nexrall-code 0.5.66 → 0.5.68

package/dist/index.js

@@ -14826,7 +14826,11 @@ var require_agentTypes = __commonJS({
         name: "security-auditor",
         description: "Read-only security auditor \u2014 hunts injection, authz, secrets, and validation flaws in a path or diff. Cannot modify files.",
         tools: READ_ONLY_TOOLS,
-        model: "pro",
+        // No `model` override — inherits the session's, same as general-purpose.
+        // A previous version forced 'pro' (Claude Opus 5) here, which silently
+        // billed Anthropic even for a session running entirely on OpenAI/DeepSeek/
+        // Qwen. Consistency with the main conversation's provider matters more
+        // than defaulting every specialist to a fixed tier.
         source: "builtin",
         prompt: [
           "You are a security auditor. You find real, exploitable flaws \u2014 not style issues.",
@@ -14861,7 +14865,10 @@ var require_agentTypes = __commonJS({
         lightPrompt: true,
         description: "Fast read-only codebase explorer \u2014 locates files, symbols, and call sites and reports concise findings. Use to keep bulk searching out of the main context. Cannot modify files.",
         tools: READ_ONLY_TOOLS,
-        model: "turbo",
+        // No `model` override — inherits the session's. A previous version forced
+        // 'turbo' (Claude Sonnet 5) on the theory that bulk search doesn't need a
+        // frontier model, but that silently billed Anthropic regardless of which
+        // provider the user actually selected for the conversation.
         source: "builtin",
         prompt: [
           "You map code. You NEVER modify anything.",
@@ -14886,7 +14893,9 @@ var require_agentTypes = __commonJS({
         name: "planner",
         description: "Read-only planning agent \u2014 researches a change and returns a concrete step-by-step implementation plan with risks and affected files. Cannot modify files.",
         tools: RESEARCH_TOOLS,
-        model: "pro",
+        // No `model` override — inherits the session's, for the same reason as
+        // explorer/security-auditor above: a fixed alias silently billed Anthropic
+        // regardless of the user's chosen provider.
         source: "builtin",
         prompt: [
           "You produce implementation plans. You NEVER modify files \u2014 planning and doing are separate steps,",
@@ -14948,7 +14957,8 @@ var require_agentTypes = __commonJS({
         name: "devops-advisor",
         description: "Read-only CI/CD, container, and infrastructure advisor \u2014 diagnoses pipelines, Dockerfiles, and k8s manifests and proposes concrete fixes as a diff. Never applies changes.",
         tools: RESEARCH_TOOLS,
-        model: "pro",
+        // No `model` override — inherits the session's, for the same reason as the
+        // other specialists above.
         source: "builtin",
         prompt: [
           "You are an infrastructure and delivery advisor. You DIAGNOSE and PROPOSE. You never apply changes.",
@@ -17143,14 +17153,17 @@ ${tail}`;
       let summary = "";
       try {
         const reply = await (0, client_1.streamChat)([{ role: "user", content: [{ type: "text", text: summaryPrompt }] }], {
-          // COST: the summariser is a mechanical "bullet-point this transcript" task —
-          // its quality is indistinguishable across model tiers, so there is no reason
-          // to run it on the user's (possibly expensive) tier. Force 'turbo' (Sonnet 5):
-          // it is the cheapest 1M-context tier (in $2.0/1M vs Opus $2.5, Fable $3.0;
-          // out $15/1M vs $25 / $50), so this is always ≤ the user's cost, and its 1M
-          // window comfortably holds the transcript (capped at ~170K tokens by
-          // transcriptOf) even on the largest sessions.
-          model: "turbo",
+          // Run on the SAME model as the actual conversation. A previous version
+          // forced 'turbo' (Claude Sonnet 5) here on the theory that a mechanical
+          // "bullet-point this transcript" task doesn't need the user's tier — but
+          // that silently billed Anthropic (and made a real network call to a
+          // provider the user may not have configured/paid for) even when the
+          // whole session was running on OpenAI/DeepSeek/Qwen. Whatever the user
+          // is already paying for is used for compaction too, so there is never a
+          // surprise charge on a different provider. Falls back to the same
+          // default as the main loop (see `runAgentLoop`) only when no model was
+          // set at all.
+          model: options.model ?? "turbo",
           mode: "ask",
           // summariser must not call tools; ask-mode discourages action
           env: options.env,
@@ -17190,10 +17203,10 @@ Continue the work from here.` }] });
       try {
         await (0, memory_1.compactMemoryIfNeeded)(scope, options.workDir, async (prompt2) => {
           const reply = await (0, client_1.streamChat)([{ role: "user", content: [{ type: "text", text: prompt2 }] }], {
-            // Same reasoning as autoCompactMessages' summariser: a mechanical
-            // consolidation task, always run on the cheapest 1M-context tier
-            // regardless of the user's chosen model for the actual conversation.
-            model: "turbo",
+            // Same reasoning as autoCompactMessages' summariser: run on the same
+            // model as the actual conversation rather than forcing a fixed tier
+            // (which silently billed Anthropic regardless of the user's provider).
+            model: options.model ?? "turbo",
             mode: "ask",
             env: options.env,
             clientType: options.clientType,
@@ -64751,6 +64764,21 @@ var SELECTABLE_MODELS = [
   "deepseek-v4-flash",
   "qwen3.7-max"
 ];
+var MODEL_COST_MULTIPLIER = {
+  "claude-sonnet-5": 1.5,
+  "claude-opus-5": 3,
+  "claude-fable-5": 5,
+  "gpt-5.4": 2,
+  "gpt-5.4-mini": 0.5,
+  "gpt-4.1": 1,
+  "deepseek-v4-pro": 0.1,
+  "deepseek-v4-flash": 0.05,
+  "qwen3.7-max": 1
+};
+function modelWithCostHint(id) {
+  const mult = 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"]);
 var NO_PDF_MODELS = /* @__PURE__ */ new Set([
   "gpt-5.4",
@@ -65815,7 +65843,7 @@ ${dirList}`;
           const requested = arg.trim();
           if (!requested) {
             console.log(source_default.dim(`  Current: ${source_default.cyan(resolveModelLabel(modelAlias))}`));
-            console.log(source_default.dim(`  Options: ${SELECTABLE_MODELS.join(" \xB7 ")}`));
+            console.log(source_default.dim(`  Options: ${SELECTABLE_MODELS.map(modelWithCostHint).join(" \xB7 ")}`));
           } else if (SELECTABLE_MODELS.includes(normaliseModelId(requested))) {
             modelAlias = normaliseModelId(requested);
             console.log(source_default.green(`  Model \u2192 ${source_default.bold(resolveModelLabel(modelAlias))}`));

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nexrall-code",
-  "version": "0.5.66",
+  "version": "0.5.68",
   "description": "Nexrall Code — AI coding assistant for your terminal (headless agent for scripts, CI and automation)",
   "keywords": [
     "ai",
@@ -39,7 +39,7 @@
     "release": "node build.js && node scripts/upload-release.cjs"
   },
   "dependencies": {
-    "@nexrall/code-core": "^1.4.38",
+    "@nexrall/code-core": "^1.4.39",
     "chalk": "^5.3.0",
     "commander": "^12.0.0",
     "diff": "^5.2.0",