Approvals & Run Budgets

term2 balances velocity with security by giving you fine-grained control over which operations execute automatically and which require interactive human approval.

Auto-Approval Modes (/auto-approve)

Configure shell auto-approval using /auto-approve or shell.autoApproveMode. The CLI --auto-approve flag is separate and applies only to non-interactive prompts.

Mode	Command	Behavior	Safety Guarantee
off (Default)	/auto-approve off	Shell commands prompt for interactive confirmation. Valid local workspace edits can apply without a prompt; file tools have separate boundary and authorization checks.	Shell confirmation under the default interactive policy.
advisory	/auto-approve advisory	Prompts for confirmation on every command, accompanied by a real-time safety explanation.	High safety with consequence analysis.
auto	/auto-approve auto	Opt-in smart approval mode. Automatically approves safe read-only operations and workspace file modifications. Risky commands (e.g. rm -rf, git reset --hard, network requests, system path edits) are strictly gated for confirmation.	Balanced velocity. Eliminates repetitive prompts on benign actions while intercepting destructive commands.
always	/auto-approve always	Completely unattended execution mode. Every tool runs without prompts except ask_user. Selecting always automatically disables the shell sandbox (sandbox.enabled = false).	For isolated, disposable environments only.

Sandbox Interaction: always and sandbox.enabled are mutually exclusive: selecting always disables the shell sandbox, and enabling the sandbox (/sandbox or sandbox.enabled = true) immediately demotes always to auto.

Run Budgets (agent.runBudget)

Run budgets track spending, unpriced tokens, active time, and stall evidence. The default escalation is warn: these thresholds are advisory, not hard spending or execution caps.

• Spending Cap (agent.runBudget.maxUsdMicros): Total priced API cost ceiling for a run (default: $5.00 / 5,000,000 micros).
• Token Cap (agent.runBudget.maxUnpricedTokens): Unpriced token ceiling (default: 5,000,000 tokens).
• Active Time Cap (agent.runBudget.maxActiveTimeMs): Active execution time ceiling, excluding user idle wait time (default: 1 hour / 3,600,000 ms).
• Turn Backstop (agent.runBudget.turnBackstop): High turn-count safeguard to detect looping subagents (default: 150 turns).
• Budget Escalation (agent.runBudget.escalation): warn (default) reports advisory evidence. contain pauses at exhaustion while warning and stall evidence remain advisory. pause also stops on warning/stall evidence. disabled emits no budget evidence.

In non-interactive runs, a budget check-in exits with code 2 and a resume hint. --auto-approve approves tools only; it never grants budget extensions. With --json, the pause is reported as run_budget_paused.