dev/docs/auto-drive.md

Auto Drive

What Auto Drive is, how to start it, and how it behaves in Every Code.

Start points

• TUI: /auto <goal>. If you omit the goal and there is recent history, Code proposes one for you. /auto settings jumps straight to the Auto Drive pane.
• CLI: code exec --auto "<goal>" or code exec "/auto <goal>". A goal is required when launching headless.
• Precondition: Full Auto mode (danger-full-access + approval=never) must be selected in the TUI; otherwise you’ll see a warning and Auto Drive will not start.

Goal handling

• Images passed on the CLI are attached before the first turn.
• If you supply no goal and there is nothing in history to derive from, Auto Drive will not start.

How it runs

• Each turn Auto Drive drafts a plan, prepares commands, optionally assigns agents, and waits for your confirmation (or the countdown) before running.
• The transcript is kept in memory and compacted automatically; you’ll see a notice if history was trimmed.
• If an AUTO_AGENTS.md exists, its guidance is applied to the run alongside any AGENTS.md rules.

Agents

• Auto Drive can spawn helper agents during a turn. Toggle this with agents_enabled in Auto Drive settings.
• Outside a git repo, Auto Drive forces those agents to run read-only to avoid unintended writes.

Observer

• A lightweight watchdog reviews the run every auto_drive_observer_cadence turns (default 5). If it spots trouble it surfaces guidance in the banner. Set the cadence to 0 to disable.

Sandbox and approvals

• TUI: needs danger-full-access plus approval_policy=never so turns aren’t blocked by prompts.
• CLI: --auto forces approvals off; add --full-auto (or --dangerously-bypass-approvals-and-sandbox) if you want Auto Drive to make edits and run commands.

Continue and countdown modes

• continue_mode: immediate, ten-seconds (default), sixty-seconds, manual.
• In countdown modes, the Auto Drive card shows a timer; Enter proceeds early, Esc reopens the draft, 0 auto-submits.
• Manual mode pauses after each prepared prompt until you confirm.

Stop and pause

• Press Esc while Auto Drive is active to pause or stop (context-dependent). Countdown modes show this hint in the footer.
• Approval dialogs never capture Esc; it always reaches Auto Drive.

Review, QA, diagnostics

• review_enabled (default true) can insert a review gate; the card shows “Awaiting review.”
• qa_automation_enabled and cross_check_enabled (default true) allow diagnostics and cross-check turns before continuing.
• auto_resolve_review_attempts limits how many times Auto Drive will auto-resolve review feedback (default 5).

Crash diagnostics

• Set CODEX_TUI_RECORD_SESSION=1 to persist a per-session JSONL trace of inbound TUI events, including Auto Drive coordinator decisions, countdowns, token metrics, and panic records.
• Optionally set CODEX_TUI_SESSION_LOG_PATH=/absolute/path/session.jsonl to choose the output file; otherwise Code writes under its normal log directory (typically ~/.code/debug_logs/, and code-dev defaults to ~/.code/debug_logs/code-dev/).
• When the TUI panics, Code now records the panic payload and captured backtrace in both the session JSONL log and the regular error log.

Models

• Defaults: model gpt-5.1, reasoning effort high.
• Toggle “use chat model” in settings to reuse your current chat model/effort instead of the dedicated Auto Drive model.

UI surfaces

• Auto Drive card shows status (Ready, Waiting, Thinking, Running, Awaiting review, Failed/Stopped), the goal, action log, token/time counters, countdown, and a celebration on success.
• Bottom pane header mirrors status and shows hints (Ctrl+S settings, Esc stop, whether agents/diagnostics are on).

Resume and persistence

• History is kept in memory; there’s no Auto Drive–specific history file. If trimmed, you’ll see a note.
• You can resume a session as usual; Auto Drive can derive a goal from restored history.
• CLI --output-last-message still works here if you only need the final reply.

Settings (config.toml)

• Top-level keys: auto_drive_use_chat_model (default false), auto_drive_observer_cadence (default 5).
• [auto_drive] defaults: review_enabled=true, agents_enabled=true, qa_automation_enabled=true, cross_check_enabled=true, observer_enabled=true, coordinator_routing=true, continue_mode="ten-seconds", model="gpt-5.1", model_reasoning_effort="high", auto_resolve_review_attempts=5.
• All of these can be changed from /auto settings in the TUI or directly in config.toml.

Tips

• Stay in the TUI if you want countdowns and visual status; use code exec --auto for CI or scripted flows.
• If Auto Drive stops because it couldn’t derive a goal, rerun /auto <goal> with a short, specific instruction.
• Turn off agents in /auto settings if you want a single-model run.