ACP journeys check

Run the command

The command is not public yet.

Point the command at the command line that starts your agent in ACP mode:

./run.sh --agent-dir ./my-agent -- /agent/bin/my-agent --acp

The journeys run in a Docker container with no network. An agent that needs its model API gets a network with --network, and its login with --login or --env.

The command writes a result file. It has the name and version that your agent gives in its initialize answer, the date, each journey with its log, and a SHA-256 of the logs.

The journeys

Rondar has 11 user journeys for each agent runtime. The command runs the 7 journeys that a real agent with a real login can run:

1. A turn to its end (turn_to_its_end)
2. A follow-up in the same session (follow_up)
3. A follow-up after a restart (resume_after_restart)
4. An image and a file (an_image_and_a_file)
5. Stop a running turn (stop)
6. A Plan turn to a saved plan (plan_turn_to_a_saved_plan)
7. Plan, then regular, then Plan again (switch_between_plan_and_regular)

These journeys are not in the command:

• rondar_tools_are_reachable: It asks the CLI to call a tool outside the Plan tools. A real model decides for itself which tools it calls.
• identical_replies_end_both_turns: A real CLI cannot be told to print the same bytes two times.
• old_container_gets_what_the_turn_needs: It needs a container that an old release of Rondar made.
• a_reply_right_after_a_turn_is_delivered: It tests a race in Rondar's worker and delivery, not in the CLI.

What your ACP agent must do to pass gives each journey, what it sends and what it expects.

What Rondar checks

Rondar does not run your agent. It checks the result file that you send:

1. The file has every journey of the current version.
2. The SHA-256 of the logs matches the logs.
3. Rondar runs the checks of each journey on its log again. A journey passes only when its log shows it.
4. A person at Rondar reads the result before it is public.

A result page shows one run of one agent on its date. Rondar does not score agents and does not compare them.

Send a result

The result is private until a person at Rondar approves it. Then it has a public page with each log, and a badge.

Send a result