OmniRoute/_ideia/_fetch/1808.json

{"author":{"id":"MDQ6VXNlcjE0NDAzMDM=","is_bot":false,"login":"matteoantoci","name":"Matteo Antoci"},"body":"### Problem\n\nWhen auto-combo scores and selects candidates, it doesn't consider whether each model's context window can fit the request. A large request can be routed to a model with a small context window, which immediately fails and wastes a retry slot before falling back.\n\n### Details\n\n- The auto strategy already filters candidates by tool-calling support (when the request includes tools), falling back to the full pool if all are filtered out\n- Token estimation (`estimateTokens()` in `contextManager.ts`) and context limit lookup (`getModelContextLimit()` in `modelCapabilities.ts`) already exist but are only used during context compression after a model is chosen\n- The `estimatedInputTokens` field exists in `routerStrategy.ts` but is never populated\n- When this happens, the user sees extra latency from the failed attempt plus the fallback retry\n\n### Reproduction\n\n1. Configure a combo with models that have different context window sizes (e.g., 16K and 128K)\n2. Send a request with a large prompt that exceeds the smaller model's context window\n3. Observe: the request is routed to the small model, fails with \"input too long\", then falls back to the larger model\n4. Expected: the small model is excluded from candidates before scoring runs\n\n### Environment\n\nObserved on v3.7.5. The context window information is available but not used during auto-combo candidate selection.","comments":[],"createdAt":"2026-04-30T09:20:52Z","labels":[],"number":1808,"title":"[Feature] Auto-combo can route to models with insufficient context window","url":"https://github.com/diegosouzapw/OmniRoute/issues/1808"}