zelari-code 2.37.1 → 2.37.2

package/dist/cli/main.bundled.js

@@ -38416,7 +38416,7 @@ var CORE_VERSION;
 var init_version = __esm({
   "packages/core/dist/version.js"() {
     "use strict";
-    CORE_VERSION = "2.37.1";
+    CORE_VERSION = "2.37.2";
   }
 });
 
@@ -48283,6 +48283,7 @@ function frozenProfile(input) {
   Object.freeze(input.buildRecovery);
   Object.freeze(input.sampling);
   Object.freeze(input.compaction);
+  if (input.stream) Object.freeze(input.stream);
   return Object.freeze(input);
 }
 function resolveHarnessProfile(model, providerId) {
@@ -48355,6 +48356,18 @@ var init_capabilities = __esm({
       buildRecovery: { forceToolChoice: true, maxForcedTurns: 1 },
       sampling: { temperature: 0.7 },
       compaction: { ...SHARED_COMPACTION },
+      // Copied from Grok Build (`~/.grok/config.toml` + user-guide):
+      //   inference_idle_timeout_secs = 600
+      //   max_retries = 8
+      //   xAI SDK timeout = 3600s on reasoning models
+      // grok-4.6 xhigh reasons with hidden tokens + SSE keep-alives and no
+      // content for minutes; a 5-min useful-token idle kills BUILD.
+      stream: {
+        idleMs: 6e5,
+        firstTokenIdleMs: 6e5,
+        maxMs: 36e5,
+        maxRetries: 8
+      },
       profile: "grok"
     });
     MINIMAX_M3_CAPS = frozenProfile({
@@ -50979,18 +50992,40 @@ var openai_compatible_exports = {};
 __export(openai_compatible_exports, {
   PROVIDER_CONNECT_TIMEOUT_MS: () => PROVIDER_CONNECT_TIMEOUT_MS,
   PROVIDER_ENDPOINTS: () => PROVIDER_ENDPOINTS,
+  PROVIDER_FIRST_TOKEN_IDLE_MS: () => PROVIDER_FIRST_TOKEN_IDLE_MS,
   PROVIDER_STREAM_IDLE_MS: () => PROVIDER_STREAM_IDLE_MS,
   PROVIDER_STREAM_MAX_MS: () => PROVIDER_STREAM_MAX_MS,
   dataUriFromImage: () => dataUriFromImage,
+  glmModelLooksVision: () => glmModelLooksVision,
+  isGlmCodingEndpoint: () => isGlmCodingEndpoint,
+  isTextOnlyContentRejection: () => isTextOnlyContentRejection,
   modelSupportsVision: () => modelSupportsVision,
   openaiCompatibleProvider: () => openaiCompatibleProvider,
   parseCachedPromptTokens: () => parseCachedPromptTokens,
   providerConfigFor: () => providerConfigFor,
   providerFromEnv: () => providerFromEnv,
   readChunkWithTimeout: () => readChunkWithTimeout,
+  resetTextOnlyVisionMemory: () => resetTextOnlyVisionMemory,
   resolveActiveProvider: () => resolveActiveProvider2,
-  resolveBaseUrl: () => resolveBaseUrl
-});
+  resolveBaseUrl: () => resolveBaseUrl,
+  resolveStreamTimeouts: () => resolveStreamTimeouts
+});
+function resolveStreamTimeouts(capabilities) {
+  const envIdle = process.env.ZELARI_PROVIDER_STREAM_IDLE_MS ?? process.env.ZELARI_PROVIDER_TIMEOUT_MS;
+  const envFirst = process.env.ZELARI_PROVIDER_FIRST_TOKEN_IDLE_MS;
+  const envMax = process.env.ZELARI_PROVIDER_STREAM_MAX_MS;
+  const envRetries = process.env.ZELARI_PROVIDER_MAX_RETRIES;
+  const idleMs = envIdle ? PROVIDER_STREAM_IDLE_MS : capabilities.stream?.idleMs ?? PROVIDER_STREAM_IDLE_MS;
+  const firstRaw = envFirst ? PROVIDER_FIRST_TOKEN_IDLE_MS : capabilities.stream?.firstTokenIdleMs ?? PROVIDER_FIRST_TOKEN_IDLE_MS;
+  const maxMs = envMax ? PROVIDER_STREAM_MAX_MS : capabilities.stream?.maxMs ?? PROVIDER_STREAM_MAX_MS;
+  const maxRetries = envRetries ? MAX_RETRIES : capabilities.stream?.maxRetries ?? MAX_RETRIES;
+  return {
+    idleMs,
+    firstTokenIdleMs: Math.max(firstRaw, idleMs),
+    maxMs,
+    maxRetries
+  };
+}
 function abortableSleep(ms, signal) {
   return new Promise((resolve9) => {
     if (signal?.aborted) return resolve9();
@@ -51009,6 +51044,12 @@ function isTimeoutAbortMessage(msg) {
   const m = msg.toLowerCase();
   return m.includes("aborted due to timeout") || m.includes("timeout") || m.includes("the operation was aborted");
 }
+function formatStreamIdleError(elapsedMs, budgetMs, kind2) {
+  const elapsedS = Math.max(1, Math.round(elapsedMs / 1e3));
+  const budgetS = Math.max(1, Math.round(budgetMs / 1e3));
+  const why = kind2 === "keep-alive" ? "no content tokens \u2014 keep-alive frames don't count" : "no tokens";
+  return `Provider stream idle for ${elapsedS}s of ${budgetS}s (${why}). The model/gateway stalled \u2014 try again or switch model. Override with ZELARI_PROVIDER_STREAM_IDLE_MS or ZELARI_PROVIDER_FIRST_TOKEN_IDLE_MS.`;
+}
 async function readChunkWithTimeout(reader, opts) {
   if (opts.signal?.aborted) {
     throw new Error("aborted");
@@ -51022,9 +51063,7 @@ async function readChunkWithTimeout(reader, opts) {
   }
   const idleElapsed = now - opts.lastUsefulAt();
   if (idleElapsed >= opts.idleMs) {
-    throw new Error(
-      `Provider stream idle for ${Math.round(idleElapsed / 1e3)}s (no content tokens \u2014 keep-alive frames don't count). The model/gateway stalled \u2014 try again or switch model. Override with ZELARI_PROVIDER_STREAM_IDLE_MS.`
-    );
+    throw new Error(formatStreamIdleError(idleElapsed, opts.idleMs, "keep-alive"));
   }
   const waitMs = Math.min(opts.idleMs - idleElapsed, remaining);
   let idleTimer;
@@ -51034,11 +51073,8 @@ async function readChunkWithTimeout(reader, opts) {
       reader.read(),
       new Promise((_, reject) => {
         idleTimer = setTimeout(() => {
-          reject(
-            new Error(
-              `Provider stream idle for ${Math.round(waitMs / 1e3)}s (no tokens). The model/gateway stalled \u2014 try again or switch model. Override with ZELARI_PROVIDER_STREAM_IDLE_MS.`
-            )
-          );
+          const elapsed = Date.now() - opts.lastUsefulAt();
+          reject(new Error(formatStreamIdleError(elapsed, opts.idleMs, "silence")));
         }, waitMs);
         if (opts.signal) {
           onAbort = () => reject(new Error("aborted"));
@@ -51060,9 +51096,52 @@ function backoffDelay(attempt, retryAfterHeader) {
   }
   return Math.min(BACKOFF_BASE_MS * 2 ** attempt, BACKOFF_CAP_MS);
 }
-function modelSupportsVision(_model) {
-  const force = process.env.ZELARI_VISION;
-  return !(force === "0" || force === "false" || force === "off");
+function visionMemoryKey(model, probe) {
+  return `${probe?.providerId ?? ""}::${model}::${probe?.baseUrl ?? ""}`;
+}
+function resetTextOnlyVisionMemory() {
+  textOnlyVisionMemory.clear();
+}
+function isGlmCodingEndpoint(baseUrl) {
+  if (!baseUrl) return false;
+  return /\/coding(\/|$)/i.test(baseUrl);
+}
+function glmModelLooksVision(model) {
+  const n = model.trim().toLowerCase();
+  if (!n) return false;
+  if (/(?:^|[-_/.])(?:vl|vision)(?:[-_/.]|$)/.test(n)) return true;
+  return /^glm[-_.]?[\w.]*v(?:-|$)/.test(n);
+}
+function glmModelIsTextOnly(model) {
+  const n = model.trim().toLowerCase();
+  if (!n.startsWith("glm")) return false;
+  return !glmModelLooksVision(n);
+}
+function isTextOnlyContentRejection(status, body) {
+  if (status !== 400) return false;
+  if (!/messages\.content\.type/i.test(body)) return false;
+  return /allowed values/i.test(body) || /\[\s*['"]text['"]\s*\]/.test(body) || /取值范围/.test(body) || /is invalid/i.test(body);
+}
+function messagesHaveImageUrl(messages) {
+  for (const m of messages) {
+    const content = m.content;
+    if (!Array.isArray(content)) continue;
+    for (const part of content) {
+      if (part && typeof part === "object" && part.type === "image_url") {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+function modelSupportsVision(model, probe) {
+  const force = (process.env.ZELARI_VISION ?? "").trim().toLowerCase();
+  if (force === "0" || force === "false" || force === "off") return false;
+  if (force === "1" || force === "true" || force === "on") return true;
+  if (textOnlyVisionMemory.has(visionMemoryKey(model, probe))) return false;
+  if (isGlmCodingEndpoint(probe?.baseUrl)) return false;
+  if (glmModelIsTextOnly(model)) return false;
+  return true;
 }
 function dataUriFromImage(img) {
   return `data:${img.mime};base64,${img.dataBase64}`;
@@ -51182,29 +51261,38 @@ function positiveEnvInt(name) {
 function openaiCompatibleProvider(config2) {
   return async function* (params) {
     const capabilities = capabilitiesFor(params.model, config2.providerId);
-    const vision = modelSupportsVision(params.model);
+    const streamTimeouts = resolveStreamTimeouts(capabilities);
+    const visionProbe = {
+      providerId: config2.providerId,
+      baseUrl: config2.baseUrl
+    };
+    let vision = modelSupportsVision(params.model, visionProbe);
     const msgsIn = params.messages;
-    let toolRunImages = [];
-    const messages = msgsIn.flatMap((m, i) => {
-      const cacheable = !(m.role === "user" && m.images && m.images.length > 0);
-      if (cacheable) {
-        const cached2 = messageMappingCache.get(m);
-        if (cached2) return [cached2];
-      }
-      const mapped = mapAgentMessage(m, vision);
-      if (cacheable) messageMappingCache.set(m, mapped);
-      if (m.role === "tool") {
-        if (m.images && m.images.length > 0) toolRunImages.push(...m.images);
-        const next = msgsIn[i + 1];
-        const runEnds = !next || next.role !== "tool";
-        if (runEnds && vision && toolRunImages.length > 0) {
-          const followUp = imagesFollowUpMessage(toolRunImages);
-          toolRunImages = [];
-          return [mapped, followUp];
-        }
-      }
-      return [mapped];
-    });
+    const wireMessages = (visionFlag) => {
+      let toolRunImages = [];
+      return msgsIn.flatMap((m, i) => {
+        const cacheable = visionFlag && !(m.role === "user" && m.images && m.images.length > 0);
+        if (cacheable) {
+          const cached2 = messageMappingCache.get(m);
+          if (cached2) return [cached2];
+        }
+        const mapped = mapAgentMessage(m, visionFlag);
+        if (cacheable) messageMappingCache.set(m, mapped);
+        if (m.role === "tool") {
+          if (m.images && m.images.length > 0) toolRunImages.push(...m.images);
+          const next = msgsIn[i + 1];
+          const runEnds = !next || next.role !== "tool";
+          if (runEnds && visionFlag && toolRunImages.length > 0) {
+            const followUp = imagesFollowUpMessage(toolRunImages);
+            toolRunImages = [];
+            return [mapped, followUp];
+          }
+          if (runEnds) toolRunImages = [];
+        }
+        return [mapped];
+      });
+    };
+    let messages = wireMessages(vision);
     const generation = params.generation;
     const body = {
       // Use `params.model` (per-call override from AgentHarness, e.g. for
@@ -51254,6 +51342,7 @@ function openaiCompatibleProvider(config2) {
       const recoveryAttempt = generation?.recoveryAttempt ?? 1;
       const forceRecoveryTool = generation?.toolChoice === "required" && capabilities.buildRecovery.forceToolChoice && recoveryAttempt <= capabilities.buildRecovery.maxForcedTurns;
       body.tool_choice = forceRecoveryTool ? "required" : "auto";
+      if (config2.providerId === "glm") body.tool_stream = true;
     }
     const headers3 = {
       "Content-Type": "application/json",
@@ -51268,7 +51357,7 @@ function openaiCompatibleProvider(config2) {
     let response;
     let lastErrText = "";
     let lastStatus = 0;
-    for (let attempt = 0; attempt <= MAX_RETRIES; attempt += 1) {
+    for (let attempt = 0; attempt <= streamTimeouts.maxRetries; attempt += 1) {
       if (params.signal?.aborted) {
         yield { kind: "error", message: "aborted" };
         return;
@@ -51309,7 +51398,7 @@ function openaiCompatibleProvider(config2) {
           yield { kind: "error", message: "aborted" };
           return;
         }
-        const maxAttempts = isTimeoutAbortMessage(lastErrText) ? Math.min(1, MAX_RETRIES) : MAX_RETRIES;
+        const maxAttempts = isTimeoutAbortMessage(lastErrText) ? Math.min(1, streamTimeouts.maxRetries) : streamTimeouts.maxRetries;
         if (attempt < maxAttempts) {
           await abortableSleep(backoffDelay(attempt, null), params.signal);
           continue;
@@ -51320,7 +51409,14 @@ function openaiCompatibleProvider(config2) {
       if (response.ok && response.body) break;
       lastStatus = response.status;
       lastErrText = await response.text().catch(() => "");
-      if (!RETRYABLE_STATUSES.has(response.status) || attempt >= MAX_RETRIES) break;
+      if (vision && isTextOnlyContentRejection(lastStatus, lastErrText) && messagesHaveImageUrl(messages)) {
+        textOnlyVisionMemory.add(visionMemoryKey(params.model, visionProbe));
+        vision = false;
+        messages = wireMessages(false);
+        body.messages = messages;
+        continue;
+      }
+      if (!RETRYABLE_STATUSES.has(response.status) || attempt >= streamTimeouts.maxRetries) break;
       const retryAfter = response.headers.get("retry-after");
       await abortableSleep(backoffDelay(attempt, retryAfter), params.signal);
     }
@@ -51366,17 +51462,19 @@ function openaiCompatibleProvider(config2) {
       }
       toolCallAccumulator.clear();
     };
-    const streamDeadline = Date.now() + PROVIDER_STREAM_MAX_MS;
+    const streamDeadline = Date.now() + streamTimeouts.maxMs;
     let lastUsefulAt = Date.now();
+    let emittedUseful = false;
     const markUseful = () => {
       lastUsefulAt = Date.now();
+      emittedUseful = true;
     };
     try {
       while (true) {
         let chunk;
         try {
           chunk = await readChunkWithTimeout(reader, {
-            idleMs: PROVIDER_STREAM_IDLE_MS,
+            idleMs: emittedUseful ? streamTimeouts.idleMs : streamTimeouts.firstTokenIdleMs,
             deadlineMs: streamDeadline,
             signal: params.signal,
             lastUsefulAt: () => lastUsefulAt
@@ -51431,17 +51529,33 @@ function openaiCompatibleProvider(config2) {
                 }
               };
             }
-            if (typeof delta?.content === "string" && delta.content.length > 0) {
+            const content = delta?.content;
+            if (typeof content === "string" && content.length > 0) {
               markUseful();
-              yield { kind: "text", delta: delta.content };
+              yield { kind: "text", delta: content };
+            } else if (Array.isArray(content)) {
+              for (const part of content) {
+                if (!part || typeof part !== "object") continue;
+                const p3 = part;
+                const text = typeof p3.text === "string" ? p3.text : "";
+                if (!text) continue;
+                markUseful();
+                const partType = typeof p3.type === "string" ? p3.type : "text";
+                yield {
+                  kind: partType === "reasoning" || partType === "thinking" ? "thinking" : "text",
+                  delta: text
+                };
+              }
             }
-            const reasoning = delta?.reasoning_content ?? delta?.reasoning;
-            if (typeof reasoning === "string" && reasoning.length > 0) {
+            const reasoningRaw = delta?.reasoning_content ?? delta?.reasoning ?? delta?.thinking;
+            const reasoning = typeof reasoningRaw === "string" ? reasoningRaw : reasoningRaw && typeof reasoningRaw === "object" && typeof reasoningRaw.text === "string" ? reasoningRaw.text : reasoningRaw && typeof reasoningRaw === "object" && typeof reasoningRaw.content === "string" ? reasoningRaw.content : "";
+            if (reasoning.length > 0) {
               markUseful();
               yield { kind: "thinking", delta: reasoning };
             }
             const details = delta?.reasoning_details;
-            if (Array.isArray(details)) {
+            if (Array.isArray(details) && details.length > 0) {
+              markUseful();
               for (const d of details) {
                 if (!d || typeof d !== "object") continue;
                 const t = d.text;
@@ -51461,6 +51575,7 @@ function openaiCompatibleProvider(config2) {
               }
             }
             if (Array.isArray(delta?.tool_calls)) {
+              markUseful();
               for (const tc of delta.tool_calls) {
                 const idx = tc.index ?? 0;
                 const existing = toolCallAccumulator.get(idx) ?? {
@@ -51527,7 +51642,7 @@ async function providerConfigFor(providerId) {
     ...extraFromStored(providerId)
   };
 }
-var RETRYABLE_STATUSES, MAX_RETRIES, BACKOFF_BASE_MS, BACKOFF_CAP_MS, PROVIDER_CONNECT_TIMEOUT_MS, PROVIDER_STREAM_IDLE_MS, PROVIDER_STREAM_MAX_MS, PROVIDER_ENDPOINTS, messageMappingCache;
+var RETRYABLE_STATUSES, MAX_RETRIES, BACKOFF_BASE_MS, BACKOFF_CAP_MS, PROVIDER_CONNECT_TIMEOUT_MS, PROVIDER_STREAM_IDLE_MS, PROVIDER_FIRST_TOKEN_IDLE_MS, PROVIDER_STREAM_MAX_MS, textOnlyVisionMemory, PROVIDER_ENDPOINTS, messageMappingCache;
 var init_openai_compatible = __esm({
   "src/cli/provider/openai-compatible.ts"() {
     "use strict";
@@ -51553,11 +51668,17 @@ var init_openai_compatible = __esm({
       const n = raw ? Number.parseInt(raw, 10) : 3e5;
       return Number.isFinite(n) && n >= 15e3 ? n : 3e5;
     })();
+    PROVIDER_FIRST_TOKEN_IDLE_MS = (() => {
+      const raw = process.env.ZELARI_PROVIDER_FIRST_TOKEN_IDLE_MS;
+      const n = raw ? Number.parseInt(raw, 10) : 6e5;
+      return Number.isFinite(n) && n >= 15e3 ? n : 6e5;
+    })();
     PROVIDER_STREAM_MAX_MS = (() => {
       const raw = process.env.ZELARI_PROVIDER_STREAM_MAX_MS;
       const n = raw ? Number.parseInt(raw, 10) : 18e5;
       return Number.isFinite(n) && n >= 6e4 ? n : 18e5;
     })();
+    textOnlyVisionMemory = /* @__PURE__ */ new Set();
     PROVIDER_ENDPOINTS = {
       "openai-compatible": "https://api.x.ai/v1",
       "minimax": "https://api.minimax.io/v1",