pentesting 0.51.1 → 0.52.1

package/dist/main.js

@@ -180,11 +180,15 @@ var MEMORY_LIMITS = {
   WORKING_MEMORY_MAX_ENTRIES: 20,
   /** Maximum events in episodic memory (session timeline) */
   EPISODIC_MEMORY_MAX_EVENTS: 100,
-  /** Consecutive failures before SWITCH VECTOR warning */
+  /** Consecutive failures before prompting to change approach or parameters */
   CONSECUTIVE_FAIL_THRESHOLD: 3,
   /** Maximum learned techniques in DynamicTechniqueLibrary */
   DYNAMIC_TECHNIQUES_MAX: 50,
-  /** Number of leading words to match for duplicate command detection */
+  /** Confidence decay per failure for DynamicTechniqueLibrary learned techniques */
+  TECHNIQUE_FAILURE_DECAY: 30,
+  /** Auto-prune threshold: techniques below this confidence are discarded */
+  TECHNIQUE_PRUNE_THRESHOLD: 10,
+  /** @deprecated Superseded by fingerprint-based matching in extractFingerprint(). Kept for reference. */
   COMMAND_MATCH_WORDS: 3,
   /** Maximum unverified techniques to show in prompt */
   PROMPT_UNVERIFIED_TECHNIQUES: 10,
@@ -345,7 +349,7 @@ var ORPHAN_PROCESS_NAMES = [
 
 // src/shared/constants/agent.ts
 var APP_NAME = "Pentest AI";
-var APP_VERSION = "0.51.1";
+var APP_VERSION = "0.52.1";
 var APP_DESCRIPTION = "Autonomous Penetration Testing AI Agent";
 var LLM_ROLES = {
   SYSTEM: "system",
@@ -3144,6 +3148,57 @@ var AttackGraph = class {
 // src/shared/utils/agent-memory.ts
 import { existsSync as existsSync4, readFileSync as readFileSync3, writeFileSync as writeFileSync4 } from "fs";
 import { join as join3 } from "path";
+function extractFingerprint(tool, command) {
+  const cmd = command || "";
+  let effectiveTool = tool.toLowerCase();
+  if (effectiveTool === "run_cmd" || effectiveTool === "run_background") {
+    const firstWord = cmd.trim().split(/\s+/)[0];
+    if (firstWord && !firstWord.startsWith("-")) {
+      effectiveTool = firstWord.toLowerCase();
+    }
+  }
+  const targetMatch = cmd.match(
+    /(?::\/\/|@)([\w.\-]+(?::\d+)?)|\b(\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3})(?::\d+)?\b/
+  );
+  const target = targetMatch?.[1] || targetMatch?.[2] || "";
+  let wordlistMatch = cmd.match(/(?:-[wPUL]\s+)([^\s]+)/);
+  if (!wordlistMatch) {
+    wordlistMatch = cmd.match(/(?:--(?:wordlist|passwords?|usernames?)(?:=|\s+))([^\s]+)/i);
+  }
+  const wordlist = wordlistMatch?.[1] || "";
+  const portMatch = cmd.match(/(?:-p\s+|--port(?:=|\s+))(\S+)/);
+  const port = portMatch?.[1] || "";
+  const flagPatterns = [
+    /--level(?:=|\s+)(\S+)/i,
+    // sqlmap
+    /--risk(?:=|\s+)(\S+)/i,
+    // sqlmap
+    /-s([VSTCAUOPMX]+)/,
+    // nmap scan type
+    /--script(?:=|\s+)(\S+)/i,
+    // nmap scripts
+    /-[el]\s+(\S+)/i,
+    // hydra login/password single
+    /--method(?:=|\s+)(\S+)/i
+    // HTTP method
+  ];
+  const flags = flagPatterns.map((p) => {
+    const m = cmd.match(p);
+    return m?.[0]?.trim() || "";
+  }).filter(Boolean).join(" ");
+  return { tool: effectiveTool, target, wordlist, flags, port };
+}
+function fingerprintsMatch(a, b) {
+  return a.tool === b.tool && a.target === b.target && a.wordlist === b.wordlist && a.flags === b.flags && a.port === b.port;
+}
+function formatFingerprint(fp) {
+  const parts = [fp.tool];
+  if (fp.target) parts.push(`\u2192${fp.target}`);
+  if (fp.wordlist) parts.push(`[wordlist:${fp.wordlist.split("/").pop()}]`);
+  if (fp.port) parts.push(`[port:${fp.port}]`);
+  if (fp.flags) parts.push(`[${fp.flags}]`);
+  return parts.join(" ");
+}
 var WorkingMemory = class {
   entries = [];
   maxEntries = MEMORY_LIMITS.WORKING_MEMORY_MAX_ENTRIES;
@@ -3156,20 +3211,29 @@ var WorkingMemory = class {
       context,
       importance
     });
-    if (this.entries.length > this.maxEntries) {
-      this.entries.sort((a, b) => {
-        const aScore = a.importance * 0.7 + (1 - (Date.now() - a.timestamp) / 6e5) * 0.3;
-        const bScore = b.importance * 0.7 + (1 - (Date.now() - b.timestamp) / 6e5) * 0.3;
-        return bScore - aScore;
-      });
-      this.entries = this.entries.slice(0, this.maxEntries);
-    }
+    this.pruneIfNeeded();
   }
   /**
-   * Record a failed attempt to avoid repeating it.
+   * Record a failed attempt with structured fingerprint.
+   *
+   * WHY: Fingerprint-based recording ensures that "hydra with rockyou.txt"
+   * and "hydra with darkweb2017.txt" are stored as distinct attempts.
+   * The LLM can see what parameter combinations have been tried and pick new ones.
    */
   recordFailure(tool, command, error) {
-    this.add("failure", `FAILED: ${tool} \u2192 ${command.slice(0, DISPLAY_LIMITS.COMMAND_PREVIEW)} \u2192 ${error.slice(0, DISPLAY_LIMITS.ERROR_PREVIEW)}`, 0.8, { tool, command });
+    const fp = extractFingerprint(tool, command);
+    const fpLabel = formatFingerprint(fp);
+    const entry = {
+      id: `wm_${Date.now()}_${Math.random().toString(36).slice(2, 6)}`,
+      timestamp: Date.now(),
+      category: "failure",
+      content: `FAILED: ${fpLabel} \u2192 ${error.slice(0, DISPLAY_LIMITS.ERROR_PREVIEW)}`,
+      context: { tool, command },
+      importance: 0.8,
+      fingerprint: fp
+    };
+    this.entries.push(entry);
+    this.pruneIfNeeded();
   }
   /**
    * Record a successful action for reference.
@@ -3178,16 +3242,53 @@ var WorkingMemory = class {
     this.add("success", `SUCCESS: ${tool} \u2192 ${command.slice(0, DISPLAY_LIMITS.COMMAND_PREVIEW)}`, 0.6, { tool, result: result2.slice(0, DISPLAY_LIMITS.OUTPUT_SUMMARY) });
   }
   /**
-   * Check if a similar command has already failed.
+   * Check if the EXACT same attack attempt (all parameters identical) has failed.
+   *
+   * WHY: Only blocks truly duplicate attempts. Different wordlists, flags, or
+   * ports are treated as new attacks and allowed through.
    */
-  hasFailedBefore(command) {
-    const lower = command.toLowerCase().split(/\s+/).slice(0, MEMORY_LIMITS.COMMAND_MATCH_WORDS).join(" ");
+  hasFailedBefore(command, tool) {
+    const effectiveTool = tool || command.split(/\s+/)[0] || "";
+    const fp = extractFingerprint(effectiveTool, command);
     return this.entries.find(
-      (e) => e.category === "failure" && e.content.toLowerCase().includes(lower)
+      (e) => e.category === "failure" && e.fingerprint != null && fingerprintsMatch(e.fingerprint, fp)
     );
   }
   /**
-   * Get count of consecutive failures (for vector switch detection).
+   * Get all previous attempts (success & failure) for a specific tool+target vector.
+   * Returns the full history so the LLM can see what parameter combinations were tried.
+   *
+   * TODO: Wire up in tools.ts to pass vector history to strategist for smarter retries.
+   */
+  getAttemptsForVector(tool, target) {
+    const lowerTool = tool.toLowerCase();
+    return this.entries.filter((e) => {
+      if (e.fingerprint) {
+        const matchTool = e.fingerprint.tool === lowerTool;
+        const matchTarget = !target || e.fingerprint.target.includes(target);
+        return matchTool && matchTarget;
+      }
+      return String(e.context.tool || "").toLowerCase() === lowerTool;
+    });
+  }
+  /** Internal prune helper (used by both add() and recordFailure()) */
+  pruneIfNeeded() {
+    if (this.entries.length > this.maxEntries) {
+      this.entries.sort((a, b) => {
+        const aScore = a.importance * 0.7 + (1 - (Date.now() - a.timestamp) / 6e5) * 0.3;
+        const bScore = b.importance * 0.7 + (1 - (Date.now() - b.timestamp) / 6e5) * 0.3;
+        return bScore - aScore;
+      });
+      this.entries = this.entries.slice(0, this.maxEntries);
+    }
+  }
+  /**
+   * Get count of consecutive failures on the SAME attack vector.
+   *
+   * WHY: Switching from "hydra with wordlist A" to "hydra with wordlist B"
+   * is a legitimate new approach. The consecutive-failure counter should
+   * only trigger when every recent attempt is against the same tool+target
+   * with no variation producing success. This prevents premature vector abandonment.
    */
   getConsecutiveFailures() {
     let count = 0;
@@ -3197,6 +3298,26 @@ var WorkingMemory = class {
     }
     return count;
   }
+  /**
+   * Get consecutive failures for a specific vector (tool+target).
+   * Returns count of sequential failures where all attempts used the same tool+target.
+   *
+   * TODO: Wire up in tools.ts for per-vector threshold checks alongside global consecutive count.
+   */
+  getConsecutiveVectorFailures(tool, target) {
+    const lowerTool = tool.toLowerCase();
+    let count = 0;
+    for (let i = this.entries.length - 1; i >= 0; i--) {
+      const e = this.entries[i];
+      if (e.category !== "failure") break;
+      if (e.fingerprint && e.fingerprint.tool === lowerTool && (!target || e.fingerprint.target.includes(target))) {
+        count++;
+      } else {
+        break;
+      }
+    }
+    return count;
+  }
   /**
    * Format for prompt injection.
    */
@@ -3207,14 +3328,20 @@ var WorkingMemory = class {
     const insights = this.entries.filter((e) => e.category === "insight" || e.category === "discovery");
     const lines = ["<working-memory>"];
     if (failures.length > 0) {
-      lines.push(`\u26A0\uFE0F FAILED ATTEMPTS (${failures.length} \u2014 DO NOT REPEAT):`);
+      lines.push(`\u26A0\uFE0F FAILED ATTEMPTS (${failures.length} \u2014 DO NOT REPEAT EXACT SAME PARAMS):`);
       for (const f of failures.slice(-DISPLAY_LIMITS.RECENT_FAILURES)) {
-        lines.push(`  \u2717 ${f.content}`);
+        const fp = f.fingerprint;
+        if (fp) {
+          lines.push(`  \u2717 ${formatFingerprint(fp)} \u2192 ${f.content.split("\u2192").pop()?.trim() || ""}`);
+        } else {
+          lines.push(`  \u2717 ${f.content}`);
+        }
       }
+      lines.push(...this.buildAttackCoverageLines(failures));
     }
     const consecutiveFails = this.getConsecutiveFailures();
     if (consecutiveFails >= MEMORY_LIMITS.CONSECUTIVE_FAIL_THRESHOLD) {
-      lines.push(`\u{1F534} ${consecutiveFails} CONSECUTIVE FAILURES \u2014 SWITCH ATTACK VECTOR NOW`);
+      lines.push(`\u{1F534} ${consecutiveFails} CONSECUTIVE FAILURES \u2014 consider changing approach or parameters`);
     }
     if (successes.length > 0) {
       lines.push(`\u2705 RECENT SUCCESSES (${successes.length}):`);
@@ -3231,6 +3358,29 @@ var WorkingMemory = class {
     lines.push("</working-memory>");
     return lines.join("\n");
   }
+  /**
+   * Build ATTACK COVERAGE lines grouped by tool+target.
+   * WHY: Shows the LLM which parameter combinations have been tried per vector,
+   * so it can pick genuinely untried variations instead of repeating.
+   */
+  buildAttackCoverageLines(failures) {
+    const vectorMap = /* @__PURE__ */ new Map();
+    for (const f of failures) {
+      if (f.fingerprint) {
+        const key = `${f.fingerprint.tool}\u2192${f.fingerprint.target}`;
+        const detail = f.fingerprint.wordlist ? f.fingerprint.wordlist.split("/").pop() || "" : f.fingerprint.flags || "default";
+        const existing = vectorMap.get(key) || [];
+        if (!existing.includes(detail)) existing.push(detail);
+        vectorMap.set(key, existing);
+      }
+    }
+    if (vectorMap.size === 0) return [];
+    const lines = [`  \u{1F4CA} ATTACK COVERAGE (tried variations):`];
+    for (const [vector, variations] of vectorMap) {
+      lines.push(`    ${vector}: tried [${variations.join(", ")}] \u2014 try DIFFERENT params`);
+    }
+    return lines;
+  }
   getEntries() {
     return [...this.entries];
   }
@@ -3467,6 +3617,25 @@ var DynamicTechniqueLibrary = class {
       }
     }
   }
+  /**
+   * Record that a technique failed in practice — decrease confidence.
+   *
+   * WHY: Web search results can contain outdated, wrong, or irrelevant techniques.
+   * If the agent tries a learned technique and it fails, we must degrade its
+   * confidence so the agent stops wasting time on bad advice.
+   * After enough failures the technique is auto-pruned entirely.
+   */
+  recordFailure(techniqueSubstring) {
+    if (!techniqueSubstring || techniqueSubstring.length < 3) return;
+    const lower = techniqueSubstring.toLowerCase();
+    this.techniques = this.techniques.filter((t) => {
+      if (t.technique.toLowerCase().includes(lower)) {
+        t.confidence -= MEMORY_LIMITS.TECHNIQUE_FAILURE_DECAY;
+        return t.confidence > MEMORY_LIMITS.TECHNIQUE_PRUNE_THRESHOLD;
+      }
+      return true;
+    });
+  }
   /**
    * Get techniques relevant to a specific service/technology.
    */
@@ -8924,6 +9093,7 @@ var ToolRegistry = class {
       this.state.workingMemory.recordSuccess(toolCall.name, command, result2.output || "");
     } else {
       this.state.workingMemory.recordFailure(toolCall.name, command, result2.error || "Unknown error");
+      this.state.dynamicTechniques.recordFailure(command);
     }
     if (result2.success && result2.output) {
       const structured = autoExtractStructured(toolCall.name, result2.output);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pentesting",
-  "version": "0.51.1",
+  "version": "0.52.1",
   "description": "Autonomous Penetration Testing AI Agent",
   "type": "module",
   "main": "dist/main.js",
@@ -19,7 +19,7 @@
     "dev:tsx": "tsx src/platform/tui/main.tsx",
     "build": "tsup",
     "start": "node dist/main.js",
-    "test": "mkdir -p /tmp/pentesting-vitest && TMPDIR=/tmp/pentesting-vitest vitest run",
+    "test": "mkdir -p .pentesting && TMPDIR=.pentesting vitest run",
     "test:watch": "vitest",
     "lint": "tsc --noEmit",
     "prepublishOnly": "npm run build",
@@ -29,7 +29,7 @@
     "release:minor": "npm version minor && npm run build && npm run publish:token",
     "release:major": "npm version major && npm run build && npm run publish:token",
     "release:docker": "docker buildx build --platform linux/amd64,linux/arm64 -t agnusdei1207/pentesting:latest --push . && docker system prune -af",
-    "check": "TMPDIR=/tmp npm run test && npm run build && npm run release:docker && bash test.sh"
+    "check": "npm run test && npm run build && npm run release:docker && bash test.sh"
   },
   "repository": {
     "type": "git",