natureco-cli 5.64.4 → 5.64.6

package/CHANGELOG.md

@@ -2,6 +2,27 @@
 
 All notable changes to NatureCo CLI will be documented in this file.
 
+## [5.64.6] - 2026-07-13 — Typed MiniMax tool arguments
+
+### Fixed
+- MiniMax XML tool parameters are now coerced from text to their declared JSON Schema types before validation and execution.
+- `computer_use_loop` now accepts XML values such as `<parameter name="maxSteps">30</parameter>` as the number `30`, so visible GUI automation starts instead of failing with `expected number, got string`.
+- Conversion also covers integer, boolean, object, and array parameters while leaving invalid values untouched for normal schema validation.
+
+### Tests
+- Added regressions for valid type conversion, invalid-value preservation, and the exact `computer_use_loop` `maxSteps` failure reported on macOS.
+
+## [5.64.5] - 2026-07-13 — Deterministic visible-browser recovery
+
+### Fixed
+- Added `computer_use_loop` to the full-mode GUI tool set and documented it as the single authority for visible multi-step desktop tasks.
+- Clarified that the headless `browser` tool only accepts `open`, `screenshot`, `evaluate`, and `html`; it does not support `navigate`, `click`, or `type` and every call requires a URL.
+- After a verified GUI loop fails, the agent now stops the current turn instead of fanning out into blind screenshots, binary `read_file`, invalid headless-browser actions, or brittle AppleScript window/tab index enumeration.
+- GUI failures now include the last concrete verification/action error instead of only returning `Goal was not verified`.
+
+### Tests
+- Added regression coverage proving an unverified GUI loop ends the agentic turn immediately and preserves the failure reason.
+
 ## [5.64.4] - 2026-07-13 — Unified MiniMax media routing
 
 ### Added

package/README.md

@@ -51,6 +51,8 @@ natureco code
 
 | Version | Highlights |
 |---------|-----------|
+| **v5.64.6** | **Reliable MiniMax tool calls:** XML parameters are converted to their declared schema types, fixing `computer_use_loop` failures such as `maxSteps: expected number, got string`. |
+| **v5.64.5** | **Deterministic GUI recovery:** visible web tasks use one verified GUI loop; failures expose their concrete reason and stop blind browser/AppleScript fallback chains. |
 | **v5.64.4** | **Unified MiniMax media:** the existing MiniMax key now powers `MiniMax-VL-01` analysis, `image-01` generation, `MiniMax-Hailuo-2.3` video, and verified GUI vision—no second key required. |
 | **v5.64.3** | **Evidence-based GUI completion:** desktop actions require a changed screen plus independent visual evidence before success. MiniMax screenshots use the Token Plan VLM with the existing provider key. |
 | **v5.64.2** | **Reliable macOS GUI automation:** tool names stay visible in REPL transcripts, visual tasks use a verified screenshot loop, MiniMax GUI routing avoids duplicate paths, and failed/unverified actions no longer claim success. |

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "natureco-cli",
-  "version": "5.64.4",
+  "version": "5.64.6",
   "description": "Terminal-native AI agent CLI with bilingual TR/EN UI, multi-agent orchestration, persistent memory, secure tools and messaging integrations.",
   "bin": {
     "natureco": "bin/natureco.js"

package/src/tools/agentic-runner.js

@@ -112,7 +112,7 @@ const AGENT_EXEC_BLOCK_PATTERNS = [
 ];
 // Full modda ("agentExec: full") acilan computer-use araclari: tarayici otomasyonu,
 // uygulama ac/kapat, GUI (tikla/yaz/ekran goruntusu). Safe modda (varsayilan) KAPALI.
-const FULL_TOOLS = ['browser', 'browser_use', 'mac_app_open', 'mac_app_quit', 'social_open', 'computer_use', 'macos_screenshot'];
+const FULL_TOOLS = ['browser', 'browser_use', 'mac_app_open', 'mac_app_quit', 'social_open', 'computer_use', 'computer_use_loop', 'macos_screenshot'];
 function agentExecAllowed(command, opts) {
   if (opts && opts.full) return true;
   if (String(process.env.NATURECO_AGENT_EXEC || '').toLowerCase() === 'full') return true;
@@ -288,6 +288,37 @@ function buildFeedback(norm, res) {
   return body ? `${head}:\n${body}` : head;
 }
 
+function coerceArgsForSchema(args, schema) {
+  if (!args || typeof args !== 'object' || !schema || typeof schema !== 'object') return { ...(args || {}) };
+  const properties = schema.properties || {};
+  const result = { ...args };
+
+  for (const [key, property] of Object.entries(properties)) {
+    const value = result[key];
+    if (typeof value !== 'string' || !property) continue;
+    const types = Array.isArray(property.type) ? property.type : [property.type];
+    const trimmed = value.trim();
+
+    if (types.includes('integer') && /^[-+]?\d+$/.test(trimmed)) {
+      const number = Number(trimmed);
+      if (Number.isSafeInteger(number)) result[key] = number;
+    } else if (types.includes('number') && trimmed !== '' && Number.isFinite(Number(trimmed))) {
+      result[key] = Number(trimmed);
+    } else if (types.includes('boolean') && /^(true|false)$/i.test(trimmed)) {
+      result[key] = trimmed.toLowerCase() === 'true';
+    } else if (types.includes('object') || types.includes('array')) {
+      try {
+        const parsed = JSON.parse(trimmed);
+        if ((types.includes('array') && Array.isArray(parsed)) ||
+            (types.includes('object') && parsed && typeof parsed === 'object' && !Array.isArray(parsed))) {
+          result[key] = parsed;
+        }
+      } catch {}
+    }
+  }
+  return result;
+}
+
 /**
  * Tek bir parse edilmis cagriyi calistir.
  * Donus: { records: [...], feedback: '...' }
@@ -389,9 +420,11 @@ async function executeCall(call, opts = {}) {
     records.push({ tool: norm, status: 'error', error: 'execute yok' });
     return { records, feedback: `${norm} HATA: execute fonksiyonu yok` };
   }
+  const schema = mod.inputSchema || mod.parameters || (mod.default && (mod.default.inputSchema || mod.default.parameters));
+  const typedArgs = coerceArgsForSchema(args, schema);
   const res = await executeThroughGateway({
     toolName: norm,
-    args,
+    args: typedArgs,
     resolveTool: () => mod,
     execute: fn,
     normalizeSuccess: value => value,
@@ -401,7 +434,7 @@ async function executeCall(call, opts = {}) {
   const ok = typeof res === 'string' ? true : (res && res.success !== false);
   const status = ok ? 'done' : 'error';
   const feedback = buildFeedback(norm, res);
-  records.push({ tool: norm, status, args: sanitizeArgs(args), result: res, error: ok ? undefined : res.error });
+  records.push({ tool: norm, status, args: sanitizeArgs(typedArgs), result: res, error: ok ? undefined : res.error });
   return { records, feedback };
 }
 
@@ -439,6 +472,16 @@ async function runAgentic({ callModel, systemPrompt, historyMessages, task, tool
       allRecords.push(...records);
       feedbacks.push(feedback);
     }
+    // A verified GUI loop is the authority for visible multi-step desktop work.
+    // If it fails, do not let the text model fan out into blind screenshots,
+    // binary read_file calls, headless browser guesses, or brittle AppleScript
+    // tab enumeration in the same turn. Surface the evidence failure instead;
+    // a user-requested retry starts a clean new turn.
+    const guiFailure = allRecords.find(record => record.tool === 'computer_use_loop' && record.status === 'error');
+    if (guiFailure) {
+      finalReply = `GUI görevi doğrulanamadı: ${guiFailure.error || guiFailure.result?.error || 'bilinmeyen hata'}`;
+      break;
+    }
     messages.push({
       role: 'user',
       content: '<tool_results>\n' + feedbacks.join('\n') + '\n</tool_results>\nGorev tamamlandiysa ARAC CAGIRMADAN tek cumlelik ozet yaz. Devam gerekiyorsa sonraki araci cagir.',
@@ -452,4 +495,4 @@ async function runAgentic({ callModel, systemPrompt, historyMessages, task, tool
   return { records: allRecords, reply: finalReply, iterations };
 }
 
-module.exports = { parseAgenticCalls, stripProtocolTokens, executeCall, runAgentic, expandHome, makeStreamFilter, makeSanitizeStream, agentExecAllowed, buildFeedback, TOOL_ALIASES, DEFAULT_ALLOWED };
+module.exports = { parseAgenticCalls, stripProtocolTokens, executeCall, runAgentic, expandHome, makeStreamFilter, makeSanitizeStream, agentExecAllowed, buildFeedback, coerceArgsForSchema, TOOL_ALIASES, DEFAULT_ALLOWED };

package/src/tools/computer_use_loop.js

@@ -285,7 +285,9 @@ async function loop(goal, maxSteps) {
   }
 
   if (!completed) {
-    return { success: false, error: 'Goal was not verified', goal, totalSteps: steps.length, steps };
+    const lastFailure = [...steps].reverse().find(step => step.error);
+    const detail = lastFailure?.error || 'No visible completion evidence was produced';
+    return { success: false, error: 'Goal was not verified: ' + detail, goal, totalSteps: steps.length, steps };
   }
   return { success: true, verified: true, evidence: completionEvidence.evidence, confidence: completionEvidence.confidence, goal, totalSteps: steps.length, steps };
 }

package/src/tools/workflow.js

@@ -243,8 +243,10 @@ async function workflow(params) {
           '\n\nTAM MOD ACIK — su araclara da ERISIMIN VAR; o an ne gerekiyorsa dogrudan cagir:',
           '- mac_app_open: macOS uygulamasi ac. parametre: appName (orn. "WhatsApp", "Google Chrome", "Spotify")',
           '- mac_app_quit: macOS uygulamasi kapat. parametre: appName',
-          '- browser: HEADLESS tarayici otomasyonu (icerik cek/screenshot — kullaniciya GORUNMEZ). parametreler: action, url, script',
-          '- computer_use: GUI otomasyonu. parametreler: action ("screenshot"/"click"/"type"/"keypress"/"scroll"), x, y, text, key',
+          '- browser: HEADLESS ve ETKILESIMSIZ icerik araci; kullaniciya GORUNMEZ. SADECE action=open|screenshot|evaluate|html; HER CAGRI url ister. navigate/click/type DESTEKLEMEZ.',
+          '- browser_use: ayri bulut/CLI tarayici servisi; yalniz kurulu ve hazirsa kullan. Acik Chrome penceresini kontrol etmek icin kullanma.',
+          '- computer_use: GUI otomasyonu. parametreler: action ("screenshot"/"click"/"type"/"keypress"/"scroll"), x, y, text, key',
+          '- computer_use_loop: GORUNUR, cok-adimli GUI icin TEK tercih. p: goal, maxSteps?. Kendi screenshot→vision→action→dogrulama dongusunu yapar.',
           '- social_open: muzik/video/sosyal ac. parametreler: query, platform (spotify/youtube...)',
           '- macos_screenshot: ekran goruntusu al',
           '\nGORUNUR ACMA (onemli): Kullanici "kendi tarayicimda ac / gorunur ac / dinlemek/izlemek istiyorum" derse `browser` (headless, gorunmez) DEGIL, GORUNUR ac: macOS bash ile `open "https://..."` ya da `open -a "Google Chrome" "https://..."`; Windows `start "" "https://..."`. Uygulama icin `open -a WhatsApp` / mac_app_open. Muzik/video icin dogrudan YouTube/Spotify URL\'sini `open` ile ac.',
@@ -300,7 +302,7 @@ async function workflow(params) {
         '- Dosya degistirmeden ONCE read_file ile oku, edit_file ile hedefli degistir. Yerini bilmiyorsan file_search/grep_search ile kesfet.',
         '- Kod yazinca/degistirince bash ile calistir/test et, hata varsa duzelt. Coklu dosya = her biri icin AYRI write_file. Hep TAM yol; "masaustu" = ' + desktop + '.',
         '- Arac sonuclari <tool_results> icinde doner; is bitince arac cagirmadan tek cumlelik ozet yaz. Basit sohbette arac cagirma, dogrudan yanitla.',
-        execFull ? '- Kullanici uygulama ac / tarayici kontrol / muzik cal / ekran / GUI istediyse yukaridaki computer-use araclarini KULLAN — "yapamam/engellendi" DEME, dogrudan ilgili araci cagir. Gorsel geri bildirim gerektiren cok adimli GUI gorevlerinde tekil screenshot/click yerine computer_use_loop kullan; basariyi arac dogrulamadan tamamlandi deme.' : '',
+        execFull ? '- Kullanici uygulama ac / tarayici kontrol / muzik cal / ekran / GUI istediyse yukaridaki computer-use araclarini KULLAN. Gorsel geri bildirim gerektiren cok adimli GORUNUR GUI gorevinde SADECE computer_use_loop kullan; once tekil screenshot alip read_file ile PNG okuma, browser ile click/navigate deneme veya AppleScript ile pencere/tab indeksleri tarama. computer_use_loop hata verirse AYNI TURDA kor alternatif denemeler yapma; hata nedenini durustce bildir. Basariyi arac kanitlamadan tamamlandi deme.' : '',
         fullToolsBlock,
         treeDigest ? ('\n\nBILDIGIN KALICI HAFIZA (bu kullaniciya ait, onceki oturumlardan hatirladiklarin; kullaniciya ozel bir sey sorulursa ONCE BUNU KULLAN — dosya arama, uydurma):\n' + treeDigest) : '',
         treeIndex ? ('\n\nHafiza agaci yapisi (yukarida olmayan detay icin memory_tree(action:read/search) ile ilgili kok/dali oku):\n' + treeIndex) : '',