nexrall-code 0.5.104 → 0.5.106

package/dist/index.js

@@ -103400,7 +103400,9 @@ var require_loop = __commonJS({
       };
     }();
     Object.defineProperty(exports2, "__esModule", { value: true });
-    exports2.VERIFY_CMD_RE = exports2.WRITE_TOOL_NAMES = exports2.ToolNotAllowedError = exports2.AGENT_MEMORY_TOOL_SCHEMA = exports2.AGENT_MEMORY_TOOL = exports2._stallLimits = exports2.bashNeedsRepoLock = void 0;
+    exports2.VERIFY_CMD_RE = exports2.WRITE_TOOL_NAMES = exports2.ToolNotAllowedError = exports2.AGENT_MEMORY_TOOL_SCHEMA = exports2.AGENT_MEMORY_TOOL = exports2._stallLimits = exports2._emptyTurnRetry = exports2.bashNeedsRepoLock = void 0;
+    exports2.emptyTurnBackoffMs = emptyTurnBackoffMs;
+    exports2.shouldRetryEmptyTurn = shouldRetryEmptyTurn;
     exports2.errorRoundSignature = errorRoundSignature;
     exports2.executeAgentMemoryWrite = executeAgentMemoryWrite;
     exports2.stopReasonNotice = stopReasonNotice;
@@ -103557,6 +103559,17 @@ var require_loop = __commonJS({
     var HARD_ITERATIONS_CAP = Infinity;
     var STALL_LIMIT = 8;
     var REPEAT_STALL_LIMIT = 12;
+    var EMPTY_TURN_RETRY_LIMIT = 3;
+    var EMPTY_TURN_RETRY_BASE_MS = 1e3;
+    function emptyTurnBackoffMs(attempt) {
+      return EMPTY_TURN_RETRY_BASE_MS * Math.pow(2, Math.max(0, attempt));
+    }
+    function shouldRetryEmptyTurn(stopReason, attemptsSoFar) {
+      if (stopReason === "max_tokens")
+        return false;
+      return attemptsSoFar < EMPTY_TURN_RETRY_LIMIT;
+    }
+    exports2._emptyTurnRetry = { EMPTY_TURN_RETRY_LIMIT, EMPTY_TURN_RETRY_BASE_MS };
     function errorRoundSignature(errored) {
       return errored.map(({ name, error }) => `${name}:${String(error).slice(0, 200)}`).sort().join("|");
     }
@@ -103600,7 +103613,7 @@ var require_loop = __commonJS({
           return null;
         case "empty-response":
           return `
-\u26A0\uFE0F The model returned an empty response, so nothing was done. This is usually a transient upstream hiccup \u2014 send "continue" to retry.
+\u26A0\uFE0F The model returned an empty response ${EMPTY_TURN_RETRY_LIMIT} times in a row, so nothing was done. This is usually a transient upstream hiccup that the agent retries by itself; it did not clear this time. Send "continue" to try again, or switch model with /model if it persists.
 `;
         case "output-limit":
           return `
@@ -104300,8 +104313,7 @@ ${partial}` : "",
       // modelRegistry.js's own TODO(unverified-by-400): DeepSeek accepts an
       // oversized max_completion_tokens without rejecting it, so there was no 400
       // to measure the ceiling from the way the OpenAI rows above were).
-      "deepseek-v4-pro": 1048576,
-      "deepseek-v4-flash": 1048576,
+      "deepseek-flash": 1048576,
       // Qwen (DashScope), by real model id (documented max input, same caveat).
       "qwen3.7-max": 991800,
       // Z.ai (GLM), by real model id. contextWindow is documented (Z.ai/
@@ -104763,6 +104775,7 @@ ${options.nexrallMd}` : "") : options.nexrallMd;
       let stalledRepeatError = null;
       let budget = maxIterations;
       let iteration = 0;
+      let emptyTurnRetries = 0;
       let filesMutatedSinceVerify = false;
       let ranVerificationCmd = false;
       let verificationNudgeSent = false;
@@ -104954,10 +104967,24 @@ ${options.nexrallMd}` : "") : options.nexrallMd;
               messages.push({ role: "user", content: [{ type: "text", text }] });
               continue;
             }
+            if (shouldRetryEmptyTurn(assistantMessage.stopReason, emptyTurnRetries)) {
+              const waitMs = emptyTurnBackoffMs(emptyTurnRetries);
+              emptyTurnRetries++;
+              options.onRetry?.(emptyTurnRetries, EMPTY_TURN_RETRY_LIMIT, "The model returned an empty response \u2014 retrying automatically");
+              await new Promise((r2) => setTimeout(r2, waitMs));
+              if (options.abortSignal?.aborted) {
+                stopReason = "aborted";
+                break;
+              }
+              options.onRetryResolved?.();
+              iteration--;
+              continue;
+            }
             runSimpleHooks(hooks.PostMessageComplete, options.workDir);
             stopReason = assistantMessage.stopReason === "max_tokens" ? "output-limit" : "empty-response";
             break;
           }
+          emptyTurnRetries = 0;
           const { stopReason: _stopReason, ...historyMessage } = assistantMessage;
           messages.push(historyMessage);
           const serverSideResultIds = new Set(assistantMessage.content.filter((b) => b.type === "tool_result").map((b) => b.tool_use_id).filter((id) => !!id));
@@ -153272,8 +153299,7 @@ var MODEL_LABELS = {
   "gpt-5.4": "GPT-5.4",
   "gpt-5.4-mini": "GPT-5.4 Mini",
   "gpt-4.1": "GPT-4.1",
-  "deepseek-v4-pro": "DeepSeek V4 Pro",
-  "deepseek-v4-flash": "DeepSeek V4 Flash",
+  "deepseek-flash": "DeepSeek Flash",
   "qwen3.7-max": "Qwen3.7 Max",
   "glm-5.3": "GLM 5.3"
 };
@@ -153295,8 +153321,7 @@ var SELECTABLE_MODELS = [
   "gpt-5.4",
   "gpt-5.4-mini",
   "gpt-4.1",
-  "deepseek-v4-pro",
-  "deepseek-v4-flash",
+  "deepseek-flash",
   "qwen3.7-max",
   "glm-5.3"
 ];
@@ -153314,8 +153339,7 @@ var MODEL_COST_MULTIPLIER = {
   "gpt-5.4": 2,
   "gpt-5.4-mini": 0.5,
   "gpt-4.1": 1,
-  "deepseek-v4-pro": 0.1,
-  "deepseek-v4-flash": 0.05,
+  "deepseek-flash": 0.1,
   "qwen3.7-max": 1,
   // $1.40/$4.40 per 1M in/out (docs.z.ai, 2026-08-30) vs gpt-4.1's $2.00/$8.00
   // (migration 113) — roughly 0.6x on a blended 75/25 in/out turn. Rounded to
@@ -153327,7 +153351,7 @@ function modelWithCostHint(id) {
   const mult = (0, import_code_core4.liveCostMultiplierFor)(id, MODEL_COST_MULTIPLIER[id]);
   return mult === void 0 ? id : `${id} (${mult}x)`;
 }
-var NO_VISION_MODELS = /* @__PURE__ */ new Set(["deepseek-v4-pro", "deepseek-v4-flash", "qwen3.7-max", "glm-5.3"]);
+var NO_VISION_MODELS = /* @__PURE__ */ new Set(["qwen3.7-max", "glm-5.3"]);
 var NO_PDF_MODELS = /* @__PURE__ */ new Set([
   "gpt-5.6-sol",
   "gpt-5.6-terra",
@@ -153336,6 +153360,7 @@ var NO_PDF_MODELS = /* @__PURE__ */ new Set([
   "gpt-5.4-mini",
   "gpt-4.1",
   "gpt-4o-mini",
+  "deepseek-flash",
   ...NO_VISION_MODELS
 ]);
 function needsVisionSidecar(model) {
@@ -153366,8 +153391,7 @@ var MODEL_EFFORT_STYLE = {
   "gpt-5.4-mini": "openai_gated",
   "gpt-4.1": "none",
   "gpt-4o-mini": "none",
-  "deepseek-v4-pro": "deepseek",
-  "deepseek-v4-flash": "deepseek",
+  "deepseek-flash": "deepseek",
   "qwen3.7-max": "qwen",
   "glm-5.3": "glm"
   // Everything else (all claude-* ids) falls through to 'anthropic' below.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nexrall-code",
-  "version": "0.5.104",
+  "version": "0.5.106",
   "description": "Nexrall Code — AI coding assistant for your terminal (headless agent for scripts, CI and automation)",
   "keywords": [
     "ai",