Non-Interactive Mode

When you provide a prompt argument on the command line, term2 executes non-interactively without launching the terminal UI:

term2 [options] [prompt...]

Basic Non-Interactive Queries

Execute one-off queries with stdout output:

term2 "Explain the purpose of package.json in this project"

A new positional-prompt session without --auto-approve uses builtin:bare by default (unless --profile is specified or a persisted Mentor or Orchestrator profile takes precedence). Tool execution is disabled without --auto-approve, independently of the selected profile. --auto-approve opts out of that implicit Bare selection and enables tools. It applies GREEN/YELLOW heuristic policy; it is not the same as /auto-approve always, and RED shell commands are still refused.

The output streams directly to standard output. Non-error diagnostic logs are printed to standard error.

Enabling Tool Execution (--auto-approve)

To permit the agent to run tools (such as reading files, writing changes, or executing commands), pass --auto-approve:

term2 --auto-approve "Run pnpm typecheck and fix any trivial lint errors"

Output Formatting & Scripting

Quiet Mode (-q, --quiet)

Suppress non-error diagnostics on stderr to cleanly capture tool output in shell variables:

SUMMARY=$(term2 -q "Explain what a TODO comment is")

JSON Mode (--json)

Emit newline-delimited JSON (NDJSON) events on standard output for machine consumption:

term2 --json "Explain common authentication security risks"

The conversation event stream can emit these type values: approval_required, background_check_in_due, background_shell_completed, background_shell_output, background_shell_started, codex_rate_limits, command_message, context_compaction_completed, context_compaction_failed, context_compaction_started, cost_update, error, final, reasoning_delta, retry, retry_exhausted, run_budget, subagent_approval_required, subagent_command_message, subagent_completed, subagent_interrupted, subagent_question, subagent_retry, subagent_run_budget, subagent_started, subagent_streaming_text, subagent_streaming_tool, subagent_text_turn, subagent_tool_started, subagent_transferred, subagent_usage_update, text_delta, tool_call_streaming_delta, tool_dispatched, tool_recovery, tool_started, usage_update, and user_message_consumed_for_abort. Non-interactive JSON also emits approval_rejected for a refusal and completed after success.

Streaming Reasoning Deltas (--show-reasoning)

Stream thinking/reasoning deltas from reasoning-capable models (e.g. o1, o3, Grok, Claude with thinking) to standard error:

term2 --show-reasoning "Solve this mathematical recurrence relation"

Minimal-Prompt Execution (--lite)

--lite (or -l) is a compatibility alias for --profile bare. Bare uses a minimal prompt with project instructions included; it retains codebase tools and exposes selected capabilities through run_code. Tools still require --auto-approve in non-interactive mode:

term2 --lite "Explain the difference between disk usage and apparent file size"

Pauses and Exit Codes

• 0: Successful completion.
• 1: Failure. A tool refusal can still be followed by successful completion; inspect stderr or JSON approval_rejected rather than relying only on the exit code.
• 2: A recoverable context-limit or run-budget pause. Completed work and session history are retained; use the printed --resume <session> hint after reviewing the limit.

With --json, context-limit pauses emit context_paused with a code of request_input_limit, context_compaction_hard_fit, or provider_context_overflow. Budget check-ins emit run_budget_paused, including budget evidence and recovery guidance. --auto-approve never grants more run budget. The default budget escalation, warn, remains advisory.