improved a lot of things

packages/ai-layer/src/client.ts

@@ -2,9 +2,9 @@ import Anthropic from '@anthropic-ai/sdk';
 
 // ── Model constants ───────────────────────────────────────────────────────────
 
-// claude-opus-4-5 is the current stable Opus 4 release.
-// Extended thinking is enabled in gateway.ts with budget_tokens.
-export const MODEL = 'claude-opus-4-5' as const;
+// claude-opus-4-7 uses adaptive thinking (thinking.type = 'adaptive').
+// It does NOT accept temperature, top_p, or top_k — those are omitted in gateway.ts.
+export const MODEL = 'claude-opus-4-7' as const;
 
 // ── Singleton client ──────────────────────────────────────────────────────────
 // The client is created once and shared. API key is injected from the server

packages/ai-layer/src/features/agent-qa.ts

@@ -40,7 +40,7 @@ export async function answerAgentQuestion(
 
   let parsed: unknown;
   try {
-    parsed = JSON.parse(response.text);
+    parsed = JSON.parse(response.text.replace(/^```(?:json)?\s*/i, "").replace(/\s*```\s*$/, "").trim());
   } catch (err) {
     throw new Error(
       `Design agent returned unparseable JSON: ${String(err)}. ` +

packages/ai-layer/src/features/artboard-query.ts

@@ -38,9 +38,11 @@ export async function queryCrossArtboard(
 
   // Returning empty results on parse failure is indistinguishable from "no match".
   // Throw so the UI can show an error rather than a misleading empty state.
+  // Strip markdown code fences that the model occasionally adds despite instructions.
+  const rawText = response.text.replace(/^```(?:json)?\s*/i, '').replace(/\s*```\s*$/, '').trim();
   let parsed: unknown;
   try {
-    parsed = JSON.parse(response.text);
+    parsed = JSON.parse(rawText);
   } catch (err) {
     throw new Error(
       `Artboard query returned unparseable JSON: ${String(err)}. ` +

packages/ai-layer/src/features/completion-zone.ts

@@ -60,7 +60,7 @@ export async function fillCompletionZone(
     });
 
     try {
-      const parsed = JSON.parse(response.text);
+      const parsed = JSON.parse(response.text.replace(/^```(?:json)?\s*/i, "").replace(/\s*```\s*$/, "").trim());
       return { proposedTree: parsed.proposedTree as unknown, description: String(parsed.description ?? ''), raw: response };
     } catch (parseErr) {
       lastError = new Error(

packages/ai-layer/src/features/drift-report.ts

@@ -70,7 +70,7 @@ export async function generateDriftReport(
   // compliance feature. Throw so the caller can surface the error clearly.
   let parsed: unknown;
   try {
-    parsed = JSON.parse(response.text);
+    parsed = JSON.parse(response.text.replace(/^```(?:json)?\s*/i, "").replace(/\s*```\s*$/, "").trim());
   } catch (err) {
     throw new Error(
       `Drift report returned unparseable JSON: ${String(err)}. ` +

packages/ai-layer/src/gateway.ts

@@ -73,10 +73,8 @@ export interface GatewayRequest {
   messages: Anthropic.Messages.MessageParam[];
   system?: Anthropic.Messages.TextBlockParam[];
   maxTokens?: number;
-  // NOTE: temperature is intentionally omitted — extended thinking rejects
-  // temperature, top_p, and top_k with a 400 error.
-  // NOTE: when thinking is enabled, max_tokens must be >= 16_000 and
-  // budget_tokens must be < max_tokens. We enforce this in complete().
+  // NOTE: temperature is intentionally omitted — claude-opus-4-7 with
+  // adaptive thinking rejects temperature, top_p, and top_k with a 400 error.
 }
 
 export interface GatewayResponse {
@@ -103,15 +101,10 @@ export class AIGateway {
     let lastError: Error | null = null;
     for (let attempt = 0; attempt < this.maxRetries; attempt++) {
       try {
-        // Extended thinking requires max_tokens >= 16_000.
-        // budget_tokens must be strictly less than max_tokens.
-        const maxTokens = Math.max(req.maxTokens ?? 4096, 16_000);
-        const budgetTokens = Math.floor(maxTokens * 0.8);
-
         const response = await this.client.messages.create({
           model: MODEL,
-          max_tokens: maxTokens,
-          thinking: { type: 'enabled', budget_tokens: budgetTokens },
+          max_tokens: req.maxTokens ?? 4096,
+          thinking: { type: 'adaptive' },
           ...(req.system !== undefined ? { system: req.system } : {}),
           messages: req.messages,
         });