@consultancy-outfit/ai-draftsman-cli 1.1.1 → 1.1.2

package/dist/index.js

@@ -719,6 +719,19 @@ var VALID_TOOL_NAME = /^[a-zA-Z0-9_.-]+$/;
 var INVALID_TOOL_NAME = "invalid_tool";
 var MAX_RETRIES = 4;
 var MAX_WAIT_SECONDS = 90;
+var UPSTREAM_BASE_WAIT_SECONDS = 5;
+var UPSTREAM_RATE_LIMIT = /RESOURCE_EXHAUSTED|RATE_LIMIT|\b429\b|quota exceeded|too many requests/i;
+var UpstreamRateLimit = class extends Error {
+};
+function isUpstreamRateLimit(message, code) {
+  return /^(resource_exhausted|rate_limit(_exceeded)?)$/i.test(code ?? "") || UPSTREAM_RATE_LIMIT.test(message);
+}
+function upstreamRateLimitError(message) {
+  return new CliError(
+    `The model provider is rate limiting requests (${message.replace(/\s+/g, " ").trim()}). Retries did not get through. Wait a minute and try again, or pick another model with /agents.`,
+    6
+  );
+}
 function retryDelaySeconds(response, body, attempt) {
   const header = Number(response.headers.get("retry-after"));
   if (Number.isFinite(header) && header > 0) return header;
@@ -760,26 +773,40 @@ async function streamChat(config, request, options = {}) {
     stream_options: { include_usage: true },
     ...request.tools?.length ? { tools: request.tools, tool_choice: "auto" } : {}
   });
-  let response;
+  const waitOut = async (seconds, who) => {
+    const pause = options.sleep ?? sleep;
+    for (let left = seconds; left > 0; left--) {
+      options.onStatus?.(`Rate limited by ${who} \xB7 retrying in ${left}s (esc to cancel)`);
+      await pause(1e3, options.signal);
+    }
+    options.onStatus?.(void 0);
+  };
+  let upstreamRetries = 0;
   for (let attempt = 0; ; attempt++) {
-    response = await fetch(new URL("/v1/chat/completions", config.baseUrl), {
+    const response = await fetch(new URL("/v1/chat/completions", config.baseUrl), {
       method: "POST",
       headers,
       body,
       signal: options.signal
     });
-    if (response.status !== 429) break;
-    const text2 = await response.text().catch(() => "");
-    const wait = retryDelaySeconds(response, text2, attempt);
-    if (attempt >= MAX_RETRIES || wait > MAX_WAIT_SECONDS) throw rateLimitError(text2, wait);
-    const pause = options.sleep ?? sleep;
-    for (let left = wait; left > 0; left--) {
-      options.onStatus?.(`Rate limited by the gateway \xB7 retrying in ${left}s (esc to cancel)`);
-      await pause(1e3, options.signal);
+    if (response.status === 429) {
+      const text2 = await response.text().catch(() => "");
+      const wait = retryDelaySeconds(response, text2, attempt);
+      if (attempt >= MAX_RETRIES || wait > MAX_WAIT_SECONDS) throw rateLimitError(text2, wait);
+      await waitOut(wait, "the gateway");
+      continue;
+    }
+    try {
+      if (!response.ok) throw await httpError(response);
+      return await readCompletion(response, options);
+    } catch (error) {
+      if (!(error instanceof UpstreamRateLimit)) throw error;
+      if (upstreamRetries >= MAX_RETRIES) throw upstreamRateLimitError(error.message);
+      await waitOut(UPSTREAM_BASE_WAIT_SECONDS * 2 ** upstreamRetries++, "the model provider");
     }
-    options.onStatus?.(void 0);
   }
-  if (!response.ok) throw await httpError(response);
+}
+async function readCompletion(response, options) {
   if (!response.body) throw new CliError("The server sent an empty response.", 8);
   let content = "";
   let finishReason = "stop";
@@ -798,7 +825,12 @@ async function streamChat(config, request, options = {}) {
       continue;
     }
     if (chunk.error) {
-      throw new CliError(chunk.error.message ?? "The model request failed.", chunk.error.code === "resource_exhausted" ? 6 : 8);
+      const message = chunk.error.message ?? "The model request failed.";
+      if (isUpstreamRateLimit(message, chunk.error.code)) {
+        if (!content && calls.size === 0) throw new UpstreamRateLimit(message);
+        throw upstreamRateLimitError(message);
+      }
+      throw new CliError(message, 8);
     }
     if (chunk.usage) usage = chunk.usage;
     const choice2 = chunk.choices?.[0];
@@ -865,6 +897,7 @@ async function httpError(response) {
   } catch {
   }
   const detail = message || response.statusText;
+  if (response.status !== 401 && response.status !== 402 && /RESOURCE_EXHAUSTED/i.test(text2)) return new UpstreamRateLimit(detail);
   if (response.status === 401) return new CliError("Your API key was rejected. Run /login to sign in again.", 2);
   if (response.status === 402) return new CliError(`Spend cap reached: ${detail}`, 7);
   if (response.status === 403) return new CliError(`Not allowed: ${detail}`, 4);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@consultancy-outfit/ai-draftsman-cli",
-  "version": "1.1.1",
+  "version": "1.1.2",
   "description": "AI Draftsman standalone CLI",
   "type": "module",
   "files": [