From 50ec6aac499356fdd23608da55cc843e1284ed53 Mon Sep 17 00:00:00 2001
From: Anthony Umeh <anthonyumeh@gmail.com>
Date: Sat, 22 Aug 2026 20:17:47 -0500
Subject: [PATCH] fix(runtime-host-bridge): strip provider prefix from
 admission-probe model id

Catalog model ids are provider-prefixed (deepseek/deepseek-v4-flash), but
OpenAI-compatible vendors expect the bare upstream id (deepseek-v4-flash)
on the wire. The chat-completions admission probe sent the prefixed form,
so vendors returned 400 which was misclassified as vendor-down, leaving the
endpoint permanently provider-unavailable and ineligible for benchmarking.

Strip the leading provider segment before probing, mirroring
resolveProviderLocalModelId in the OpenAI execution adapter.
---
 .../src/remote-health-probe.ts                   | 16 +++++++++++++++-
 .../test/remote-endpoint-admission-probe.test.ts |  2 +-
 2 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/role-model-router/apps/runtime-host-bridge/src/remote-health-probe.ts b/role-model-router/apps/runtime-host-bridge/src/remote-health-probe.ts
index 3e977b8f..d49e7b74 100644
--- a/role-model-router/apps/runtime-host-bridge/src/remote-health-probe.ts
+++ b/role-model-router/apps/runtime-host-bridge/src/remote-health-probe.ts
@@ -138,6 +138,20 @@ export function buildChatCompletionsProbeUrl(apiBase: string): string {
   return trimmed.endsWith("/v1") ? `${trimmed}/chat/completions` : `${trimmed}/v1/chat/completions`;
 }
 
+/**
+ * Resolve the wire model id used by the chat-completions admission probe.
+ *
+ * Catalog model ids are provider-prefixed (e.g. ``deepseek/deepseek-v4-flash``),
+ * but OpenAI-compatible vendors expect the bare upstream id on the wire
+ * (``deepseek-v4-flash``). Sending the prefixed form makes the vendor reject
+ * the probe with a 400 that we previously misclassified as ``vendor-down``.
+ * Mirrors ``resolveProviderLocalModelId`` in the OpenAI execution adapter.
+ */
+function resolveProbeModelId(modelId: string): string {
+  const trimmed = modelId.trim();
+  return trimmed.includes("/") ? trimmed.split("/").slice(1).join("/") : trimmed;
+}
+
 function isTimeoutError(error: unknown): boolean {
   return (
     error instanceof Error &&
@@ -337,7 +351,7 @@ export async function probeRemoteEndpointAdmission(
 
   const probeUrl = buildChatCompletionsProbeUrl(context.apiBase);
   const body = {
-    model: context.modelId,
+    model: resolveProbeModelId(context.modelId),
     messages: [{ role: "user", content: "role-model admission readiness probe" }],
     max_tokens: 1,
     stream: false,
diff --git a/role-model-router/apps/runtime-host-bridge/test/remote-endpoint-admission-probe.test.ts b/role-model-router/apps/runtime-host-bridge/test/remote-endpoint-admission-probe.test.ts
index 59713d91..562c5f51 100644
--- a/role-model-router/apps/runtime-host-bridge/test/remote-endpoint-admission-probe.test.ts
+++ b/role-model-router/apps/runtime-host-bridge/test/remote-endpoint-admission-probe.test.ts
@@ -41,7 +41,7 @@ describe("remote endpoint admission probes", () => {
       }),
     );
     expect(JSON.parse(String(calls[0]?.init?.body))).toEqual({
-      model: "deepseek/deepseek-v4-flash",
+      model: "deepseek-v4-flash",
       messages: [{ role: "user", content: "role-model admission readiness probe" }],
       max_tokens: 1,
       stream: false,