zelari-code 2.37.1 → 2.37.3

package/dist/cli/main.bundled.js

@@ -28993,6 +28993,9 @@ var init_AgentHarness = __esm({
       maxToolLoopIterations;
       maxToolLoopHardCap;
       cancelled = false;
+      /** Host-supplied cancel cause (`turn_timeout` vs user Stop). */
+      cancelReason;
+      cancelEventEmitted = false;
       activeController = null;
       queue = [];
       /**
@@ -29347,13 +29350,28 @@ ${shared.content}`,
        * After cancel() returns, the harness should be discarded by the
        * caller (the run() generator finishes after the current turn ends).
        * For mid-stream interrupt + new-prompt injection, see Task C.3.2.
+       *
+       * `reason` is optional host context: `turn_timeout` is the Desktop sidecar
+       * idle watchdog (not a user Stop). The error event message must not claim
+       * the user cancelled when they did not.
        */
-      cancel() {
+      cancel(reason) {
         if (this.cancelled)
           return;
         this.cancelled = true;
+        if (reason)
+          this.cancelReason = reason;
         this.activeController?.abort();
       }
+      buildCancelEvent() {
+        this.cancelEventEmitted = true;
+        const watchdog = this.cancelReason === "turn_timeout";
+        return createBrainEvent("error", this.sessionId, {
+          severity: "cancelled",
+          message: watchdog ? "Turn cancelled: Desktop idle watchdog saw no events (silent model thinking or a tentacle with no tools). This was not a user Stop." : "Run cancelled by user.",
+          code: watchdog ? "turn_timeout" : "cancelled"
+        });
+      }
       /** Current size of the queued user-prompt buffer. */
       get queueLength() {
         return this.queue.length;
@@ -29398,6 +29416,7 @@ ${shared.content}`,
       async *run() {
         const startTime = Date.now();
         this.activeController = new AbortController();
+        this.cancelEventEmitted = false;
         this.toolCallCache = /* @__PURE__ */ new Map();
         this.toolCallCounts = /* @__PURE__ */ new Map();
         this.textToolReentries = 0;
@@ -29643,6 +29662,11 @@ ${shared.content}`,
           } catch {
           }
         }
+        if (this.cancelled && !this.cancelEventEmitted) {
+          const cancelEvent = this.buildCancelEvent();
+          this.emit(cancelEvent);
+          yield cancelEvent;
+        }
         const agentEnd = createBrainEvent("agent_end", this.sessionId, {
           reason: hadError ? "error" : this.cancelled ? "cancelled" : "completed",
           durationMs: Date.now() - startTime,
@@ -29754,11 +29778,7 @@ ${shared.content}`,
           let lastLoopCheckLen = 0;
           for await (const delta of stream) {
             if (this.cancelled) {
-              const cancelEvent = createBrainEvent("error", this.sessionId, {
-                severity: "cancelled",
-                message: "Run cancelled by user.",
-                code: "cancelled"
-              });
+              const cancelEvent = this.buildCancelEvent();
               this.emit(cancelEvent);
               yield cancelEvent;
               break;
@@ -38416,7 +38436,7 @@ var CORE_VERSION;
 var init_version = __esm({
   "packages/core/dist/version.js"() {
     "use strict";
-    CORE_VERSION = "2.37.1";
+    CORE_VERSION = "2.37.3";
   }
 });
 
@@ -43086,6 +43106,43 @@ var init_metrics2 = __esm({
   }
 });
 
+// src/cli/tools/tentacleHeartbeat.ts
+function resolveTentacleHeartbeatMs(env = process.env) {
+  const raw = env.ZELARI_TENTACLE_HEARTBEAT_MS?.trim();
+  if (raw === "0" || raw === "off") return 0;
+  if (!raw) return TENTACLE_HEARTBEAT_DEFAULT_MS;
+  const n = Number.parseInt(raw, 10);
+  return Number.isFinite(n) && n >= 0 ? n : TENTACLE_HEARTBEAT_DEFAULT_MS;
+}
+function formatTentacleElapsed(ms) {
+  const totalSec = Math.max(0, Math.floor(ms / 1e3));
+  const m = Math.floor(totalSec / 60);
+  const s = totalSec % 60;
+  return m > 0 ? `${m}m ${s}s` : `${s}s`;
+}
+function tentacleHeartbeatCaption(elapsedMs) {
+  return `reasoning \xB7 ${formatTentacleElapsed(elapsedMs)}`;
+}
+function startTentacleHeartbeat(onBeat, opts) {
+  const intervalMs = opts?.intervalMs ?? resolveTentacleHeartbeatMs();
+  if (intervalMs <= 0) return () => {
+  };
+  const now = opts?.now ?? Date.now;
+  const started = now();
+  const timer = setInterval(() => {
+    onBeat(tentacleHeartbeatCaption(now() - started));
+  }, intervalMs);
+  timer.unref?.();
+  return () => clearInterval(timer);
+}
+var TENTACLE_HEARTBEAT_DEFAULT_MS;
+var init_tentacleHeartbeat = __esm({
+  "src/cli/tools/tentacleHeartbeat.ts"() {
+    "use strict";
+    TENTACLE_HEARTBEAT_DEFAULT_MS = 15e3;
+  }
+});
+
 // src/cli/kraken/tentacle.ts
 var tentacle_exports = {};
 __export(tentacle_exports, {
@@ -46875,36 +46932,54 @@ ${taskUserContent}`,
       emitActivity({ type: "agent_tool", agentId: liveId, toolCallId: ev.toolCallId, tool: startedTools.get(ev.toolCallId) ?? "unknown", status: ev.isError ? "failed" : "completed", durationMs: ev.durationMs, ts: Date.now() });
     }
   };
-  let { result, error: error51, aborted: aborted2, usage, toolTrace } = await runSubAgent(harness, {
-    ...opts.signal ? { signal: opts.signal } : {},
-    onEvent: onHarnessEvent
+  const stopHeartbeat = startTentacleHeartbeat((caption) => {
+    emitActivity({
+      type: "agent_status",
+      agentId: liveId,
+      status: "running",
+      message: caption,
+      ts: Date.now()
+    });
   });
-  if (!aborted2 && !result && sub.fallback && sub.fallback.model !== sub.model) {
-    const { isUnknownModelError: isUnknownModelError2 } = await Promise.resolve().then(() => (init_krakenModel(), krakenModel_exports));
-    if (isUnknownModelError2(error51)) {
-      emitPhase(`model ${sub.model} unavailable \u2014 retrying with ${sub.fallback.model}`);
-      const retryConfig = {
-        ...config2,
-        model: sub.fallback.model,
-        provider: sub.fallback.provider,
-        providerStream: sub.fallback.providerStream
-      };
-      try {
-        harness = deps.harnessFactory ? deps.harnessFactory(retryConfig) : new (await Promise.resolve().then(() => (init_harness(), harness_exports))).AgentHarness(retryConfig);
-        const retry = await runSubAgent(harness, {
-          ...opts.signal ? { signal: opts.signal } : {},
-          onEvent: onHarnessEvent
-        });
-        result = retry.result;
-        error51 = retry.error;
-        aborted2 = retry.aborted;
-        usage = retry.usage;
-        toolTrace = retry.toolTrace;
-        sub = { ...sub, model: sub.fallback.model, provider: sub.fallback.provider };
-      } catch (err) {
-        error51 = err instanceof Error ? err.message : String(err);
+  let result;
+  let error51;
+  let aborted2;
+  let usage;
+  let toolTrace;
+  try {
+    ({ result, error: error51, aborted: aborted2, usage, toolTrace } = await runSubAgent(harness, {
+      ...opts.signal ? { signal: opts.signal } : {},
+      onEvent: onHarnessEvent
+    }));
+    if (!aborted2 && !result && sub.fallback && sub.fallback.model !== sub.model) {
+      const { isUnknownModelError: isUnknownModelError2 } = await Promise.resolve().then(() => (init_krakenModel(), krakenModel_exports));
+      if (isUnknownModelError2(error51)) {
+        emitPhase(`model ${sub.model} unavailable \u2014 retrying with ${sub.fallback.model}`);
+        const retryConfig = {
+          ...config2,
+          model: sub.fallback.model,
+          provider: sub.fallback.provider,
+          providerStream: sub.fallback.providerStream
+        };
+        try {
+          harness = deps.harnessFactory ? deps.harnessFactory(retryConfig) : new (await Promise.resolve().then(() => (init_harness(), harness_exports))).AgentHarness(retryConfig);
+          const retry = await runSubAgent(harness, {
+            ...opts.signal ? { signal: opts.signal } : {},
+            onEvent: onHarnessEvent
+          });
+          result = retry.result;
+          error51 = retry.error;
+          aborted2 = retry.aborted;
+          usage = retry.usage;
+          toolTrace = retry.toolTrace;
+          sub = { ...sub, model: sub.fallback.model, provider: sub.fallback.provider };
+        } catch (err) {
+          error51 = err instanceof Error ? err.message : String(err);
+        }
       }
     }
+  } finally {
+    stopHeartbeat();
   }
   const durationMs = Date.now() - started;
   if (aborted2) {
@@ -46914,7 +46989,7 @@ ${taskUserContent}`,
       agent,
       thoroughness,
       description: args.description,
-      detail: "cancelled: node timeout",
+      detail: "cancelled by parent",
       model: sub.model,
       worktree: worktree?.path ?? null,
       durationMs,
@@ -46926,7 +47001,7 @@ ${taskUserContent}`,
       detail: "cancelled",
       durationMs
     });
-    return { ok: false, agent, error: "task: sub-agent cancelled (node timeout)", cancelled: true };
+    return { ok: false, agent, error: "task: sub-agent cancelled by parent", cancelled: true };
   }
   if (!result) {
     if (worktree) await cleanupKrakenWorktree(worktree);
@@ -47204,6 +47279,7 @@ var init_taskTool = __esm({
     init_verifyReport();
     init_metrics2();
     init_dist();
+    init_tentacleHeartbeat();
     TASK_TOOL_TIMEOUT_MS = 27e5;
     EXPLORE_PROMPT = [
       "You are a focused EXPLORE tentacle of Kraken (parent super-agent).",
@@ -48283,6 +48359,7 @@ 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 +48432,18 @@ 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 +51068,40 @@ 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 +51120,12 @@ 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 +51139,7 @@ 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 +51149,8 @@ async function readChunkWithTimeout(reader, opts) {
       reader.read(),
       new Promise((_, reject) => {
         idleTimer = setTimeout(() => {
-          reject(
-            new Error(
-              `Provider stream idle for ${Math.round(waitMs / 1e3)}s (no tokens). The model/gateway stalled \u2014 try again or switch model. Override with ZELARI_PROVIDER_STREAM_IDLE_MS.`
-            )
-          );
+          const elapsed = Date.now() - opts.lastUsefulAt();
+          reject(new Error(formatStreamIdleError(elapsed, opts.idleMs, "silence")));
         }, waitMs);
         if (opts.signal) {
           onAbort = () => reject(new Error("aborted"));
@@ -51060,9 +51172,52 @@ function backoffDelay(attempt, retryAfterHeader) {
   }
   return Math.min(BACKOFF_BASE_MS * 2 ** attempt, BACKOFF_CAP_MS);
 }
-function modelSupportsVision(_model) {
-  const force = process.env.ZELARI_VISION;
-  return !(force === "0" || force === "false" || force === "off");
+function visionMemoryKey(model, probe) {
+  return `${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 +51337,38 @@ 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 +51418,7 @@ 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 +51433,7 @@ 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 +51474,7 @@ 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 +51485,14 @@ 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 +51538,19 @@ 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 +51605,33 @@ 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 +51651,7 @@ 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 +51718,7 @@ 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 +51744,17 @@ 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",
@@ -67393,7 +67590,7 @@ var init_controlBridge = __esm({
 function attachHeadlessLiveCancel(opts) {
   const abort = new AbortController();
   const controlQueue = new RuntimeControlQueue();
-  const cancel = () => {
+  const cancel = (_reason) => {
     if (!abort.signal.aborted) abort.abort();
     return true;
   };
@@ -67951,10 +68148,10 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
   })() : void 0;
   const unregisterLiveTurnControl = process.env.ZELARI_SERVE_HARNESS === "1" ? registerLiveTurnControl({
     queue: controlQueue,
-    cancel: () => {
+    cancel: (reason) => {
       const cancelHook = harnessHolder.cancel;
       if (!cancelHook) return false;
-      cancelHook();
+      cancelHook(reason);
       return true;
     }
   }) : void 0;
@@ -68240,7 +68437,7 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
         memoryContextChars: 2e3
       } : {}
     });
-    harnessHolder.cancel = () => harness.cancel();
+    harnessHolder.cancel = (reason) => harness.cancel(reason);
     const readBuildProgress = () => {
       const getter = harness.getBuildProgress;
       return typeof getter === "function" ? getter.call(harness) : { mutationsAttempted: 0, mutationsSucceeded: 0 };
@@ -68427,11 +68624,11 @@ async function runOneTurn(opts, provider, model, providerStream, extras) {
     process.stdout.write(pass.textBuffer.join(""));
   }
   process.stdout.write("");
-  if (pass.finalReason !== "error" && opts.output === "json" && wantWrites && pass.successfulWrites === 0) {
+  if (pass.finalReason !== "error" && pass.finalReason !== "cancelled" && opts.output === "json" && wantWrites && pass.successfulWrites === 0) {
     emitEvent({ type: "log", message: "[headless] BUILD failed: zero successful mutations after liveness recovery" });
   }
   try {
-    const closeStatus = pass.finalReason === "error" ? "error" : strictExit !== 0 ? "stopped" : "completed";
+    const closeStatus = pass.finalReason === "error" ? "error" : pass.finalReason === "cancelled" ? "cancelled" : strictExit !== 0 ? "stopped" : "completed";
     await spine.close(closeStatus);
   } catch {
   }
@@ -72765,7 +72962,8 @@ function startHarnessServer(options = {}) {
           };
         }
         if (isCancel) {
-          const delivered = live.cancel();
+          const reason = typeof params.reason === "string" && params.reason.trim().length > 0 ? params.reason.trim() : void 0;
+          const delivered = live.cancel(reason);
           if (delivered) {
             write(JSON.stringify(controlAppliedEvent(controlId, "cancel", "cancel")));
           }