CLI Quickstart¶

Get one agent to a verified result before adding orchestration. This guide uses the codex alias; substitute claude if that is the CLI you authenticated.

1. Run one agent¶

From a directory you want the agent to inspect:

li agent codex "Describe this directory in one concise paragraph." --cwd .

Success has two observable signals: a non-empty response prints to stdout, and LionAGI prints a [to resume] command containing the saved branch ID. --cwd sets the working directory for the CLI-backed provider.

If the provider fails before responding, run codex --version and codex login directly, then rerun li doctor.

2. Continue the conversation¶

Continue the most recently used branch:

li agent -c "Turn that into three bullets."

Or use the explicit branch ID from the resume hint:

li agent -r <branch-id> "Name one file I should read first and explain why."

For bounded work, set a hard wall-clock deadline:

li agent -c "Finish with the most important caveat." --timeout 120

LionAGI adds the deadline to the agent's prompt and terminates the run when the limit is reached. Add --resume-on-timeout when you want one automatic resume attempt after an agent timeout:

li agent -c "Complete the review." --timeout 120 --resume-on-timeout

3. Fan out independent work¶

Use fan-out when several workers can answer independently:

li o fanout codex \
  "Review this repository from correctness and maintainability perspectives." \
  --cwd . -n 2 --with-synthesis --save ./lion-results/fanout

The saved directory contains worker_1.md, worker_2.md, and, because --with-synthesis is set, synthesis.md. The command also records run state under ~/.lionagi/runs/.

4. Preview a dependency-aware flow¶

A flow asks an orchestrator to plan work with dependency edges. Preview the plan before spending worker turns:

li o flow codex \
  "Inspect this repository, identify one documentation gap, then propose a fix." \
  --cwd . --max-ops 4 --dry-run

--dry-run prints the planned agents, operations, dependencies, and model resolution without executing the worker graph. The exact plan is model-driven, so its node names and count may vary within the --max-ops cap.

Execute the same task after the preview looks appropriate:

li o flow codex \
  "Inspect this repository, identify one documentation gap, then propose a fix." \
  --cwd . --max-ops 4 --save ./lion-results/flow

Flow artifacts are written below the --save directory. The durable manifest, branch snapshots, stream buffers, and checkpoint remain under the run's ~/.lionagi/runs/<run-id>/ directory.

What you have now¶

• A resumable single-agent branch.
• Parallel worker artifacts from fan-out.
• A previewed and executed dependency-aware flow.
• Durable local run state for inspection and recovery.

Next, learn the full agent → fan-out → flow → playbook progression, or see how to monitor and recover durable work.