@meshxdata/fops 0.1.36 → 0.1.37

package/CHANGELOG.md

@@ -1,3 +1,25 @@
+# Changelog
+
+All notable changes to @meshxdata/fops (Foundation Operator CLI) are documented here.
+
+## [0.1.37] - 2026-03-06
+
+- **Azure plugin & VM management:** Major refactor of the Azure plugin with improved VM lifecycle management, non-interactive VM support, better state handling, and expanded test coverage. Includes Azure-specific fixes, token refresh improvements, and enhanced preflight checks.
+
+- **Agent & retrieval improvements:** Added Reciprocal Rank Fusion (RRF) to knowledge and skill retrieval, improved Jaccard scoring, expanded agent context, and reduced context size (100k → 10k) via compaction. Various skills updates and documentation improvements for knowledge tools.
+
+- **Predictions & ML:** Introduced support for local predictions, added prediction warnings, and early MLflow integration work.
+
+- **CLI & TUI improvements:** Fixed multiple TUI/UI issues (status per tab, chat display, compute display), improved `up/down/bootstrap` flows, added plan mode, preserved environment variables with `sudo`, and enhanced sequential startup and container reconciliation.
+
+- **Docker & multi-arch:** Added multi-arch Spark Dockerfile, improved multi-arch support, ensured OPA inclusion in compose, and replaced `localhost` with `127.0.0.1` for better compatibility.
+
+- **Storage, compute & data reliability:** Improved `compute_run` readiness checks (CSV header validation, table probes), added storage read enhancements, strengthened checksum handling, fixed corrupted Iceberg metadata cleanup, and improved storage defaults and defensive checks.
+
+- **Plugins & integrations:** Added TTYD plugin, improved compose/fops file plugin, GitHub login and credential handling fixes, Cloudflare CIDR updates, Kafka reconciliation improvements, and registry enhancements.
+
+- **Maintenance & deps:** Dependency bumps (e.g., tar, ajv, hono), test fixes and expansions, repo splits/refactors, packer cleanup, and general stability improvements.
+
 ## [0.1.36] - 2026-03-05
 
 - **Azure & Cloud Enhancements**

package/fops.mjs

@@ -37,6 +37,19 @@ function loadProjectEnv() {
         }
         process.env[key] = value;
       }
+      // Save the directory of the found .env as projectRoot so future runs from any
+      // directory (e.g. ~) can find credentials via findProjectRoot() in client.js.
+      try {
+        const home = process.env.HOME || process.env.USERPROFILE || "";
+        const fopsJsonPath = path.join(home, ".fops.json");
+        const foundRoot = path.dirname(envPath);
+        let cfg = {};
+        try { cfg = JSON.parse(fs.readFileSync(fopsJsonPath, "utf8")); } catch {}
+        if (cfg.projectRoot !== foundRoot) {
+          cfg.projectRoot = foundRoot;
+          fs.writeFileSync(fopsJsonPath, JSON.stringify(cfg, null, 2) + "\n");
+        }
+      } catch { /* non-fatal */ }
       break; // use the first .env found
     } catch { /* ignore read errors */ }
   }
@@ -48,15 +61,32 @@ loadProjectEnv();
 // whenever ONNX was loaded; otherwise it uses regular process.exit.
 const _realExit = process.exit.bind(process);
 function safeExit(code) {
-  // Defer hard exit when native addons (ONNX, better-sqlite3) were used so C++ destructors
-  // don't run during exit and trigger mutex/abort.
-  if (globalThis.__onnxLoaded || globalThis.__onnxDone || globalThis.__sqliteStoreUsed) {
-    process.exitCode = code;
+  if (globalThis.__exitScheduled) return;
+  globalThis.__exitScheduled = true;
+
+  process.exitCode = code ?? 0;
+
+  if (globalThis.__onnxLoaded) {
+    // ONNX C++ destructors call abort() — bypass them with SIGKILL.
+    try { fs.closeSync(2); } catch {}
     globalThis.__onnxDone = true;
-    globalThis.__sqliteStoreUsed = false;
+    if (code === 0 || code == null) {
+      // Unref'd: let event loop drain naturally; SIGKILL only if handles linger.
+      setTimeout(() => process.kill(process.pid, "SIGKILL"), 400).unref();
+    } else {
+      // Ref'd: ensure we exit even if event loop is stuck.
+      setTimeout(() => process.kill(process.pid, "SIGKILL"), 100);
+    }
     return;
   }
-  _realExit(code);
+
+  // ONNX was never loaded — safe to exit normally without SIGKILL.
+  // Use reallyExit to bypass lingering handles (Ink, MCP) without crashing.
+  if (typeof process.reallyExit === "function") {
+    process.reallyExit(code ?? 0);
+  } else {
+    _realExit(code ?? 0);
+  }
 }
 // Intercept all process.exit() calls so plugin code (which can't import safeExit)
 // also avoids the ONNX C++ destructor abort.
@@ -141,16 +171,9 @@ program.parseAsync().then(() => {
   // Without this, commands hang because MCP connections / ink refs keep the event loop alive.
   const code = typeof process.exitCode === "number" ? process.exitCode : 0;
   setTimeout(() => {
-    if (globalThis.__onnxDone) return;
+    if (globalThis.__exitScheduled) return;
     safeExit(code);
   }, 200);
-
-  // Fallback: if ONNX native handles keep the event loop alive after safeExit
-  // set the exit code, force-terminate via SIGKILL (avoids atexit / C++ destructor crash).
-  // The timer is unref'd so it won't itself prevent a clean exit.
-  if (globalThis.__onnxLoaded || globalThis.__onnxDone) {
-    setTimeout(() => process.kill(process.pid, "SIGKILL"), 300).unref();
-  }
 }).catch((err) => {
   console.error("Command failed:", err?.message || err);
   if (err?.stack) console.error(err.stack);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@meshxdata/fops",
-  "version": "0.1.36",
+  "version": "0.1.37",
   "description": "CLI to install and manage data mesh platforms",
   "keywords": [
     "fops",