micro-models-agent 0.29.1 → 0.29.2

package/dist/main.js

@@ -2304,6 +2304,7 @@ var init_en = __esm(() => {
     "error.response_blocked": "Response blocked: {reason}",
     "error.max_iters": "Max iterations ({max}) reached",
     "error.empty_response": "Model returned an empty response after retries",
+    "error.audit_failed": "Task could not be verified as complete: {summary}",
     "error.grep_failed": "Grep failed: {message}",
     "error.search_failed": "Search failed: {message}",
     "error.fetch_failed": "Fetch failed: {message}",
@@ -2448,6 +2449,8 @@ Command: {command}`,
     "plan.step_status": "Step {step}: {status}",
     "plan.no_active": "No active plan",
     "plan.step_not_found": "Step not found",
+    "plan.step_already_done": "Step {step} is already marked done. Stop re-marking it — either continue with the next step or investigate why progress appears stuck.",
+    "plan.order_blocked": 'Cannot mark step {step} done: step {first} ("{desc}") is not finished yet. Complete earlier steps first, or mark step {first} status=skipped if it is not needed.',
     "plan.show_header": "Plan status:",
     "plan.show_empty": "(plan has no steps)",
     "plan.list_header": "Plans:",
@@ -2870,6 +2873,7 @@ var init_ru = __esm(() => {
     "error.response_blocked": "Ответ заблокирован: {reason}",
     "error.max_iters": "Достигнут максимум итераций ({max})",
     "error.empty_response": "Модель вернула пустой ответ после повторных попыток",
+    "error.audit_failed": "Задача не может быть подтверждена как выполненная: {summary}",
     "error.grep_failed": "Ошибка grep: {message}",
     "error.search_failed": "Ошибка поиска: {message}",
     "error.fetch_failed": "Ошибка загрузки: {message}",
@@ -3013,6 +3017,8 @@ var init_ru = __esm(() => {
     "plan.step_status": "Шаг {step}: {status}",
     "plan.no_active": "Нет активного плана",
     "plan.step_not_found": "Шаг не найден",
+    "plan.step_already_done": "Шаг {step} уже отмечен выполненным. Прекратите повторную отметку — переходите к следующему шагу или выясните, почему прогресс заблокирован.",
+    "plan.order_blocked": "Нельзя отметить шаг {step} выполненным: шаг {first} («{desc}») ещё не завершён. Сначала завершите предыдущие шаги или отметьте шаг {first} status=skipped, если он не нужен.",
     "plan.show_header": "Статус плана:",
     "plan.show_empty": "(в плане нет шагов)",
     "plan.list_header": "Планы:",
@@ -7469,7 +7475,7 @@ function emptyCliRunHint(command, output, code) {
     return null;
   return "the command exited 0 but printed NOTHING to stdout. If this should run a CLI program, the file probably has no entry point: read it with read_file and check the code actually calls its main function with command-line arguments (e.g. main(process.argv[2])) and prints results with console.log.";
 }
-var BASH_GRACE_MS = 5000, SPAWN_SETTLE_MS = 100, bashGraceMs, UNIX_TO_WIN_HINTS, UNIX_TO_WIN_TRANSLATE, NEVER_TOOL_CALLS, CLI_FILE_RUN_RE, bashTool;
+var BASH_GRACE_MS = 5000, SPAWN_SETTLE_MS = 100, bashGraceMs, FAILING_FIRST_WORDS, HARD_BLOCK_THRESHOLD = 3, UNIX_TO_WIN_HINTS, UNIX_TO_WIN_TRANSLATE, NEVER_TOOL_CALLS, CLI_FILE_RUN_RE, bashTool;
 var init_bash = __esm(() => {
   init_command_validator();
   init_audit_log();
@@ -7478,6 +7484,7 @@ var init_bash = __esm(() => {
   init_processes();
   init_i18n();
   bashGraceMs = BASH_GRACE_MS;
+  FAILING_FIRST_WORDS = new Map;
   UNIX_TO_WIN_HINTS = {
     ls: "Use the list_dir tool instead.",
     pwd: "Use the file_info tool instead.",
@@ -7629,13 +7636,22 @@ ${output2}`;
           if (!output2 && code !== 0) {
             output2 = `(exit code ${code})`;
           }
-          if (platform2() === "win32" && code !== 0) {
+          if (platform2() === "win32") {
             const firstWord = command.trim().split(/\s+/)[0]?.split(/[\\/]/).pop();
-            const hint = firstWord ? UNIX_TO_WIN_HINTS[firstWord] : undefined;
-            if (hint) {
-              output2 = `${output2}
+            if (firstWord && firstWord in UNIX_TO_WIN_HINTS) {
+              if (code === 0) {
+                FAILING_FIRST_WORDS.delete(firstWord);
+              } else {
+                const failures = (FAILING_FIRST_WORDS.get(firstWord) || 0) + 1;
+                FAILING_FIRST_WORDS.set(firstWord, failures);
+                if (failures >= HARD_BLOCK_THRESHOLD) {
+                  output2 = `STOP using "${firstWord}" — it does not work in this cmd.exe shell and has failed ${failures} times in a row. ${UNIX_TO_WIN_HINTS[firstWord]}`;
+                } else {
+                  output2 = `${output2}
 
-Hint: "${firstWord}" may not work on Windows. ${hint}`;
+Hint: "${firstWord}" may not work on Windows. ${UNIX_TO_WIN_HINTS[firstWord]}`;
+                }
+              }
             }
           }
           if (securityConfig?.logCommands) {
@@ -10392,6 +10408,7 @@ ${t("tool.truncated", { tokens: Math.ceil(removedChars / 2) })}`;
     this.abortController = new AbortController;
     let iteration = 0;
     let lastText = "";
+    let lastForcedCompactionIteration = -FORCED_COMPACTION_COOLDOWN;
     let hallucinationRetries = 0;
     let lastToolSignature = "";
     let apiPromptTokens = 0;
@@ -10405,6 +10422,8 @@ ${t("tool.truncated", { tokens: Math.ceil(removedChars / 2) })}`;
     let emptyResponseRetries = 0;
     const MAX_EMPTY_RESPONSE_RETRIES = 2;
     let emptyResponseExhausted = false;
+    let auditFailed = false;
+    let lastAuditSummary = "";
     let repeatedToolCount = 0;
     const MAX_REPEATED_TOOL_CALLS = 2;
     const allToolsForBudget = toolExecutor.getToolDefinitions(this.deps.toolTags);
@@ -10431,7 +10450,8 @@ ${t("tool.truncated", { tokens: Math.ceil(removedChars / 2) })}`;
       const currentTokens = contextManager.getEstimatedTokens();
       const budget2 = contextManager.getBudget();
       const quality = contextManager.getQuality();
-      if (quality < QUALITY_TRIGGER_THRESHOLD && contextManager.getCompactionCount() > 0) {
+      if (quality < QUALITY_TRIGGER_THRESHOLD && contextManager.getCompactionCount() > 0 && iteration - lastForcedCompactionIteration >= FORCED_COMPACTION_COOLDOWN) {
+        lastForcedCompactionIteration = iteration;
         contextManager.compact();
         logger.warn(`Low context quality (${quality}%) — forced compaction`);
         slog.logCompaction(`quality-triggered compaction (${quality}% < ${QUALITY_TRIGGER_THRESHOLD}%), iteration ${iteration}`, iteration, currentTokens, budget2.history);
@@ -10794,8 +10814,10 @@ ${warnLine}
             });
             slog.logAudit(audit.summary, iteration);
             auditRetries++;
+            lastAuditSummary = audit.summary;
             if (auditRetries >= MAX_AUDIT_RETRIES || iteration >= config.maxToolIterations - 1) {
-              logger.warn(`Final audit still incomplete after ${auditRetries} retries — finishing anyway`);
+              logger.warn(`Final audit still incomplete after ${auditRetries} retries — reporting failure`);
+              auditFailed = true;
               break;
             }
             continue;
@@ -10823,9 +10845,9 @@ ${warnLine}
       };
     }
     return {
-      success: emptyResponseExhausted ? false : true,
+      success: emptyResponseExhausted || auditFailed ? false : true,
       text: lastText,
-      error: emptyResponseExhausted ? t("error.empty_response") : undefined,
+      error: emptyResponseExhausted ? t("error.empty_response") : auditFailed ? t("error.audit_failed", { summary: lastAuditSummary }) : undefined,
       iterationCount: iteration,
       contextUsed: tokensUsed,
       contextLimit: budget.history,
@@ -10890,7 +10912,7 @@ ${warnLine}
     });
   }
 }
-var TOOL_RESULT_MAX_TOKENS_RATIO = 0.3, TOOL_RESULT_ABSOLUTE_MAX_CHARS = 15000, QUALITY_TRIGGER_THRESHOLD = 40;
+var TOOL_RESULT_MAX_TOKENS_RATIO = 0.3, TOOL_RESULT_ABSOLUTE_MAX_CHARS = 15000, QUALITY_TRIGGER_THRESHOLD = 40, FORCED_COMPACTION_COOLDOWN = 3;
 var init_agent = __esm(() => {
   init_i18n();
   init_colors();
@@ -10993,7 +11015,8 @@ class ContextManager {
   getQuality() {
     const usedTokens = this.getEstimatedTokens();
     const tokenLoad = Math.max(0, 1 - usedTokens / this.budget.history);
-    const compactionLoss = Math.max(0, 1 - this.compactionCount * 0.15);
+    const COMPACTION_PENALTY_FLOOR = 0.5;
+    const compactionLoss = Math.max(COMPACTION_PENALTY_FLOOR, 1 - this.compactionCount * 0.15);
     const msgCount = this.messages.length || 1;
     const errorDensity = Math.max(0, 1 - Math.min(1, this.errorFacts.length / msgCount));
     const freshness = Math.max(0, 1 - this.iterationsSinceCompaction / COMPACTION_INTERVAL);
@@ -11139,6 +11162,9 @@ ${lines.join(`
     const newErrors = [];
     for (const msg of turns) {
       const content = getMessageText(msg.content);
+      if (msg.role === "user" && (content.startsWith("<system-summary>") || content.includes("[Compressed:"))) {
+        continue;
+      }
       if (msg.role === "tool") {
         for (const pattern of filePatterns) {
           for (const match of content.matchAll(pattern)) {
@@ -15592,7 +15618,34 @@ ${display}`,
             };
           }
           if (action === "update" && args.step && this.tracker) {
-            this.tracker.updateStepStatus(Number(args.step), args.status || "done");
+            const stepId = Number(args.step);
+            const target = this.tracker.getStep(stepId);
+            if (!target) {
+              return { success: false, output: t("plan.step_not_found") };
+            }
+            const status = String(args.status || "done");
+            if (status === "done" && target.status === "done") {
+              return {
+                success: false,
+                output: t("plan.step_already_done", {
+                  step: String(stepId)
+                })
+              };
+            }
+            if (status === "done") {
+              const blockers = this.tracker.getPlan().steps.filter((s) => s.id < stepId && s.status !== "done" && s.status !== "skipped");
+              if (blockers.length > 0) {
+                return {
+                  success: false,
+                  output: t("plan.order_blocked", {
+                    step: String(stepId),
+                    first: String(blockers[0].id),
+                    desc: blockers[0].description
+                  })
+                };
+              }
+            }
+            this.tracker.updateStepStatus(stepId, status);
             if (args.note)
               this.tracker.addNote(Number(args.step), String(args.note));
             this.tracker.syncCurrentStep();
@@ -17016,8 +17069,8 @@ class LspClient {
         }
       });
       return await diagPromise;
-    } catch {
-      return [];
+    } catch (e) {
+      throw e instanceof Error ? e : new Error(String(e));
     } finally {
       await this.shutdown();
     }
@@ -17249,7 +17302,15 @@ ${items}`;
 [LSP warnings]:
 ${items}`;
           }
-        } catch {}
+        } catch (e) {
+          const msg = e instanceof Error ? e.message : String(e);
+          try {
+            _ctx.logger?.warn(`LSP check failed for ${filePath}: ${msg}`);
+          } catch {}
+          result.output += `
+
+[LSP check failed]: ${msg}`;
+        }
       }
     };
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "micro-models-agent",
-  "version": "0.29.1",
+  "version": "0.29.2",
   "description": "Micro Models Agent (MMA) — LLM agent harness for small models (Qwen3.5-9B, 32K-64K context)",
   "type": "module",
   "bin": {