micro-models-agent 0.16.1 → 0.16.2

package/dist/core/agent.js

@@ -113,6 +113,8 @@ export class Agent {
         let apiPromptTokens = 0;
         let apiCompletionTokens = 0;
         const MAX_HALLUCINATION_RETRIES = 3;
+        let consecutiveToolFailures = 0;
+        const MAX_CONSECUTIVE_TOOL_FAILURES = 5;
         while (iteration < config.maxToolIterations) {
             iteration++;
             pluginManager.runOnBeforeThink({
@@ -246,6 +248,7 @@ export class Agent {
                     })),
                 });
                 const summaries = [];
+                let anyToolFailed = false;
                 for (const call of toolCalls) {
                     this.setScope();
                     const startTime = Date.now();
@@ -258,6 +261,8 @@ export class Agent {
                     slog.logToolCall(call, iteration);
                     const result = await toolExecutor.execute(call);
                     const duration = Date.now() - startTime;
+                    if (!result.success)
+                        anyToolFailed = true;
                     pluginManager.runOnToolEnd({ iteration, logger }, { id: call.id, name: call.name, arguments: call.arguments }, result, duration);
                     if (result.display) {
                         onMeta?.("\n" + result.display + "\n");
@@ -292,6 +297,22 @@ export class Agent {
                         logger.debug("Context compacted after tool result");
                     }
                 }
+                if (anyToolFailed) {
+                    consecutiveToolFailures++;
+                }
+                else {
+                    consecutiveToolFailures = 0;
+                }
+                if (consecutiveToolFailures >= MAX_CONSECUTIVE_TOOL_FAILURES) {
+                    const recoveryMsg = t("exec.consecutive_failures_recovery", { count: consecutiveToolFailures });
+                    logger.warn(`Consecutive tool failures: ${consecutiveToolFailures}`);
+                    const taskSnippet = input.length > 200 ? input.slice(0, 200) + "..." : input;
+                    const taskReminder = t("exec.task_reminder", { task: taskSnippet });
+                    contextManager.addMessage({
+                        role: "user",
+                        content: `<system-summary>${recoveryMsg}\n${taskReminder}</system-summary>`,
+                    });
+                }
                 contextManager.addMessage({
                     role: "user",
                     content: `<system-summary>${summaries.join("\n")}</system-summary>`,

package/dist/i18n/en.json

@@ -375,6 +375,8 @@
     "exec.tool_errors": "Tool {tool} failed {count} times. Suggest alternative approach.",
     "exec.tool_errors_recovery": "[Tool {tool} failed {count} times. Stop using this tool and try a different approach or explain the issue to the user.]",
     "exec.repetitive_tool": "[You called {tool} {count} times in a row with the same arguments and result. Stop repeating — try a different approach: create files, change arguments, or explain the problem to the user.]",
+    "exec.consecutive_failures_recovery": "[{count} consecutive tool calls failed. You are going in circles — stop trying commands that keep failing. Explain the problem to the user, or try a completely different approach: write files directly, use a different tool, or break the task into smaller parts.]",
+    "exec.task_reminder": "[Remember the original task: {task}. Do not switch to unrelated tasks. If you cannot complete the original task, explain why to the user and ask for guidance.]",
     "exec.plan_warning": "[\u26a0 Warning: you are working on step {step} - \"{description}\". Your {tool} call touches files not mentioned in the current step. Focus on the current step or update the plan via plan update.]",
     "exec.off_track": "[\u26a0 Outside plan: step {stepId} is \"{description}\", but you're using {tool} on a different path. Explain or return.]",
     "exec.audit_pass": "[\u2713] Task complete: {done}/{total} steps done, {files} files verified",

package/dist/i18n/ru.json

@@ -375,6 +375,8 @@
     "exec.tool_errors": "Инструмент {tool} сбоил {count} раз. Предложите другой подход.",
     "exec.tool_errors_recovery": "[Инструмент {tool} сбоил {count} раз. Прекратите использовать этот инструмент и попробуйте другой подход или объясните проблему пользователю.]",
     "exec.repetitive_tool": "[Вы вызвали {tool} {count} раз подряд с одинаковыми аргументами и результатом. Прекратите повторять — попробуйте другой подход: создайте файлы, измените аргументы или объясните пользователю проблему.]",
+    "exec.consecutive_failures_recovery": "[{count} последовательных вызовов тулов завершились ошибкой. Вы ходите по кругу — прекратите пытаться выполнить команды, которые постоянно падают. Объясните проблему пользователю или попробуйте совершенно другой подход: напишите файлы напрямую, используйте другой инструмент или разбейте задачу на части.]",
+    "exec.task_reminder": "[Напоминание о задаче: {task}. Не переключайтесь на посторонние задачи. Если не можете выполнить задачу — объясните почему и спросите совета.]",
     "exec.plan_warning": "[⚠ Внимание: вы работаете над шагом {step} — \"{description}\". Ваш вызов {tool} касается файлов, не упомянутых в текущем шаге. Сосредоточьтесь на текущем шаге или обновите план через plan update.]",
     "exec.off_track": "[⚠ Вне плана: шаг {stepId} — \"{description}\", но вы используете {tool} для другого пути. Объясните или вернитесь.]",
     "exec.audit_pass": "[✓] Задача выполнена: {done}/{total} шагов, {files} файлов проверено",

package/dist/modules/execution/module.js

@@ -271,7 +271,10 @@ export class ExecutionModule {
             },
             onAfterTool: (ctx, call, result) => {
                 if (!result.success) {
-                    this.stuckDetector.recordToolError("tool");
+                    this.stuckDetector.recordToolError(call.name);
+                }
+                else {
+                    this.stuckDetector.recordToolSuccess();
                 }
                 if (this.tracker &&
                     result.success &&

package/dist/modules/execution/stuck-detector.js

@@ -9,6 +9,8 @@ export class StuckDetector {
     recentToolCalls = [];
     maxRecentCalls = 10;
     repetitionThreshold = 3;
+    consecutiveFailures = 0;
+    lastFailedTool = '';
     constructor(threshold = 8, errorThreshold = 3) {
         this.threshold = threshold;
         this.errorThreshold = errorThreshold;
@@ -31,6 +33,12 @@ export class StuckDetector {
     }
     recordToolError(toolName) {
         this.toolErrors.set(toolName, (this.toolErrors.get(toolName) || 0) + 1);
+        this.consecutiveFailures++;
+        this.lastFailedTool = toolName;
+    }
+    recordToolSuccess() {
+        this.consecutiveFailures = 0;
+        this.lastFailedTool = '';
     }
     setCurrentStep(stepId, description) {
         if (stepId !== this.currentStepId) {
@@ -63,6 +71,12 @@ export class StuckDetector {
         }
         return count >= this.repetitionThreshold;
     }
+    hasConsecutiveFailures() {
+        return this.consecutiveFailures >= this.errorThreshold;
+    }
+    getConsecutiveFailuresCount() {
+        return this.consecutiveFailures;
+    }
     getRepetitiveToolMessage() {
         if (!this.hasRepetitiveToolCalls())
             return '';
@@ -83,6 +97,9 @@ export class StuckDetector {
         if (this.isStuck()) {
             return t('exec.stuck_recovery', { iterations: this.iterationsOnCurrentStep });
         }
+        if (this.hasConsecutiveFailures()) {
+            return t('exec.consecutive_failures_recovery', { count: this.consecutiveFailures });
+        }
         if (this.hasRepetitiveToolCalls()) {
             return this.getRepetitiveToolMessage();
         }
@@ -109,5 +126,7 @@ export class StuckDetector {
     reset() {
         this.iterationsOnCurrentStep = 0;
         this.toolErrors.clear();
+        this.consecutiveFailures = 0;
+        this.lastFailedTool = '';
     }
 }

package/dist/modules/security/command-validator.js

@@ -19,6 +19,8 @@ export const DEFAULT_BASH_CONFIG = DEFAULT_SECURITY_CONFIG?.bash || FALLBACK_BAS
  *   "NODE_ENV=prod node app.js"   → "node"
  *   "env PATH=/x rm -rf /"        → "rm"
  *   "cmd=rm; $cmd -rf /"          → "cmd" (variable indirection — not expanded)
+ *   "powershell Remove-Item x"    → "Remove-Item"
+ *   "cmd /c del /s /q"            → "del"
  */
 function extractBaseCommand(trimmed) {
     const tokens = trimmed.split(/\s+/);
@@ -30,7 +32,8 @@ function extractBaseCommand(trimmed) {
             i++;
             continue;
         }
-        if (tok === "sudo" || tok === "env" || tok === "command" || tok === "exec" || tok === "nohup") {
+        if (tok === "sudo" || tok === "env" || tok === "command" || tok === "exec" || tok === "nohup" ||
+            tok === "powershell" || tok === "pwsh" || tok === "cmd") {
             i++;
             continue;
         }

package/dist/tools/bash.js

@@ -6,6 +6,7 @@ import { runCommand, processRegistry, isLongRunningCommand } from '../modules/pr
 import { t } from '../i18n/index';
 import { platform } from 'os';
 const BASH_TIMEOUT_MS = 120_000;
+const MAX_PREVIEW_LINES = 25;
 function adaptCommandForWindows(command) {
     if (platform() !== 'win32')
         return command;
@@ -121,6 +122,10 @@ export const bashTool = {
             if (securityConfig?.logCommands) {
                 logBashCommand(ctx.sessionId, sanitizeCommandForLog(command), res.code === 0, `Working directory: ${workdir}, Output length: ${res.stdout.length}`);
             }
+            const lines = output.split('\n');
+            if (lines.length > MAX_PREVIEW_LINES) {
+                output = lines.slice(0, MAX_PREVIEW_LINES).join('\n') + `\n... (${lines.length - MAX_PREVIEW_LINES} more lines)`;
+            }
             return { success: res.code === 0, output };
         }
         catch (e) {

package/dist/tools/glob-tool.js

@@ -1,8 +1,9 @@
 import { globSync } from 'fs';
 import { t } from '../i18n/index';
+const MAX_PREVIEW_LINES = 25;
 export const globTool = {
     name: 'glob',
-    description: 'Search for files matching a glob pattern. Uses standard glob syntax (e.g., **/*.ts, src/**/*.test.ts).',
+    description: `Search for files matching a glob pattern. Shows up to ${MAX_PREVIEW_LINES} results by default. Uses standard glob syntax (e.g., **/*.ts, src/**/*.test.ts).`,
     tags: ['file', 'code', 'research'],
     parameters: {
         type: 'object',
@@ -14,6 +15,12 @@ export const globTool = {
     handler: async (ctx, args) => {
         const pattern = String(args.pattern);
         const results = globSync(pattern, { cwd: ctx.baseDir });
-        return { success: true, output: results.length > 0 ? results.join('\n') : t('file.no_matches') };
+        if (results.length === 0) {
+            return { success: true, output: t('file.no_matches') };
+        }
+        const truncated = results.length > MAX_PREVIEW_LINES
+            ? `\n... (${results.length - MAX_PREVIEW_LINES} more files)`
+            : '';
+        return { success: true, output: results.slice(0, MAX_PREVIEW_LINES).join('\n') + truncated };
     },
 };

package/dist/tools/grep-tool.js

@@ -2,9 +2,16 @@ import { execFileSync, execSync } from 'child_process';
 import { resolve } from 'path';
 import { t } from '../i18n/index';
 import { logBashCommand } from '../modules/security/audit-log';
+const MAX_PREVIEW_LINES = 25;
+function truncateLines(output) {
+    const lines = output.split('\n');
+    if (lines.length <= MAX_PREVIEW_LINES)
+        return output;
+    return lines.slice(0, MAX_PREVIEW_LINES).join('\n') + `\n... (${lines.length - MAX_PREVIEW_LINES} more lines)`;
+}
 export const grepTool = {
     name: 'grep',
-    description: 'Search file contents using a regular expression. Uses ripgrep (rg) if available, otherwise falls back to grep -r.',
+    description: `Search file contents using a regular expression. Shows up to ${MAX_PREVIEW_LINES} matching lines by default. Uses ripgrep (rg) if available, otherwise falls back to grep -r.`,
     tags: ['file', 'code', 'research'],
     parameters: {
         type: 'object',
@@ -32,7 +39,7 @@ export const grepTool = {
                 maxBuffer: 1024 * 1024,
                 cwd: ctx.baseDir,
             });
-            return { success: true, output: output || t('file.no_matches') };
+            return { success: true, output: truncateLines(output || t('file.no_matches')) };
         }
         catch (e) {
             if (e.status === 1)
@@ -45,7 +52,7 @@ export const grepTool = {
                     maxBuffer: 1024 * 1024,
                     cwd: ctx.baseDir,
                 });
-                return { success: true, output: output || t('file.no_matches') };
+                return { success: true, output: truncateLines(output || t('file.no_matches')) };
             }
             catch (e2) {
                 if (e2.status === 1)

package/dist/tools/list-dir.js

@@ -3,9 +3,10 @@ import { resolve } from 'path';
 import { t } from '../i18n/index';
 import { isPathInScope } from '../modules/security/path-validator';
 import { safeResolvePath } from './path-utils';
+const MAX_PREVIEW_LINES = 25;
 export const listDirTool = {
     name: 'list_dir',
-    description: 'List files and directories in a given path.',
+    description: `List files and directories in a given path. Shows up to ${MAX_PREVIEW_LINES} entries by default.`,
     tags: ['file'],
     parameters: {
         type: 'object',
@@ -35,6 +36,12 @@ export const listDirTool = {
             const full = resolve(resolved, e);
             return statSync(full).isDirectory() ? `${e}/` : e;
         });
-        return { success: true, output: lines.join('\n') || t('file.empty') };
+        if (lines.length === 0) {
+            return { success: true, output: t('file.empty') };
+        }
+        const truncated = lines.length > MAX_PREVIEW_LINES
+            ? `\n... (${lines.length - MAX_PREVIEW_LINES} more entries)`
+            : '';
+        return { success: true, output: lines.slice(0, MAX_PREVIEW_LINES).join('\n') + truncated };
     },
 };

package/dist/tools/process-log.js

@@ -1,14 +1,15 @@
 import { processRegistry } from '../modules/processes';
 import { t } from '../i18n/index';
+const DEFAULT_TAIL = 25;
 export const processLogTool = {
     name: 'process_log',
-    description: 'Show the buffered output of a background process started via bash. Use after starting a dev server to verify it came up without errors, and while it runs to check its state.',
+    description: `Show the buffered output of a background process started via bash. Shows the last ${DEFAULT_TAIL} lines by default. Use after starting a dev server to verify it came up without errors, and while it runs to check its state.`,
     tags: ['shell'],
     parameters: {
         type: 'object',
         properties: {
             id: { type: 'string', description: 'Process id from bash output or process_list' },
-            tail: { type: 'number', description: 'Number of trailing lines to show (default: all buffered lines, max 300)' },
+            tail: { type: 'number', description: `Number of trailing lines to show (default: ${DEFAULT_TAIL}, max 300)` },
         },
         required: ['id'],
     },
@@ -23,7 +24,7 @@ export const processLogTool = {
             : entry.status === 'exited'
                 ? t('proc.status_exited')
                 : t('proc.status_killed');
-        const tail = typeof args.tail === 'number' ? args.tail : undefined;
+        const tail = typeof args.tail === 'number' ? args.tail : DEFAULT_TAIL;
         const log = processRegistry.getLog(id, tail);
         if (!log) {
             return {

package/dist/tools/read-file.js

@@ -6,7 +6,7 @@ import { DEFAULT_SECURITY_CONFIG } from "../config/security";
 import { logSecurityBlock } from "../modules/security/audit-log";
 import { safeResolvePath } from "./path-utils";
 /** Default number of lines returned when the caller omits `limit`. */
-const DEFAULT_LIMIT = 50;
+const DEFAULT_LIMIT = 25;
 export const readFileTool = {
     name: "read_file",
     description: `Read a file from the filesystem. Reads up to ${DEFAULT_LIMIT} lines at a time by default; use offset to page through large files.`,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "micro-models-agent",
-  "version": "0.16.1",
+  "version": "0.16.2",
   "description": "Micro Models Agent (MMA) — LLM agent harness for small models (Qwen3.5-9B, 32K-64K context)",
   "type": "module",
   "bin": {