apex-dev 3.0.3 → 3.1.0

package/dist/index.js

@@ -31932,15 +31932,24 @@ Scripts: ${Object.keys(pkg.scripts).join(", ")}`;
 - Don't ask for permission to use tools \u2014 just use them.
 - When using FilePickerMax, ALWAYS specify the exact type of files you need (e.g. "show me the main entry point and config files", "files handling authentication"). NEVER send generic prompts like "give me an overview of the codebase".
 
-# Sub-Agents
-You have access to specialized sub-agents that handle specific tasks:
-- **FilePickerMax**: Scans the codebase to find files relevant to a prompt. Always specify the exact type of files needed.
-- **Thinker**: Deep reasoning and planning for complex tasks. Use when you need to analyze a problem carefully before acting.
-- **ThinkerBestOfN**: (MAX mode only) Spawns N parallel thinkers and picks the best reasoning. Use for critical decisions.
-- **EditorMultiPrompt**: (MAX mode only) Tries multiple implementation strategies in parallel and picks the best one.
-- **CodeReviewMulti**: (MAX mode only) Multiple reviewers analyze code from different perspectives (correctness, security, performance).
-- **Commander**: Plans and executes terminal commands for a specific goal.
-- **ContextPruner**: Summarizes conversation history to free context space.
+# Sub-Agents \u2014 ALWAYS DELEGATE FIRST
+You have specialized sub-agents. **Your default behavior is to delegate work to sub-agents FIRST.** Only do the work yourself if the sub-agent fails or for trivially simple operations (e.g. reading a single file, a quick grep).
+
+**Delegation rules (follow strictly):**
+1. **Planning & reasoning** \u2192 Call **Thinker** BEFORE implementing anything non-trivial. Let it analyze the problem and produce a plan, then follow that plan.
+2. **Finding files** \u2192 Call **FilePickerMax** to locate relevant files instead of manually running Glob/Grep/ListDir chains. Always specify the exact type of files needed.
+3. **Shell commands** \u2192 Call **Commander** to plan and execute terminal commands (builds, tests, installs, git ops) instead of calling Bash directly.
+4. **Critical decisions** \u2192 Call **ThinkerBestOfN** (MAX mode) to get multiple perspectives before deciding.
+5. **Important code changes** \u2192 Call **EditorMultiPrompt** (MAX mode) to try multiple implementation strategies.
+6. **Code review** \u2192 **CodeReviewMulti** (MAX mode) runs automatically; don't duplicate its work.
+7. **Context management** \u2192 **ContextPruner** summarizes long conversations automatically.
+
+**Fallback rule:** If a sub-agent call fails or returns an error, THEN do the work yourself using basic tools (Read, Edit, Bash, Grep, etc.). Never give up \u2014 always fall back to manual if delegation fails.
+
+**Do NOT skip delegation** just because you think you can do it faster. Sub-agents provide better results through specialization. The only exceptions are:
+- Reading a single known file path (just use Read)
+- A single targeted grep for a known pattern (just use Grep)
+- Answering a question from memory/context (just respond)
 
 # Modes
 Current mode: ${getMode()}
@@ -33437,7 +33446,7 @@ var require_agent = __commonJS((exports, module2) => {
     const lower = text.toLowerCase();
     return EXPLORATION_KEYWORDS.some((keyword) => lower.includes(keyword));
   }
-  var COMPLEX_TASK_KEYWORDS = [
+  var CODING_TASK_KEYWORDS = [
     "refactor",
     "redesign",
     "architect",
@@ -33445,8 +33454,8 @@ var require_agent = __commonJS((exports, module2) => {
     "overhaul",
     "implement",
     "build",
-    "create a system",
-    "design pattern",
+    "create",
+    "design",
     "add feature",
     "new feature",
     "integrate",
@@ -33454,19 +33463,93 @@ var require_agent = __commonJS((exports, module2) => {
     "multiple files",
     "across files",
     "full stack",
-    "database schema",
-    "api endpoint",
+    "database",
+    "api",
+    "endpoint",
     "authentication",
     "complex",
     "tricky",
     "challenging",
-    "difficult"
+    "difficult",
+    "fix",
+    "bug",
+    "error",
+    "broken",
+    "crash",
+    "issue",
+    "problem",
+    "update",
+    "change",
+    "modify",
+    "edit",
+    "write",
+    "rewrite",
+    "add",
+    "remove",
+    "delete",
+    "move",
+    "rename",
+    "replace",
+    "optimize",
+    "improve",
+    "enhance",
+    "extend",
+    "upgrade",
+    "set up",
+    "setup",
+    "configure",
+    "install",
+    "connect",
+    "debug",
+    "troubleshoot",
+    "investigate",
+    "diagnose",
+    "test",
+    "spec",
+    "coverage"
   ];
-  function isComplexTask(text) {
+  function isCodingTask(text) {
     const lower = text.toLowerCase();
     const wordCount = text.split(/\s+/).length;
-    const keywordMatch = COMPLEX_TASK_KEYWORDS.some((kw) => lower.includes(kw));
-    return keywordMatch || wordCount > 40;
+    const keywordMatch = CODING_TASK_KEYWORDS.some((kw) => lower.includes(kw));
+    return keywordMatch || wordCount > 20;
+  }
+  var COMMAND_TASK_KEYWORDS = [
+    "run",
+    "execute",
+    "start",
+    "stop",
+    "restart",
+    "deploy",
+    "build",
+    "compile",
+    "bundle",
+    "test",
+    "lint",
+    "format",
+    "install",
+    "uninstall",
+    "upgrade",
+    "npm",
+    "bun",
+    "yarn",
+    "git commit",
+    "git push",
+    "git pull",
+    "git merge",
+    "git rebase",
+    "docker",
+    "container",
+    "serve",
+    "dev server",
+    "check status",
+    "disk space",
+    "process",
+    "kill"
+  ];
+  function isCommandTask(text) {
+    const lower = text.toLowerCase();
+    return COMMAND_TASK_KEYWORDS.some((kw) => lower.includes(kw)) && !isCodingTask(text);
   }
   async function handleUserInput(userInput) {
     isProcessing = true;
@@ -33492,23 +33575,42 @@ ${pickerResult}`
       }
     }
     const mode = getMode();
-    if (mode === "max" && isComplexTask(userInput)) {
-      const thinkId = store.addMessage({ role: "system", content: "Deep thinking...", label: "\uD83E\uDDE0 Thinker" });
+    if (isCommandTask(userInput)) {
+      const cmdId = store.addMessage({ role: "system", content: "Delegating to Commander...", label: "\u26A1 Commander" });
+      try {
+        const cmdResult = await executeTool("Commander", { prompt: userInput }, (partial) => {
+          store.updateMessage(cmdId, { content: partial, label: "\u26A1 Commander" });
+        });
+        store.updateMessage(cmdId, { content: cmdResult, label: "\u26A1 Commander" });
+        session.conversationHistory.push({
+          role: "assistant",
+          content: `[Commander result]
+${cmdResult}`
+        });
+      } catch (err) {
+        store.updateMessage(cmdId, { content: `Commander failed (falling back to main agent): ${err.message}` });
+      }
+    }
+    if (isCodingTask(userInput) && !isCommandTask(userInput)) {
+      const thinkLabel = mode === "max" ? "\uD83E\uDDE0 Deep Thinker (MAX)" : "\uD83E\uDDE0 Thinker";
+      const thinkId = store.addMessage({ role: "system", content: "Planning approach...", label: thinkLabel });
       try {
-        const thinkerResult = await executeTool("Thinker", { prompt: userInput }, (partial) => {
-          store.updateMessage(thinkId, { content: partial, label: "\uD83E\uDDE0 Thinker" });
+        const toolName = mode === "max" ? "ThinkerBestOfN" : "Thinker";
+        const toolArgs = mode === "max" ? { prompt: userInput, n: 3 } : { prompt: userInput };
+        const thinkerResult = await executeTool(toolName, toolArgs, (partial) => {
+          store.updateMessage(thinkId, { content: partial, label: thinkLabel });
         });
-        store.updateMessage(thinkId, { content: thinkerResult, label: "\uD83E\uDDE0 Thinker" });
+        store.updateMessage(thinkId, { content: thinkerResult, label: thinkLabel });
         session.conversationHistory.push({
           role: "assistant",
-          content: `[Thinker analysis]
+          content: `[Thinker analysis \u2014 follow this plan]
 ${thinkerResult}`
         });
       } catch (err) {
-        store.updateMessage(thinkId, { content: `Thinker failed: ${err.message}` });
+        store.updateMessage(thinkId, { content: `Thinker failed (falling back to main agent): ${err.message}` });
       }
     }
-    if (mode === "max" && session.conversationHistory.length > 20) {
+    if (session.conversationHistory.length > 20) {
       try {
         await executeTool("ContextPruner", {}, null);
       } catch {}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "apex-dev",
-  "version": "3.0.3",
+  "version": "3.1.0",
   "description": "Apex AI - a friendly agentic coding assistant for the terminal",
   "main": "dist/index.js",
   "bin": {