fix(inference): trust explicit custom switch metadata

docs/reference/commands-nemohermes.mdx

@@ -1519,7 +1519,7 @@ Under the `nemohermes` alias, it uses the registered Hermes sandbox when exactly
 By default, the command syncs the default registered sandbox.
 
 ```bash
-nemohermes inference set --provider <provider> --model <model> [--sandbox <name>] [--no-verify]
+nemohermes inference set --provider <provider> --model <model> [--sandbox <name>] [--no-verify] [--endpoint-url <url>] [--credential-env <ENV>] [--inference-api <api>]
 ```
 
 Pass both `--provider` and `--model` when you want NemoClaw to update the OpenShell inference route and sync the selected sandbox's agent config.
@@ -1530,6 +1530,7 @@ If the in-sandbox config sync fails, NemoClaw keeps the gateway and registry ali
 
 Supported provider names are `nvidia-prod`, `nvidia-nim`, `nvidia-router`, `openai-api`, `anthropic-prod`, `compatible-anthropic-endpoint`, `gemini-api`, `compatible-endpoint`, `hermes-provider`, `ollama-local`, and `vllm-local`.
 Use `--no-verify` only when OpenShell cannot verify the provider at switch time but you have already confirmed the provider and credential.
+When switching to `compatible-endpoint` or `compatible-anthropic-endpoint` from a different provider family, pass `--endpoint-url` with the trusted custom provider URL so NemoClaw can persist durable rebuild metadata. `--credential-env` and `--inference-api` may also be supplied for the compatible provider metadata; supported API values are `openai-completions`, `anthropic-messages`, and `openai-responses`.
 
 ### `nemohermes setup`
 

docs/reference/commands.mdx

@@ -1874,7 +1874,7 @@ Under the `nemohermes` alias, it uses the registered Hermes sandbox when exactly
 By default, the command syncs the default registered sandbox.
 
 ```bash
-$$nemoclaw inference set --provider <provider> --model <model> [--sandbox <name>] [--no-verify]
+$$nemoclaw inference set --provider <provider> --model <model> [--sandbox <name>] [--no-verify] [--endpoint-url <url>] [--credential-env <ENV>] [--inference-api <api>]
 ```
 
 Pass both `--provider` and `--model` when you want NemoClaw to update the OpenShell inference route and sync the selected sandbox's agent config.
@@ -1885,6 +1885,7 @@ If the in-sandbox config sync fails, NemoClaw keeps the gateway and registry ali
 
 Supported provider names are `nvidia-prod`, `nvidia-nim`, `nvidia-router`, `openai-api`, `anthropic-prod`, `compatible-anthropic-endpoint`, `gemini-api`, `compatible-endpoint`, `hermes-provider`, `ollama-local`, and `vllm-local`.
 Use `--no-verify` only when OpenShell cannot verify the provider at switch time but you have already confirmed the provider and credential.
+When switching to `compatible-endpoint` or `compatible-anthropic-endpoint` from a different provider family, pass `--endpoint-url` with the trusted custom provider URL so NemoClaw can persist durable rebuild metadata. `--credential-env` and `--inference-api` may also be supplied for the compatible provider metadata; supported API values are `openai-completions`, `anthropic-messages`, and `openai-responses`.
 
 ### `$$nemoclaw setup`
 

src/commands/global-oclif-command-adapters.test.ts

@@ -231,6 +231,12 @@ describe("global oclif command adapters", () => {
         "--sandbox",
         "alpha",
         "--no-verify",
+        "--endpoint-url",
+        "https://example.test/v1",
+        "--credential-env",
+        "COMPATIBLE_API_KEY",
+        "--inference-api",
+        "openai-completions",
       ],
       rootDir,
     );
@@ -240,6 +246,9 @@ describe("global oclif command adapters", () => {
       model: "nvidia/nemotron-3-super-120b-a12b",
       sandboxName: "alpha",
       noVerify: true,
+      endpointUrl: "https://example.test/v1",
+      credentialEnv: "COMPATIBLE_API_KEY",
+      inferenceApi: "openai-completions",
     });
   });
 

src/commands/inference/set.ts

@@ -25,7 +25,7 @@ export default class InferenceSetCommand extends NemoClawCommand {
   static description =
     "Update the OpenShell inference route and sync the running OpenClaw or Hermes sandbox config.";
   static usage = [
-    "inference set --provider <provider> --model <model> [--sandbox <name>] [--no-verify]",
+    "inference set --provider <provider> --model <model> [--sandbox <name>] [--no-verify] [--endpoint-url <url>] [--credential-env <ENV>] [--inference-api <api>]",
   ];
   static examples = [
     "<%= config.bin %> inference set --provider nvidia-prod --model nvidia/nemotron-3-super-120b-a12b",
@@ -41,6 +41,17 @@ export default class InferenceSetCommand extends NemoClawCommand {
     "no-verify": Flags.boolean({
       description: "Pass --no-verify through to openshell inference set",
     }),
+    "endpoint-url": Flags.string({
+      description: "Trusted endpoint URL to persist when switching to a compatible custom provider",
+    }),
+    "credential-env": Flags.string({
+      description:
+        "Trusted credential env name to persist when switching to a compatible custom provider",
+    }),
+    "inference-api": Flags.string({
+      description:
+        "Trusted API family to persist for compatible custom providers (openai-completions, anthropic-messages, openai-responses)",
+    }),
   };
 
   public async run(): Promise<void> {
@@ -55,6 +66,9 @@ export default class InferenceSetCommand extends NemoClawCommand {
         model: flags.model,
         sandboxName: flags.sandbox ?? null,
         noVerify: flags["no-verify"] === true,
+        endpointUrl: flags["endpoint-url"] ?? null,
+        credentialEnv: flags["credential-env"] ?? null,
+        inferenceApi: flags["inference-api"] ?? null,
       });
     } catch (error) {
       if (error instanceof InferenceSetError) {

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

@@ -693,6 +693,158 @@ describe("runInferenceSet", () => {
     expect(deps.calls.updateSandbox).not.toHaveBeenCalled();
   });
 
+  it("rejects Anthropic Messages metadata for OpenAI-compatible endpoint switches", async () => {
+    const deps = createDeps({
+      config: { agents: { defaults: { model: { primary: "inference/nvidia/model-a" } } } },
+      entry: {
+        name: "alpha",
+        agent: "openclaw",
+        provider: "nvidia-prod",
+        model: "nvidia/model-a",
+      },
+      session: baseSession({
+        provider: "nvidia-prod",
+        model: "nvidia/model-a",
+        endpointUrl: "https://integrate.api.nvidia.com/v1",
+        credentialEnv: "NVIDIA_INFERENCE_API_KEY",
+      }),
+    });
+
+    await expect(
+      runInferenceSet(
+        {
+          provider: "compatible-endpoint",
+          model: "mock-openai-model",
+          noVerify: true,
+          endpointUrl: "https://compatible.example/v1",
+          credentialEnv: "COMPATIBLE_API_KEY",
+          inferenceApi: "anthropic-messages",
+        },
+        deps,
+      ),
+    ).rejects.toThrow(
+      /inference-api for 'compatible-endpoint' must be one of: openai-completions, openai-responses/,
+    );
+
+    expect(deps.calls.runOpenshell).not.toHaveBeenCalled();
+    expect(deps.calls.updateSandbox).not.toHaveBeenCalled();
+  });
+
+  it("preserves explicit inference API through the final registry and session sync", async () => {
+    const config: ConfigObject = {
+      agents: { defaults: { model: { primary: "inference/nvidia/model-a" } } },
+      models: { providers: { inference: { api: "openai-completions", models: [] } } },
+    };
+    const deps = createDeps({
+      config,
+      entry: {
+        name: "alpha",
+        agent: "openclaw",
+        provider: "nvidia-prod",
+        model: "nvidia/model-a",
+      },
+      session: baseSession({
+        provider: "nvidia-prod",
+        model: "nvidia/model-a",
+        endpointUrl: "https://integrate.api.nvidia.com/v1",
+        credentialEnv: "NVIDIA_INFERENCE_API_KEY",
+        preferredInferenceApi: "openai-completions",
+      }),
+    });
+
+    await runInferenceSet(
+      {
+        provider: "compatible-endpoint",
+        model: "mock-responses-model",
+        noVerify: true,
+        endpointUrl: "https://compatible.example/v1",
+        credentialEnv: "COMPATIBLE_API_KEY",
+        inferenceApi: "openai-responses",
+      },
+      deps,
+    );
+
+    expect(config.models).toMatchObject({
+      providers: {
+        inference: {
+          api: "openai-responses",
+          models: [{ id: "mock-responses-model", name: "inference/mock-responses-model" }],
+        },
+      },
+    });
+    expect(deps.calls.updateSandbox.mock.calls.at(-1)).toEqual([
+      "alpha",
+      expect.objectContaining({
+        provider: "compatible-endpoint",
+        model: "mock-responses-model",
+        endpointUrl: "https://compatible.example/v1",
+        credentialEnv: "COMPATIBLE_API_KEY",
+        preferredInferenceApi: "openai-responses",
+      }),
+    ]);
+    expect(deps.getSession()).toMatchObject({
+      provider: "compatible-endpoint",
+      model: "mock-responses-model",
+      endpointUrl: "https://compatible.example/v1",
+      credentialEnv: "COMPATIBLE_API_KEY",
+      preferredInferenceApi: "openai-responses",
+    });
+  });
+
+  it("accepts explicit compatible Anthropic endpoint metadata for provider-family switches", async () => {
+    const config: ConfigObject = {
+      agents: { defaults: { model: { primary: "inference/nvidia/model-a" } } },
+      models: { providers: { inference: { api: "openai-completions", models: [] } } },
+    };
+    const deps = createDeps({
+      config,
+      entry: {
+        name: "alpha",
+        agent: "openclaw",
+        provider: "nvidia-prod",
+        model: "nvidia/model-a",
+      },
+      session: baseSession({
+        provider: "nvidia-prod",
+        model: "nvidia/model-a",
+        endpointUrl: "https://integrate.api.nvidia.com/v1",
+        credentialEnv: "NVIDIA_INFERENCE_API_KEY",
+      }),
+    });
+
+    await runInferenceSet(
+      {
+        provider: "compatible-anthropic-endpoint",
+        model: "mock-anthropic-model",
+        noVerify: true,
+        endpointUrl: "http://host.openshell.internal:18767/",
+        credentialEnv: "COMPATIBLE_ANTHROPIC_API_KEY",
+        inferenceApi: "anthropic-messages",
+      },
+      deps,
+    );
+
+    expect(deps.calls.updateSandbox.mock.calls.at(-1)).toEqual([
+      "alpha",
+      expect.objectContaining({
+        provider: "compatible-anthropic-endpoint",
+        model: "mock-anthropic-model",
+        endpointUrl: "http://host.openshell.internal:18767",
+        credentialEnv: "COMPATIBLE_ANTHROPIC_API_KEY",
+        preferredInferenceApi: "anthropic-messages",
+        nimContainer: null,
+      }),
+    ]);
+    expect(deps.getSession()).toMatchObject({
+      provider: "compatible-anthropic-endpoint",
+      model: "mock-anthropic-model",
+      endpointUrl: "http://host.openshell.internal:18767",
+      credentialEnv: "COMPATIBLE_ANTHROPIC_API_KEY",
+      preferredInferenceApi: "anthropic-messages",
+      nimContainer: null,
+    });
+  });
+
   it("preserves same-provider Bedrock Runtime adapter routing for OpenClaw switches", async () => {
     const config: ConfigObject = {
       agents: {