Parent epic: #2141

Phase 3: Agent loop integration

Tasks

• Insert triage check in process_user_message_inner() after prepare_context(), before process_response()
• Pattern: similar to guardrail check (mod.rs:3118) — async pre-call gate
• Pass context token count from ContextBudget to TriageRouter::assess()
• On tier decision: swap provider via existing provider_override: Arc<RwLock<Option<AnyProvider>>> slot
• Context window hard override: if accumulated tokens > target provider's context_window() × 0.8, auto-escalate
• Status message via channel.send_status("evaluating complexity...")
• Fallback: if triage call fails → use default provider, log warning
• Reset provider_override after response (so next turn gets fresh triage)

Key files

• crates/zeph-core/src/agent/mod.rs (process_user_message_inner)
• crates/zeph-core/src/agent/builder.rs (triage provider injection)

Acceptance criteria

• Simple prompt ("hello") routes to cheap provider in multi-provider setup
• Complex prompt routes to expensive provider
• Triage provider failure falls back to default without error to user
• Context overflow auto-escalates to provider with larger window