Hollow

Adaptive Model Routing

Static complexity→model tiers ignore what actually happens when you send tasks. This release adds observation-driven routing that learns from real performance data.

What's new

agents/adaptive_router.py — AdaptiveRouter with:

• observe(model, complexity, duration_ms, tokens_out, success) — EMA update per (model, complexity) pair
• score(model, complexity) — composite score: 0.5×success_rate + 0.3×throughput + 0.2×latency_inv
• recommend(complexity, candidates) — highest-scoring candidate with ≥MIN_OBSERVATIONS (5)
• add_override() / remove_override() / resolve_override() — hard routing rules by agent/role/complexity
• EMA α=0.15 (≈13 observations to weight recent ≥ old). Stats persisted to disk.

Routing hierarchy:

1. Hard override (admin-set, bypasses scoring)
2. Adaptive score (≥5 observations for this complexity)
3. VRAM affinity (v0.9.0 — prefer loaded model)
4. Static tier default

Integration: scheduler calls observe() directly after every _run_task and _run_task_streaming completion (success and failure paths). Override specificity: agent_id > role > complexity-only > global.

5 new API routes: /routing/stats, /routing/recommend/{c}, /routing/overrides, POST /routing/override, DELETE /routing/override/{id}

4 new MCP tools: routing_stats, routing_recommend, routing_override, routing_overrides_list

7 integration tests — structural tests + 1 Ollama-dependent test verifying stats populate after task completion.

Design notes

Execution of 'adaptive' implies time series. EMA with α=0.15 weights the last ~13 observations equally to all history — responsive to recent drift without overfitting to single outliers. MIN_OBSERVATIONS=5 prevents cold-start noise from corrupting routing decisions before enough data exists.

Requires: scheduler (v0.9.0) as observation source.