@columenlabs/lumen-ai 0.1.110 → 0.1.112

package/dist/assistant/engine.js

@@ -155,11 +155,11 @@ function openaiStreamError(code, plan) {
         return 'OpenAI declined to answer this message.';
     return `OpenAI could not answer (${code || 'no reason given'}). It is logged; try again or pick another model.`;
 }
-/** Set once a plan backend refuses `max_output_tokens`, for the rest of the isolate. */
-let planCapRefused = false;
-const refusesCap = (err) => err instanceof OpenAI.APIError &&
-    err.status === 400 &&
-    /max_output_tokens/.test(`${err.message} ${err.error?.param ?? ''}`);
+/** A tool's schema without JSON Schema's `$schema` line, which zod writes and no provider needs. */
+const withoutMeta = (schema) => {
+    const { $schema: _, ...rest } = schema ?? {};
+    return rest;
+};
 /** Whether a response's output has words or a call to show. */
 const shows = (output) => output.some((b) => b.type === 'function_call' ||
     (b.type === 'message' && b.content.some((c) => c.type === 'output_text' && c.text)));
@@ -564,7 +564,10 @@ export class Assistant {
         return [...fns.entries()].map(([name, f]) => ({
             name,
             description: `${f.draft ? "DRAFT ONLY: this call changes nothing. It puts a card on the person's screen that they Save or Dismiss. " : f.write ? 'WRITES AT ONCE: this call makes the change now; the person can undo it. ' : ''}${f.def.title ? `${f.def.title}. ` : ''}${f.def.description ?? ''}`.slice(0, 4000),
-            input_schema: { type: 'object', ...f.def.inputSchema },
+            input_schema: {
+                type: 'object',
+                ...withoutMeta(f.def.inputSchema),
+            },
         }));
     }
     // ---------------------------------------------------------------- the turn
@@ -1115,11 +1118,11 @@ export class Assistant {
             .filter((m) => m.role !== 'assistant' || m.content.replace('…(cut)', '').trim() !== '');
         let answer = '';
         let force = o.first?.type === 'tool' ? { type: 'function', name: o.first.name } : undefined;
-        // A plan sign-in takes function tools only inside a namespace. Its output cap is sent too,
-        // unless the backend refused one before (then it keeps its own).
-        const sent = (only, cap) => {
+        // A plan sign-in takes function tools only inside a namespace, and no output cap
+        // (`max_output_tokens` is unsupported there: developers.openai.com/siwc, Preview limitations).
+        const sent = (only) => {
             const fns = only ? fnTools.filter((t) => t.name === only) : fnTools;
-            const out = cap ? { max_output_tokens: session.tier.out ?? DEFAULT_OUT } : {};
+            const out = plan ? {} : { max_output_tokens: session.tier.out ?? DEFAULT_OUT };
             if (!plan)
                 return { tools: fns, ...out };
             return {
@@ -1164,20 +1167,20 @@ export class Assistant {
                 }));
                 return new AssistantError(openaiStreamError(er.code || er.type || '', plan), 'upstream');
             };
-            // One call of this round; asked again (fresh stream state) when the plan refuses the cap or
-            // spends its whole output on reasoning.
-            const ask = (cap, eff) => {
+            // One call of this round; asked again (fresh stream state) when it spends its whole output
+            // on reasoning.
+            const ask = (eff) => {
                 partial = '';
                 seen = {};
                 items = [];
-                return askOnce(cap, eff);
+                return askOnce(eff);
             };
-            const askOnce = (cap, eff) => (async () => {
+            const askOnce = (eff) => (async () => {
                 const stream = await client.responses.create({
                     model: session.model.id,
                     instructions,
                     input,
-                    ...sent(only, cap),
+                    ...sent(only),
                     store: false,
                     stream: true,
                     include: ['reasoning.encrypted_content'],
@@ -1220,24 +1223,13 @@ export class Assistant {
                 }
                 throw err;
             });
-            let cap = !plan || !planCapRefused;
-            let res;
-            try {
-                res = await ask(cap, effort);
-            }
-            catch (err) {
-                // A plan backend that refuses an output cap keeps its own: ask once more without it.
-                if (!(plan && cap && refusesCap(err)))
-                    throw err;
-                planCapRefused = true;
-                cap = false;
-                console.error(JSON.stringify({ event: 'assistant_openai_cap_refused', model: session.model.id }));
-                res = await ask(cap, effort);
-            }
+            let res = await ask(effort);
             // A reasoning model can spend the whole output on thinking and say nothing: ask once more
-            // with less thinking before giving up.
+            // with less thinking before giving up. One that stopped before writing anything is not asked
+            // again: less thinking cannot help it.
             if (res.status === 'incomplete' &&
                 res.incomplete_details?.reason === 'max_output_tokens' &&
+                (res.usage?.output_tokens ?? 0) > 0 &&
                 !partial &&
                 !shows(res.output?.length ? res.output : items) &&
                 effort !== 'low') {
@@ -1245,12 +1237,11 @@ export class Assistant {
                     event: 'assistant_openai_retry',
                     model: session.model.id,
                     plan,
-                    cap,
                     why: 'no_words',
                     usage: res.usage ?? null,
                     events: seen,
                 }));
-                res = await ask(cap, 'low');
+                res = await ask('low');
             }
             if (res.status === 'failed') {
                 if (partial)
@@ -1301,11 +1292,13 @@ export class Assistant {
                     parts: output.flatMap((b) => b.type === 'message' ? b.content.map((c) => c.type) : []),
                 }));
                 if (!answer)
-                    throw new AssistantError(res.incomplete_details?.reason === 'max_output_tokens'
-                        ? 'OpenAI spent the whole answer thinking and wrote nothing. Try again, or pick a smaller tier.'
-                        : res.incomplete_details?.reason
-                            ? `OpenAI stopped before answering (${res.incomplete_details.reason}).`
-                            : 'OpenAI sent no answer.', 'upstream');
+                    throw new AssistantError(res.incomplete_details?.reason === 'max_output_tokens' && !res.usage?.output_tokens
+                        ? 'OpenAI stopped before reading the message. Try again, or pick another model for this tier.'
+                        : res.incomplete_details?.reason === 'max_output_tokens'
+                            ? 'OpenAI spent the whole answer thinking and wrote nothing. Try again, or pick a smaller tier.'
+                            : 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)'

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@columenlabs/lumen-ai",
-  "version": "0.1.110",
+  "version": "0.1.112",
   "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": "0.1.110",
-    "@columenlabs/lumen-ui-types": "0.1.110",
+    "@columenlabs/lumen-ops": "0.1.112",
+    "@columenlabs/lumen-ui-types": "0.1.112",
     "openai": "7.32.0",
     "zod": "^4.3.6"
   },