Agent and Command modes

The default Agent plans across model steps, invokes tools, receives real results, and continues. Partial streaming tool arguments never execute. Broken tool-call JSON yields non-executable failures with at most two regeneration attempts rather than guessed arguments; rejected calls still consume budget.

nl2sh --mode command "Show memory usage" generates one command, then classifies/confirms/executes it. --dry-run only displays it. Agent suits multi-stage feedback; !command runs a command explicitly written by the user.

Preset	Steps	Tool calls	Active time
fast	20	40	10 minutes
normal	50	100	30 minutes
deep	100	200	60 minutes

Explicit max_agent_steps, max_tool_calls, and max_task_execution_time_secs override presets; the example file explicitly sets Normal limits. hard_max_agent_steps defaults to 200 and remains bounded by the system ceiling. Approval waits exclude active time; cancellation/timeouts retain observed statistics.

Repeated identical command results eventually prevent further execution. Stalls force replanning before termination; approaching the step budget requests convergence. History evicts complete oldest turns, preserving system instructions, the current task, and complete tool rounds. max_context_turns remains a hard bound.

Model/tool/log/display limits are independent; see configuration. Budgets and modes do not authorize changes or bypass approvals.