fix: retry 429 with backoff; switch Tier B/C to claude-sonnet-4-6

- VertexOpenAIClient: retry on 429/503 up to 4 times with exponential
  backoff (2s/4s/8s/16s + jitter), respects Retry-After header
- Tier B/C default: zai-org/glm-5-maas → claude-sonnet-4-6 (much higher
  rate limits, still Vertex MaaS)
- /agent/execute: accept continueTask param to run a follow-up within
  the original task context without starting a fresh session

Made-with: Cursor

src/server.ts

@@ -359,13 +359,14 @@ app.post('/webhook/gitea', (req: Request, res: Response) => {
 const activeSessions = new Map<string, { stopped: boolean }>();
 
 app.post('/agent/execute', async (req: Request, res: Response) => {
-    const { sessionId, projectId, appName, appPath, giteaRepo, task } = req.body as {
+    const { sessionId, projectId, appName, appPath, giteaRepo, task, continueTask } = req.body as {
         sessionId?: string;
         projectId?: string;
         appName?: string;
         appPath?: string;
         giteaRepo?: string;
         task?: string;
+        continueTask?: string; // if set, appended as follow-up to the original task
     };
 
     if (!sessionId || !projectId || !appPath || !task) {
@@ -418,8 +419,14 @@ app.post('/agent/execute', async (req: Request, res: Response) => {
         return;
     }
 
+    // If continuing a previous task, combine into a single prompt so the agent
+    // understands what was already attempted.
+    const effectiveTask = continueTask
+        ? `Original task: ${task}\n\nFollow-up instruction: ${continueTask}`
+        : task!;
+
     // Run the streaming agent loop (fire and forget)
-    runSessionAgent(agentConfig, task, ctx, {
+    runSessionAgent(agentConfig, effectiveTask, ctx, {
         sessionId,
         projectId,
         vibnApiUrl,