fix: keep metadata reuse scoped to agent turns

src/agents/agent-command.ts

@@ -20,7 +20,6 @@ import {
 import { formatErrorMessage } from "../infra/errors.js";
 import { buildOutboundSessionContext } from "../infra/outbound/session-context.js";
 import { createSubsystemLogger } from "../logging/subsystem.js";
-import { setCurrentPluginMetadataSnapshot } from "../plugins/current-plugin-metadata-snapshot.js";
 import { loadManifestMetadataSnapshot } from "../plugins/manifest-contract-eligibility.js";
 import {
   isSubagentSessionKey,
@@ -403,11 +402,6 @@ async function prepareAgentCommandExecution(
     workspaceDir,
     env: process.env,
   });
-  setCurrentPluginMetadataSnapshot(manifestMetadataSnapshot, {
-    config: cfg,
-    env: process.env,
-    workspaceDir,
-  });
   const manifestPlugins = manifestMetadataSnapshot.plugins;
   const configuredModel = resolveConfiguredModelRef({
     cfg,

src/agents/model-selection-shared.ts

@@ -134,6 +134,8 @@ export function inferUniqueProviderFromConfiguredModels(
           manifestPlugins: params.manifestPlugins,
         });
         if (
+          modelId === model ||
+          normalizeLowercaseStringOrEmpty(modelId) === normalized ||
           normalizedModelId === model ||
           normalizeLowercaseStringOrEmpty(normalizedModelId) === normalized
         ) {
@@ -604,7 +606,9 @@ export function resolveConfiguredModelRef(
         manifestPlugins: params.manifestPlugins,
       });
       if (inferredProvider) {
-        return { provider: inferredProvider, model: trimmed };
+        return normalizeModelRef(inferredProvider, trimmed, {
+          manifestPlugins: params.manifestPlugins,
+        });
       }
 
       const safeTrimmed = sanitizeModelWarningValue(trimmed);

src/agents/model-selection.test.ts

@@ -695,6 +695,25 @@ describe("model-selection", () => {
       ).toBe("qwen-dashscope");
     });
 
+    it("infers provider from raw configured ids when manifest policies add prefixes", () => {
+      const cfg = {
+        models: {
+          providers: {
+            nvidia: {
+              models: [{ id: "llama-fast" }],
+            },
+          },
+        },
+      } as unknown as OpenClawConfig;
+
+      expect(
+        inferUniqueProviderFromConfiguredModels({
+          cfg,
+          model: "llama-fast",
+        }),
+      ).toBe("nvidia");
+    });
+
     it("infers Google provider from canonicalized configured provider catalogs", () => {
       const cfg = {
         models: {
@@ -1655,6 +1674,34 @@ describe("model-selection", () => {
       }
     });
 
+    it("normalizes bare configured default model strings with manifest policies", () => {
+      const cfg = {
+        agents: {
+          defaults: {
+            model: { primary: "llama-fast" },
+          },
+        },
+        models: {
+          providers: {
+            nvidia: {
+              models: [{ id: "llama-fast" }],
+            },
+          },
+        },
+      } as unknown as OpenClawConfig;
+
+      const result = resolveConfiguredModelRef({
+        cfg,
+        defaultProvider: "openai",
+        defaultModel: "gpt-5.4",
+      });
+
+      expect(result).toEqual({
+        provider: "nvidia",
+        model: "nvidia/llama-fast",
+      });
+    });
+
     it("prefers slash-form aliases for configured default models", () => {
       const cfg = {
         agents: {