micro-models-agent 0.24.3 → 0.24.5

package/dist/main.js

@@ -2125,9 +2125,9 @@ var init_defaults = __esm(() => {
       offloadKvCacheToGpu: true
     },
     retry: {
-      maxRetries: 3,
-      baseDelay: 1000,
-      maxDelay: 30000
+      maxRetries: 5,
+      baseDelay: 5000,
+      maxDelay: 120000
     },
     maxToolIterations: 1000,
     stuckThreshold: 8,
@@ -4435,12 +4435,12 @@ class OpenAICompatProvider {
     };
     this.rateLimiter = createRateLimiter(config.rateLimits);
   }
-  async* chat(messages, tools) {
+  async* chat(messages, tools, signal) {
     if (!this.rateLimiter.canMakeRequest()) {
       throw new Error(`Rate limit exceeded: ${this.rateLimiter.getConfig().maxRequestsPerMinute} requests per minute`);
     }
     this.rateLimiter.recordRequest();
-    const streamResult = this.doStream(messages, tools);
+    const streamResult = this.doStream(messages, tools, signal);
     let hasToolCall = false;
     let hasText = false;
     let reasoningAcc = "";
@@ -4455,13 +4455,13 @@ class OpenAICompatProvider {
       yield chunk;
     }
     if (!hasToolCall && !hasText) {
-      const fallback = await this.doNonStreaming(messages, tools);
+      const fallback = await this.doNonStreaming(messages, tools, signal);
       for (const chunk of fallback) {
         yield chunk;
       }
     }
   }
-  async* doStream(messages, tools) {
+  async* doStream(messages, tools, externalSignal) {
     const body = {
       model: this.model,
       messages,
@@ -4488,6 +4488,15 @@ class OpenAICompatProvider {
     const controller = new AbortController;
     const totalTimeoutMs = 120000;
     const timeoutId = setTimeout(() => controller.abort(), totalTimeoutMs);
+    if (externalSignal) {
+      if (externalSignal.aborted) {
+        controller.abort();
+      } else {
+        externalSignal.addEventListener("abort", () => controller.abort(), {
+          once: true
+        });
+      }
+    }
     const response = await this.fetchWithRetry(`${this.config.baseUrl}/chat/completions`, {
       method: "POST",
       headers,
@@ -4591,7 +4600,7 @@ class OpenAICompatProvider {
       reader.releaseLock();
     }
   }
-  async doNonStreaming(messages, tools) {
+  async doNonStreaming(messages, tools, externalSignal) {
     const body = {
       model: this.model,
       messages,
@@ -4616,11 +4625,15 @@ class OpenAICompatProvider {
       headers["Authorization"] = `Bearer ${this.config.apiKey}`;
     }
     try {
-      const response = await this.fetchWithRetry(`${this.config.baseUrl}/chat/completions`, {
+      const fetchInit = {
         method: "POST",
         headers,
         body: JSON.stringify(body)
-      });
+      };
+      if (externalSignal) {
+        fetchInit.signal = externalSignal;
+      }
+      const response = await this.fetchWithRetry(`${this.config.baseUrl}/chat/completions`, fetchInit);
       if (!response.ok) {
         const errorText = await response.text();
         throw new Error(t("error.llm_api", {
@@ -4884,7 +4897,8 @@ function runCommand(command, options = {}) {
     cwd,
     callId,
     timeoutMs = 120000,
-    maxBuffer = 10 * 1024 * 1024
+    maxBuffer = 10 * 1024 * 1024,
+    onOutput
   } = options;
   return new Promise((resolve) => {
     let timedOut = false;
@@ -4903,6 +4917,7 @@ function runCommand(command, options = {}) {
     const decodeChunk = (chunk) => decoder ? decoder.decode(chunk, { stream: true }) : chunk.toString("utf-8");
     child.stdout?.on("data", (chunk) => {
       const text = decodeChunk(chunk);
+      onOutput?.(text);
       const next = stdoutRef.value.length + text.length;
       if (next > maxBuffer) {
         stdoutRef.value = stdoutRef.value.slice(stdoutRef.value.length + text.length - maxBuffer) + text;
@@ -4912,6 +4927,7 @@ function runCommand(command, options = {}) {
     });
     child.stderr?.on("data", (chunk) => {
       const text = decodeChunk(chunk);
+      onOutput?.(text);
       const next = stderrRef.value.length + text.length;
       if (next > maxBuffer) {
         stderrRef.value = stderrRef.value.slice(stderrRef.value.length + text.length - maxBuffer) + text;
@@ -4945,7 +4961,13 @@ function runCommand(command, options = {}) {
       clearTimeout(timer);
       if (callId)
         activeChildren.delete(callId);
-      resolve({ stdout: stdoutRef.value, stderr: stderrRef.value, code, signal, timedOut });
+      resolve({
+        stdout: stdoutRef.value,
+        stderr: stderrRef.value,
+        code,
+        signal,
+        timedOut
+      });
     });
   });
 }
@@ -4954,11 +4976,61 @@ var init_runner = __esm(() => {
   activeChildren = new Map;
 });
 
+// src/tools/cache.ts
+class ToolResultCache {
+  cache = new Map;
+  maxEntries;
+  defaultTtlMs;
+  constructor(maxEntries = 50, defaultTtlMs = 60000) {
+    this.maxEntries = maxEntries;
+    this.defaultTtlMs = defaultTtlMs;
+  }
+  get(key) {
+    const entry = this.cache.get(key);
+    if (!entry)
+      return;
+    if (Date.now() > entry.expiresAt) {
+      this.cache.delete(key);
+      return;
+    }
+    this.cache.delete(key);
+    this.cache.set(key, entry);
+    return entry.value;
+  }
+  set(key, value, ttlMs) {
+    if (this.cache.size >= this.maxEntries) {
+      const firstKey = this.cache.keys().next().value;
+      if (firstKey !== undefined) {
+        this.cache.delete(firstKey);
+      }
+    }
+    this.cache.set(key, {
+      value,
+      expiresAt: Date.now() + (ttlMs ?? this.defaultTtlMs)
+    });
+  }
+  has(key) {
+    return this.get(key) !== undefined;
+  }
+  clear() {
+    this.cache.clear();
+  }
+  get size() {
+    return this.cache.size;
+  }
+}
+function hashKey(toolName, args) {
+  const payload = toolName + JSON.stringify(args);
+  const hash = Bun.hash(payload).toString(16).slice(0, 16);
+  return hash;
+}
+
 // src/tools/executor.ts
 class ToolExecutor {
   registry;
   ctx;
   pluginManager;
+  cache = new ToolResultCache(50, CACHE_TTL_MS);
   constructor(registry, ctx, pluginManager) {
     this.registry = registry;
     this.ctx = ctx;
@@ -4973,6 +5045,18 @@ class ToolExecutor {
         toolCallId: call.id
       };
     }
+    if (CACHEABLE_TOOLS.has(call.name)) {
+      const key = hashKey(call.name, call.arguments);
+      const cached = this.cache.get(key);
+      if (cached !== undefined) {
+        this.ctx.logger.debug(`Tool ${call.name}: CACHE HIT`);
+        return {
+          success: true,
+          output: cached,
+          toolCallId: call.id
+        };
+      }
+    }
     for (const plugin of this.pluginManager.getAllPlugins()) {
       if (plugin.onBeforeTool) {
         try {
@@ -5043,6 +5127,11 @@ class ToolExecutor {
         }
       }
     }
+    if (result.success && CACHEABLE_TOOLS.has(call.name)) {
+      const key = hashKey(call.name, call.arguments);
+      this.cache.set(key, result.output);
+      this.ctx.logger.debug(`Tool ${call.name}: CACHED`);
+    }
     this.ctx.logger.debug(`Tool ${call.name}: ${result.success ? "OK" : "FAIL"}`);
     return result;
   }
@@ -5052,6 +5141,9 @@ class ToolExecutor {
   getRegistry() {
     return this.registry;
   }
+  getContext() {
+    return this.ctx;
+  }
   setScope(scope) {
     this.ctx.scope = scope;
   }
@@ -5059,10 +5151,11 @@ class ToolExecutor {
     this.ctx.llmProvider = provider;
   }
 }
-var TOOL_EXECUTION_TIMEOUT_MS = 60000;
+var TOOL_EXECUTION_TIMEOUT_MS = 60000, CACHEABLE_TOOLS, CACHE_TTL_MS = 60000;
 var init_executor = __esm(() => {
   init_runner();
   init_i18n();
+  CACHEABLE_TOOLS = new Set(["web_fetch"]);
 });
 
 // src/modules/security/path-validator.ts
@@ -5593,7 +5686,7 @@ function safeResolvePath(baseDir, userPath) {
 var init_path_utils = () => {};
 
 // src/tools/preview.ts
-var MAX_PREVIEW_LINES = 15;
+var MAX_PREVIEW_LINES = 100;
 
 // src/tools/read-file.ts
 import { readFileSync as readFileSync5, existsSync as existsSync9 } from "fs";
@@ -6957,8 +7050,14 @@ var init_bash = __esm(() => {
       type: "object",
       properties: {
         command: { type: "string", description: "Shell command to execute" },
-        workdir: { type: "string", description: "Working directory (default: baseDir)" },
-        background: { type: "boolean", description: "Start the command in the background and return immediately with a process id (default: auto-detect long-running commands)" }
+        workdir: {
+          type: "string",
+          description: "Working directory (default: baseDir)"
+        },
+        background: {
+          type: "boolean",
+          description: "Start the command in the background and return immediately with a process id (default: auto-detect long-running commands)"
+        }
       },
       required: ["command"]
     },
@@ -7005,7 +7104,8 @@ ${t("proc.manage_hint", {
         const res = await runCommand(command, {
           cwd: workdir,
           callId: ctx.activeCallId,
-          timeoutMs: BASH_TIMEOUT_MS
+          timeoutMs: BASH_TIMEOUT_MS,
+          onOutput: ctx.onMeta
         });
         const parts = [res.stdout.trimEnd(), res.stderr.trimEnd()].filter(Boolean);
         let output = parts.join(`
@@ -9065,6 +9165,16 @@ var init_store = __esm(() => {
 
 // src/core/agent.ts
 import { join as join10 } from "path";
+function canRunInParallel(calls) {
+  if (calls.length <= 1)
+    return false;
+  for (const call of calls) {
+    if (SIDE_EFFECT_TOOLS.has(call.name) || WRITE_TOOLS.has(call.name)) {
+      return false;
+    }
+  }
+  return true;
+}
 
 class Agent {
   deps;
@@ -9232,7 +9342,7 @@ ${t("tool.truncated", { tokens: Math.ceil(removedChars / 2) })}`;
       const textChunks = [];
       this.emitPhase(iteration, "thinking", onPhase);
       try {
-        for await (const chunk of llmProvider.chat(history, allTools)) {
+        for await (const chunk of llmProvider.chat(history, allTools, this.abortController?.signal)) {
           if (this.shutdownRequested)
             break;
           if (chunk.type === "text" && chunk.content) {
@@ -9327,7 +9437,7 @@ ${t("tool.truncated", { tokens: Math.ceil(removedChars / 2) })}`;
         });
         const summaries = [];
         let anyToolFailed = false;
-        for (const call of toolCalls) {
+        const executeToolCall = async (call) => {
           this.setScope();
           const startTime = Date.now();
           pluginManager.runOnToolCall({
@@ -9338,11 +9448,28 @@ ${t("tool.truncated", { tokens: Math.ceil(removedChars / 2) })}`;
           onTool?.({ type: "start", tool: call.name, args: call.arguments });
           slog.logToolCall(call, iteration);
           const tokensBeforeTool = contextManager.getEstimatedTokens();
+          const toolCtx = this.deps.toolExecutor?.getContext?.();
+          const prevOnMeta = toolCtx?.onMeta;
+          if (toolCtx) {
+            toolCtx.onMeta = onMeta;
+          }
           const result = await toolExecutor.execute(call, this.abortController?.signal);
+          if (toolCtx) {
+            toolCtx.onMeta = prevOnMeta;
+          }
           const duration = Date.now() - startTime;
+          pluginManager.runOnToolEnd({ iteration, logger }, { id: call.id, name: call.name, arguments: call.arguments }, result, duration);
+          return { call, result, duration, tokensBefore: tokensBeforeTool };
+        };
+        const processToolResult = (entry) => {
+          const {
+            call,
+            result,
+            duration,
+            tokensBefore: tokensBeforeTool
+          } = entry;
           if (!result.success)
             anyToolFailed = true;
-          pluginManager.runOnToolEnd({ iteration, logger }, { id: call.id, name: call.name, arguments: call.arguments }, result, duration);
           if (result.display) {
             onMeta?.(`
 ` + result.display + `
@@ -9386,6 +9513,17 @@ ${t("tool.truncated", { tokens: Math.ceil(removedChars / 2) })}`;
             contextManager.compact();
             logger.debug("Context compacted after tool result");
           }
+        };
+        if (canRunInParallel(toolCalls)) {
+          const entries = await Promise.all(toolCalls.map(executeToolCall));
+          for (const entry of entries) {
+            processToolResult(entry);
+          }
+        } else {
+          for (const call of toolCalls) {
+            const entry = await executeToolCall(call);
+            processToolResult(entry);
+          }
         }
         if (anyToolFailed) {
           consecutiveToolFailures++;
@@ -9596,7 +9734,7 @@ ${taskReminder}</system-summary>`
     });
   }
 }
-var TOOL_RESULT_MAX_TOKENS_RATIO = 0.3, TOOL_RESULT_ABSOLUTE_MAX_CHARS = 15000;
+var TOOL_RESULT_MAX_TOKENS_RATIO = 0.3, TOOL_RESULT_ABSOLUTE_MAX_CHARS = 15000, WRITE_TOOLS, SIDE_EFFECT_TOOLS;
 var init_agent = __esm(() => {
   init_i18n();
   init_colors();
@@ -9604,6 +9742,14 @@ var init_agent = __esm(() => {
   init_processes();
   init_agent_moe();
   init_store();
+  WRITE_TOOLS = new Set([
+    "write_file",
+    "edit_file",
+    "delete_file",
+    "move_file",
+    "create_dir"
+  ]);
+  SIDE_EFFECT_TOOLS = new Set(["bash", "subagent", "pipeline_run"]);
 });
 
 // src/modules/context/manager.ts
@@ -15309,6 +15455,7 @@ function buildSystemInfo(config, baseDir, profileCompressed) {
     lines.push(``, `Windows environment — use Windows-compatible commands:`, `- Use "dir" instead of "ls". Use "dir /b" for bare listing.`, `- Use "type" or "Get-Content" instead of "cat".`, `- Use "cd" instead of "pwd". Use "echo %cd%" to print working directory.`, `- Use "copy" instead of "cp", "move" instead of "mv", "del" instead of "rm".`, `- Do not use "mkdir -p" — Windows mkdir creates intermediate dirs by default. Use the create_dir tool instead.`, `- Use forward slashes (/) or escaped backslashes (\\\\) in file paths.`, `- Your working directory is: ${baseDir}`);
   }
   lines.push(``, `Bash tool rules:`, `- Use the "workdir" parameter to run commands in a specific directory. Prefer workdir over "cd dir && cmd" chaining.`, `- Run one command per tool call. Split multi-step shell operations into separate bash calls.`, `- Background processes (dev servers, watchers, long-running npm install): use the "background: true" parameter or the tool will auto-detect and background them. Check output with process_log.`);
+  lines.push(``, `Parallel tool calls:`, `- When you need to read multiple files, call read_file for ALL of them in a SINGLE response (one tool_calls array with multiple read_file calls). They will execute in parallel — much faster than reading one at a time.`, `- Example: to read 3 files, return one assistant message with 3 read_file tool_calls. Do NOT read them one by one in separate iterations.`, `- This works for: read_file, glob, grep, file_info, list_dir — any read-only operations.`, `- Do NOT parallelize write operations (write_file, edit_file, delete_file, bash) — these run sequentially.`);
   lines.push(``, `=== DEVELOPMENT RULES — follow these strictly ===`, ``, `1. DEPENDENCIES FIRST: Before writing any source code, ALWAYS install project dependencies (e.g., "npm install", "pip install -r requirements.txt", "cargo build", "go mod tidy"). Verify the package manager's lock file or dependency directory exists. Never write code that imports/uses packages that aren't installed yet.`, `2. TOOLKIT/FWK FIRST: If the task specifies a framework or UI library, initialize and configure it BEFORE writing application code. Run its project init command first, then add components/modules. Never write your own version of what the framework already provides.`, `3. ONE STEP AT A TIME: Follow the plan sequentially. Complete step N before starting step N+1. When a step is done: verify the deliverables exist and have real content (not empty), then call "plan update step=N status=done". Do not redo completed work.`, `4. VERIFY YOUR WORK: After creating/modifying files, verify they exist on disk. After installing dependencies, verify the package manager completed successfully. After any command, check its output for errors. Don't assume operations succeeded.`, `5. NO PREMATURE WORK: Do not create files for future steps. Do not add imports/references to packages or modules that haven't been installed yet. Do not reference files or components that don't exist yet. Build incrementally — one layer at a time.`, `6. WHEN STUCK: If a command fails 2+ times, STOP and try a different approach. Write files directly instead of using commands. Ask the user for help. Never repeat the same failing command more than twice.`, ``, `=== PLAN QUALITY RULES — your plan MUST follow these ===`, ``, `- Each step must describe CONCRETE deliverables: exact filenames with paths, exact packages to install, exact CLI commands to run. Avoid vague steps — be specific.`, `- A step like "Настройка проекта" or "Setup the project" is too vague — describe what exactly needs to be configured or set up.`, `- A step like "Создать src/components/Header.tsx с навигацией и логотипом, добавить в src/App.tsx импорт <Header />" is GOOD.`, `- Include file extensions (.tsx, .css, .json) and directory paths. Every step must mention at least one file or command.`, `- The plan must cover EVERYTHING needed: init → deps → framework setup → code → verification.`, `- Number of steps: 5-8 for a typical task. Too few means you're being vague. Too many means you're over-splitting.`);
   if (config.autoPlan) {
     lines.push(``, `Plan rule (MANDATORY): For ANY task that requires creating files, installing packages, or multiple actions — you MUST create a plan using the "plan" tool BEFORE starting work. Each step must describe a concrete deliverable (specific files to create, packages to install, commands to run). Do not combine unrelated work into one step.`);
@@ -26122,6 +26269,7 @@ class Repl {
 ${t("repl.interrupt")}
 `));
               this.agent.shutdown();
+              setTimeout(() => process.exit(1), 3000);
             }
             this.lastEscTime = 0;
             return;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "micro-models-agent",
-  "version": "0.24.3",
+  "version": "0.24.5",
   "description": "Micro Models Agent (MMA) — LLM agent harness for small models (Qwen3.5-9B, 32K-64K context)",
   "type": "module",
   "bin": {