feat(ai): make omniroute primary reviewer model, qwen-3.6-fast fallback

.github/workflows/agent-pr-review.yaml

@@ -38,14 +38,16 @@ jobs:
           github_token: ${{ github.token }}
           ai_base_url: http://litellm.ai.svc.cluster.local/v1
           ai_api_key: ${{ secrets.LITELLM_API_KEY }}
-          ai_model: qwen-3.6-fast # thinking-off litellm twin (see litellm configmap)
-          # Empty omits the field (action default sends 0.1); SGLang runs sampling_defaults=model,
-          # so the model's generation_config applies (temp 1.0, top_k 20, top_p 0.95) — Qwen's
-          # shipped sampling; near-greedy decoding is the documented repetition-loop trigger.
-          ai_temperature: ""
+          # omniroute routes to a larger free-tier model than the self-hosted 27B
+          # (see litellm routerSettings.fallbacks); falls back to qwen-3.6-fast
+          # automatically if the free provider fails or times out.
+          ai_model: omniroute
+          # omniroute's proxied models don't share SGLang's sampling_defaults, so
+          # an empty temperature has no known-good fallback here — pin explicitly.
+          ai_temperature: "0.2"
           ai_response_format: json_schema
-          # This GPU does ~10 tok/s with ~155-220s cold prefills near ~100K ctx (validated), so the
-          # action's 300s default per-request timeout is too tight. Raise it.
+          # Covers both omniroute (observed 2-74s) and, on fallback, qwen-3.6-fast's
+          # ~10 tok/s cold prefill on this GPU (155-220s near ~100K ctx, validated).
           ai_request_timeout_sec: "600"
           verdict_policy: findings_severity_gated
           inline_findings: "true"

kubernetes/apps/ai/litellm/instance/models.yaml

@@ -64,7 +64,8 @@ spec:
   proxyRef: litellm
   params:
     # "auto" is OmniRoute's zero-config smart-routing model id — it builds a
-    # virtual combo from whichever free/keyed providers are connected.
+    # virtual combo from whichever free/keyed providers are connected and
+    # picks the best-scoring one live per request (quota/health/cost/taskFit).
     model: openai/auto
     apiBase: http://omniroute.ai.svc.cluster.local:20128/v1
     # REQUIRE_API_KEY=false internally; LiteLLM's OpenAI provider still needs

kubernetes/apps/ai/litellm/instance/proxy.yaml

@@ -54,6 +54,12 @@ spec:
     # already burned 900s only deepens the contention that caused it.
     retry_policy:
       InternalServerErrorRetries: 2
+    # omniroute's free providers have no SLA and can fail/time out without
+    # warning; fall back to the self-hosted (slower but always-available)
+    # qwen-3.6-fast alias so a flaky free provider doesn't take out the PR
+    # review action.
+    fallbacks:
+      - omniroute: ["qwen-3.6-fast"]
   route:
     hostnames:
       - "litellm.${SECRET_DOMAIN}"