Route Ox Alpha traffic with overflow from OpenRouter to Nous and Venice #56

Project

Stress testing Ox Alpha

Outcome

Keep every harness lane saturated for the free window by failing over between providers that serve the same stealth/ox-alpha slug.

Lanes

1. OpenRouter (primary). stealth/ox-alpha, free, single stealth provider, observed 18-26 TPS.
2. Nous Portal API. Verified live at https://inference-api.nousresearch.com/v1 with the same stealth/ox-alpha slug and $0/$0 pricing. OpenAI-compatible chat completions; reasoning is mandatory with default effort max (max, high, low supported); tools supported; context 1,048,576; max output 131,072. Rate limits follow the Portal tier: Free 50 RPM / 500K TPM, Plus 400 RPM / 4M TPM, Super 800 RPM / 8M TPM, Ultra 1600 RPM / 16M TPM.
3. Venice API. Unverified for Ox Alpha. Probe /models before wiring; drop the lane if absent.

Scope

• One provider-adapter interface behind the OpenAI-compatible request shape so lanes differ only by base URL, key, and budget.
• Health tracking per lane: 429 rate, p95 latency, tokens per second. Overflow triggers on sustained 429s or latency SLO breach, not single errors.
• Per-lane budgets derived from measured limits in the rate-limit issue; a lane that exhausts its budget stops receiving dispatches until reset.
• Tag every completion with the lane that served it so throughput reports can attribute usage.

Non-goals

• No multi-account rotation against one provider's terms of service.
• No paid x402 fallback during the free window.