Delegate CLI

Run the target CLI's live help before invoking it. Give each delegated call one outcome, only the essential context, and a request for a concise response.

Claude Code

• Treat the Claude Pro allowance as scarce: make one fresh, non-interactive call per task, and avoid resumed sessions, exploratory calls, and follow-ups.
• Use the strongest available Opus model with high effort.

Codex

• Treat the Codex allowance as scarce: make one fresh non-interactive codex exec call per task, and avoid resumed sessions, exploratory calls, and follow-ups.
• Keep the model and reasoning-effort defaults from ~/.codex/config.toml; do not override them.
• For a second opinion, critique, or explanation, run with the read-only sandbox and forbid edits.

OpenCode

• Read the refreshed verbose model catalog, then use opencode-go/glm-5.3-flash. If it is unavailable, use the strongest GLM Flash model from the OpenCode Go provider.
• Use the high reasoning level, not the model's highest available level.
• For a second opinion, critique, or explanation, send an ask-mode prompt with the relevant context and exact question, and forbid edits.