fix(inference): propagate anthropic wire mode to Hermes sandbox config sync

src/lib/actions/inference-set.test.ts

@@ -295,6 +295,58 @@ describe("patchHermesInferenceConfig", () => {
     });
     expect(config.terminal).toEqual({ backend: "local" });
   });
+
+  it("propagates the anthropic wire mode for anthropic-prod routes", () => {
+    const config: ConfigObject = {
+      model: {
+        default: "nvidia-routed",
+        provider: "custom",
+        base_url: "https://inference.local/v1",
+      },
+    };
+
+    const result = patchHermesInferenceConfig(config, "anthropic-prod", "claude-sonnet-4-6");
+
+    expect(result.changed).toBe(true);
+    expect(config.model).toEqual({
+      default: "claude-sonnet-4-6",
+      provider: "custom",
+      base_url: "https://inference.local",
+      api_mode: "anthropic_messages",
+    });
+  });
+
+  it("propagates the anthropic wire mode for compatible-anthropic-endpoint routes", () => {
+    const config: ConfigObject = { model: {} };
+
+    patchHermesInferenceConfig(config, "compatible-anthropic-endpoint", "claude-sonnet-4-6");
+
+    expect(config.model).toEqual({
+      default: "claude-sonnet-4-6",
+      provider: "custom",
+      base_url: "https://inference.local",
+      api_mode: "anthropic_messages",
+    });
+  });
+
+  it("clears a stale anthropic wire mode when switching to an OpenAI-wire route", () => {
+    const config: ConfigObject = {
+      model: {
+        default: "claude-sonnet-4-6",
+        provider: "custom",
+        base_url: "https://inference.local",
+        api_mode: "anthropic_messages",
+      },
+    };
+
+    patchHermesInferenceConfig(config, "hermes-provider", "openai/gpt-5.4-mini");
+
+    expect(config.model).toEqual({
+      default: "openai/gpt-5.4-mini",
+      provider: "custom",
+      base_url: "https://inference.local/v1",
+    });
+  });
 });
 
 describe("runInferenceSet", () => {

src/lib/actions/inference-set.ts

@@ -270,6 +270,19 @@ export function patchHermesInferenceConfig(
   modelConfig.default = model;
   modelConfig.base_url = route.inferenceBaseUrl;
   modelConfig.provider = "custom";
+  // Hermes "custom" providers default to the OpenAI chat-completions wire and
+  // pick other transports only from explicit config or base-url heuristics —
+  // "https://inference.local/v1" matches none. Anthropic-wire routes
+  // (anthropic-prod, compatible-anthropic-endpoint) must therefore propagate
+  // the API mode, or the agent POSTs /chat/completions at an anthropic-type
+  // provider and the gateway L7 policy denies every inference call. Clear the
+  // key on non-anthropic routes so a stale wire mode never survives a provider
+  // switch (mirrors Hermes's own stale-api_mode cleanup on provider changes).
+  if (route.inferenceApi === "anthropic-messages") {
+    modelConfig.api_mode = "anthropic_messages";
+  } else {
+    delete modelConfig.api_mode;
+  }
 
   return { changed: before !== JSON.stringify(config), route };
 }

src/lib/inference/config.test.ts

@@ -271,6 +271,46 @@ describe("getSandboxInferenceConfig", () => {
     });
   });
 
+  it("maps anthropic-prod to the anthropic wire behind inference.local", () => {
+    expect(getSandboxInferenceConfig("claude-sonnet-4-6", "anthropic-prod")).toEqual({
+      providerKey: "anthropic",
+      primaryModelRef: "anthropic/claude-sonnet-4-6",
+      inferenceBaseUrl: "https://inference.local",
+      inferenceApi: "anthropic-messages",
+      inferenceCompat: null,
+    });
+  });
+
+  it("defaults compatible-anthropic-endpoint to the anthropic wire when no API preference is baked", () => {
+    expect(getSandboxInferenceConfig("claude-sonnet-4-6", "compatible-anthropic-endpoint")).toEqual(
+      {
+        providerKey: "anthropic",
+        primaryModelRef: "anthropic/claude-sonnet-4-6",
+        inferenceBaseUrl: "https://inference.local",
+        inferenceApi: "anthropic-messages",
+        inferenceCompat: null,
+      },
+    );
+  });
+
+  it("keeps compatible-anthropic-endpoint on the OpenAI wire only for an explicit preference", () => {
+    expect(
+      getSandboxInferenceConfig(
+        "claude-sonnet-4-6",
+        "compatible-anthropic-endpoint",
+        "openai-completions",
+      ),
+    ).toEqual({
+      providerKey: MANAGED_PROVIDER_ID,
+      primaryModelRef: `${MANAGED_PROVIDER_ID}/claude-sonnet-4-6`,
+      inferenceBaseUrl: INFERENCE_ROUTE_URL,
+      inferenceApi: "openai-completions",
+      inferenceCompat: {
+        supportsStore: false,
+      },
+    });
+  });
+
   it("maps OpenAI-compatible endpoints to the managed inference provider", () => {
     expect(getSandboxInferenceConfig("deepseek-ai/DeepSeek-V4-Flash", "compatible-endpoint"))
       .toEqual({

src/lib/inference/config.ts

@@ -194,7 +194,16 @@ export function getSandboxInferenceConfig(
       break;
     case "anthropic-prod":
     case "compatible-anthropic-endpoint":
-      if (provider === "compatible-anthropic-endpoint" && inferenceApi === "openai-completions") {
+      // Only an EXPLICIT openai-completions preference keeps an
+      // Anthropic-compatible endpoint on the OpenAI wire. The fallback default
+      // ("openai-completions" when preferredInferenceApi is null) must not:
+      // callers without a baked preference — notably the Hermes config sync,
+      // which never passes one — would otherwise speak the OpenAI protocol at
+      // an anthropic-type provider and be denied by the gateway L7 policy.
+      if (
+        provider === "compatible-anthropic-endpoint" &&
+        preferredInferenceApi === "openai-completions"
+      ) {
         providerKey = MANAGED_PROVIDER_ID;
         primaryModelRef = `${MANAGED_PROVIDER_ID}/${model}`;
         inferenceCompat = {