Headless, Streaming, and Auto-Mode

Session Management

Claude Code conversations persist as sessions that can be resumed later, rather than being lost when a terminal closes. /clear resets the conversation context to keep it focused -- useful when accumulated history has stopped being relevant to the current task (a manual antidote to context rot).

The Three Modes

These three axes are independent: a run can be headless and streaming at once (a CI job consuming stream-JSON output), and auto-mode is an autonomy setting orthogonal to whether the session is interactive.

Auto-mode is powerful precisely because it removes confirmation checkpoints -- which is why it should be paired with settings.json permission rules and hooks, not used unguarded. Removing human-in-the-loop confirmation raises the stakes of any single action Claude takes, so the deterministic controls that would otherwise be backed up by a confirmation prompt need to be enforced structurally instead.

The Best-Practice Loop

Anthropic's operational guidance is to work in a gather context -> make a plan -> act -> verify loop: understand the task and codebase, form a plan, execute it, and use tests/builds as ground-truth verification rather than trusting the model's own claim of success. Tighten this loop over time with custom slash commands and hooks rather than accumulating long, brittle ad-hoc prompts.

Common exam traps

• Confusing headless mode with auto-mode. Headless mode is about *interactivity* (no TUI, suitable for CI/scripting); auto-mode is about *autonomy* (fewer confirmations before acting). An exam item may describe "running in a CI pipeline" (headless) as a distractor for a question actually asking about auto-mode, or vice versa.
• Assuming streaming mode implies headless, or that headless implies auto-mode -- the three are independent settings, not a single spectrum.
• Using auto-mode without corresponding settings.json permission rules/hooks, on the assumption that reduced confirmation is safe by default.