@pentoshi/clai 3.6.0 → 3.7.3

package/dist/agent/task-evidence.js

@@ -1,6 +1,19 @@
 import { homedir } from "node:os";
 import { join } from "node:path";
 import { getActiveProjectRoot } from "./project-root.js";
+/** Rehydrate the live task ledger from durable plan evidence after a resume. */
+export function ledgerFromTaskEvidence(taskId, evidence) {
+    return {
+        taskId,
+        successWorkCount: evidence?.successWorkCount ?? 0,
+        ...(evidence ?? {}),
+    };
+}
+/** Store only serializable evidence on the plan; the task id is already its key. */
+export function taskEvidenceFromLedger(ledger) {
+    const { taskId: _taskId, ...evidence } = ledger;
+    return evidence;
+}
 export function isMetaPlanTool(name) {
     return name === "plan.create" || name === "task.update" || name === "agent.handoff";
 }
@@ -8,23 +21,256 @@ export function isMetaPlanTool(name) {
 export function isEvidenceWorkTool(name) {
     return !isMetaPlanTool(name);
 }
-export function recordTaskWorkSuccess(ledger, taskId, toolName) {
-    if (!taskId || !isEvidenceWorkTool(toolName))
+export function isPentestPlanKind(kind) {
+    return (kind ?? "").toLowerCase() === "pentest";
+}
+export function classifyTaskTitle(title, opts) {
+    const t = title.toLowerCase();
+    const pentest = isPentestPlanKind(opts?.planKind);
+    // Pentest plans: never map security work onto coding verify/implement gates.
+    if (pentest) {
+        if (/\b(report|write.?up|finding|summar|document|residual)\b/.test(t)) {
+            return "report";
+        }
+        if (/\b(exploit|poc|payload|privesc|lateral|inject|bruteforce|sqlmap|listener|reverse\s+shell|c2)\b/.test(t)) {
+            return "exploit";
+        }
+        if (/\b(recon|enumerat|fingerprint|osint|whois|dns|nmap|discover|probe|scan|fuzz|content\s+discover|subdomain|port)\b/.test(t)) {
+            return "recon";
+        }
+        if (/\b(explore|inspect|list|read|survey|map)\b/.test(t)) {
+            return "explore";
+        }
+        return "generic";
+    }
+    if (/\b(dev\s*server|run\s+dev|start\s+.*server|localhost|shell\.start|leave\s+.*running|probe|verify\s+in\s+browser)\b/.test(t) ||
+        (/\b(start|run)\b/.test(t) && /\b(server|dev|app|verify)\b/.test(t))) {
+        return "verify";
+    }
+    if (looksLikeInstallTaskTitle(title))
+        return "install";
+    if (looksLikeScaffoldTaskTitle(title))
+        return "scaffold";
+    if (/\b(implement|integrate|feature|rewrite|component|todo|persist|localstorage|styling|styles?|ui|page)\b/.test(t) ||
+        // endpoint/route only count as implement for product apps, not bare "API routes" recon phrasing
+        (/\b(endpoint|route)\b/.test(t) &&
+            /\b(implement|build|add|create|feature|component|ui|page)\b/.test(t))) {
+        return "implement";
+    }
+    if (/\b(recon|enumerat|fingerprint|osint|whois|dns|nmap|discover)\b/.test(t)) {
+        return "recon";
+    }
+    if (/\b(exploit|poc|payload|privesc|lateral|inject|bruteforce|sqlmap)\b/.test(t)) {
+        return "exploit";
+    }
+    if (/\b(report|write.?up|finding|summar|document)\b/.test(t)) {
+        return "report";
+    }
+    if (/\b(explore|inspect|list|read|survey|map)\b/.test(t) ||
+        /\bcheck\b.+\b(exists?|empty|directory|folder|path)\b/.test(t) ||
+        /\b(exists?|empty)\b.+\b(directory|folder|path|project)\b/.test(t) ||
+        // "Check Node.js and npm availability", "verify tools present", etc.
+        /\bcheck\b.+\b(node|npm|pnpm|yarn|bun|python|availability|available|installed|present|toolchain|tools?)\b/.test(t) ||
+        /\b(availability|available|installed|present)\b.+\b(node|npm|toolchain|tools?)\b/.test(t) ||
+        /\bverify\b.+\b(tools?|node|npm|setup|environment|prereq)\b/.test(t)) {
+        return "explore";
+    }
+    return "generic";
+}
+/**
+ * Whether a successful tool can satisfy a task of the given class.
+ * Used to retroactively credit preflight work (tool.check before in_progress).
+ */
+export function toolFitsTaskClass(toolName, taskTitle, opts) {
+    if (!isEvidenceWorkTool(toolName))
+        return false;
+    const cls = classifyTaskTitle(taskTitle, { planKind: opts?.planKind });
+    const s = opts?.signals;
+    switch (cls) {
+        case "explore":
+            return (toolName === "tool.check" ||
+                toolName === "sysinfo" ||
+                toolName === "fs.list" ||
+                toolName === "fs.read" ||
+                toolName === "fs.search" ||
+                toolName === "net.context" ||
+                toolName === "shell.exec" // version probes, which/where
+            );
+        case "scaffold":
+            return (Boolean(s?.scaffoldOk) ||
+                toolName === "fs.write" ||
+                toolName === "fs.writeMany" ||
+                toolName === "shell.exec" ||
+                toolName === "shell.start");
+        case "install":
+            return Boolean(s?.installOk) || toolName === "pkg.install";
+        case "implement":
+            return (Boolean(s?.sourceWrite || s?.featureWrite) ||
+                toolName === "fs.write" ||
+                toolName === "fs.writeMany" ||
+                toolName === "fs.edit" ||
+                toolName === "fs.replaceLines" ||
+                toolName === "fs.append");
+        case "verify":
+            return (Boolean(s?.devServerStart ||
+                s?.localHttpProbeOk ||
+                s?.serverReady ||
+                s?.portListening) ||
+                toolName === "shell.start" ||
+                toolName === "shell.tail" ||
+                toolName === "http.fetch" ||
+                toolName === "web.fetch");
+        case "recon":
+            return (Boolean(s?.remoteReconOk) ||
+                /^(dns\.lookup|whois\.lookup|http\.fetch|web\.fetch|net\.scan|pentest\.recon|net\.pingSweep|net\.context|tool\.batch)$/.test(toolName));
+        case "exploit":
+            return Boolean(s?.remoteActiveTestOk) || isEvidenceWorkTool(toolName);
+        default:
+            return true;
+    }
+}
+/**
+ * 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 function absorbLooseWorkIntoLedger(ledger, taskId, taskTitle, loose, opts) {
+    if (!taskId || loose.length === 0)
         return ledger;
-    if (!ledger || ledger.taskId !== taskId) {
-        return { taskId, successWorkCount: 1, lastOkTool: toolName };
+    let led = ledger;
+    for (const item of loose) {
+        if (!toolFitsTaskClass(item.toolName, taskTitle, {
+            planKind: opts?.planKind,
+            signals: item.signals,
+        })) {
+            continue;
+        }
+        led = recordTaskWorkSuccess(led, taskId, item.toolName, item.signals);
     }
-    return {
-        taskId,
-        successWorkCount: ledger.successWorkCount + 1,
-        lastOkTool: toolName,
-    };
+    return led;
+}
+export function recordTaskWorkSuccess(ledger, taskId, toolName, signals) {
+    if (!taskId || !isEvidenceWorkTool(toolName))
+        return ledger;
+    const base = !ledger || ledger.taskId !== taskId
+        ? { taskId, successWorkCount: 1, lastOkTool: toolName }
+        : {
+            ...ledger,
+            successWorkCount: ledger.successWorkCount + 1,
+            lastOkTool: toolName,
+        };
+    if (signals?.sourceWrite)
+        base.sawSourceWrite = true;
+    if (signals?.featureWrite)
+        base.sawFeatureWrite = true;
+    if (signals?.installOk)
+        base.sawInstallOk = true;
+    if (signals?.scaffoldOk)
+        base.sawScaffoldOk = true;
+    if (signals?.devServerStart)
+        base.sawDevServerStart = true;
+    if (signals?.localHttpProbeOk)
+        base.sawLocalHttpProbeOk = true;
+    if (signals?.serverReady)
+        base.sawServerReady = true;
+    if (signals?.portListening)
+        base.sawPortListening = true;
+    if (signals?.remoteReconOk)
+        base.sawRemoteReconOk = true;
+    if (signals?.remoteActiveTestOk)
+        base.sawRemoteActiveTestOk = true;
+    return base;
 }
 export function openTaskLedger(taskId) {
     return { taskId, successWorkCount: 0 };
 }
-export function canMarkTaskDone(ledger, taskId) {
-    if (!ledger || ledger.taskId !== taskId) {
+/** Local app runtime is proven by any strong independent signal. */
+export function hasLocalRuntimeProof(evidence) {
+    return Boolean(evidence?.sawDevServerStart ||
+        evidence?.sawLocalHttpProbeOk ||
+        evidence?.sawServerReady ||
+        evidence?.sawPortListening);
+}
+export function hasRemoteWorkProof(evidence) {
+    return Boolean(evidence?.sawRemoteReconOk || evidence?.sawRemoteActiveTestOk);
+}
+/** Dev-server job log shows ready + local URL. */
+export function isServerReadyOutput(output) {
+    if (!output)
+        return false;
+    const hasUrl = /\bhttps?:\/\/(?:localhost|127\.0\.0\.1|\[::1\])(?::\d+)?\b/i.test(output) ||
+        /➜\s*Local:\s*https?:\/\//i.test(output);
+    const hasReady = /\b(?:ready in|VITE\s+v?\d|compiled successfully|started server|listening on|Local:\s*http)\b/i.test(output);
+    return hasUrl || hasReady;
+}
+/** lsof/ss (or similar) shows a process LISTENing on a local port. */
+export function isPortListeningOutput(command, output) {
+    if (!output || !/\bLISTEN(?:ING)?\b/i.test(output))
+        return false;
+    const cmd = command.toLowerCase();
+    if (/\b(lsof|ss\b|netstat)\b/.test(cmd))
+        return true;
+    // Output-only: node/… LISTEN with a localhost/port line
+    return (/\b(?:localhost|127\.0\.0\.1|\[::1\]|\*:)\S*\d+/i.test(output) ||
+        /\bTCP\b.+\bLISTEN/i.test(output));
+}
+/**
+ * Leave/keep-running observational tasks may inherit already-proven runtime.
+ * "for user to test" must NOT block inheritance (prior brittle exclusion of "test").
+ */
+export function isRuntimeObservationTask(title) {
+    const t = title.toLowerCase();
+    const leaveKeep = /\bleave\s+(?:it\s+)?running\b|\bkeep\s+(?:it\s+)?running\b|\bserver\s+(?:is\s+)?running\b|\bfor\s+(?:the\s+)?user\s+to\s+test\b/.test(t);
+    if (!leaveKeep)
+        return false;
+    // Primary action is still start/launch/probe → not pure observation
+    if (/\b(?:start|launch|run)\s+(?:the\s+)?(?:dev\s*)?server\b/.test(t) ||
+        /\b(?:probe|curl)\s+(?:localhost|http)\b/.test(t)) {
+        // Combined "start … leave running" is verify work, not pure observation
+        if (/\b(start|run|launch)\b/.test(t) && /\b(dev\s*server|npm\s+run\s+dev|shell\.start)\b/.test(t)) {
+            return false;
+        }
+    }
+    // Titles that are ONLY leave/keep running (optionally "for user to test")
+    if (/\bleave\s+(?:it\s+)?running\b|\bkeep\s+(?:it\s+)?running\b|\bfor\s+(?:the\s+)?user\s+to\s+test\b/.test(t)) {
+        // If title is primarily leave-running without requiring a fresh start as the main verb first
+        if (!/^(?:start|run|launch)\b/.test(t.trim()))
+            return true;
+        // "Leave server running for user to test" — leave is the head intent
+        if (/^leave\b|^keep\b/.test(t.trim()))
+            return true;
+    }
+    return (/\bserver\s+(?:is\s+)?running\b/.test(t) &&
+        !/\b(?:start|run|launch)\b/.test(t));
+}
+/** Report / residual documentation that can inherit prior remote evidence. */
+export function isRemoteObservationTask(title) {
+    const t = title.toLowerCase();
+    if (/\b(report|write.?up|summar|document|residual|findings?\s+summary)\b/.test(t)) {
+        return true;
+    }
+    if (/\bleave\s+(?:the\s+)?(?:listener|job|tunnel)\s+running\b|\bkeep\s+(?:the\s+)?listener\b/.test(t)) {
+        return true;
+    }
+    return false;
+}
+/**
+ * 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 function canMarkTaskDone(ledger, taskId, opts) {
+    const title = opts?.taskTitle ?? "";
+    const planKind = opts?.planKind;
+    const pentest = isPentestPlanKind(planKind);
+    const cls = title ? classifyTaskTitle(title, { planKind }) : "generic";
+    const leaveRunningIntent = isRuntimeObservationTask(title) ||
+        /\bleave\s+(?:it\s+)?running\b|\bkeep\s+(?:it\s+)?running\b|\bfor\s+(?:the\s+)?user\s+to\s+test\b/i.test(title);
+    const inheritedCompletion = (cls === "scaffold" && Boolean(opts?.existingProject)) ||
+        (!pentest && leaveRunningIntent && Boolean(opts?.runtimeVerified)) ||
+        (pentest &&
+            isRemoteObservationTask(title) &&
+            Boolean(opts?.remoteWorkVerified));
+    if ((!ledger || ledger.taskId !== taskId) && !inheritedCompletion) {
         return {
             ok: false,
             reason: `Cannot mark [${taskId}] done: no work ledger for this task. ` +
@@ -32,18 +278,113 @@ export function canMarkTaskDone(ledger, taskId) {
                 `inspect the tool result, and only then mark done if you are satisfied.`,
         };
     }
-    if (ledger.successWorkCount < 1) {
+    if ((ledger?.successWorkCount ?? 0) < 1 && !inheritedCompletion) {
         return {
             ok: false,
             reason: `Cannot mark [${taskId}] done: no successful tool result observed since it went in_progress. ` +
                 `Do the work, wait for the tool output, verify it succeeded, and only mark done when satisfied. ` +
-                (ledger.lastOkTool
+                (ledger?.lastOkTool
                     ? ""
                     : "Example: fs.write / shell.exec / shell.start must return ok first."),
         };
     }
+    const evidence = ledger && ledger.taskId === taskId ? ledger : openTaskLedger(taskId);
+    // Feature implement gate is coding-only.
+    if (!pentest && cls === "implement" && opts?.featureAppRequired) {
+        const featureOk = Boolean(evidence.sawFeatureWrite) || Boolean(opts.sessionFeatureSeen);
+        if (!featureOk) {
+            return {
+                ok: false,
+                reason: `Cannot mark [${taskId}] done: this is a feature/implement task for a product app, ` +
+                    `but no real feature code was written yet (starter boilerplate does not count). ` +
+                    `Replace the default starter with the requested feature (e.g. todo add/list/toggle), then mark done.`,
+            };
+        }
+    }
+    if (!pentest && cls === "install" && !evidence.sawInstallOk) {
+        if (evidence.successWorkCount < 1) {
+            return {
+                ok: false,
+                reason: `Cannot mark [${taskId}] done: install task needs a successful package install ` +
+                    `(npm/pnpm/yarn/bun install, etc.) under this task.`,
+            };
+        }
+    }
+    // Local app runtime: multi-signal proof. Never apply to pentest plans.
+    if (!pentest && (cls === "verify" || leaveRunningIntent)) {
+        const wantsServer = leaveRunningIntent ||
+            /\b(dev\s*server|localhost|shell\.start|run\s+dev|leave\s+.*running)\b/i.test(title);
+        if (wantsServer) {
+            const proof = hasLocalRuntimeProof(evidence) || Boolean(opts?.runtimeVerified);
+            if (!proof) {
+                return {
+                    ok: false,
+                    reason: `Cannot mark [${taskId}] done: local app runtime not proven yet. ` +
+                        `Accept any of: shell.start (dev server), shell.tail ready+URL, port LISTEN (lsof/ss), ` +
+                        `or successful localhost GET probe. If an earlier plan task already proved the server, ` +
+                        `confirm it is still alive once then mark done — do not thrash ports with restarts.`,
+                };
+            }
+        }
+    }
+    // Pentest recon: need real remote recon, not local fs noise alone.
+    if (pentest && cls === "recon") {
+        const remoteOk = hasRemoteWorkProof(evidence) || Boolean(opts?.remoteWorkVerified);
+        if (!remoteOk) {
+            // Allow if last tool was a recon tool name even without typed flag (soft)
+            const last = evidence.lastOkTool ?? "";
+            const lastLooksRecon = /^(dns\.lookup|whois\.lookup|http\.fetch|web\.fetch|net\.scan|pentest\.recon|net\.pingSweep|tool\.batch)$/.test(last);
+            if (!lastLooksRecon) {
+                return {
+                    ok: false,
+                    reason: `Cannot mark [${taskId}] done: recon needs successful remote evidence ` +
+                        `(dns/http/net.scan/pentest.recon/whois against the target), not local-only tools.`,
+                };
+            }
+        }
+    }
     return { ok: true };
 }
+/** True when a tool call is remote recon (not local app tooling). */
+export function isRemoteReconToolCall(call) {
+    if (call.name === "dns.lookup" ||
+        call.name === "whois.lookup" ||
+        call.name === "net.scan" ||
+        call.name === "pentest.recon" ||
+        call.name === "net.pingSweep" ||
+        call.name === "net.context") {
+        return true;
+    }
+    if (call.name === "http.fetch" || call.name === "web.fetch") {
+        const url = typeof call.args.url === "string" ? call.args.url : "";
+        // Localhost is coding probe, not remote recon
+        if (/^(?:https?:\/\/)?(?:localhost|127\.0\.0\.1|\[::1\])(?::|\/|$)/i.test(url)) {
+            return false;
+        }
+        return Boolean(url);
+    }
+    if (call.name === "shell.exec" || call.name === "shell.start") {
+        const cmd = commandOf(call);
+        return /\b(?:nmap|masscan|ffuf|gobuster|feroxbuster|nikto|nuclei|httpx|subfinder|amass|dig|whois|whatweb)\b/i.test(cmd);
+    }
+    return false;
+}
+/** Active/offensive test (not passive recon). */
+export function isRemoteActiveTestCall(call) {
+    if (call.name === "http.fetch") {
+        const method = String(call.args.method ?? "GET").toUpperCase();
+        if (method !== "GET" && method !== "HEAD" && method !== "OPTIONS") {
+            return true;
+        }
+        const url = typeof call.args.url === "string" ? call.args.url : "";
+        return /\b(?:union\s+select|<\s*script|javascript:|file:\/\/|169\.254\.169\.254)\b/i.test(url);
+    }
+    if (call.name === "shell.exec" || call.name === "shell.start") {
+        const cmd = commandOf(call);
+        return /\b(?:sqlmap|hydra|nikto|nuclei|msfconsole|exploit|payload|reverse\s+shell)\b/i.test(cmd);
+    }
+    return false;
+}
 function commandOf(call) {
     return typeof call.args.command === "string" ? call.args.command : "";
 }
@@ -70,7 +411,7 @@ export function isDevServerCall(call) {
     }
     return (/\bnpm\s+run\s+dev\b|\byarn\s+dev\b|\bpnpm\s+(run\s+)?dev\b|\bbun\s+(run\s+)?dev\b|\bnext\s+dev\b|\bnuxt\s+dev\b|\brails\s+s(?:erver)?\b|\bdjango(-admin)?\s+runserver\b/i.test(cmd) || /(?:^|[;&|]\s*)vite(?:\s|$)/i.test(cmd));
 }
-/** Preflight / inspect tools allowed before the model opens a task. */
+/** Inspect tools allowed without opening a plan task. */
 export function isPlanPreflightTool(name) {
     return (name === "tool.check" ||
         name === "sysinfo" ||
@@ -79,16 +420,45 @@ export function isPlanPreflightTool(name) {
         name === "fs.search" ||
         name === "net.context");
 }
+/** Read-only recon tools that may run without an in_progress task on pentest plans. */
+export function isReadOnlyReconTool(name) {
+    return (isPlanPreflightTool(name) ||
+        name === "tool.batch" ||
+        name === "dns.lookup" ||
+        name === "whois.lookup" ||
+        name === "http.fetch" ||
+        name === "web.fetch" ||
+        name === "web.search" ||
+        name === "net.scan" ||
+        name === "pentest.recon" ||
+        name === "wordlist.find");
+}
 /**
- * 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 function isBatchSoftFailTool(name) {
+    if (name === "plan.create" || name === "task.update")
+        return true;
+    if (name === "tool.batch" || name === "tool.check")
+        return true;
+    if (isReadOnlyReconTool(name))
+        return true;
+    if (name === "net.pingSweep" ||
+        name === "shell.jobs" ||
+        name === "shell.tail" ||
+        name === "sysinfo") {
+        return true;
+    }
+    return false;
+}
+/** Tools allowed on a coding build turn before plan.create exists. */
 export function isBuildPrePlanAllowedTool(name) {
     return (name === "plan.create" ||
         isPlanPreflightTool(name) ||
         name === "tool.batch" ||
         name === "fs.search" ||
-        // Research before planning (docs/examples) is fine; it does not mutate the project
         name === "web.search" ||
         name === "web.fetch");
 }
@@ -120,8 +490,47 @@ export function toolStallBudgetMs(call) {
         isLongQuietInstallOrScaffoldCommand(cmd)) {
         return 15 * 60_000; // create-next-app + npm install often 2–10+ min with quiet stretches
     }
+    // Network tools should not sit silent for a full minute — their own
+    // per-request timeouts are shorter; 45s is a safety net for hung sockets.
+    if (call.name === "web.search" || call.name === "web.fetch" || call.name === "http.fetch") {
+        return 45_000;
+    }
     return 60_000;
 }
+/**
+ * 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 function toolHardBudgetMs(call) {
+    const cmd = typeof call.args.command === "string" ? call.args.command : "";
+    if (call.name === "pkg.install")
+        return 20 * 60_000;
+    if ((call.name === "shell.exec" || call.name === "shell.start") &&
+        isLongQuietInstallOrScaffoldCommand(cmd)) {
+        return 20 * 60_000;
+    }
+    if (call.name === "web.search") {
+        // search (15s) + up to 3 fetchTop pages (30s each) + margin
+        const fetchTop = typeof call.args.fetchTop === "number" && Number.isFinite(call.args.fetchTop)
+            ? Math.max(0, Math.min(3, Math.floor(call.args.fetchTop)))
+            : 0;
+        return 15_000 + fetchTop * 30_000 + 15_000;
+    }
+    if (call.name === "web.fetch")
+        return 45_000;
+    if (call.name === "http.fetch")
+        return 90_000;
+    if (call.name === "net.scan" || call.name === "pentest.recon")
+        return 15 * 60_000;
+    if (call.name === "shell.start")
+        return 120_000; // should background quickly
+    // Default hard cap so a silent hung tool cannot freeze the session forever.
+    return 5 * 60_000;
+}
+/** Grace period after abort before force-settling a hung tool promise. */
+export const TOOL_ABORT_GRACE_MS = 2_500;
 /** Harmless version/which probes models use instead of tool.check before planning. */
 export function isReadOnlyVersionProbeCommand(cmd) {
     const c = cmd.trim();
@@ -170,67 +579,18 @@ export function looksLikeScaffoldTaskTitle(title) {
 export function isPackageInstallCommand(cmd) {
     return /\bnpm\s+i(nstall)?\b|\byarn\s+install\b|\bpnpm\s+i(nstall)?\b|\bbun\s+install\b|\bpip\s+install\b|\bpoetry\s+install\b|\bcargo\s+build\b|\bcomposer\s+install\b|\bbundle\s+install\b|\bgo\s+mod\s+tidy\b/.test(cmd);
 }
-/**
- * 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 function workOutOfScopeForTask(taskTitle, call, opts) {
-    if (isMetaPlanTool(call.name))
-        return undefined;
-    const t = taskTitle.toLowerCase();
-    const isRunVerifyTask = /\b(dev\s*server|run\s+dev|start\s+.*server|localhost|shell\.start|leave\s+.*running|probe|verify\s+in\s+browser)\b/.test(t) ||
-        (/\b(start|run)\b/.test(t) && /\b(server|dev|app|verify)\b/.test(t));
-    if (isRunVerifyTask)
-        return undefined;
-    const isInstallOnly = looksLikeInstallTaskTitle(taskTitle);
-    const isScaffoldOnly = looksLikeScaffoldTaskTitle(taskTitle);
-    const isImplementOnly = !isInstallOnly &&
-        !isScaffoldOnly &&
-        (/\b(implement|integrate|feature|rewrite|write\s+src|add\s+\w+\s+(component|module|handler|endpoint|page|route|model)|persist|localstorage|styling|styles?)\b/.test(t) ||
-            (/\b(component|module|handler|endpoint|ui|page|todo)\b/.test(t) &&
-                !/\b(install|scaffold|dev\s*server|create\s+project)\b/.test(t)));
-    // Install always OK under scaffold or install tasks
-    if ((isScaffoldOnly || isInstallOnly) &&
-        call.name === "shell.exec" &&
-        isPackageInstallCommand(commandOf(call))) {
-        return undefined;
-    }
-    if ((isInstallOnly || isScaffoldOnly || isImplementOnly) &&
-        (isDevServerCall(call) || isLocalHttpProbe(call))) {
-        return (`Out of scope for current task "${taskTitle}": do not start the dev server or probe localhost yet. ` +
-            `Finish and VERIFY this task's own work, mark it done only when satisfied, then open the run/verify task with task.update in_progress.`);
-    }
-    if (isImplementOnly &&
-        call.name === "shell.exec" &&
-        isPackageInstallCommand(commandOf(call))) {
-        const planTitles = opts?.planTaskTitles;
-        const planHasInstallTask = !planTitles ||
-            planTitles.length === 0 ||
-            planTitles.some((title) => looksLikeInstallTaskTitle(title));
-        if (planHasInstallTask) {
-            return (`Out of scope for current task "${taskTitle}": run package install under the install task, not while implementing files.`);
-        }
-        // No install task in the plan — allow install so work is not stuck.
-        return undefined;
-    }
-    return undefined;
-}
 /**
  * Choose which pending plan task should own this tool call (auto-start).
  * Prevents npm install from being attributed to "localStorage" / style tasks.
  */
-export function pickPendingTaskForToolCall(pending, call, allTitles) {
+export function pickPendingTaskForToolCall(pending, call, _allTitles) {
     if (pending.length === 0)
         return undefined;
-    const titles = allTitles ?? pending.map((p) => p.title);
     const cmd = commandOf(call);
     const score = (title) => {
-        if (workOutOfScopeForTask(title, call, { planTaskTitles: titles })) {
-            return -100;
-        }
-        let s = 1; // any in-scope pending task is viable
+        // Heuristics rank likely ownership only. They must never make a pending
+        // task ineligible or authorize/block the underlying tool call.
+        let s = 1;
         if (isPackageInstallCommand(cmd)) {
             if (looksLikeInstallTaskTitle(title))
                 s += 20;
@@ -255,6 +615,19 @@ export function pickPendingTaskForToolCall(pending, call, allTitles) {
                 s += 12;
             }
         }
+        // Keyword overlap (SSRF task ↔ /api/og-image fetch, fuzz ↔ /api/*, …)
+        const blob = `${call.name} ${cmd} ${JSON.stringify(call.args ?? {})}`.toLowerCase();
+        const words = title
+            .toLowerCase()
+            .split(/[^a-z0-9]+/)
+            .filter((w) => w.length > 3);
+        let hits = 0;
+        for (const w of words) {
+            if (blob.includes(w))
+                hits += 1;
+        }
+        if (hits > 0)
+            s += Math.min(12, hits * 3);
         return s;
     };
     let best;
@@ -266,9 +639,7 @@ export function pickPendingTaskForToolCall(pending, call, allTitles) {
             bestScore = sc;
         }
     }
-    if (!best || bestScore < 0)
-        return undefined;
-    return best;
+    return best ?? pending[0];
 }
 /**
  * User asked for a product feature app (todo, blog, …), not "just scaffold framework X".
@@ -321,33 +692,90 @@ export function extractWritePathsFromCall(call) {
     }
     return paths;
 }
+/** Default framework starter / marketing copy — not the product feature. */
+export function looksLikeStarterBoilerplate(content) {
+    const c = content.slice(0, 8_000);
+    return (/\bWelcome to (?:Vite|React|Next\.js|Create React App|Nuxt|SvelteKit|Angular)\b/i.test(c) ||
+        /\bGet started by editing\b/i.test(c) ||
+        /\bEdit\s+`?(?:src\/)?App\.(?:tsx?|jsx?)`?\s+and save to (?:test|reload)/i.test(c) ||
+        /\blogos?\/(?:react|vite|next)\.svg\b/i.test(c) ||
+        /\bnext\.js logo\b/i.test(c) ||
+        /https?:\/\/(?:react\.dev|vitejs\.dev|nextjs\.org)\/(?:learn|docs)?/i.test(c) && /\b(?:Deploy|Learn|Templates|create-next-app)\b/i.test(c));
+}
+/** Content that looks like real product UI/logic for common feature apps. */
+export function looksLikeFeatureProductCode(content) {
+    const c = content.slice(0, 12_000);
+    if (looksLikeStarterBoilerplate(c))
+        return false;
+    // Interactive product signals (todo/crud/auth/etc.)
+    if (/\b(?:useState|useReducer|createContext|signal\(|createSignal|ref\()\b/.test(c) &&
+        /\b(?:todo|todos|onClick|onSubmit|addTodo|toggle|complete|delete|remove|login|signup|password|fetch\(|axios|localStorage)\b/i.test(c)) {
+        return true;
+    }
+    if (/\b(?:add|create|toggle|delete|remove|update)\b/i.test(c) &&
+        /\b(?:todo|item|task|post|note|user|cart)\b/i.test(c) &&
+        /\b(?:map\s*\(|filter\s*\(|set[A-Z]\w+|useState)\b/.test(c)) {
+        return true;
+    }
+    // API/backend feature
+    if (/\b(?:app\.(?:get|post|put|patch|delete)|router\.(?:get|post)|@(?:Get|Post|Put|Delete)|def\s+\w+\(.*request)/i.test(c) &&
+        !looksLikeStarterBoilerplate(c)) {
+        return true;
+    }
+    return false;
+}
+function extractWriteContentsFromCall(call) {
+    const out = [];
+    if ((call.name === "fs.write" ||
+        call.name === "fs.append" ||
+        call.name === "fs.replaceLines") &&
+        typeof call.args.content === "string") {
+        out.push(call.args.content);
+    }
+    if (call.name === "fs.edit" && typeof call.args.newText === "string") {
+        out.push(call.args.newText);
+    }
+    if (call.name === "fs.writeMany" && Array.isArray(call.args.files)) {
+        for (const f of call.args.files) {
+            if (f &&
+                typeof f === "object" &&
+                typeof f.content === "string") {
+                out.push(f.content);
+            }
+        }
+    }
+    return out;
+}
 /**
  * True when a successful tool call implements product feature code (not only scaffold).
+ * Starter boilerplate alone does not count.
  */
 export function isFeatureImplementationCall(call) {
     const paths = extractWritePathsFromCall(call);
-    if (paths.some(isAppSourcePath))
+    const contents = extractWriteContentsFromCall(call);
+    if (contents.some(looksLikeFeatureProductCode))
+        return true;
+    // Content present but only starter → not feature
+    if (contents.length > 0 &&
+        contents.every((c) => looksLikeStarterBoilerplate(c) || c.trim().length < 40)) {
+        return false;
+    }
+    // Path-only signal: source write without inspectable content still counts
+    // (e.g. truncated args) but not lockfiles/config alone.
+    if (paths.some(isAppSourcePath) && contents.length === 0)
         return true;
-    // Hand-written multi-file setup without official scaffolder
-    if (call.name === "fs.writeMany" && paths.length >= 2)
+    if (paths.some(isAppSourcePath) &&
+        contents.some((c) => !looksLikeStarterBoilerplate(c) && c.trim().length >= 80)) {
         return true;
+    }
+    // Hand-written multi-file setup with non-starter content
+    if (call.name === "fs.writeMany" &&
+        paths.length >= 2 &&
+        contents.some((c) => !looksLikeStarterBoilerplate(c))) {
+        return true;
+    }
     return false;
 }
-/**
- * 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 function incompleteFeatureBeforeServerMessage(userPrompt, sawFeatureImpl, call) {
-    if (!userAskedForFeatureApp(userPrompt))
-        return undefined;
-    if (sawFeatureImpl)
-        return undefined;
-    if (!isDevServerCall(call) && !(call.name === "shell.start"))
-        return undefined;
-    return (`Feature incomplete: the user asked for a product app (not just a blank scaffold). ` +
-        `Implement the requested feature first (real UI/state or API — replace starter boilerplate), ` +
-        `THEN start the server. Do not shell.start / npm run dev until the feature is written.`);
-}
 /**
  * User said work belongs on the Desktop (or similar). Returns absolute path.
  */