@firstlovecenter/ai-chat 0.5.0 → 0.6.1

package/dist/server/index.d.cts

@@ -1,5 +1,5 @@
-import { S as SystemBlock, T as ToolSchema, a as ToolContext, b as ToolDefinition, P as PresentPayload, c as PersistencePort, A as AuthPort, d as ScopePort, e as ToolsPort, V as VertexPort, L as LoggerPort } from '../types-CDKxdzQc.cjs';
-export { f as AiSettings, g as AiSettingsPatch, h as AppendMessageInput, i as AuthFail, j as AuthOk, k as AuthResult, B as Block, C as ChartSpec, l as ChatMessage, m as ChatMessageRole, n as ChatSession, o as CreateSessionInput, p as ListSessionsOpts, q as TERMINAL_TOOL_NAME, r as ToolResult, s as err, t as ok } from '../types-CDKxdzQc.cjs';
+import { S as SystemBlock, T as ToolSchema, a as ToolContext, b as ToolDefinition, P as PresentPayload, c as PersistencePort, A as AuthPort, d as ScopePort, e as ToolsPort, V as VertexPort, L as LoggerPort } from '../types-CQntnyDJ.cjs';
+export { f as AiSettings, g as AiSettingsPatch, h as AppendMessageInput, i as AuthFail, j as AuthOk, k as AuthResult, B as Block, C as ChartSpec, l as ChatMessage, m as ChatMessageRole, n as ChatSession, o as CreateSessionInput, p as ListSessionsOpts, q as TERMINAL_TOOL_NAME, r as ToolResult, s as err, t as ok } from '../types-CQntnyDJ.cjs';
 import { GoogleAuth } from 'google-auth-library';
 export { GoogleAuth } from 'google-auth-library';
 import 'zod';
@@ -398,7 +398,7 @@ declare function createChatSessionsRoutes<S>(ctx: ChatSessionsRouteCtx<S>): {
 /**
  * `/api/admin/ai-settings` route factory — global AI configuration (super_admin only).
  *
- * Three patchable fields on the singleton settings row:
+ * Five patchable fields on the singleton settings row:
  *   - `tool_provider` — vendor that drives the agent tool loop. Validated
  *     against the registered `toolProviders` registry passed in via ctx.
  *   - `gcp_location` — the Vertex region every provider call hits. Stays
@@ -408,6 +408,15 @@ declare function createChatSessionsRoutes<S>(ctx: ChatSessionsRouteCtx<S>): {
  *     against the `chatInterfaces` registry passed in via ctx (the actual
  *     registry lives in `@firstlovecenter/ai-chat/ui` so the host wires it through;
  *     the route stays free of UI imports).
+ *   - `max_output_tokens` — caps the agent loop's per-turn output AND each
+ *     narrator's prose pass. Bounded `[256, 64000]` — anything below 256 can't
+ *     fit a useful response, anything above 64000 exceeds the headroom of any
+ *     model currently routed through Vertex.
+ *   - `role_prompt` — admin-editable persona string. Empty string and the
+ *     explicit JSON `null` both clear the override back to the host's static
+ *     `configureAiChat({ rolePrompt })` fallback; we canonicalise an empty/
+ *     whitespace-only string to `null` on write so the "no override" state has
+ *     a single representation in storage.
  *
  * Wire format is snake_case to preserve byte-for-byte parity with the
  * host route the package replaces — existing host UIs keep working

package/dist/server/index.d.ts

@@ -1,5 +1,5 @@
-import { S as SystemBlock, T as ToolSchema, a as ToolContext, b as ToolDefinition, P as PresentPayload, c as PersistencePort, A as AuthPort, d as ScopePort, e as ToolsPort, V as VertexPort, L as LoggerPort } from '../types-CDKxdzQc.js';
-export { f as AiSettings, g as AiSettingsPatch, h as AppendMessageInput, i as AuthFail, j as AuthOk, k as AuthResult, B as Block, C as ChartSpec, l as ChatMessage, m as ChatMessageRole, n as ChatSession, o as CreateSessionInput, p as ListSessionsOpts, q as TERMINAL_TOOL_NAME, r as ToolResult, s as err, t as ok } from '../types-CDKxdzQc.js';
+import { S as SystemBlock, T as ToolSchema, a as ToolContext, b as ToolDefinition, P as PresentPayload, c as PersistencePort, A as AuthPort, d as ScopePort, e as ToolsPort, V as VertexPort, L as LoggerPort } from '../types-CQntnyDJ.js';
+export { f as AiSettings, g as AiSettingsPatch, h as AppendMessageInput, i as AuthFail, j as AuthOk, k as AuthResult, B as Block, C as ChartSpec, l as ChatMessage, m as ChatMessageRole, n as ChatSession, o as CreateSessionInput, p as ListSessionsOpts, q as TERMINAL_TOOL_NAME, r as ToolResult, s as err, t as ok } from '../types-CQntnyDJ.js';
 import { GoogleAuth } from 'google-auth-library';
 export { GoogleAuth } from 'google-auth-library';
 import 'zod';
@@ -398,7 +398,7 @@ declare function createChatSessionsRoutes<S>(ctx: ChatSessionsRouteCtx<S>): {
 /**
  * `/api/admin/ai-settings` route factory — global AI configuration (super_admin only).
  *
- * Three patchable fields on the singleton settings row:
+ * Five patchable fields on the singleton settings row:
  *   - `tool_provider` — vendor that drives the agent tool loop. Validated
  *     against the registered `toolProviders` registry passed in via ctx.
  *   - `gcp_location` — the Vertex region every provider call hits. Stays
@@ -408,6 +408,15 @@ declare function createChatSessionsRoutes<S>(ctx: ChatSessionsRouteCtx<S>): {
  *     against the `chatInterfaces` registry passed in via ctx (the actual
  *     registry lives in `@firstlovecenter/ai-chat/ui` so the host wires it through;
  *     the route stays free of UI imports).
+ *   - `max_output_tokens` — caps the agent loop's per-turn output AND each
+ *     narrator's prose pass. Bounded `[256, 64000]` — anything below 256 can't
+ *     fit a useful response, anything above 64000 exceeds the headroom of any
+ *     model currently routed through Vertex.
+ *   - `role_prompt` — admin-editable persona string. Empty string and the
+ *     explicit JSON `null` both clear the override back to the host's static
+ *     `configureAiChat({ rolePrompt })` fallback; we canonicalise an empty/
+ *     whitespace-only string to `null` on write so the "no override" state has
+ *     a single representation in storage.
  *
  * Wire format is snake_case to preserve byte-for-byte parity with the
  * host route the package replaces — existing host UIs keep working

package/dist/server/index.js

@@ -650,7 +650,7 @@ async function* streamClaudeNarration(opts) {
   });
   const stream = await client.messages.stream({
     model: opts.modelId,
-    max_tokens: 400,
+    max_tokens: opts.maxTokens,
     system: NARRATIVE_SYSTEM,
     messages: [{ role: "user", content: buildNarrativeUserMessage(opts.input) }]
   });
@@ -710,7 +710,7 @@ async function* streamGeminiNarration(opts) {
           parts: [{ text: buildNarrativeUserMessage(opts.input) }]
         }
       ],
-      generationConfig: { maxOutputTokens: 400, temperature: 0 }
+      generationConfig: { maxOutputTokens: opts.maxTokens, temperature: 0 }
     })
   });
   if (!res.ok || !res.body) {
@@ -786,7 +786,7 @@ async function* streamGrokNarration(opts) {
     },
     body: JSON.stringify({
       model: opts.modelId,
-      max_tokens: 400,
+      max_tokens: opts.maxTokens,
       stream: true,
       messages: [
         { role: "system", content: NARRATIVE_SYSTEM3 },
@@ -1009,7 +1009,8 @@ data: ${JSON.stringify(data)}
               ctx: toolContext,
               tools: tools.tools,
               systemBlocks,
-              provider
+              provider,
+              maxOutputTokens: aiSettings.maxOutputTokens
             });
             if (!agentResult.ok) {
               persistedError = agentResult.error;
@@ -1035,6 +1036,7 @@ data: ${JSON.stringify(data)}
                     projectId: vertex.projectId,
                     location: aiSettings.gcpLocation,
                     modelId: narratorModelId,
+                    maxTokens: aiSettings.maxOutputTokens,
                     input: {
                       question,
                       structured,
@@ -1320,7 +1322,7 @@ function createAgentVercelRoutes(ctx) {
           messages: [{ role: "user", content: question }],
           tools: vercelTools,
           maxSteps: 12,
-          maxTokens: 4096,
+          maxTokens: aiSettings.maxOutputTokens,
           onFinish: async ({ text }) => {
             try {
               let blocks = presentPayload?.blocks ?? [];
@@ -1609,6 +1611,8 @@ function createChatSessionsRoutes(ctx) {
 
 // src/server/routes/admin-settings.ts
 var VALID_LOCATIONS = ["us-east5", "global"];
+var MIN_MAX_OUTPUT_TOKENS = 256;
+var MAX_MAX_OUTPUT_TOKENS = 64e3;
 function isStringRecord(v) {
   return typeof v === "object" && v !== null && !Array.isArray(v);
 }
@@ -1623,6 +1627,8 @@ function toWire(settings) {
     tool_provider: settings.toolProvider,
     gcp_location: settings.gcpLocation,
     chat_interface: settings.chatInterface,
+    max_output_tokens: settings.maxOutputTokens,
+    role_prompt: settings.rolePrompt,
     updated_at: settings.updatedAt ? settings.updatedAt.toISOString() : null,
     updated_by_user_id: settings.updatedByUserId
   };
@@ -1702,11 +1708,29 @@ function createAdminSettingsRoutes(ctx) {
       }
       patch.chatInterface = v;
     }
-    if (patch.toolProvider === void 0 && patch.gcpLocation === void 0 && patch.chatInterface === void 0) {
+    if ("max_output_tokens" in body) {
+      const v = body.max_output_tokens;
+      if (typeof v !== "number" || !Number.isInteger(v) || v < MIN_MAX_OUTPUT_TOKENS || v > MAX_MAX_OUTPUT_TOKENS) {
+        return jsonResponse({ error: "invalid_max_output_tokens" }, 400);
+      }
+      patch.maxOutputTokens = v;
+    }
+    if ("role_prompt" in body) {
+      const v = body.role_prompt;
+      if (v === null) {
+        patch.rolePrompt = null;
+      } else if (typeof v === "string") {
+        const trimmed = v.trim();
+        patch.rolePrompt = trimmed === "" ? null : trimmed;
+      } else {
+        return jsonResponse({ error: "invalid_role_prompt" }, 400);
+      }
+    }
+    if (patch.toolProvider === void 0 && patch.gcpLocation === void 0 && patch.chatInterface === void 0 && patch.maxOutputTokens === void 0 && !("rolePrompt" in patch)) {
       return jsonResponse(
         {
           error: "empty_patch",
-          message: "Body must set at least one of tool_provider, gcp_location, chat_interface."
+          message: "Body must set at least one of tool_provider, gcp_location, chat_interface, max_output_tokens, role_prompt."
         },
         400
       );
@@ -1731,16 +1755,27 @@ function configureAiChat(opts) {
   ];
   const getProvider = (id) => toolProviders2.find((p) => p.id === id) ?? getToolProvider(id);
   const chatInterfaces = opts.chatInterfaces ?? BUILTIN_CHAT_INTERFACE_IDS.map((id) => ({ id }));
-  const tools = opts.rolePrompt ? {
+  const staticRolePrompt = opts.rolePrompt;
+  const tools = {
     tools: opts.tools.tools,
     async buildSystemBlocks(ctx) {
       const inner = await opts.tools.buildSystemBlocks(ctx);
-      const rolePrompt = opts.rolePrompt;
-      const role = typeof rolePrompt === "function" ? await rolePrompt(ctx) : rolePrompt;
-      if (!role || !role.trim()) return inner;
+      let role = null;
+      try {
+        const settings = await opts.persistence.getAiSettings();
+        if (settings.rolePrompt && settings.rolePrompt.trim()) {
+          role = settings.rolePrompt;
+        }
+      } catch {
+      }
+      if (!role && staticRolePrompt) {
+        const resolved = typeof staticRolePrompt === "function" ? await staticRolePrompt(ctx) : staticRolePrompt;
+        if (resolved && resolved.trim()) role = resolved;
+      }
+      if (!role) return inner;
       return [{ text: role, cached: true }, ...inner];
     }
-  } : opts.tools;
+  };
   const runAgentBound = async ({
     question,
     ctx,