Automatic model: the right model for each task

Each tier’s model and effort are set under Settings → Token savings → Automatic model, separately for Claude and Codex.

• Simple: copy, styling, renames, one-line fixes. Claude defaults to Haiku.
• Normal: an ordinary feature or bug fix. Claude defaults to Sonnet at medium effort. New agents start here.
• Hard: architecture, refactors, security, performance, tough debugging. Claude defaults to Opus at high effort.
• Codex keeps the CLI’s own model; only its effort changes between low, medium and high.

Clear cases are sorted by free rules: short requests like “typo”, “rename” or “padding” are simple; broad work like “refactor”, “migration”, “security” or “from scratch”, or requests over 2,500 characters with many bullets, are hard. When the rules aren’t sure, Claude Haiku is asked for a one-word rating; if it doesn’t answer in 6 seconds, the rule’s answer stands.

• Short follow-ups like “yes”, “ok” or “go on” keep the current tier.
• Phrases like “still broken”, “didn’t work”, “same error” or the Turkish “hâlâ çalışmıyor” move the agent up one tier.
• The pane reads “Otomatik · model”, and an activity line shows the tier and the reason on each switch.

When the model or effort changes, AgentVera restarts the CLI resuming the same conversation, then sends your message. Claude’s /model command isn’t used because it would also change your default model. Prompts typed straight into a Claude terminal are routed too: the prompt is held, the agent reopens on the new model and the prompt is resent within a few seconds. Prompts typed straight into a Codex terminal aren’t routed; messages sent from the app are routed for both.