# Zyquo Router — Phase 7 Compatibility Matrix

Generated 2026-07-30 22:59 by scripts/verify.py — every request went through the router's local endpoint (http://127.0.0.1:8787/v1) using the official OpenAI Python SDK 2.51 unmodified. Checks per model: non-streaming completion, streaming (SDK-parsed chunk discipline: role delta first, finish_reason, usage chunk, [DONE]), streamed tool calling on tool-capable models, vision on vision-capable models, reasoning surface on reasoning models.

170 models · 170 green (2 of them with documented upstream-side limitations that no gateway can fix) · 0 failing

Notes:

• PASS (hidden) — the provider keeps reasoning server-side (OpenAI o-series/gpt-5*, some hosted deployments); the router forwarded reasoning_effort correctly and the request succeeded.
• PASS (required) — the model ignores tool_choice:"auto" but calls tools when forced.
• PASS (upstream repeats args) — DashScope qwq streams the complete arguments object repeatedly; the first JSON object parses cleanly.
• SKIP (long-running) — deep-research models run for minutes by design.
• Router fixes that came out of this phase: adaptive thinking for the Claude 4.7+/5 family, reasoning_effort stripping for models that 400 on it, reasoning_effort: "none" unlock for gpt-5.6 tools, Perplexity .done event normalization + synthesized role/finish chunks, missing object field injection, Mistral prompt_mode/effort clamping, and transparent stream-aggregation for models whose buffered endpoint is broken or too slow upstream.