Fleet Doctrine — Model Routing

Aliases

• opus → anthropic/claude-opus-4-6
• sonnet → anthropic/claude-sonnet-4-6
• codex → openai-codex/gpt-5.3-codex
• Gemini: google/gemini-3-pro-preview
• Grok: xai/grok-4

Routing Rules

Opus — Commander

When: Main session, orchestration, security decisions, financial tasks, reviewing other models' output, anything high-stakes or ambiguous.

Never waste on: Routine crons, simple lookups, email summaries, templated tasks.

Codex — Chief Engineer

When: Big coding tasks (refactors, new features, full repo work), PR reviews, debugging complex issues, checking other models' code output.

Spawn as: sessions_spawn(model: "codex", task: "...") or sub-agent with --model codex.

Pairs with: Grok for parallel work or second opinions on code.

Sonnet — Workhorse

When: Cron jobs, email briefings, admissions reports, routine admin, quick lookups, drafts, form letters, anything repetitive or templated.

Default for: All crons unless the task requires reasoning.

Gemini — Creative & Vision

When: Image generation, analysing long documents (1M context), visual tasks, when a different perspective helps.

Best at: Processing massive context windows, multimodal input.

Grok — Fast Backup

When: Parallel work alongside Codex, speed-over-depth tasks, sanity-checking other models' output, when you need a quick second opinion.

Good for: Lightweight code reviews, fast research, draft generation.

Decision Flow

1. Is it security, finance, or high-stakes? → Opus
2. Is it a big coding task or repo work? → Codex (with Grok as backup)
3. Does it need image gen or 1M+ context? → Gemini
4. Is it routine/scheduled/templated? → Sonnet
5. Need a fast second opinion? → Grok
6. Not sure? → Opus decides, then delegates.

Fallback Rule

If a model is unavailable on your instance, fall back to your default model (typically Sonnet). The doctrine describes intent — use the best model you have access to that matches the category. When in doubt, do the task with what you've got rather than failing.

Anti-Patterns

• Don't use Opus for email summaries or cron jobs
• Don't use Sonnet for complex multi-step reasoning
• Don't use Codex for non-coding tasks
• Don't spawn multiple models on the same task unless deliberately seeking a second opinion