defineIsolatedTask
Purpose and behavior
Section titled “Purpose and behavior”Declare a task that runs dispatch() at each attempt with the repository, provider, agent and options returned by request; each attempt allocates and closes its own sandbox. The output is the full DispatchResult with its resume() and fork() methods, so a checkpointed run must wrap the task in a defineTask() that returns JSON. cache is rejected.
Complete example and detailed rules.
Parameters and properties
Section titled “Parameters and properties”optionsRequiredOmit<TaskOptions<DispatchResult<T>>, "cache" | "perform"> & IsolatedTaskOptions<T>Task scheduling settings and request factory selecting a separate repository and sandbox for each attempt.options.retryOptionalRetry | undefinedRetry policy for failed attempts; without it the task runs once. A retried attempt repeats its side effects.options.gateOptionalWorkflowGate | undefinedPersisted approval or pause definition; execution requires a checkpoint and a matching trusted decision.options.keyRequiredstringUnique key in the workflow, matching [A-Za-z0-9][A-Za-z0-9._-]*. Records, events and checkpoints identify the task by it.options.afterOptionalreadonly Task<unknown>[] | undefinedTasks that must be done before this one starts, default none; only these can be read with context.value().options.interactionOptionalTaskInteraction | undefinedOpt-in durable human-input contract; requires checkpointed scheduling and cannot be combined with a gate.options.conditionOptional((context: TaskContext) => boolean | Promise<boolean>) | undefinedEvaluated with attempt 0 before the task runs, including when a later start() resumes it; false ends the task as skipped, which skips its dependents.options.timeoutMsOptionalnumber | undefinedDeadline in milliseconds for each attempt, a positive integer up to 2147483647. Expiry aborts context.signal and fails the attempt, which retry may repeat.options.requestRequired(context: TaskContext) => IsolatedTaskRequest<T> | Promise<IsolatedTaskRequest<T>>Build the repository, sandbox provider, agent and dispatch options for each attempt; may be async.options.quotaResumeOptionalQuotaResumePolicy | undefinedAfter a quota pause, continue the captured conversation in the new dispatch (continue, default) or start a new one (restart). Automatically integrated workspaces then start from the interrupted branch; uncommitted changes stay in its retained worktree.