[FEATURE] Sub-agent model selector with cost/benefit guidance

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Problem

When launching sub-agents via the Agent tool, there is no guidance
on which model to use. The default behavior leads users to use
Sonnet or Opus for all sub-agents, including simple extraction
tasks that could run on Haiku at 20x lower cost.

## Proposed solution
Add a model selection recommendation system when launching sub-agents:

• Show a simple decision guide based on task type
• Default to Haiku for extraction/search/grep tasks
• Default to Sonnet for code generation tasks
• Suggest Opus only for complex synthesis/architecture tasks

## Example decision matrix that could be built-in:
| Task type | Recommended model |
|----------------------------------|-------------------|
| Read files, grep, extract data | Haiku |
| Write/fix localized code | Sonnet |
| Synthesize large documents | Opus |

## Real-world impact
In a session with ~25 sub-agents for a codebase audit, applying
this rule would have reduced token consumption by ~40-50%,
directly extending the usable session length.

## Current workaround
Users can add this rule manually to CLAUDE.md, but it should be
built-in default behavior.

Proposed Solution

Choix du modele pour les sous-agents (OBLIGATOIRE)

Quand tu lances des sous-agents via le tool Agent, tu DOIS choisir le modele adapte a la tache.
Ne jamais utiliser Opus par defaut — toujours raisonner en ratio benefice/tokens.

Regle de decision :

| Tache | Modele | Pourquoi |
|-------|--------|----------|
| Lire des fichiers, extraire des donnees, grep, recherche | haiku | Rapide, pas besoin de raisonnement |
| Reformater, convertir, lister, inventorier | haiku | Mecanique, pas de jugement requis |
| Corriger un bug precis, ecrire du code localise | sonnet | Competence technique suffisante |
| Auditer un fichier, analyser du code | sonnet | Bon rapport qualite/cout |
| Synthesiser de grandes quantites de donnees | sonnet | Capacite de synthese adequate |
| Architecture, decisions complexes, reorganisation globale | opus | Jugement superieur justifie le cout |
| Redaction d'un document de reference complet | opus | Coherence et qualite sur la longueur |

Principe general :

• Les taches d'extraction/recherche → toujours Haiku (20x moins cher qu'Opus)
• Les taches de production de code → Sonnet par defaut
• Monter a Opus seulement si la tache requiert du jugement architectural ou une synthese complexe

Exemple concret :
Si on me demande d'auditer 10 fichiers et de produire un rapport :

• 10 agents Haiku lisent les fichiers en parallele et extraient les donnees brutes
• 1 agent Sonnet (ou Opus si > 50 fichiers) synthetise le rapport final

→ Resultat identique, cout reduit de 80%

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

_No response_

Additional Context

_No response_