Configuration

Runtimes

A runtime is the agent CLI that does the work. Claude Code and Codex are built in. Anything else can run as a command.

runtime: "claude:claude-opus-5-5:high"

CLI, model and effort

Written as cli:model:effort. The CLI is claude or codex. opencode, for local models through Ollama, is planned and not supported yet.

runtime: "claude:claude-opus-5-5:high"
runtime: "codex:gpt-6.1-sol:high"
runtime: "opencode:ollama/qwen3.8:high"   # planned, not supported yet

Alternatives

Give a list. The first one that is installed and allowed runs.

runtime: ["claude:claude-opus-5-5:high", "codex:gpt-6.1-sol:high"]

What runs

The prompt, made of the assignment context and the agent's instructions, arrives on stdin. Every run starts a fresh session.

$ claude --print --output-format stream-json --verbose --model MODEL --effort EFFORT
$ codex exec --json --model MODEL --config model_reasoning_effort="EFFORT"

Usage limits

Nothing to configure. When a run reports a usage limit, the launcher stops starting runs on that CLI until the limit resets, plus one minute. Other CLIs keep working, and the item spends no attempt. Restarting the launcher clears the pause.

claude usage limit reached; pausing claude runs until 2026-10-04T12:01:00Z

Not allowed

runtime-args must not change the model or effort, resume a session, set Claude's --output-format or Codex's --ephemeral. ub-agents check rejects them.