Configuration

Permissions

ub-agents adds no permission flags. The CLIs start in their default headless modes and usually cannot edit files or push. Grant each role what it needs with runtime-args.

implementer:
  runtime: "codex:gpt-6.1-sol:high"
  runtime-args: [--sandbox, danger-full-access]

Starter permissions

In a terminal, ub-agents init asks once whether to turn on the examples below for all four starter agents. Otherwise they stay commented out in ub-agents.yaml.

Codex

Full access, including commits in private worktrees. Codex's workspace-write sandbox cannot commit there, because their Git metadata lives in the main checkout.

runtime-args: [--sandbox, danger-full-access]

Claude Code

File edits plus the commands the role needs, with no prompts. Commands not listed in --allowedTools are denied, so add your project's check commands.

runtime-args: [--permission-mode, acceptEdits, --permission-prompts, none,
               --allowedTools, "Bash(git *)", "Bash(gh *)",
               "Bash(npm test)",
               "Bash({report_command} report *)", "Bash({report_command} read *)",
               --add-dir, "{scratch}"]

{report_command} becomes the launcher's own ub-agents command. A rule like Bash(ub-agents *) would not match it, because Claude matches the command text including its path. Agents with a retrospectives board also need the retrospective rule.

"Bash({report_command} report *)"
"Bash({report_command} read *)"
"Bash({report_command} retrospective *)"

Scratch

{scratch} becomes the run's private scratch directory, outside the worktree. Quote it, or YAML reads it as a mapping.

runtime-args: [--add-dir, "{scratch}"]

runtime-args apply to every runtime in a list. Use CLI-specific flags only on agents with a single runtime.

Check

ub-agents doctor warns about runtime agents without runtime-args. It does not test whether the arguments grant enough.

$ ub-agents doctor