fix(service-ai): resolve current object for AI chat across languages (#1815)

* fix(service-ai): resolve current object for AI chat across languages

The console assistant reported "can't find the X object" when asked to
analyse the object on the current page — most visibly for non-English
prompts. Three compounding gaps:

- The keyword SchemaRetriever tokeniser dropped all CJK text, so a
  Chinese request yielded zero terms and "No matching objects".
- describe_object/list_objects are cloud-only, leaving query_data's
  keyword match as the only object-discovery path in the open edition.
- Nothing fed the current object's schema to the agent, so "this object"
  couldn't be resolved without a lucky keyword hit.

Changes:
- SchemaRetriever.tokenise() now emits CJK single-char + bigram terms.
- AgentRuntime.buildContextSchemaMessages() injects the current object's
  schema into the system prompt; both chat routes call it.
- ToolExecutionContext gains currentObjectName/currentViewName; routes
  thread them through and query_data falls back to the current object
  when keyword retrieval is empty.

Tests cover CJK retrieval, schema injection (4 cases), and the
query_data current-object fallback (2 cases). 81 service-ai tests pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(service-ai): add buildContextSchemaMessages to the quota-test mock runtime

The mock AgentRuntime in agent-chat-quota.test.ts didn't implement the new
buildContextSchemaMessages method, so the route's await on it threw and the
handler returned 500 instead of 200. Mirror the real runtime in the mock.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Jack Zhuang <277994282+os-zhuang@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: 3 people

packages/services/service-ai/src/__tests__/agent-chat-quota.test.ts

@@ -120,6 +120,7 @@ function mockAgentRuntime(): AgentRuntime {
     })),
     resolveActiveSkills: vi.fn(async () => []),
     buildSystemMessages: vi.fn(() => [{ role: 'system', content: 'sys' }]),
+    buildContextSchemaMessages: vi.fn(async () => []),
     buildRequestOptions: vi.fn(() => ({})),
     listAgents: vi.fn(async () => []),
   } as unknown as AgentRuntime;

packages/services/service-ai/src/__tests__/chatbot-features.test.ts

@@ -639,6 +639,38 @@ describe('AgentRuntime', () => {
     });
   });
 
+  describe('buildContextSchemaMessages', () => {
+    it('injects the current object schema so "this object" resolves without a tool', async () => {
+      (metadataService.getObject as any).mockResolvedValue({
+        name: 'showcase_task',
+        label: 'Task',
+        fields: { id: { type: 'text' }, subject: { type: 'text', label: 'Subject' } },
+      });
+      const messages = await runtime.buildContextSchemaMessages({ objectName: 'showcase_task' });
+      expect(metadataService.getObject).toHaveBeenCalledWith('showcase_task');
+      expect(messages).toHaveLength(1);
+      expect(messages[0].role).toBe('system');
+      expect(messages[0].content).toContain('currently viewing the "showcase_task" object');
+      expect(messages[0].content).toContain('### showcase_task');
+      expect(messages[0].content).toContain('subject: text');
+    });
+
+    it('returns nothing when no object is in context', async () => {
+      expect(await runtime.buildContextSchemaMessages(undefined)).toEqual([]);
+      expect(await runtime.buildContextSchemaMessages({})).toEqual([]);
+    });
+
+    it('returns nothing when the object cannot be resolved', async () => {
+      (metadataService.getObject as any).mockResolvedValue(undefined);
+      expect(await runtime.buildContextSchemaMessages({ objectName: 'nope' })).toEqual([]);
+    });
+
+    it('degrades gracefully when metadata lookup throws', async () => {
+      (metadataService.getObject as any).mockRejectedValue(new Error('boom'));
+      expect(await runtime.buildContextSchemaMessages({ objectName: 'showcase_task' })).toEqual([]);
+    });
+  });
+
   describe('buildRequestOptions', () => {
     it('should derive model config from agent', () => {
       const options = runtime.buildRequestOptions(DATA_CHAT_AGENT, []);

packages/services/service-ai/src/__tests__/schema-retriever.test.ts

@@ -68,6 +68,18 @@ describe('SchemaRetriever', () => {
     expect(hits).toEqual([]);
   });
 
+  it('tokenises CJK queries so a Chinese label still matches', async () => {
+    const cjkTask: ObjectShape = {
+      name: 'showcase_task',
+      label: '任务',
+      pluralLabel: '任务',
+      fields: { id: { type: 'text' }, 标题: { type: 'text', label: '标题' } },
+    };
+    const r = new SchemaRetriever(mockMetadata([cjkTask, accountObject]));
+    const hits = await r.retrieve('帮我分析任务对象');
+    expect(hits[0]?.object.name).toBe('showcase_task');
+  });
+
   it('respects limit option', async () => {
     const r = new SchemaRetriever(
       mockMetadata([taskObject, accountObject, unrelatedObject]),

packages/services/service-ai/src/agent-runtime.ts

@@ -9,6 +9,7 @@ import type {
 import type { Agent, Skill } from '@objectstack/spec/ai';
 import { AgentSchema } from '@objectstack/spec/ai';
 import { SkillRegistry, type SkillContext } from './skill-registry.js';
+import { SchemaRetriever, type ObjectShape } from './schema-retriever.js';
 import { DEFAULT_DATA_AGENT_NAME } from './agents/index.js';
 
 /**
@@ -191,6 +192,49 @@ export class AgentRuntime {
     return [{ role: 'system' as const, content: parts.join('\n') }];
   }
 
+  /**
+   * Build an extra system message carrying the schema of the object the user
+   * is currently viewing ({@link AgentChatContext.objectName}).
+   *
+   * `buildSystemMessages` only states the object's *name*; it can't fetch the
+   * schema because it is synchronous. This async companion looks the object up
+   * and renders a compact field snippet so the agent can answer "analyse /
+   * describe this object" questions with ZERO tool calls — important on the
+   * open-framework deployment where `describe_object` isn't registered, and
+   * for non-English prompts that the keyword-based {@link SchemaRetriever}
+   * can't otherwise resolve to an English-named object.
+   *
+   * Returns an empty array when there is no current object or it can't be
+   * resolved — callers append the result, so a miss simply injects nothing.
+   */
+  async buildContextSchemaMessages(context?: AgentChatContext): Promise<ModelMessage[]> {
+    const objectName = context?.objectName;
+    if (!objectName) return [];
+
+    let raw: unknown;
+    try {
+      raw = await this.metadataService.getObject(objectName);
+    } catch {
+      return [];
+    }
+    if (!raw || typeof raw !== 'object' || !(raw as ObjectShape).name) return [];
+
+    const snippet = SchemaRetriever.renderSnippet([{ object: raw as ObjectShape, score: 1 }]);
+    if (!snippet) return [];
+
+    return [
+      {
+        role: 'system' as const,
+        content:
+          `The user is currently viewing the "${objectName}" object. When they refer to ` +
+          '"this object", "the current object", or its label without naming another one, ' +
+          `assume they mean "${objectName}". Use the schema below to answer questions about ` +
+          'its structure directly, and as the target object name for data-query tools.\n\n' +
+          snippet,
+      },
+    ];
+  }
+
   /**
    * Derive {@link AIRequestOptions} from an agent definition.
    *

packages/services/service-ai/src/routes/agent-routes.ts

@@ -189,6 +189,11 @@ export function buildAgentRoutes(
           // Build system messages from agent instructions + UI context + skills
           const systemMessages = agentRuntime.buildSystemMessages(agent, chatContext, activeSkills);
 
+          // Inject the schema of the object the user is currently viewing so
+          // "analyse / describe this object" works without a lookup tool and
+          // regardless of the prompt's language.
+          systemMessages.push(...(await agentRuntime.buildContextSchemaMessages(chatContext)));
+
           // Resolve agent model/tools + skill tools → request options
           const agentOptions = agentRuntime.buildRequestOptions(
             agent,
@@ -234,6 +239,13 @@ export function buildAgentRoutes(
                     typeof chatContext?.environmentId === 'string'
                       ? chatContext.environmentId
                       : undefined,
+                  // The object/view the user has open — lets built-in data
+                  // tools fall back to "this object" when the request doesn't
+                  // name one (ADR-aligned with the schema injection above).
+                  currentObjectName:
+                    typeof chatContext?.objectName === 'string' ? chatContext.objectName : undefined,
+                  currentViewName:
+                    typeof chatContext?.viewName === 'string' ? chatContext.viewName : undefined,
                 }
               : undefined,
           };

packages/services/service-ai/src/routes/assistant-routes.ts

@@ -197,6 +197,9 @@ export function buildAssistantRoutes(
           }
 
           const systemMessages = agentRuntime.buildSystemMessages(agent, context, activeSkills);
+          // Inject the current object's schema so "describe this object" works
+          // without a lookup tool and regardless of prompt language.
+          systemMessages.push(...(await agentRuntime.buildContextSchemaMessages(context)));
           const agentOptions = agentRuntime.buildRequestOptions(
             agent,
             aiService.toolRegistry.getAll(),
@@ -237,6 +240,10 @@ export function buildAssistantRoutes(
                     typeof body.conversationId === 'string' ? body.conversationId : undefined,
                   environmentId:
                     typeof context.environmentId === 'string' ? context.environmentId : undefined,
+                  currentObjectName:
+                    typeof context.objectName === 'string' ? context.objectName : undefined,
+                  currentViewName:
+                    typeof context.viewName === 'string' ? context.viewName : undefined,
                 }
               : undefined,
           };

packages/services/service-ai/src/schema-retriever.ts

@@ -181,12 +181,33 @@ export interface FieldShape {
 
 // ── internal helpers ──────────────────────────────────────────────
 
-/** Lower-case alphanumeric tokens of length ≥ 2 (English stop-words excluded). */
+/**
+ * Tokenise a query into match terms.
+ *
+ * Latin/digit runs split on any non-alphanumeric (including underscores) so
+ * `todo_task` tokenises to ['todo', 'task'] and matches snake_case names.
+ *
+ * CJK text carries no word boundaries, so a `[a-z0-9]+` scan drops it entirely
+ * — a question like "分析任务对象" would yield zero terms and surface a
+ * misleading "no matching objects". To keep CJK queries scoreable against
+ * CJK object/field labels, every ideograph is emitted as a single-char term
+ * plus each adjacent bigram (so "任务" matches a label containing "任务").
+ */
 function tokenise(query: string): string[] {
-  // Split on any non-alphanumeric (including underscores) so `todo_task`
-  // tokenises to ['todo', 'task'] and matches snake_case object names.
-  const raw = query.toLowerCase().match(/[a-z0-9]+/g) ?? [];
-  return raw.filter(t => t.length >= 2 && !STOPWORDS.has(t));
+  const lower = query.toLowerCase();
+  const latin = (lower.match(/[a-z0-9]+/g) ?? []).filter(
+    t => t.length >= 2 && !STOPWORDS.has(t),
+  );
+  const tokens = [...latin];
+  // CJK Unified Ideographs (+ Ext-A, compatibility) and Japanese kana.
+  const cjkRuns = lower.match(/[぀-ヿ㐀-䶿一-鿿豈-﫿]+/g) ?? [];
+  for (const run of cjkRuns) {
+    for (let i = 0; i < run.length; i++) {
+      tokens.push(run[i]);
+      if (i + 1 < run.length) tokens.push(run.slice(i, i + 2));
+    }
+  }
+  return tokens;
 }
 
 const STOPWORDS = new Set([

packages/services/service-ai/src/tools/query-data.tool.test.ts

@@ -95,6 +95,61 @@ describe('query_data — federated query timeout (P4)', () => {
     expect(out.records[0].id).toBe('o1');
   });
 
+  it('falls back to the current object when keyword retrieval finds nothing', async () => {
+    const currentObject = {
+      name: 'showcase_task',
+      label: 'Task',
+      fields: { id: { type: 'text' }, subject: { type: 'text' } },
+    };
+    let findObject: string | undefined;
+    const ctx: QueryDataToolContext = {
+      ai: {
+        generateObject: async () => ({
+          object: {
+            objectName: 'showcase_task',
+            whereJson: null,
+            fields: null,
+            orderBy: null,
+            limit: null,
+          } satisfies QueryPlan,
+        }),
+      } as never,
+      metadata: {
+        // Catalogue is non-empty, but a Chinese request tokenises to terms
+        // that don't match the English name/label → zero keyword hits.
+        listObjects: async () => [currentObject],
+        getObject: async (name: string) => (name === 'showcase_task' ? currentObject : undefined),
+      } as never,
+      dataEngine: {
+        find: async (objectName: string) => {
+          findObject = objectName;
+          return [{ id: 't1', subject: 'Build the thing' }];
+        },
+      } as never,
+    };
+    const handler = createQueryDataHandler(ctx);
+    const out = JSON.parse(
+      (await handler({ request: '分析这个对象的数据' }, { currentObjectName: 'showcase_task' } as never)) as string,
+    );
+    expect(out.error).toBeUndefined();
+    expect(findObject).toBe('showcase_task');
+    expect(out.count).toBe(1);
+  });
+
+  it('still errors when retrieval is empty and no current object is supplied', async () => {
+    const ctx: QueryDataToolContext = {
+      ai: { generateObject: async () => ({ object: {} as QueryPlan }) } as never,
+      metadata: {
+        listObjects: async () => [{ name: 'account', label: 'Account', fields: {} }],
+        getObject: async () => undefined,
+      } as never,
+      dataEngine: { find: async () => [] } as never,
+    };
+    const handler = createQueryDataHandler(ctx);
+    const out = JSON.parse((await handler({ request: '分析这个对象' })) as string);
+    expect(out.error).toMatch(/No matching objects in metadata/);
+  });
+
   it('does not wrap managed (non-external) objects in a timeout', async () => {
     const managedObject = {
       name: 'task',

packages/services/service-ai/src/tools/query-data.tool.ts

@@ -9,7 +9,7 @@ import type {
   ModelMessage,
 } from '@objectstack/spec/contracts';
 import type { ExecutionContext } from '@objectstack/spec/kernel';
-import { SchemaRetriever } from '../schema-retriever.js';
+import { SchemaRetriever, type SchemaHit } from '../schema-retriever.js';
 import type { ToolHandler, ToolRegistry, ToolExecutionContext } from './tool-registry.js';
 
 /** See `data-tools.ts#buildEngineContext` — duplicated here to keep
@@ -161,6 +161,35 @@ export function createQueryDataHandler(ctx: QueryDataToolContext): ToolHandler {
   const maxLimit = ctx.maxLimit ?? 100;
   const externalTimeoutFallback = ctx.externalQueryTimeoutMs ?? 30_000;
 
+  /** Load a single object definition by exact name, mirroring the dual-source
+   *  lookup (MetadataManager → ObjectQL SchemaRegistry via protocol) so the
+   *  current-object fallback can see system objects too. Never throws. */
+  const loadObjectByName = async (name: string): Promise<SchemaHit['object'] | undefined> => {
+    try {
+      const direct = await ctx.metadata.getObject?.(name);
+      if (direct && typeof direct === 'object' && (direct as { name?: string }).name) {
+        return direct as SchemaHit['object'];
+      }
+    } catch {
+      // fall through to protocol enumeration
+    }
+    if (ctx.protocol?.getMetaItems) {
+      try {
+        const all = await ctx.protocol.getMetaItems({ type: 'object' });
+        const arr = Array.isArray(all)
+          ? all
+          : (all && typeof all === 'object' && Array.isArray((all as { items?: unknown }).items)
+            ? (all as { items: unknown[] }).items
+            : []);
+        const found = (arr as Array<{ name?: string }>).find(o => o?.name === name);
+        if (found) return found as SchemaHit['object'];
+      } catch {
+        // ignore — caller treats undefined as "not found"
+      }
+    }
+    return undefined;
+  };
+
   /** Resolve a federated object's per-query timeout (datasource-declared,
    *  else the tool fallback). Never throws — degrades to the fallback. */
   const resolveExternalTimeout = async (datasource: string): Promise<number> => {
@@ -190,7 +219,14 @@ export function createQueryDataHandler(ctx: QueryDataToolContext): ToolHandler {
     }
 
     // 1. Schema retrieval
-    const hits = await retriever.retrieve(request);
+    let hits = await retriever.retrieve(request);
+    // Fallback: when keyword retrieval finds nothing (e.g. a non-English
+    // request, or one that says "this object" without naming it), target the
+    // object the user is currently viewing if the UI supplied one.
+    if (hits.length === 0 && execCtx?.currentObjectName) {
+      const current = await loadObjectByName(execCtx.currentObjectName);
+      if (current) hits = [{ object: current, score: 1 }];
+    }
     if (hits.length === 0) {
       return JSON.stringify({
         error:

packages/spec/src/contracts/ai-service.ts

@@ -402,6 +402,16 @@ export interface ToolExecutionContext {
     messageId?: string;
     /** Active environment (multi-tenant project) id, if known. */
     environmentId?: string;
+    /**
+     * Object the user is currently viewing in the UI (e.g. the list/detail
+     * page they have open). Built-in data tools use this as a fallback target
+     * when the user refers to "this object" / "the current object" and the
+     * free-text request doesn't name one explicitly — so a question phrased in
+     * any language still resolves to the right object without a keyword match.
+     */
+    currentObjectName?: string;
+    /** View the user is currently viewing, if known. */
+    currentViewName?: string;
     /** Distributed-trace id for cross-service correlation. */
     traceId?: string;
     /**