mirror of
https://github.com/diegosouzapw/OmniRoute.git
synced 2026-08-20 22:22:57 +03:00
fd489c9594
Deep combo-flow audit (all combo types x 14 strategies x resilience x proxy), TDD-validated. Three production fixes: 1. chatCore executor-level emergency budget hop REMOVED: it executed the emergency provider's request with the FAILING provider's credentials (e.g. the OpenAI API key sent to integrate.api.nvidia.com) on every non-stream 402/quota-429 - a cross-provider credential leak that also never succeeded upstream (foreign keys are always rejected). Emergency fallback is now orchestrated exclusively by the routing layer (src/sse/handlers/chat.ts), which resolves credentials FOR the emergency provider via account selection. 2. chat.ts emergency hop gated with !comboName: combos are the operator's fallback policy (target-level orchestration + global fallback #689), and the per-target hop burned extra upstream calls against exhausted providers (#1731) and bypassed context-relay handoff injection on account switches (the switch happened inside the detour with comboStrategy=null, silently swapping the model for gpt-oss-120b). 3. /v1/messages/count_tokens now resolves the connection proxy and wraps the provider-side count in runWithProxyContext - it previously ran with no proxy context at all, going DIRECT regardless of configured proxies. safeResolveProxy failures are now logged at warn (silent proxy bypass was debug-only). Tests: restores the five #1731 fast-skip integration tests skipped since v3.8.2 (rewritten against the current routing policy, 8/8), adds combo-strategy-fallbacks.test.ts (11 tests: fill-first/p2c/random/ cost-optimized/strict-random fallback paths, HALF_OPEN recovery, strategy normalization), adds proxy-context-passthrough.test.ts (per-target combo proxy independence + count_tokens proxy), and realigns the three tests that asserted the removed leaking behavior.