feat(service-ai): ADR-0040 — unify the platform assistant (data + authoring registers) (#1735)

The default data_chat agent becomes THE platform assistant:

- skills: + metadata_authoring + solution_design (registered by the cloud
  AI Studio plugin; deployments without it degrade gracefully to data-only —
  the skill registry ignores unresolved names).
- instructions: intent preamble — classify build/change vs data intent FIRST,
  apply that register's discipline, never mix registers or narrate failures.
- persona: label 'Assistant', role 'Business Application Assistant'.
- model: temperature 0.2 (authoring determinism wins).
- guardrails: blocklist union MINUS alter_schema/drop_table (the build
  register IS schema work, and it is already draft-gated per ADR-0033);
  60s execution budget for whole-app builds.
- planning: react ×10 with replan (builds take more steps than answers).

Motivation (staging 2026-06-11): the data-only persona accepted a whole-app
build — tools are registry-global — without the authoring disciplines, hand-
rolled the schema, collided with leftovers, and narrated its debugging to a
business user. ADR-0040: the end user never picks an agent; one assistant
carries both registers and the skills system does per-intent behavior.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: os-zhuang

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

@@ -566,7 +566,7 @@ describe('AgentRuntime', () => {
 
       expect(metadataService.get).toHaveBeenCalledWith('agent', 'data_chat');
       expect(agent?.name).toBe('data_chat');
-      expect(agent?.role).toBe('Business Data Analyst');
+      expect(agent?.role).toBe('Business Application Assistant');
     });
 
     it('should return undefined for unknown agent', async () => {
@@ -587,7 +587,7 @@ describe('AgentRuntime', () => {
       const messages = runtime.buildSystemMessages(DATA_CHAT_AGENT);
       expect(messages).toHaveLength(1);
       expect(messages[0].role).toBe('system');
-      expect(messages[0].content).toContain('helpful data assistant');
+      expect(messages[0].content).toContain('assistant for this business application platform');
     });
 
     it('should include context when provided', () => {
@@ -625,7 +625,7 @@ describe('AgentRuntime', () => {
     it('should derive model config from agent', () => {
       const options = runtime.buildRequestOptions(DATA_CHAT_AGENT, []);
       expect(options.model).toBe('gpt-4');
-      expect(options.temperature).toBe(0.3);
+      expect(options.temperature).toBe(0.2);
       expect(options.maxTokens).toBe(4096);
     });
 
@@ -667,7 +667,7 @@ describe('AgentRuntime', () => {
       ]);
       const agents = await runtime.listAgents();
       expect(agents).toHaveLength(2);
-      expect(agents[0]).toEqual({ name: 'data_chat', label: 'Data Assistant', role: 'Business Data Analyst' });
+      expect(agents[0]).toEqual({ name: 'data_chat', label: 'Assistant', role: 'Business Application Assistant' });
       expect(agents[1]).toEqual({ name: 'metadata_assistant', label: 'Metadata Assistant', role: 'Schema Architect' });
     });
 
@@ -998,14 +998,14 @@ describe('Agent Routes', () => {
 describe('DATA_CHAT_AGENT', () => {
   it('should be a valid agent definition', () => {
     expect(DATA_CHAT_AGENT.name).toBe('data_chat');
-    expect(DATA_CHAT_AGENT.role).toBe('Business Data Analyst');
+    expect(DATA_CHAT_AGENT.role).toBe('Business Application Assistant');
     expect(DATA_CHAT_AGENT.active).toBe(true);
     expect(DATA_CHAT_AGENT.visibility).toBe('global');
   });
 
   it('should reference the data_explorer skill (capability bundle moved to skill metadata)', () => {
     expect(DATA_CHAT_AGENT.tools ?? []).toHaveLength(0);
-    expect(DATA_CHAT_AGENT.skills).toEqual(['data_explorer', 'actions_executor']);
+    expect(DATA_CHAT_AGENT.skills).toEqual(['data_explorer', 'actions_executor', 'metadata_authoring', 'solution_design']);
   });
 
   it('should have guardrails configured', () => {

packages/services/service-ai/src/agents/data-chat-agent.ts

@@ -12,7 +12,7 @@ import type { Agent } from '@objectstack/spec/ai';
  *   - identity (name / label / role)
  *   - persona (system prompt)
  *   - model + safety config
- *   - skills attached → `skills: ['data_explorer']`
+ *   - skills attached → `skills: [...]` (ADR-0040: data + authoring)
  *
  * To grant data-exploration powers to a different agent, just add
  * `data_explorer` to its `skills[]`. To revoke globally, set the
@@ -40,36 +40,61 @@ export const DEFAULT_DATA_AGENT_NAME = 'data_chat';
 
 export const DATA_CHAT_AGENT: Agent = {
   name: DEFAULT_DATA_AGENT_NAME,
-  label: 'Data Assistant',
-  role: 'Business Data Analyst',
-  instructions: `You are a helpful data assistant that helps users explore and understand their business data through natural language.
+  label: 'Assistant',
+  role: 'Business Application Assistant',
+  // ADR-0040 — the unified platform assistant. End users never pick an
+  // agent; this one persona answers BOTH registers, and the FIRST job of
+  // every turn is classifying which register the user is in. The per-register
+  // disciplines (plan-first blueprints, draft semantics, no failure
+  // narration, data-query guidance) live in the attached skills.
+  instructions: `You are the assistant for this business application platform. You can both ANSWER QUESTIONS about the user's data and BUILD or CHANGE the application itself (objects, fields, views, dashboards, whole apps).
+
+INTENT FIRST — before acting, classify the request:
+- BUILD/CHANGE intent ("build…", "create an app/object/field…", "add/change/remove …", "建/做一个…系统/应用/字段"): follow the solution-design and metadata-authoring disciplines from your skills — plan-first for whole systems, drafts are not live, verify after building, and never narrate tool errors or internal retries to the user; present outcomes, not your debugging.
+- DATA intent ("how many…", "show/list…", "查/统计/看一下…"): use the data-exploration tools and answer concisely with real numbers.
+Never mix the registers in one reply: a build turn reports what was built and its verification status; a data turn answers the question.
 
 Always answer in the same language the user is using. Detailed tool-usage guidance is supplied by the skills attached to this agent.`,
 
   model: {
     provider: 'openai',
     model: 'gpt-4',
-    temperature: 0.3,
+    // The stricter of the merged personas: authoring needs determinism.
+    temperature: 0.2,
     maxTokens: 4096,
   },
 
-  // Capability bundle lives on the skill; the agent only references it.
-  // `data_explorer` = read side, `actions_executor` = write side.
-  skills: ['data_explorer', 'actions_executor'],
+  // Capability bundles live on skills; the agent only references them.
+  // `data_explorer`/`actions_executor` = the data register;
+  // `metadata_authoring`/`solution_design` = the build register (ADR-0040).
+  // The authoring skills are registered by the cloud AI Studio plugin — on
+  // deployments without it these references simply don't resolve and the
+  // assistant gracefully degrades to data-only (the skill registry ignores
+  // unknown names).
+  skills: ['data_explorer', 'actions_executor', 'metadata_authoring', 'solution_design'],
 
   active: true,
   visibility: 'global',
 
   guardrails: {
     maxTokensPerInvocation: 8192,
-    maxExecutionTimeSec: 30,
-    blockedTopics: ['delete_records', 'drop_table', 'alter_schema'],
+    // Whole-app builds (blueprint + per-artifact drafting + verification)
+    // legitimately run past the old 30s data-answer budget.
+    maxExecutionTimeSec: 60,
+    // Union of both personas' blocklists MINUS the ones that contradict the
+    // build register: `alter_schema`/`drop_table` were the data-only agent's
+    // way of refusing schema work, but authoring IS schema work — and it is
+    // already draft-gated (ADR-0033: nothing is live until publish, and
+    // destructive changes carry their own warning + HITL). What remains is
+    // genuinely off-limits in both registers.
+    blockedTopics: ['delete_records', 'drop_database', 'raw_sql', 'system_tables'],
   },
 
   planning: {
     strategy: 'react',
-    maxIterations: 5,
-    allowReplan: false,
+    // Builds take more steps than data answers (blueprint → drafts → verify).
+    maxIterations: 10,
+    allowReplan: true,
   },
 
   memory: {