Onboarding template overhaul + 6 evaluation-driven iterations

builds/typescript/adapters/index.ts

@@ -18,10 +18,13 @@ export function createModelAdapter(
     preferences.active_provider_profile?.trim() ||
     adapterConfig.default_provider_profile?.trim() ||
     "";
+  // Model resolution: per-provider default takes priority. If no per-provider
+  // default exists, use the adapter profile's built-in model — never fall back
+  // to the global default_model, since model IDs are provider-specific.
   const providerModel = activeProfile
     ? preferences.provider_default_models?.[activeProfile]?.trim()
     : undefined;
-  const preferenceModel = (providerModel ?? preferences.default_model).trim();
+  const preferenceModel = providerModel ?? "";
   const useAdapterModel =
     preferenceModel.length === 0 ||
     (preferenceModel === legacyBootstrapModel && selectedAdapterConfig.model !== legacyBootstrapModel);

builds/typescript/client_web/src/api/useGatewayChat.ts

@@ -181,7 +181,9 @@ export function useGatewayChat(options: UseGatewayChatOptions = {}): {
       conversationIdRef.current = restored.conversationId;
       backgroundStates.delete(cacheKey);
     } else {
-      setMessages(externalMessages);
+      // For draft conversations (no external ID), always start empty — externalMessages
+      // may be stale from the previous project's history that hasn't cleared yet.
+      setMessages(externalConversationId ? externalMessages : EMPTY_MESSAGES);
       setIsLoading(false);
       setError(null);
       setToolStatus(null);

builds/typescript/gateway/server.ts

@@ -30,13 +30,13 @@ import {
   savePreferences,
 } from "../config.js";
 import type { AdapterConfig, ClientMessageRequest, Preferences, RuntimeConfig } from "../contracts.js";
-import { runAgentLoop } from "../engine/loop.js";
-import { classifyProviderError } from "../engine/errors.js";
-import { formatSseEvent } from "../engine/stream.js";
-import { ToolExecutor } from "../engine/tool-executor.js";
-import { ensureGitReady } from "../git.js";
-import { auditLog } from "../logger.js";
-import { ensureAuthState, saveAuthState } from "../memory/auth-state.js";
+import { runAgentLoop } from "../engine/loop.js";
+import { classifyProviderError } from "../engine/errors.js";
+import { formatSseEvent } from "../engine/stream.js";
+import { ToolExecutor } from "../engine/tool-executor.js";
+import { ensureGitReady } from "../git.js";
+import { auditLog } from "../logger.js";
+import { ensureAuthState, saveAuthState } from "../memory/auth-state.js";
 import type { ConversationRepository } from "../memory/conversation-repository.js";
 import { MarkdownConversationStore } from "../memory/conversation-store-markdown.js";
 import { exportMemory } from "../memory/export.js";
@@ -209,15 +209,15 @@ export async function buildServer(rootDir = process.cwd()) {
 
   const app = Fastify({ logger: false, trustProxy: true });
   const approvalStore = new ApprovalStore();
-  const toolExecutor = new ToolExecutor(tools);
-  const conversations = new GatewayConversationService(createConversationRepository(runtimeConfig));
-  const projects = new GatewayProjectService(runtimeConfig.memory_root, { rootDir });
-  const skills = new GatewaySkillService(runtimeConfig.memory_root);
-  const signupRateLimiter = new FixedWindowRateLimiter(5, 5 * 60 * 1000);
-  const loginRateLimiter = new FixedWindowRateLimiter(10, 5 * 60 * 1000);
-  const refreshRateLimiter = new FixedWindowRateLimiter(30, 5 * 60 * 1000);
-  const signupBootstrapToken = process.env.PAA_AUTH_BOOTSTRAP_TOKEN?.trim();
-  const allowFirstSignupFromAnyIp = readBooleanEnv(process.env.PAA_AUTH_ALLOW_FIRST_SIGNUP_ANY_IP, false);
+  const toolExecutor = new ToolExecutor(tools);
+  const conversations = new GatewayConversationService(createConversationRepository(runtimeConfig));
+  const projects = new GatewayProjectService(runtimeConfig.memory_root, { rootDir });
+  const skills = new GatewaySkillService(runtimeConfig.memory_root);
+  const signupRateLimiter = new FixedWindowRateLimiter(5, 5 * 60 * 1000);
+  const loginRateLimiter = new FixedWindowRateLimiter(10, 5 * 60 * 1000);
+  const refreshRateLimiter = new FixedWindowRateLimiter(30, 5 * 60 * 1000);
+  const signupBootstrapToken = process.env.PAA_AUTH_BOOTSTRAP_TOKEN?.trim();
+  const allowFirstSignupFromAnyIp = readBooleanEnv(process.env.PAA_AUTH_ALLOW_FIRST_SIGNUP_ANY_IP, false);
   const persistAuthState = async (nextState: typeof authState): Promise<void> => {
     authState = await saveAuthState(runtimeConfig.memory_root, nextState);
   };
@@ -245,28 +245,28 @@ export async function buildServer(rootDir = process.cwd()) {
       return;
     }
 
-    if (!authState.account_initialized && !allowFirstSignupFromAnyIp) {
-      const signupAccess = evaluateSignupBootstrapAccess(
-        {
-          ip: request.ip,
-          headers: request.headers as Record<string, unknown>,
-        },
+    if (!authState.account_initialized && !allowFirstSignupFromAnyIp) {
+      const signupAccess = evaluateSignupBootstrapAccess(
+        {
+          ip: request.ip,
+          headers: request.headers as Record<string, unknown>,
+        },
         signupBootstrapToken
       );
       if (!signupAccess.allowed) {
         auditLog("auth.signup.denied", {
           reason: signupAccess.reason,
           ip: request.ip,
         });
-        reply.code(403).send({ error: signupAccess.reason });
-        return;
-      }
-    } else if (!authState.account_initialized && allowFirstSignupFromAnyIp) {
-      auditLog("auth.signup.bootstrap_override", {
-        reason: "allow_first_signup_any_ip",
-        ip: request.ip,
-      });
-    }
+        reply.code(403).send({ error: signupAccess.reason });
+        return;
+      }
+    } else if (!authState.account_initialized && allowFirstSignupFromAnyIp) {
+      auditLog("auth.signup.bootstrap_override", {
+        reason: "allow_first_signup_any_ip",
+        ip: request.ip,
+      });
+    }
 
     const parsed = authCredentialsSchema.safeParse(request.body);
     if (!parsed.success) {
@@ -438,25 +438,33 @@ export async function buildServer(rootDir = process.cwd()) {
     const projectId = isProjectMetadata(body.metadata) ? body.metadata.project.trim() : null;
     if (isProjectMetadata(body.metadata)) {
       await projects.attachConversation(body.metadata.project.trim(), conversationId);
-    }
-    const conversationSkillIds = conversations.getConversationSkills(conversationId) ?? [];
-    const projectSkillIds = projectId ? (await projects.getProjectSkills(projectId)) ?? [] : [];
-    const promptWithSkills = await skills.composePromptWithSkills(systemPrompt, [...projectSkillIds, ...conversationSkillIds]);
-
-    auditLog("skills.apply", {
-      conversation_id: conversationId,
-      project_id: projectId,
-      applied_skill_ids: promptWithSkills.applied,
-      missing_skill_ids: promptWithSkills.missing,
-      truncated: promptWithSkills.truncated,
-    });
-
-    const engineRequest = gatewayAdapter.buildEngineRequest({
-      conversationId,
-      correlationId: crypto.randomUUID(),
-      messages: conversations.buildConversationMessages(conversationId, promptWithSkills.prompt),
-      ...(body.metadata ? { clientMetadata: body.metadata } : {}),
-    });
+    }
+    const conversationSkillIds = conversations.getConversationSkills(conversationId) ?? [];
+    const projectSkillIds = projectId ? (await projects.getProjectSkills(projectId)) ?? [] : [];
+    const promptWithSkills = await skills.composePromptWithSkills(systemPrompt, [...projectSkillIds, ...conversationSkillIds]);
+
+    auditLog("skills.apply", {
+      conversation_id: conversationId,
+      project_id: projectId,
+      applied_skill_ids: promptWithSkills.applied,
+      missing_skill_ids: promptWithSkills.missing,
+      truncated: promptWithSkills.truncated,
+    });
+
+    // Inject project context so the AI knows which project it's operating in.
+    // Without this, the AI sees the base prompt but doesn't know which project
+    // files to read — it would read all projects and behave like BD+1.
+    const projectContext = projectId
+      ? `\n\n## Active Project\n\nYou are currently in the **${projectId}** project. Read this project's AGENT.md, spec.md, and plan.md from the documents/${projectId}/ folder. Stay focused on this domain — do not read or reference other projects unless the conversation specifically calls for cross-domain connections.`
+      : "";
+    const finalPrompt = promptWithSkills.prompt + projectContext;
+
+    const engineRequest = gatewayAdapter.buildEngineRequest({
+      conversationId,
+      correlationId: crypto.randomUUID(),
+      messages: conversations.buildConversationMessages(conversationId, finalPrompt),
+      ...(body.metadata ? { clientMetadata: body.metadata } : {}),
+    });
 
     reply.raw.writeHead(200, {
       "content-type": "text/event-stream",
@@ -990,6 +998,18 @@ export async function buildServer(rootDir = process.cwd()) {
         return;
       } else {
         nextPreferences.active_provider_profile = body.active_provider_profile;
+        // When switching providers, sync default_model to the new provider's
+        // per-provider default so display stays consistent. Model IDs are
+        // provider-specific — the global default_model should reflect the
+        // active provider's selection.
+        if (body.default_model === undefined) {
+          const newProviderModel = nextPreferences.provider_default_models?.[body.active_provider_profile];
+          const profileConfig = adapterConfig.provider_profiles?.[body.active_provider_profile];
+          const effectiveModel = newProviderModel ?? profileConfig?.model;
+          if (effectiveModel) {
+            nextPreferences.default_model = effectiveModel;
+          }
+        }
       }
     }
 
@@ -1414,34 +1434,34 @@ function serializeRefreshCookie(refreshToken: string, maxAgeSeconds: number, sec
     .join("; ");
 }
 
-function serializeRefreshCookieClear(secure: boolean): string {
-  return [
-    `${REFRESH_COOKIE_NAME}=`,
+function serializeRefreshCookieClear(secure: boolean): string {
+  return [
+    `${REFRESH_COOKIE_NAME}=`,
     "HttpOnly",
     "SameSite=Strict",
     "Path=/",
     "Max-Age=0",
     "Expires=Thu, 01 Jan 1970 00:00:00 GMT",
     secure ? "Secure" : "",
   ]
-    .filter((segment) => segment.length > 0)
-    .join("; ");
-}
-
-function readBooleanEnv(value: string | undefined, defaultValue = false): boolean {
-  const normalized = value?.trim().toLowerCase();
-  if (!normalized) {
-    return defaultValue;
-  }
-
-  return normalized === "1" || normalized === "true" || normalized === "yes" || normalized === "on";
-}
-
-function isSecureRequest(request: { headers: Record<string, unknown> }): boolean {
-  const forwardedProto = request.headers["x-forwarded-proto"];
-  if (typeof forwardedProto === "string" && forwardedProto.toLowerCase().includes("https")) {
-    return true;
-  }
+    .filter((segment) => segment.length > 0)
+    .join("; ");
+}
+
+function readBooleanEnv(value: string | undefined, defaultValue = false): boolean {
+  const normalized = value?.trim().toLowerCase();
+  if (!normalized) {
+    return defaultValue;
+  }
+
+  return normalized === "1" || normalized === "true" || normalized === "yes" || normalized === "on";
+}
+
+function isSecureRequest(request: { headers: Record<string, unknown> }): boolean {
+  const forwardedProto = request.headers["x-forwarded-proto"];
+  if (typeof forwardedProto === "string" && forwardedProto.toLowerCase().includes("https")) {
+    return true;
+  }
 
   return process.env.NODE_ENV === "production";
 }
@@ -1487,7 +1507,7 @@ class FixedWindowRateLimiter {
   }
 }
 
-function createConversationRepository(runtimeConfig: RuntimeConfig): ConversationRepository {
+function createConversationRepository(runtimeConfig: RuntimeConfig): ConversationRepository {
   switch (runtimeConfig.conversation_store) {
     case "markdown":
       return new MarkdownConversationStore(runtimeConfig.memory_root);