cap.
Console users can set monthly budgets for covered usage. These block new requests once reached, but requests already running finish and can leave spending slightly above the limit. Do not assume Console policies cover a direct connection outside Console. Check the billing route and applicable limits before leaving either feature running.
For readers still moving from OpenCode 1, RohitAI’s OpenCode 2 migration guide covers the separate question of plugin and integration readiness.