Specifying and Running Harnesses and Loops in Claude Code, Codex, and Antigravity CLI

Repeat a prompt on a fixed interval while the session stays open. Omit the interval and Claude self-paces between iterations. Omit the prompt and Claude runs an autonomous maintenance check or the prompt in .claude/loop.md (where available). Stops when the session ends or you press Esc.

Polls every 5 minutes. Claude reports each iteration without requiring manual re-prompting.

Register a session-level completion condition. Claude keeps working across turns until a separate evaluator model (defaults to Haiku) reads the transcript and confirms the condition is met. A live overlay shows elapsed time, turn count, and token usage. Clear with /goal clear.

Claude iterates until the evaluator confirms both conditions. No "keep going" prompts needed.

Decompose a large codebase change into 5–30 independent units and run each in a parallel background subagent in an isolated git worktree. Each subagent implements its unit, runs tests, and opens a pull request. Requires a git repository.

Claude decomposes by component, spawns one agent per unit, and opens a PR per unit.

Detach the current session to run as a background agent and free the terminal. Monitor with claude agents. Pass a prompt to send one final instruction before detaching. Alias: /bg.

Terminal is freed immediately. Attach later with claude agents to review output.

Create, update, list, or run routines that execute on Anthropic-managed cloud infrastructure (Managed Agents). Claude walks through setup conversationally: cron expression, timezone, and vault environment variables. Each scheduled fire starts a new agent session. Alias: /routines.

Runs on Anthropic servers without an open local session. Pairs with vault environment variables for secret injection.

Summarize the conversation so far to free context window space without losing continuity. Optionally pass focus instructions to weight the summary. Use this mid-loop before context pressure causes degraded output rather than waiting for it to fail.

Manage allow, ask, and deny rules for tool permissions. Controls which shell commands, file edits, and MCP tool calls Claude can run autonomously vs. requiring approval. Rules persist in .claude/settings.json.

The Ralph loop was a Gemini CLI extension (/ralph-loop) that ran an agent in a self-correcting cycle: act, verify via hook, feed result back, repeat up to a configurable iteration limit. It is not available in Claude Code or Antigravity CLI.

In Claude Code, the equivalent pattern is /goal with a verifiable exit condition, or a custom slash command in .claude/commands/ that embeds assess-act-verify logic. In Antigravity CLI, replicate it with hooks and subagents.

Set a persistent objective that Codex works toward until it finishes the task, pauses, or needs input. Use /plan first to shape the goal before committing to it. Progress shows above the composer; pause, resume, edit, or clear from there.

Start code review mode. A separate Codex reviewer agent inspects uncommitted changes or a base branch diff and returns prioritized findings. Use /diff to inspect exact changes before or after review.

Adjust what Codex can do without asking for approval in the current session. Use /approve to retry a recently denied action once under the current session policy.

CLI flag controlling when Codex pauses for human confirmation before running a command. Set at invocation; also configurable in config.toml as the default approval policy.

Run a single prompt non-interactively. The primary mechanism for scripting Codex in CI pipelines and external orchestrators. Combine with a system scheduler (cron, GitHub Actions, etc.) to build interval loops outside a session.

Codex spawns subagents only when explicitly asked. Each subagent does its own model and tool work; subagent workflows consume more tokens than single-agent runs. Use /agent to switch between active agent threads, and configure roles and parallelism under [agents] in config.toml.

Set a persistent objective that Antigravity works toward until the task is complete, paused, or needs input. Use /planning first to shape complex goals before committing. Progress shows above the composer. Pair with /schedule for recurrent goal execution.

Enable multi-turn plan mode for complex tasks. The agent produces a structured plan before executing, equivalent to Claude Code's /plan. Use before setting a /goal on work that spans many files or subagents.

Open the Agent Manager panel to review subagent approvals and activity. Antigravity CLI is multi-agent by default; subagents are dispatched automatically based on task complexity. Use Alt+J to switch focus to the next subagent awaiting confirmation, and Ctrl+K to approve from the fast-path alert.

Open the Task Manager panel to view shell execution logs from the current session. Use this to inspect what tool calls subagents have made and verify gate conditions (tests, linting, security scans) without switching terminals.

Switch the permission preset controlling which tool calls require approval. Four presets are available. Changes persist to settings.json. Add, edit, or remove granular rules directly from this panel (added v1.0.5).

Run a single prompt non-interactively and print the response. The primary mechanism for scripting Antigravity CLI in CI pipelines and external schedulers. Combine with cron or a scheduler for interval-based loops outside a session. Pair with --model to select a tier appropriate to the task.