@columenlabs/lumen-ai 0.1.96 → 0.1.97

package/dist/assistant/engine.js

@@ -1108,6 +1108,10 @@ export class Assistant {
             // the SDK does not know and its accumulator throws on them. The final response rides on the
             // completed, incomplete or failed event; every other event but text deltas is skipped.
             let partial = '';
+            // Event types seen and the items finished one by one: a plan sign-in's final event may carry
+            // an empty `output`, its items having come only as `response.output_item.done`.
+            const seen = {};
+            const items = [];
             const res = await (async () => {
                 const stream = await client.responses.create({
                     model: session.model.id,
@@ -1122,6 +1126,9 @@ export class Assistant {
                 }, signal ? { signal } : {});
                 let done;
                 for await (const e of stream) {
+                    seen[e.type] = (seen[e.type] ?? 0) + 1;
+                    if (e.type === 'response.output_item.done' && e.item)
+                        items.push(e.item);
                     if (e.type === 'response.output_text.delta' && typeof e.delta === 'string') {
                         partial += e.delta;
                         emit('text', { delta: e.delta });
@@ -1147,6 +1154,14 @@ export class Assistant {
                 if (partial)
                     sink.push({ kind: 'text', text: partial });
                 const code = res.error?.code ?? '';
+                console.error(JSON.stringify({
+                    event: 'assistant_openai_failed',
+                    model: session.model.id,
+                    plan,
+                    code,
+                    message: res.error?.message?.slice(0, 300) ?? null,
+                    events: seen,
+                }));
                 throw new AssistantError(/usage_limit/.test(code)
                     ? 'Your ChatGPT plan reached its usage limit for now: try again later, or pick another model.'
                     : 'OpenAI could not finish the answer.', 'upstream');
@@ -1164,14 +1179,36 @@ export class Assistant {
                 o.context?.(u.input_tokens + u.output_tokens);
             }
             o.called();
-            const cut = res.incomplete_details?.reason === 'max_output_tokens' ? ' …(cut)' : '';
-            if (cut)
-                emit('text', { delta: cut });
+            const output = (res.output?.length ? res.output : items);
+            res.output = output;
             // Raw events carry no `output_text`: the words are the output messages' text parts.
-            const words = res.output
+            const words = output
                 .flatMap((b) => (b.type === 'message' ? b.content : []))
                 .map((c) => (c.type === 'output_text' ? c.text : ''))
-                .join('');
+                .join('') || partial;
+            const hasCall = output.some((b) => b.type === 'function_call');
+            if (!words && !hasCall) {
+                // Nothing to show: say why instead of an empty or bare "cut" answer, and log the shape.
+                console.error(JSON.stringify({
+                    event: 'assistant_openai_empty',
+                    model: session.model.id,
+                    plan,
+                    status: res.status,
+                    incomplete: res.incomplete_details ?? null,
+                    usage: res.usage ?? null,
+                    events: seen,
+                    output: output.map((b) => b.type),
+                }));
+                if (!answer)
+                    throw new AssistantError(res.incomplete_details?.reason
+                        ? `OpenAI stopped before answering (${res.incomplete_details.reason}).`
+                        : 'OpenAI sent no answer.', 'upstream');
+            }
+            const cut = res.incomplete_details?.reason === 'max_output_tokens' && (words || hasCall)
+                ? ' …(cut)'
+                : '';
+            if (cut)
+                emit('text', { delta: cut });
             const text = words + cut;
             if (text) {
                 answer += (answer ? '\n\n' : '') + text;
@@ -1179,8 +1216,8 @@ export class Assistant {
                 emit('text_end', {});
             }
             // Reasoning, words and calls go back as they came, so the model keeps its own thread.
-            input.push(...res.output);
-            const calls = res.output.filter((b) => b.type === 'function_call');
+            input.push(...output);
+            const calls = output.filter((b) => b.type === 'function_call');
             if (calls.length === 0)
                 break;
             const results = await Promise.all(calls.map((c) => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@columenlabs/lumen-ai",
-  "version": "0.1.96",
+  "version": "0.1.97",
   "description": "Lumen's assistant and agent parts: a chat assistant's server (Claude tiers, tools over MCP, the web, hand off, spend caps), serving an MCP server from a Worker, read-only tool sets and metered tool calls.",
   "license": "MIT",
   "type": "module",
@@ -41,8 +41,8 @@
   },
   "dependencies": {
     "@anthropic-ai/sdk": "0.133.0",
-    "@columenlabs/lumen-ops": "0.1.96",
-    "@columenlabs/lumen-ui-types": "0.1.96",
+    "@columenlabs/lumen-ops": "0.1.97",
+    "@columenlabs/lumen-ui-types": "0.1.97",
     "openai": "7.32.0",
     "zod": "^4.3.6"
   },