manage-foreman-agentslisted

# Manage Foreman Use Claude or Codex as an implementation worker and keep the managing session responsible for scope, approvals, and final review. ## Delegate work 1. Call `doctor` before the first task in a session. Require `auth_mode: subscription` and confirm the chosen provider is ready; never request or configure an Anthropic or OpenAI API key. 2. Create a goal for multi-task outcomes. For one atomic task, call `task_create` directly. 3. Supply an absolute Git repository path, a concrete completion condition, provider, model, effort, and relevant verification commands in the prompt. Pass routine commands through `verification_commands` so Foreman executes them independently and records structured gates. Do not encode shell operators, network access, deployment, or destructive actions as verification gates. Use `task_configure` to retune a queued task; never change model or effort after it has been claimed. 4. Keep unrelated tasks in separate worktrees. Use dependencies when one task needs an accepted result from another. 5. Start the daemon only when queued work should run. For Claude, use `sonnet` with `medium` effort for routine implementation and `opus` only for unusually difficult work. For Codex, choose `gpt-5.6-terra` for balanced everyday work, `gpt-5.6-sol` for the hardest work, or `gpt-5.6-luna` for the fastest iteration. Sol and Terra support through `ultra`; Luna supports through `max`. Preserve an explicit user choice. ## Monitor and unblock Use c