opencrowd 0.15.3 → 0.15.4

package/bundle/opencrowd.js

@@ -15432,6 +15432,13 @@ var init_primitives = __esm({
 });
 
 // ../../packages/protocol/dist/outcomes.js
+function upstreamRefusalSummary(error53) {
+  const refusal = error53;
+  if (!refusal || typeof refusal !== "object" || refusal.name !== "HostedRequestError" || refusal.code !== "upstream_refused")
+    return void 0;
+  const detail = typeof refusal.message === "string" ? refusal.message.trim() : "";
+  return `The model provider refused this request before any payment, so nothing was charged${detail ? `: ${detail}` : "."} You can retry or send another message.`;
+}
 var RUN_FAILURE_MESSAGE, RUN_OUTCOMES, RunOutcomeSchema, RESUME_CAUSES, ResumeCauseSchema, RunStateSchema;
 var init_outcomes = __esm({
   "../../packages/protocol/dist/outcomes.js"() {
@@ -16101,6 +16108,18 @@ function estimateContextTokens(value) {
 function estimateModelInput(messages, tools) {
   return messages.reduce((total, message) => total + estimateContextTokens(message), estimateContextTokens(tools) + 32);
 }
+function isRequestTooLargeError(value, status) {
+  if (status === 413)
+    return true;
+  const error53 = value;
+  if (!error53 || typeof error53 !== "object")
+    return false;
+  if (["request_too_large", "payload_too_large", "request_entity_too_large", "body_too_large"].includes(String(error53.code)))
+    return true;
+  if (typeof error53.message === "string" && /request (?:body|entity|payload) (?:is )?too large|payload too large|request too large|body must be at most/i.test(error53.message))
+    return true;
+  return error53.error !== value && !!error53.error && isRequestTooLargeError(typeof error53.error === "string" ? { message: error53.error } : error53.error);
+}
 function isContextWindowError(value) {
   const error53 = value;
   if (!error53 || typeof error53 !== "object")
@@ -16111,11 +16130,12 @@ function isContextWindowError(value) {
     return true;
   return error53.error !== value && !!error53.error && isContextWindowError(typeof error53.error === "string" ? { message: error53.error } : error53.error);
 }
-var MODEL_REQUEST_MAX_BYTES, MODEL_MAX_MESSAGES, MODEL_GATEWAY_TIMEOUT_MS, MODEL_BRIDGE_TIMEOUT_MS, SERVICE_GATEWAY_TIMEOUT_MS, SERVICE_BRIDGE_TIMEOUT_MS, HOSTED_TOOL_BRIDGE_TIMEOUT_MS, HOSTED_CONTROL_TOOL_TIMEOUT_MS, SERVICE_RESPONSE_MAX_BYTES, DEFAULT_CONTEXT_WINDOW, ContextWindowExceeded;
+var MODEL_REQUEST_MAX_BYTES, MODEL_SELLER_REQUEST_MAX_BYTES, MODEL_MAX_MESSAGES, MODEL_GATEWAY_TIMEOUT_MS, MODEL_BRIDGE_TIMEOUT_MS, SERVICE_GATEWAY_TIMEOUT_MS, SERVICE_BRIDGE_TIMEOUT_MS, HOSTED_TOOL_BRIDGE_TIMEOUT_MS, HOSTED_CONTROL_TOOL_TIMEOUT_MS, SERVICE_RESPONSE_MAX_BYTES, DEFAULT_CONTEXT_WINDOW, ContextWindowExceeded;
 var init_context = __esm({
   "../../packages/protocol/dist/context.js"() {
     "use strict";
     MODEL_REQUEST_MAX_BYTES = 16 * 1024 * 1024;
+    MODEL_SELLER_REQUEST_MAX_BYTES = 5 * 1024 * 1024;
     MODEL_MAX_MESSAGES = 1e3;
     MODEL_GATEWAY_TIMEOUT_MS = 42e4;
     MODEL_BRIDGE_TIMEOUT_MS = 2 * MODEL_GATEWAY_TIMEOUT_MS + 3e4;
@@ -59738,7 +59758,7 @@ async function completeWithContext(session, messages, options) {
       options.state.retry = 0;
       return response;
     } catch (error53) {
-      const overflow = error53 instanceof ContextWindowExceeded || !options.hosted && isContextWindowError(error53);
+      const overflow = error53 instanceof ContextWindowExceeded || !options.hosted && (isContextWindowError(error53) || isRequestTooLargeError(error53));
       if (!overflow || (options.state.retry ?? 0) >= 2)
         throw error53;
       options.state.retry = (options.state.retry ?? 0) + 1;
@@ -59764,10 +59784,12 @@ async function compactContext(session, messages, options) {
   const count = (list2) => Math.ceil(list2.reduce((total, m) => total + measure(m).tokens, toolTokens) * scale);
   const tokensBefore = count(messages);
   const retry = options.state.retry ?? 0;
-  const trigger = Math.floor(limits.trigger * 0.5 ** retry);
-  const target = Math.min(limits.target, Math.floor(trigger * 0.5));
   const bytes = (list2) => list2.reduce((total, m) => total + measure(m).bytes, toolBytes);
-  const transportLimit = MODEL_REQUEST_MAX_BYTES - 64 * 1024;
+  const budget = options.maxRequestBytes ?? (options.hosted ? MODEL_SELLER_REQUEST_MAX_BYTES : MODEL_REQUEST_MAX_BYTES);
+  const byteCeiling = budget - 64 * 1024;
+  const trigger = retry ? Math.min(Math.floor(limits.trigger * 0.5 ** retry), Math.floor(tokensBefore * 0.5)) : limits.trigger;
+  const target = Math.min(limits.target, Math.floor(trigger * 0.5));
+  const transportLimit = retry ? Math.floor(Math.min(byteCeiling, bytes(messages) * 0.5)) : byteCeiling;
   if (tokensBefore <= trigger && bytes(messages) <= transportLimit && messages.length <= MODEL_MAX_MESSAGES && !retry)
     return { tokensBefore, tokensAfter: tokensBefore };
   const raw = messages.map((message) => JSON.stringify(message)).join("\n") + "\n";
@@ -59826,7 +59848,7 @@ async function compactContext(session, messages, options) {
       entry.message.content = await externalize(entry.message, entry.index);
   }
   const tokensAfter = count(current());
-  if (tokensAfter > trigger || bytes(current()) > transportLimit || list.length > MODEL_MAX_MESSAGES)
+  if (tokensAfter > Math.max(trigger, limits.trigger) || bytes(current()) > byteCeiling || list.length > MODEL_MAX_MESSAGES)
     throw new Error(`Context cannot fit after compaction: system instructions, tools and required references exceed the input budget. Original context: ${archivePath}`);
   messages.splice(0, messages.length, ...current());
   return { tokensBefore, tokensAfter, archivePath };
@@ -87133,8 +87155,8 @@ function createHostedProvider(options) {
       if (!operationId?.startsWith(`${options.runId}:llm:`))
         throw new Error("Hosted model operation scope is missing");
       const body = JSON.stringify({ runId: options.runId, sessionId: options.sessionId, operationId, messages: messages.map(({ contextArchive: _archive, ...message }) => message), tools });
-      if (Buffer.byteLength(body) > MODEL_REQUEST_MAX_BYTES)
-        throw new Error("Hosted request exceeds the transport byte limit");
+      if (Buffer.byteLength(body) > MODEL_SELLER_REQUEST_MAX_BYTES)
+        throw new ContextWindowExceeded(`Hosted request is ${Buffer.byteLength(body)} bytes, above the ${MODEL_SELLER_REQUEST_MAX_BYTES} a seller accepts`);
       const result = await new Promise((resolve9, reject) => {
         const req = request({
           socketPath: options.socketPath,
@@ -87201,10 +87223,14 @@ function rejectedRequest(status, text) {
       return new HostedRequestError("model_failed", status, String(body.message ?? "The model response failed after payment settled. Its receipt is saved."));
     if (body.paid !== false)
       return void 0;
-    if (isContextWindowError({ code: body.error, message: body.message }))
+    const upstream = body.upstream && typeof body.upstream === "object" && typeof body.upstream.code === "string" ? { status: Number(body.upstream.status) || status, code: body.upstream.code } : void 0;
+    const refusal = { code: upstream?.code ?? body.error, message: body.message };
+    if (isContextWindowError(refusal))
       return new ContextWindowExceeded(String(body.message ?? "Provider context window exceeded"));
+    if (isRequestTooLargeError(refusal, upstream?.status ?? status))
+      return new ContextWindowExceeded(String(body.message ?? "The model provider refused this request as too large"));
     const code = typeof body.error === "string" && body.error ? body.error : `http_${status}`;
-    return new HostedRequestError(code, status, typeof body.message === "string" && body.message ? body.message : `The model request was rejected (${code})`);
+    return new HostedRequestError(code, status, typeof body.message === "string" && body.message ? body.message : `The model request was rejected (${code})`, upstream);
   } catch {
     return void 0;
   }
@@ -87308,12 +87334,18 @@ var init_hosted_provider = __esm({
     HostedRequestError = class extends Error {
       code;
       status;
-      constructor(code, status, message) {
+      upstream;
+      constructor(code, status, message, upstream) {
         super(message);
         this.code = code;
         this.status = status;
+        this.upstream = upstream;
         this.name = "HostedRequestError";
       }
+      /** What the owner reads as the run's summary; undefined keeps the generic copy. */
+      get summary() {
+        return upstreamRefusalSummary(this);
+      }
     };
     HOSTED_TOOL_REQUEST_BYTES = 1024 * 1024;
     HostedToolTimeout = class extends Error {
@@ -87660,7 +87692,7 @@ var init_hosted_economy = __esm({
           throw new SpendingDeclined();
         if (reply.code === "approval_required")
           throw new RuntimePause("waiting_for_approval", purchaseId, "Spending approval required. Review the quoted amount to continue.");
-        if (reply.code === "expected_payment_required") {
+        if (reply.code === "expected_payment_required" || reply.code === "upstream_refused") {
           return this.fetchUnpaid(request2.url, method, body, request2.headers);
         }
         if (reply.code === "funding_required") {
@@ -88392,7 +88424,7 @@ var init_worker = __esm({
           } else if (error53 instanceof SpendingDeclined)
             await this.finish(runId, run, "user_stopped", error53.message);
           else
-            await this.finish(runId, run, "failed", RUN_FAILURE_MESSAGE);
+            await this.finish(runId, run, "failed", upstreamRefusalSummary(error53) ?? RUN_FAILURE_MESSAGE);
         }
       }
       /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencrowd",
-  "version": "0.15.3",
+  "version": "0.15.4",
   "description": "A CLI agent with its own wallet — discovers x402 services, pays in USDC, and completes tasks under a local budget",
   "type": "module",
   "license": "MIT",