@pentoshi/clai 3.6.0 → 3.7.3

package/dist/agent/task-analyzer.js

@@ -1,33 +1,273 @@
+import { createPlanStep } from "./task-plan.js";
+const PENTEST_RE = /\b(?:pentest|pen[\s-]?test|penetration|security\s*(?:test|audit|scan|assess)|vapt|csrf|xss|sqli|sql[\s-]?inject|rce|lfi|rfi|ssrf|idor|xxe|deserialization|brute[\s-]?force|enumerat|exploit|vulnerabilit|recon|bug[\s-]?bounty|ctf|capture[\s-]?the[\s-]?flag|red[\s-]?team|offensive|nmap|nikto|nuclei|ffuf|gobuster|sqlmap|hydra|metasploit)\b/i;
+const BUILD_RE = /\b(?:build|scaffold|create|implement|refactor|migrate|bootstrap|set\s*up|setup)\b/i;
+const APP_RE = /\b(?:app|application|project|website|api|service|cli|dashboard|todo|blog|frontend|backend)\b/i;
+const FIX_RE = /\b(?:fix|debug|broken|error|failing|crash|bug|regression|not\s+work|doesn'?t\s+work|typeerror|exception)\b/i;
+const RESEARCH_RE = /\b(?:what\s+is|how\s+(?:do|does|to)|explain|compare|difference|latest|current|who\s+is|when\s+did|why\s+(?:is|does))\b/i;
+const NETWORK_RE = /\b(?:scan|nmap|ping|subnet|cidr|local\s*network|hosts?|ports?|dns|whois|http|https|url|domain|ip\s*address)\b/i;
+const SHELL_RE = /\b(?:run|execute|install|upgrade|brew|apt|npm\s+i|docker|kubectl|chmod|chown|kill|restart|service)\b/i;
+const FS_RE = /\b(?:read|write|edit|file|directory|folder|path|rename|delete|list\s+files?|open\s+)\b/i;
 /**
- * Pentest / security keywords — these tasks are inherently multi-step and
- * always deserve the full step budget regardless of how terse the prompt is.
+ * A bounded, explicit nmap operation is not a request for a full pentest.
+ * Keep it to one net.scan call unless the user also asks for assessment,
+ * reconnaissance, enumeration, exploitation, or multiple named operations.
  */
-const PENTEST_RE = /\b(?:pentest|pen[\s-]?test|penetration|security\s*(?:test|audit|scan|assess)|csrf|xss|sqli|sql[\s-]?inject|rce|lfi|rfi|ssrf|idor|xxe|deserialization|brute[\s-]?force|enumerat|exploit|vulnerabilit|recon|bug[\s-]?bounty|ctf|capture[\s-]?the[\s-]?flag|red[\s-]?team|offensive|nmap|nikto|nuclei|ffuf|gobuster|sqlmap|hydra|metasploit)\b/i;
+export function isNarrowExplicitNmapOperation(prompt) {
+    const text = prompt.replace(/\s+/g, " ").trim();
+    if (!/\bnmap\b/i.test(text))
+        return false;
+    const explicitRun = /^(?:sudo\s+)?nmap(?:\s|$)/i.test(text) ||
+        /\b(?:run|execute|start|perform)\b[\s\S]{0,40}\bnmap\b/i.test(text) ||
+        /\bnmap\b[\s\S]{0,20}\bscan\b/i.test(text) ||
+        /\bscan\b[\s\S]{0,40}\bwith\s+nmap\b/i.test(text);
+    if (!explicitRun)
+        return false;
+    if (/\b(?:pentest|pen[\s-]?test|penetration|security\s+(?:audit|assessment|test)|vapt|recon(?:naissance)?|enumerat\w*|attack\s+surface|vulnerabilit\w*|exploit\w*|bug\s*bounty|red\s*team|threat\s+model|web\s+assessment)\b/i.test(text)) {
+        return false;
+    }
+    if (/\b(?:and|then|also)\b[\s\S]{0,40}\b(?:whois|dns|dig|http|crawl|fuzz|nikto|nuclei|sqlmap|gobuster|ffuf)\b/i.test(text)) {
+        return false;
+    }
+    return true;
+}
+function wordCount(prompt) {
+    return prompt.trim().split(/\s+/).filter(Boolean).length;
+}
+function classifyCategory(prompt) {
+    if (PENTEST_RE.test(prompt)) {
+        if (/\b(?:dns|whois|subdomain)\b/i.test(prompt))
+            return "dns";
+        if (/\b(?:whois|registrar|owner)\b/i.test(prompt))
+            return "whois";
+        if (/\b(?:dir|fuzz|gobuster|ffuf|content\s*discover)\b/i.test(prompt)) {
+            return "web-enum";
+        }
+        if (/\b(?:recon|enumerat|osint|fingerprint)\b/i.test(prompt)) {
+            return "pentest-recon";
+        }
+        return "pentest-recon";
+    }
+    if (/\b(?:whois)\b/i.test(prompt))
+        return "whois";
+    if (/\b(?:dns|dig|nslookup|resolve)\b/i.test(prompt))
+        return "dns";
+    if (NETWORK_RE.test(prompt) && /\b(?:discover|sweep|live\s*hosts?)\b/i.test(prompt)) {
+        return "network-discovery";
+    }
+    if (/\b(?:pkg|package|brew\s+install|apt\s+install|install\s+\w+)\b/i.test(prompt)) {
+        return "package";
+    }
+    if (BUILD_RE.test(prompt) || FIX_RE.test(prompt) || FS_RE.test(prompt)) {
+        return "filesystem";
+    }
+    if (SHELL_RE.test(prompt))
+        return "shell";
+    if (RESEARCH_RE.test(prompt) || prompt.trim().endsWith("?"))
+        return "answer";
+    return "other";
+}
+function estimateComplexity(prompt, category, words) {
+    if (PENTEST_RE.test(prompt))
+        return "complex";
+    if (BUILD_RE.test(prompt) && APP_RE.test(prompt))
+        return "complex";
+    if (/\b(?:multi-?file|across\s+files|refactor|migrate|end-?to-?end|full\s+stack)\b/i.test(prompt)) {
+        return "complex";
+    }
+    if (FIX_RE.test(prompt)) {
+        return words > 40 ? "complex" : "standard";
+    }
+    if (category === "answer" && words <= 20)
+        return "simple";
+    if (words <= 6 && !BUILD_RE.test(prompt) && !PENTEST_RE.test(prompt)) {
+        return "simple";
+    }
+    if (words <= 28 && !BUILD_RE.test(prompt))
+        return "standard";
+    return words > 40 || BUILD_RE.test(prompt) ? "complex" : "standard";
+}
+function shouldPlanFor(prompt, complexity, category) {
+    if (complexity === "simple")
+        return false;
+    if (category === "answer")
+        return false;
+    if (PENTEST_RE.test(prompt))
+        return true;
+    if (BUILD_RE.test(prompt) && APP_RE.test(prompt))
+        return true;
+    if (/\b(?:implement|refactor|migrate|multi-?step|several\s+files|end-?to-?end)\b/i.test(prompt)) {
+        return true;
+    }
+    return complexity === "complex" && wordCount(prompt) > 25;
+}
+function inferLikelyTools(prompt, category) {
+    const tools = new Set();
+    if (category === "answer" || RESEARCH_RE.test(prompt)) {
+        tools.add("web.search");
+        tools.add("web.fetch");
+    }
+    if (category === "filesystem" || BUILD_RE.test(prompt) || FIX_RE.test(prompt)) {
+        tools.add("fs.list");
+        tools.add("fs.read");
+        tools.add("fs.write");
+        tools.add("fs.edit");
+        tools.add("shell.exec");
+    }
+    if (BUILD_RE.test(prompt) && APP_RE.test(prompt)) {
+        tools.add("plan.create");
+        tools.add("task.update");
+        tools.add("shell.start");
+    }
+    if (PENTEST_RE.test(prompt) || category.startsWith("pentest") || category === "web-enum") {
+        tools.add("dns.lookup");
+        tools.add("whois.lookup");
+        tools.add("http.fetch");
+        tools.add("net.scan");
+        tools.add("tool.check");
+        tools.add("plan.create");
+    }
+    if (category === "network-discovery") {
+        tools.add("net.context");
+        tools.add("net.pingSweep");
+        tools.add("net.scan");
+    }
+    if (category === "dns")
+        tools.add("dns.lookup");
+    if (category === "whois")
+        tools.add("whois.lookup");
+    if (category === "package")
+        tools.add("pkg.install");
+    if (SHELL_RE.test(prompt))
+        tools.add("shell.exec");
+    if (/\b(?:pdf)\b/i.test(prompt))
+        tools.add("pdf.read");
+    if (/\b(?:ocr|image|screenshot)\b/i.test(prompt))
+        tools.add("image.ocr");
+    return [...tools];
+}
+function softSteps(prompt, category) {
+    if (PENTEST_RE.test(prompt)) {
+        return [
+            createPlanStep("Recon and fingerprint stack", "pentest-recon", {
+                toolHint: "dns.lookup / http.fetch / net.scan",
+                successCriteria: "real findings from tool output",
+            }),
+            createPlanStep("Plan from findings", "other", {
+                toolHint: "plan.create",
+                successCriteria: "durable plan with verifiable tasks",
+            }),
+            createPlanStep("Test high-value vectors", "web-enum", {
+                successCriteria: "validated evidence or clean negatives",
+            }),
+            createPlanStep("Report with evidence", "other", {
+                successCriteria: "findings with repro + severity",
+            }),
+        ];
+    }
+    if (BUILD_RE.test(prompt) && APP_RE.test(prompt)) {
+        return [
+            createPlanStep("Explore destination and stack", "filesystem", {
+                toolHint: "fs.list / fs.read",
+            }),
+            createPlanStep("Plan multi-step build", "other", {
+                toolHint: "plan.create",
+            }),
+            createPlanStep("Implement requested feature", "filesystem", {
+                toolHint: "fs.write / fs.edit",
+            }),
+            createPlanStep("Verify (build and/or live server)", "shell", {
+                toolHint: "shell.start / shell.exec",
+            }),
+        ];
+    }
+    if (FIX_RE.test(prompt)) {
+        return [
+            createPlanStep("Reproduce and capture error", "shell"),
+            createPlanStep("Localize root cause", "filesystem", {
+                toolHint: "fs.read / fs.search",
+            }),
+            createPlanStep("Apply minimal fix", "filesystem", {
+                toolHint: "fs.edit",
+            }),
+            createPlanStep("Re-run failing check", "shell"),
+        ];
+    }
+    if (category === "network-discovery") {
+        return [
+            createPlanStep("Local network context", "network-discovery", {
+                toolHint: "net.context",
+            }),
+            createPlanStep("Discover live hosts", "network-discovery", {
+                toolHint: "net.pingSweep",
+            }),
+            createPlanStep("Summarize findings", "other"),
+        ];
+    }
+    return [];
+}
+function stopWhenFor(prompt, category) {
+    if (PENTEST_RE.test(prompt)) {
+        return "Findings report with evidence; no local dev server; scope respected.";
+    }
+    if (BUILD_RE.test(prompt) && APP_RE.test(prompt)) {
+        return "Requested feature works; verified via build and/or live localhost probe; server left running when applicable.";
+    }
+    if (FIX_RE.test(prompt)) {
+        return "Original failing check now passes; root cause addressed.";
+    }
+    if (category === "answer") {
+        return "Accurate answer grounded in tools when facts are volatile.";
+    }
+    return "User request completed with tool-backed evidence.";
+}
 /**
- * Complexity estimator. Uses keyword detection for pentest/security tasks
- * (always complex) and falls back to word-count heuristic otherwise.
- * No pre-baked steps — the AI decides everything dynamically.
+ * Lightweight task analysis for step budgets and soft guidance.
+ * Does not execute or force a plan — the agent still owns decisions.
  */
 export function analyzeTask(prompt) {
-    const words = prompt.trim().split(/\s+/).length;
-    // Pentest tasks are inherently multi-step — always complex.
-    const complexity = PENTEST_RE.test(prompt)
-        ? "complex"
-        : words <= 5
-            ? "simple"
-            : words <= 30
-                ? "standard"
-                : "complex";
+    const text = prompt.replace(/\s+/g, " ").trim();
+    if (isNarrowExplicitNmapOperation(text)) {
+        return {
+            complexity: "standard",
+            shouldPlan: false,
+            category: "pentest-recon",
+            goal: text.slice(0, 100),
+            needsNetworkContext: false,
+            needsToolPreflight: true,
+            likelyTools: ["net.scan"],
+            stopWhen: "The requested nmap operation has started or completed and its canonical job ID, status, and output are reported; do not broaden scope.",
+            suggestedSteps: [],
+        };
+    }
+    const words = wordCount(text);
+    const category = classifyCategory(text);
+    const complexity = estimateComplexity(text, category, words);
+    const shouldPlan = shouldPlanFor(text, complexity, category);
+    const likelyTools = inferLikelyTools(text, category);
+    const suggestedSteps = complexity === "simple" || category === "answer" ? [] : softSteps(text, category);
     return {
         complexity,
-        shouldPlan: false,
-        category: "other",
-        goal: prompt.slice(0, 100),
-        needsNetworkContext: false,
-        needsToolPreflight: false,
-        likelyTools: [],
-        stopWhen: "",
-        suggestedSteps: [],
+        shouldPlan,
+        category,
+        goal: text.slice(0, 100),
+        needsNetworkContext: category === "network-discovery" ||
+            /\b(?:local\s*network|subnet|cidr|lan|my\s+network)\b/i.test(text),
+        needsToolPreflight: PENTEST_RE.test(text) ||
+            /\b(?:nmap|ffuf|gobuster|sqlmap|nuclei|hydra|nikto)\b/i.test(text),
+        likelyTools,
+        stopWhen: stopWhenFor(text, category),
+        suggestedSteps,
     };
 }
+/** Compact one-liner for session/system context (optional inject). */
+export function formatTaskAnalysisHint(analysis) {
+    const tools = analysis.likelyTools.length > 0
+        ? analysis.likelyTools.slice(0, 8).join(", ")
+        : "—";
+    const tracking = analysis.shouldPlan
+        ? "durable task tracking strongly recommended: create one concise plan before mutation, then keep task.update states aligned with real work"
+        : "direct execution ok";
+    return (`TASK ANALYSIS: category=${analysis.category} complexity=${analysis.complexity} (${tracking}). ` +
+        `Likely tools: ${tools}. Done when: ${analysis.stopWhen}`);
+}
 //# sourceMappingURL=task-analyzer.js.map

package/dist/agent/task-analyzer.js.map

@@ -1 +1 @@
-{"version":3,"file":"task-analyzer.js","sourceRoot":"","sources":["../../src/agent/task-analyzer.ts"],"names":[],"mappings":"AAkBA;;;GAGG;AACH,MAAM,UAAU,GACd,8UAA8U,CAAC;AAEjV;;;;GAIG;AACH,MAAM,UAAU,WAAW,CAAC,MAAc;IACxC,MAAM,KAAK,GAAG,MAAM,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC;IAChD,4DAA4D;IAC5D,MAAM,UAAU,GAAmB,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC;QACxD,CAAC,CAAC,SAAS;QACX,CAAC,CAAC,KAAK,IAAI,CAAC;YACV,CAAC,CAAC,QAAQ;YACV,CAAC,CAAC,KAAK,IAAI,EAAE;gBACX,CAAC,CAAC,UAAU;gBACZ,CAAC,CAAC,SAAS,CAAC;IAClB,OAAO;QACL,UAAU;QACV,UAAU,EAAE,KAAK;QACjB,QAAQ,EAAE,OAAO;QACjB,IAAI,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC;QAC1B,mBAAmB,EAAE,KAAK;QAC1B,kBAAkB,EAAE,KAAK;QACzB,WAAW,EAAE,EAAE;QACf,QAAQ,EAAE,EAAE;QACZ,cAAc,EAAE,EAAE;KACnB,CAAC;AACJ,CAAC"}
+{"version":3,"file":"task-analyzer.js","sourceRoot":"","sources":["../../src/agent/task-analyzer.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,cAAc,EAAE,MAAM,gBAAgB,CAAC;AAgBhD,MAAM,UAAU,GACd,mVAAmV,CAAC;AAEtV,MAAM,QAAQ,GACZ,oFAAoF,CAAC;AAEvF,MAAM,MAAM,GACV,+FAA+F,CAAC;AAElG,MAAM,MAAM,GACV,6GAA6G,CAAC;AAEhH,MAAM,WAAW,GACf,yHAAyH,CAAC;AAE5H,MAAM,UAAU,GACd,gHAAgH,CAAC;AAEnH,MAAM,QAAQ,GACZ,uGAAuG,CAAC;AAE1G,MAAM,KAAK,GACT,yFAAyF,CAAC;AAE5F;;;;GAIG;AACH,MAAM,UAAU,6BAA6B,CAAC,MAAc;IAC1D,MAAM,IAAI,GAAG,MAAM,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;IAChD,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,KAAK,CAAC;IAC1C,MAAM,WAAW,GACf,4BAA4B,CAAC,IAAI,CAAC,IAAI,CAAC;QACvC,wDAAwD,CAAC,IAAI,CAAC,IAAI,CAAC;QACnE,+BAA+B,CAAC,IAAI,CAAC,IAAI,CAAC;QAC1C,sCAAsC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACpD,IAAI,CAAC,WAAW;QAAE,OAAO,KAAK,CAAC;IAC/B,IACE,0NAA0N,CAAC,IAAI,CAAC,IAAI,CAAC,EACrO,CAAC;QACD,OAAO,KAAK,CAAC;IACf,CAAC;IACD,IACE,2GAA2G,CAAC,IAAI,CAAC,IAAI,CAAC,EACtH,CAAC;QACD,OAAO,KAAK,CAAC;IACf,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,SAAS,SAAS,CAAC,MAAc;IAC/B,OAAO,MAAM,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC;AAC3D,CAAC;AAED,SAAS,gBAAgB,CAAC,MAAc;IACtC,IAAI,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QAC5B,IAAI,8BAA8B,CAAC,IAAI,CAAC,MAAM,CAAC;YAAE,OAAO,KAAK,CAAC;QAC9D,IAAI,gCAAgC,CAAC,IAAI,CAAC,MAAM,CAAC;YAAE,OAAO,OAAO,CAAC;QAClE,IAAI,oDAAoD,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;YACtE,OAAO,UAAU,CAAC;QACpB,CAAC;QACD,IAAI,2CAA2C,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;YAC7D,OAAO,eAAe,CAAC;QACzB,CAAC;QACD,OAAO,eAAe,CAAC;IACzB,CAAC;IACD,IAAI,gBAAgB,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,OAAO,OAAO,CAAC;IAClD,IAAI,mCAAmC,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,OAAO,KAAK,CAAC;IACnE,IAAI,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,uCAAuC,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACpF,OAAO,mBAAmB,CAAC;IAC7B,CAAC;IACD,IAAI,iEAAiE,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACnF,OAAO,SAAS,CAAC;IACnB,CAAC;IACD,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACvE,OAAO,YAAY,CAAC;IACtB,CAAC;IACD,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,OAAO,OAAO,CAAC;IAC1C,IAAI,WAAW,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,GAAG,CAAC;QAAE,OAAO,QAAQ,CAAC;IAC7E,OAAO,OAAO,CAAC;AACjB,CAAC;AAED,SAAS,kBAAkB,CACzB,MAAc,EACd,QAAkB,EAClB,KAAa;IAEb,IAAI,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,OAAO,SAAS,CAAC;IAC9C,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,OAAO,SAAS,CAAC;IACnE,IACE,gFAAgF,CAAC,IAAI,CACnF,MAAM,CACP,EACD,CAAC;QACD,OAAO,SAAS,CAAC;IACnB,CAAC;IACD,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACxB,OAAO,KAAK,GAAG,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,UAAU,CAAC;IAC7C,CAAC;IACD,IAAI,QAAQ,KAAK,QAAQ,IAAI,KAAK,IAAI,EAAE;QAAE,OAAO,QAAQ,CAAC;IAC1D,IAAI,KAAK,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACrE,OAAO,QAAQ,CAAC;IAClB,CAAC;IACD,IAAI,KAAK,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,OAAO,UAAU,CAAC;IAC7D,OAAO,KAAK,GAAG,EAAE,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,UAAU,CAAC;AACtE,CAAC;AAED,SAAS,aAAa,CACpB,MAAc,EACd,UAA0B,EAC1B,QAAkB;IAElB,IAAI,UAAU,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IAC1C,IAAI,QAAQ,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IACxC,IAAI,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,OAAO,IAAI,CAAC;IACzC,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,OAAO,IAAI,CAAC;IAC9D,IACE,8EAA8E,CAAC,IAAI,CACjF,MAAM,CACP,EACD,CAAC;QACD,OAAO,IAAI,CAAC;IACd,CAAC;IACD,OAAO,UAAU,KAAK,SAAS,IAAI,SAAS,CAAC,MAAM,CAAC,GAAG,EAAE,CAAC;AAC5D,CAAC;AAED,SAAS,gBAAgB,CAAC,MAAc,EAAE,QAAkB;IAC1D,MAAM,KAAK,GAAG,IAAI,GAAG,EAAU,CAAC;IAChC,IAAI,QAAQ,KAAK,QAAQ,IAAI,WAAW,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACtD,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;QACxB,KAAK,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC;IACzB,CAAC;IACD,IAAI,QAAQ,KAAK,YAAY,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QAC9E,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QACrB,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QACrB,KAAK,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC;QACtB,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QACrB,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;IAC1B,CAAC;IACD,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACjD,KAAK,CAAC,GAAG,CAAC,aAAa,CAAC,CAAC;QACzB,KAAK,CAAC,GAAG,CAAC,aAAa,CAAC,CAAC;QACzB,KAAK,CAAC,GAAG,CAAC,aAAa,CAAC,CAAC;IAC3B,CAAC;IACD,IAAI,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,QAAQ,CAAC,UAAU,CAAC,SAAS,CAAC,IAAI,QAAQ,KAAK,UAAU,EAAE,CAAC;QACzF,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;QACxB,KAAK,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC;QAC1B,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;QACxB,KAAK,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC;QACtB,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;QACxB,KAAK,CAAC,GAAG,CAAC,aAAa,CAAC,CAAC;IAC3B,CAAC;IACD,IAAI,QAAQ,KAAK,mBAAmB,EAAE,CAAC;QACrC,KAAK,CAAC,GAAG,CAAC,aAAa,CAAC,CAAC;QACzB,KAAK,CAAC,GAAG,CAAC,eAAe,CAAC,CAAC;QAC3B,KAAK,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC;IACxB,CAAC;IACD,IAAI,QAAQ,KAAK,KAAK;QAAE,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;IAChD,IAAI,QAAQ,KAAK,OAAO;QAAE,KAAK,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC;IACpD,IAAI,QAAQ,KAAK,SAAS;QAAE,KAAK,CAAC,GAAG,CAAC,aAAa,CAAC,CAAC;IACrD,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;IACnD,IAAI,cAAc,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,KAAK,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC;IACvD,IAAI,+BAA+B,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,KAAK,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC;IACzE,OAAO,CAAC,GAAG,KAAK,CAAC,CAAC;AACpB,CAAC;AAED,SAAS,SAAS,CAAC,MAAc,EAAE,QAAkB;IACnD,IAAI,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QAC5B,OAAO;YACL,cAAc,CAAC,6BAA6B,EAAE,eAAe,EAAE;gBAC7D,QAAQ,EAAE,oCAAoC;gBAC9C,eAAe,EAAE,gCAAgC;aAClD,CAAC;YACF,cAAc,CAAC,oBAAoB,EAAE,OAAO,EAAE;gBAC5C,QAAQ,EAAE,aAAa;gBACvB,eAAe,EAAE,oCAAoC;aACtD,CAAC;YACF,cAAc,CAAC,yBAAyB,EAAE,UAAU,EAAE;gBACpD,eAAe,EAAE,uCAAuC;aACzD,CAAC;YACF,cAAc,CAAC,sBAAsB,EAAE,OAAO,EAAE;gBAC9C,eAAe,EAAE,gCAAgC;aAClD,CAAC;SACH,CAAC;IACJ,CAAC;IACD,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACjD,OAAO;YACL,cAAc,CAAC,+BAA+B,EAAE,YAAY,EAAE;gBAC5D,QAAQ,EAAE,mBAAmB;aAC9B,CAAC;YACF,cAAc,CAAC,uBAAuB,EAAE,OAAO,EAAE;gBAC/C,QAAQ,EAAE,aAAa;aACxB,CAAC;YACF,cAAc,CAAC,6BAA6B,EAAE,YAAY,EAAE;gBAC1D,QAAQ,EAAE,oBAAoB;aAC/B,CAAC;YACF,cAAc,CAAC,mCAAmC,EAAE,OAAO,EAAE;gBAC3D,QAAQ,EAAE,0BAA0B;aACrC,CAAC;SACH,CAAC;IACJ,CAAC;IACD,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACxB,OAAO;YACL,cAAc,CAAC,6BAA6B,EAAE,OAAO,CAAC;YACtD,cAAc,CAAC,qBAAqB,EAAE,YAAY,EAAE;gBAClD,QAAQ,EAAE,qBAAqB;aAChC,CAAC;YACF,cAAc,CAAC,mBAAmB,EAAE,YAAY,EAAE;gBAChD,QAAQ,EAAE,SAAS;aACpB,CAAC;YACF,cAAc,CAAC,sBAAsB,EAAE,OAAO,CAAC;SAChD,CAAC;IACJ,CAAC;IACD,IAAI,QAAQ,KAAK,mBAAmB,EAAE,CAAC;QACrC,OAAO;YACL,cAAc,CAAC,uBAAuB,EAAE,mBAAmB,EAAE;gBAC3D,QAAQ,EAAE,aAAa;aACxB,CAAC;YACF,cAAc,CAAC,qBAAqB,EAAE,mBAAmB,EAAE;gBACzD,QAAQ,EAAE,eAAe;aAC1B,CAAC;YACF,cAAc,CAAC,oBAAoB,EAAE,OAAO,CAAC;SAC9C,CAAC;IACJ,CAAC;IACD,OAAO,EAAE,CAAC;AACZ,CAAC;AAED,SAAS,WAAW,CAAC,MAAc,EAAE,QAAkB;IACrD,IAAI,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QAC5B,OAAO,sEAAsE,CAAC;IAChF,CAAC;IACD,IAAI,QAAQ,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACjD,OAAO,+GAA+G,CAAC;IACzH,CAAC;IACD,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;QACxB,OAAO,0DAA0D,CAAC;IACpE,CAAC;IACD,IAAI,QAAQ,KAAK,QAAQ,EAAE,CAAC;QAC1B,OAAO,4DAA4D,CAAC;IACtE,CAAC;IACD,OAAO,mDAAmD,CAAC;AAC7D,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,WAAW,CAAC,MAAc;IACxC,MAAM,IAAI,GAAG,MAAM,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;IAChD,IAAI,6BAA6B,CAAC,IAAI,CAAC,EAAE,CAAC;QACxC,OAAO;YACL,UAAU,EAAE,UAAU;YACtB,UAAU,EAAE,KAAK;YACjB,QAAQ,EAAE,eAAe;YACzB,IAAI,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC;YACxB,mBAAmB,EAAE,KAAK;YAC1B,kBAAkB,EAAE,IAAI;YACxB,WAAW,EAAE,CAAC,UAAU,CAAC;YACzB,QAAQ,EACN,wIAAwI;YAC1I,cAAc,EAAE,EAAE;SACnB,CAAC;IACJ,CAAC;IACD,MAAM,KAAK,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC;IAC9B,MAAM,QAAQ,GAAG,gBAAgB,CAAC,IAAI,CAAC,CAAC;IACxC,MAAM,UAAU,GAAG,kBAAkB,CAAC,IAAI,EAAE,QAAQ,EAAE,KAAK,CAAC,CAAC;IAC7D,MAAM,UAAU,GAAG,aAAa,CAAC,IAAI,EAAE,UAAU,EAAE,QAAQ,CAAC,CAAC;IAC7D,MAAM,WAAW,GAAG,gBAAgB,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;IACrD,MAAM,cAAc,GAClB,UAAU,KAAK,QAAQ,IAAI,QAAQ,KAAK,QAAQ,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;IAEpF,OAAO;QACL,UAAU;QACV,UAAU;QACV,QAAQ;QACR,IAAI,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC;QACxB,mBAAmB,EACjB,QAAQ,KAAK,mBAAmB;YAChC,uDAAuD,CAAC,IAAI,CAAC,IAAI,CAAC;QACpE,kBAAkB,EAChB,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC;YACrB,uDAAuD,CAAC,IAAI,CAAC,IAAI,CAAC;QACpE,WAAW;QACX,QAAQ,EAAE,WAAW,CAAC,IAAI,EAAE,QAAQ,CAAC;QACrC,cAAc;KACf,CAAC;AACJ,CAAC;AAED,sEAAsE;AACtE,MAAM,UAAU,sBAAsB,CAAC,QAAsB;IAC3D,MAAM,KAAK,GACT,QAAQ,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC;QAC7B,CAAC,CAAC,QAAQ,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;QAC7C,CAAC,CAAC,GAAG,CAAC;IACV,MAAM,QAAQ,GAAG,QAAQ,CAAC,UAAU;QAClC,CAAC,CAAC,0IAA0I;QAC5I,CAAC,CAAC,qBAAqB,CAAC;IAC1B,OAAO,CACL,2BAA2B,QAAQ,CAAC,QAAQ,eAAe,QAAQ,CAAC,UAAU,KAAK,QAAQ,KAAK;QAChG,iBAAiB,KAAK,gBAAgB,QAAQ,CAAC,QAAQ,EAAE,CAC1D,CAAC;AACJ,CAAC"}

package/dist/agent/task-evidence.d.ts

@@ -1,26 +1,100 @@
 import type { ToolCall } from "../types.js";
-/**
- * Tracks successful non-meta tool work under the currently open plan task.
- * Models (esp. GPT-OSS) must not mark done or jump ahead without real
- * evidence from tool results this task has already received.
- */
-export interface TaskWorkLedger {
+import type { TaskEvidence } from "../store/plan.js";
+/** Successful non-meta work under the open plan task (typed evidence for done). */
+export type TaskClass = "explore" | "scaffold" | "install" | "implement" | "verify" | "recon" | "exploit" | "report" | "generic";
+export interface TaskWorkLedger extends TaskEvidence {
     taskId: string;
-    /** Successful work tools since this task became in_progress. */
-    successWorkCount: number;
-    lastOkTool?: string | undefined;
+}
+/** Rehydrate the live task ledger from durable plan evidence after a resume. */
+export declare function ledgerFromTaskEvidence(taskId: string, evidence?: TaskEvidence | undefined): TaskWorkLedger;
+/** Store only serializable evidence on the plan; the task id is already its key. */
+export declare function taskEvidenceFromLedger(ledger: TaskWorkLedger): TaskEvidence;
+/** Optional signals recorded from a successful tool call. */
+export interface TaskWorkSignals {
+    sourceWrite?: boolean;
+    featureWrite?: boolean;
+    installOk?: boolean;
+    scaffoldOk?: boolean;
+    devServerStart?: boolean;
+    localHttpProbeOk?: boolean;
+    serverReady?: boolean;
+    portListening?: boolean;
+    remoteReconOk?: boolean;
+    remoteActiveTestOk?: boolean;
 }
 export declare function isMetaPlanTool(name: string): boolean;
 /** Tools that count as real work/verify evidence for the open task. */
 export declare function isEvidenceWorkTool(name: string): boolean;
-export declare function recordTaskWorkSuccess(ledger: TaskWorkLedger | null, taskId: string | undefined, toolName: string): TaskWorkLedger | null;
+export declare function isPentestPlanKind(kind: string | undefined): boolean;
+export declare function classifyTaskTitle(title: string, opts?: {
+    planKind?: string | undefined;
+}): TaskClass;
+/** Successful work observed this turn without a credited open task. */
+export interface LooseWorkReceipt {
+    readonly toolName: string;
+    readonly signals?: TaskWorkSignals | undefined;
+}
+/**
+ * Whether a successful tool can satisfy a task of the given class.
+ * Used to retroactively credit preflight work (tool.check before in_progress).
+ */
+export declare function toolFitsTaskClass(toolName: string, taskTitle: string, opts?: {
+    planKind?: string | undefined;
+    signals?: TaskWorkSignals | undefined;
+}): boolean;
+/**
+ * Fold turn-level loose successes into a task ledger so preflight work
+ * (tool.check / fs.list before in_progress) can satisfy explore/check tasks.
+ */
+export declare function absorbLooseWorkIntoLedger(ledger: TaskWorkLedger | null, taskId: string, taskTitle: string, loose: readonly LooseWorkReceipt[], opts?: {
+    planKind?: string | undefined;
+}): TaskWorkLedger | null;
+export declare function recordTaskWorkSuccess(ledger: TaskWorkLedger | null, taskId: string | undefined, toolName: string, signals?: TaskWorkSignals): TaskWorkLedger | null;
 export declare function openTaskLedger(taskId: string): TaskWorkLedger;
-export declare function canMarkTaskDone(ledger: TaskWorkLedger | null, taskId: string): {
+export interface CanMarkTaskDoneOpts {
+    taskTitle?: string | undefined;
+    /** User asked for a product feature app (todo/blog/…). */
+    featureAppRequired?: boolean | undefined;
+    /** Session-level flag: feature write observed any time this turn. */
+    sessionFeatureSeen?: boolean | undefined;
+    /** A real project root was discovered, so a scaffold checkpoint is already satisfied. */
+    existingProject?: boolean | undefined;
+    /** An earlier plan task already proved the local server is ready and running. */
+    runtimeVerified?: boolean | undefined;
+    /** Plan kind: coding | pentest | general — selects domain gates. */
+    planKind?: string | undefined;
+    /** Any plan task already has successful remote recon/active test evidence. */
+    remoteWorkVerified?: boolean | undefined;
+}
+/** Local app runtime is proven by any strong independent signal. */
+export declare function hasLocalRuntimeProof(evidence?: Pick<TaskEvidence, "sawDevServerStart" | "sawLocalHttpProbeOk" | "sawServerReady" | "sawPortListening"> | null): boolean;
+export declare function hasRemoteWorkProof(evidence?: Pick<TaskEvidence, "sawRemoteReconOk" | "sawRemoteActiveTestOk"> | null): boolean;
+/** Dev-server job log shows ready + local URL. */
+export declare function isServerReadyOutput(output: string): boolean;
+/** lsof/ss (or similar) shows a process LISTENing on a local port. */
+export declare function isPortListeningOutput(command: string, output: string): boolean;
+/**
+ * Leave/keep-running observational tasks may inherit already-proven runtime.
+ * "for user to test" must NOT block inheritance (prior brittle exclusion of "test").
+ */
+export declare function isRuntimeObservationTask(title: string): boolean;
+/** Report / residual documentation that can inherit prior remote evidence. */
+export declare function isRemoteObservationTask(title: string): boolean;
+/**
+ * Typed evidence gate: base rule is ≥1 successful work tool; implement/install/verify
+ * (and pentest recon) require stronger signals when classified. Existing project and
+ * already-proven runtime/remote facts satisfy corresponding observational tasks.
+ */
+export declare function canMarkTaskDone(ledger: TaskWorkLedger | null, taskId: string, opts?: CanMarkTaskDoneOpts): {
     ok: true;
 } | {
     ok: false;
     reason: string;
 };
+/** True when a tool call is remote recon (not local app tooling). */
+export declare function isRemoteReconToolCall(call: ToolCall): boolean;
+/** Active/offensive test (not passive recon). */
+export declare function isRemoteActiveTestCall(call: ToolCall): boolean;
 /** One-shot scaffolders — must NEVER be treated as "start the dev server". Stack-agnostic. */
 export declare function isScaffoldCreateCommand(cmd: string): boolean;
 /**
@@ -28,12 +102,17 @@ export declare function isScaffoldCreateCommand(cmd: string): boolean;
  * npm create which merely contain the word "vite".
  */
 export declare function isDevServerCall(call: ToolCall): boolean;
-/** Preflight / inspect tools allowed before the model opens a task. */
+/** Inspect tools allowed without opening a plan task. */
 export declare function isPlanPreflightTool(name: string): boolean;
+/** Read-only recon tools that may run without an in_progress task on pentest plans. */
+export declare function isReadOnlyReconTool(name: string): boolean;
 /**
- * Tools allowed on a coding BUILD turn before plan.create exists.
- * Explore + optional web research + plan — no scaffold/write/install freestyle.
+ * Tools whose failure must NOT cancel later siblings in the same model turn.
+ * Plan bookkeeping failures (task.update done-gate, etc.) and independent
+ * recon lookups should never wipe a whole batch of http.fetch / dns calls.
  */
+export declare function isBatchSoftFailTool(name: string): boolean;
+/** Tools allowed on a coding build turn before plan.create exists. */
 export declare function isBuildPrePlanAllowedTool(name: string): boolean;
 /**
  * Commands that legitimately sit quiet for minutes (npm install, create-next-app).
@@ -45,6 +124,18 @@ export declare function toolStallBudgetMs(call: {
     name: string;
     args: Record<string, unknown>;
 }): number;
+/**
+ * Hard wall-clock budget for a tool call. When exceeded the runner aborts
+ * and force-settles even if the underlying promise never resolves (hung
+ * socket that ignored AbortSignal). Distinct from the stall watchdog
+ * (which fires on *silence*); this fires on total elapsed time.
+ */
+export declare function toolHardBudgetMs(call: {
+    name: string;
+    args: Record<string, unknown>;
+}): number;
+/** Grace period after abort before force-settling a hung tool promise. */
+export declare const TOOL_ABORT_GRACE_MS = 2500;
 /** Harmless version/which probes models use instead of tool.check before planning. */
 export declare function isReadOnlyVersionProbeCommand(cmd: string): boolean;
 /**
@@ -59,15 +150,6 @@ export declare function codingBuildRequiresPlan(prompt: string, opts?: {
 export declare function looksLikeInstallTaskTitle(title: string): boolean;
 export declare function looksLikeScaffoldTaskTitle(title: string): boolean;
 export declare function isPackageInstallCommand(cmd: string): boolean;
-/**
- * Soft scope check: refuse run/verify tools while the open task is only
- * scaffold / implement / install.
- * When `planTaskTitles` is provided, install is only blocked on implement
- * tasks if the plan actually has a dedicated install task (otherwise allow).
- */
-export declare function workOutOfScopeForTask(taskTitle: string, call: ToolCall, opts?: {
-    planTaskTitles?: string[];
-}): string | undefined;
 /**
  * Choose which pending plan task should own this tool call (auto-start).
  * Prevents npm install from being attributed to "localStorage" / style tasks.
@@ -75,7 +157,7 @@ export declare function workOutOfScopeForTask(taskTitle: string, call: ToolCall,
 export declare function pickPendingTaskForToolCall<T extends {
     id: string;
     title: string;
-}>(pending: T[], call: ToolCall, allTitles?: string[]): T | undefined;
+}>(pending: T[], call: ToolCall, _allTitles?: string[]): T | undefined;
 /**
  * User asked for a product feature app (todo, blog, …), not "just scaffold framework X".
  * Stack-agnostic: detects intent, not a particular framework.
@@ -84,15 +166,15 @@ export declare function userAskedForFeatureApp(prompt: string): boolean;
 /** Paths that look like application source (not lockfiles / build output). */
 export declare function isAppSourcePath(path: string): boolean;
 export declare function extractWritePathsFromCall(call: ToolCall): string[];
+/** Default framework starter / marketing copy — not the product feature. */
+export declare function looksLikeStarterBoilerplate(content: string): boolean;
+/** Content that looks like real product UI/logic for common feature apps. */
+export declare function looksLikeFeatureProductCode(content: string): boolean;
 /**
  * True when a successful tool call implements product feature code (not only scaffold).
+ * Starter boilerplate alone does not count.
  */
 export declare function isFeatureImplementationCall(call: ToolCall): boolean;
-/**
- * Soft block: don't start the local server until the product feature exists.
- * Returns a message when the open work is still scaffold-only for a feature ask.
- */
-export declare function incompleteFeatureBeforeServerMessage(userPrompt: string, sawFeatureImpl: boolean, call: ToolCall): string | undefined;
 /**
  * User said work belongs on the Desktop (or similar). Returns absolute path.
  */