Task lifetime and goal continuation

Ordinary CLI work and /goal have no aggregate time or token budget. LocalAgentExecutor.chat() creates tasks with Task.timeout=None and does not copy a runtime's external benchmark deadline into an AWorld stop condition. Model, tool, hook, transport, and trajectory-finalization liveness timeouts remain separate operations. A timed-out operation is not successful task completion.

Goal continuation

A goal continues until the objective is complete, the user pauses or clears it, or an optional /goal ... --max-turns N attempt limit is reached. No maximum is imposed by default. /goal has no time-budget or deadline arguments, and old goal-state time fields do not restrict continuation or resume.

Goal state persists the attempt count, last task identity/epoch, configured agent identities, and outcome. /goal pause preserves state and interrupts execution; /goal clear interrupts and discards it. /goal resume uses the same session and rebinds recovery records to the new task and agent identities. Resume does not reset an explicitly supplied maximum attempt count.

Continuation is iterative, not recursive. Each attempt receives its own task and trajectory identity. Trusted continuation explicitly migrates the prior work ledger via carry_goal_work_state; durable resume uses the persisted source identity and resume_goal_work_state. Requirements, plans, and historical observations survive, while completion state and fresh validation evidence do not carry over to the new attempt.

The completion control plane is TaskResponse.semantic_status, completion_reason, and recoverable. Incomplete or exhausted attempts do not become success because final prose or a FINISHED message exists. An active goal continues after an incomplete result or a failed attempt, retaining the error for the next attempt. Repeated reads and no-progress advice do not impose another stop rule. Reaching the optional attempt limit preserves the unsuccessful outcome (the existing persisted status name is budget_limited). User stop always wins over a late completion or error hook.

Direct-run aworld.run.outcome.v1 also projects the bounded completion_reason. The compatibility status budget_exhausted therefore no longer hides whether termination came from task_deadline_exhausted, an agent loop limit, or another typed reserve; supervisors may display the specific reason without parsing logs or model prose.

Typed success completes a goal without a promise. When a promise is specified it must match and the control plane must also report success. Explicit --verify commands execute with real exit codes and bash pipefail; their results must be reobtained for each new attempt. Caller-provided checks remain part of the completion contract.

Library compatibility

The generic Task API retains optional caller-supplied timeout and deadline_epoch_seconds support for library consumers. Both default to None. A finite duration binds when constructed, the earliest bound wins, and a child, retry, or wall-clock rollback cannot extend it. Invalid explicit values fail closed; the historical timeout=0 sentinel is now None. This compatibility API is independent of CLI/goal execution and adds no CLI time-budget entry point.

The event runner handles Task.request_pause() and Task.request_cancel() during bootstrap and execution. Finalization joins the existing terminal snapshot under repeated cancellation instead of replacing the primary failure.

These mechanisms do not demonstrate a benchmark reward improvement. A scored result still requires a new qualified execution and the unmodified verifier. The benchmark's external protocol and scoring deadlines are unchanged.