apex-dev 3.6.0 → 3.8.0

package/dist/index.js

@@ -31674,12 +31674,12 @@ var require_openai = __commonJS((exports, module2) => {
 var require_config = __commonJS((exports, module2) => {
   var OpenAI = require_openai();
   var NVIDIA_MODEL = "z-ai/glm4.7";
-  var REVIEWER_MODEL = "nvidia/llama-3.3-nemotron-70b-instruct";
+  var REVIEWER_MODEL = "nvidia/llama-3.3-nemotron-super-49b-v1.5";
   var FILE_PICKER_MODEL = "qwen/qwen3-coder-480b-a35b-instruct";
   var THINKER_MODEL = "z-ai/glm4.7";
-  var COMMANDER_MODEL = "nvidia/llama-3.3-nemotron-70b-instruct";
-  var CONTEXT_PRUNER_MODEL = "nvidia/llama-3.3-nemotron-70b-instruct";
-  var RESEARCHER_MODEL = "nvidia/llama-3.3-nemotron-70b-instruct";
+  var COMMANDER_MODEL = "nvidia/llama-3.3-nemotron-super-49b-v1.5";
+  var CONTEXT_PRUNER_MODEL = "nvidia/llama-3.3-nemotron-super-49b-v1.5";
+  var RESEARCHER_MODEL = "nvidia/llama-3.3-nemotron-super-49b-v1.5";
   var GENERAL_AGENT_MODEL = "z-ai/glm4.7";
   var MAX_TOOL_ITERATIONS = 50;
   var MAX_OUTPUT_LEN = 12000;
@@ -31919,6 +31919,7 @@ Scripts: ${Object.keys(pkg.scripts).join(", ")}`;
 - **Do what the user asks:** If the user asks you to do something, even running a risky command, do it.
 - **If a tool fails, try again or try a different tool.** Don't give up after one attempt.
 - **Act on errors.** If the user pastes an error or stack trace, locate the source, identify root cause, and fix it. Never punt back with "try checking X."
+- **Nothing is automatic.** The agent loop is a thin shell \u2014 it only executes tool calls you explicitly make. No code review, no context pruning, no validation happens unless YOU call the corresponding tool.
 - **Use <think></think> tags for moderate reasoning.** Call the Thinker sub-agent for anything more complex.
 
 # Output Style
@@ -31947,7 +31948,7 @@ Scripts: ${Object.keys(pkg.scripts).join(", ")}`;
 
 # Sub-Agent Orchestration
 
-You have specialized sub-agents available as tools. **Your strong default should be to delegate to sub-agents** for anything beyond trivially simple operations. Sub-agents are specialists \u2014 they produce better, more thorough results than you chaining basic tools manually.
+You have specialized sub-agents available as tools. **Nothing happens automatically \u2014 you are responsible for orchestrating ALL sub-agent work through your own tool calls.** No code review, no context pruning, no validation runs unless YOU explicitly call the appropriate tool. Sub-agents are specialists \u2014 they produce better, more thorough results than you chaining basic tools manually.
 
 ## Available Sub-Agents
 
@@ -31966,8 +31967,8 @@ You have specialized sub-agents available as tools. **Your strong default should
 - **Commander** \u2014 Terminal command specialist. Plans and executes shell commands for a goal. Use instead of calling Bash directly for multi-step operations.
 
 **Review & Maintenance:**
-- **CodeReview** / **CodeReviewMulti** \u2014 Reviews code changes for bugs, security, edge cases. CodeReviewMulti runs automatically after edits.
-- **ContextPruner** \u2014 Summarizes conversation history to free context space.
+- **CodeReview** / **CodeReviewMulti** \u2014 Reviews code changes for bugs, security, edge cases. You MUST call one of these yourself after making code changes.
+- **ContextPruner** \u2014 Summarizes conversation history to free context space. Call when the conversation is getting long.
 
 ## How to Orchestrate (use your judgment)
 
@@ -31993,7 +31994,7 @@ You have specialized sub-agents available as tools. **Your strong default should
 - If checks fail, fix and re-run. If blocked, clearly state what's failing.
 
 **Phase 5 \u2014 Review:**
-- CodeReviewMulti runs automatically after edits \u2014 don't duplicate its work.
+- After making code changes, call CodeReview or CodeReviewMulti yourself to review the changes. Nothing runs automatically.
 - If the review finds issues, fix them and re-validate.
 
 ## When to Skip Sub-Agents and Act Directly
@@ -32474,7 +32475,11 @@ var require_toolExecutors = __commonJS((exports, module2) => {
           return response.choices[0]?.message?.content || response.choices[0]?.message?.reasoning_content || "";
         }
       } catch (err) {
-        if (attempt < 2 && err.status >= 400 && err.status < 500) {
+        if (err.status === 404 && params.model !== NVIDIA_MODEL && attempt < 2) {
+          params = { ...params, model: NVIDIA_MODEL };
+          continue;
+        }
+        if (attempt < 2 && (err.status === 429 || err.status >= 500)) {
           await sleep(1000 * Math.pow(2, attempt));
           continue;
         }
@@ -33763,9 +33768,7 @@ var require_agent = __commonJS((exports, module2) => {
     session.turnCount++;
     store.addMessage({ role: "user", content: userInput });
     session.conversationHistory.push({ role: "user", content: userInput });
-    const startTime2 = Date.now();
     let turnTokens = 0;
-    let turnToolCalls = 0;
     try {
       store.addMessage({ role: "divider" });
       const systemPrompt = buildSystemPrompt();
@@ -34037,7 +34040,6 @@ var require_agent = __commonJS((exports, module2) => {
             const success = !result.startsWith("Error");
             const elapsed = Date.now() - callStart;
             session.toolCallCount++;
-            turnToolCalls++;
             if (msgId) {
               store.updateMessage(msgId, {
                 detail,
@@ -34078,27 +34080,6 @@ var require_agent = __commonJS((exports, module2) => {
         store.addMessage({ role: "system", content: `\u26A0 Reached maximum tool iterations (${MAX_TOOL_ITERATIONS}). Stopping.` });
       }
       session.totalTokens += turnTokens;
-      if (session.filesModified.size > 0) {
-        const reviewPrompt = `User request: ${userInput}
-
-${turnTokens > 0 ? `Processed with ${turnTokens} tokens and ${turnToolCalls} tool calls.` : ""}`;
-        const reviewId = store.addMessage({ role: "system", content: "Multi-perspective code review...", label: "\uD83D\uDCCB Code Review" });
-        try {
-          const reviewResult = await executeTool("CodeReviewMulti", { prompt: reviewPrompt }, (partial) => {
-            store.updateMessage(reviewId, { content: partial, label: "\uD83D\uDCCB Code Review" });
-          });
-          store.updateMessage(reviewId, { content: reviewResult, label: "\uD83D\uDCCB Code Review" });
-          session.conversationHistory.push({
-            role: "assistant",
-            content: `
-
---- Code Review ---
-${reviewResult}`
-          });
-        } catch (err) {
-          store.updateMessage(reviewId, { content: `Code review failed: ${err.message}` });
-        }
-      }
     } catch (err) {
       store.clearStreaming();
       let errorMsg = `Error: ${err.message}`;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "apex-dev",
-  "version": "3.6.0",
+  "version": "3.8.0",
   "description": "Apex AI - a friendly agentic coding assistant for the terminal",
   "main": "dist/index.js",
   "bin": {