fix: SDK path resolution, debug prompt removal, improved error logging

src/server.ts

@@ -680,21 +680,7 @@ server.tool(
       "- `worklog_entry`: narrative summary (5-15 lines markdown)",
       "- `startup_text`: ready-to-paste text for next session",
       "",
-      "## Step 3: Verify AXME Code Functionality (ALPHA - MANDATORY)",
-      "",
-      "AXME Code is in alpha. Before finalizing, verify that core functionality worked correctly in this session.",
-      "Run these checks and report results to the user:",
-      "",
-      "1. **Context loading**: call `axme_status` - verify it returns data (not error)",
-      "2. **Saves**: if you saved any memories/decisions in this session, verify files exist on disk:",
-      "   - `ls .axme-code/decisions/` and `ls .axme-code/memory/` in the target storage",
-      "3. **Pending audits**: check `axme_context` output for 'Pending audits' section - report if any are stuck",
-      "4. **Safety hooks**: verify `.claude/settings.json` has PreToolUse + PostToolUse + SessionEnd hooks configured",
-      "5. **Session tracking**: verify current session meta exists: `ls .axme-code/sessions/` should have current session dir",
-      "",
-      "Report results as a table to the user. If anything is broken, flag it BEFORE calling finalize_close.",
-      "",
-      "## Step 4: Call `axme_finalize_close`",
+      "## Step 3: Call `axme_finalize_close`",
       "",
       "Pass everything in one call. MCP writes all files atomically.",
       "After it returns:",

src/tools/init.ts

@@ -174,7 +174,10 @@ export async function initProjectWithLLM(projectPath: string, opts?: {
   log(`  [${projectName}] Scanners complete, processing results...`);
   for (const settled of scanners) {
     if (settled.status === "rejected") {
-      errors.push(`LLM scan failed: ${settled.reason?.message ?? settled.reason}`);
+      const err = settled.reason;
+      const msg = err?.message ?? String(err);
+      const stack = err?.stack ? `\n${err.stack.split("\n").slice(0, 3).join("\n")}` : "";
+      errors.push(`LLM scan failed: ${msg}${stack}`);
       continue;
     }
     const val = settled.value;

src/utils/agent-options.ts

@@ -2,8 +2,22 @@
  * Shared agent query options builder for LLM scanner agents.
  */
 
+import { execSync } from "node:child_process";
+
 type Options = import("@anthropic-ai/claude-agent-sdk").Options;
 
+/** Find claude binary path. Cached after first lookup. */
+let _claudePath: string | undefined;
+function findClaudePath(): string | undefined {
+  if (_claudePath !== undefined) return _claudePath || undefined;
+  try {
+    _claudePath = execSync("which claude", { encoding: "utf-8" }).trim();
+  } catch {
+    _claudePath = "";
+  }
+  return _claudePath || undefined;
+}
+
 export type AgentRole = "scanner" | "tester" | "reviewer" | "engineer" | "architect" | "auditor";
 
 const ROLE_TOOLS: Record<AgentRole, { allowed: string[]; disallowed: string[] }> = {
@@ -41,6 +55,8 @@ export function buildAgentQueryOptions(base: {
 }, role: AgentRole): Options {
   const tools = ROLE_TOOLS[role];
 
+  const claudePath = findClaudePath();
+
   return {
     cwd: base.cwd,
     model: base.model,
@@ -49,6 +65,7 @@ export function buildAgentQueryOptions(base: {
       : { type: "preset" as const, preset: "claude_code" as const },
     settingSources: ["project"],
     ...(base.maxTurns !== undefined ? { maxTurns: base.maxTurns } : {}),
+    ...(claudePath ? { pathToClaudeCodeExecutable: claudePath } : {}),
     permissionMode: "bypassPermissions",
     allowDangerouslySkipPermissions: true,
     allowedTools: tools.allowed,