@pensar/apex 2.6.1 → 2.7.0-canary.21cb83dc

package/build/lib/{chunk-z96pbjhd.js → chunk-m4swxy5r.js}

@@ -1,6 +1,6 @@
 import {
   TargetedPentestAgent2
-} from "./chunk-1jvwqnem.js";
+} from "./chunk-gm3b9j4y.js";
 
 // src/core/runtime/agentExecution.ts
 function startAgentExecution(agent) {
@@ -13,18 +13,25 @@ function startAgentExecution(agent) {
 }
 
 // src/core/api/targetedPentest.ts
-function normalizeResult(result) {
+function normalizeResult(result, findingPolicy) {
   const r = result;
   if (!r || !Array.isArray(r.findings)) {
     console.error("runTargetedPentestAgent: consume() returned unexpected shape", { keys: result ? Object.keys(result) : null });
   }
   const findings = Array.isArray(r?.findings) ? r.findings : [];
-  console.log(`
+  const reportedFindings = Array.isArray(r?.reportedFindings) ? r.reportedFindings : [];
+  if (findingPolicy === "external") {
+    console.log(`
+Reported ${reportedFindings.length} finding(s) for external evaluation`);
+  } else {
+    console.log(`
 Found ${findings.length} vulnerabilities`);
-  console.log(`Findings: ${r?.findingsPath}`);
-  console.log(`POCs: ${r?.pocsPath}`);
+    console.log(`Findings: ${r?.findingsPath}`);
+    console.log(`POCs: ${r?.pocsPath}`);
+  }
   return {
     findings,
+    ...findingPolicy === "external" ? { reportedFindings } : {},
     findingsPath: r?.findingsPath,
     pocsPath: r?.pocsPath,
     objectiveResults: r?.objectiveResults,
@@ -41,7 +48,7 @@ function startTargetedPentestAgent2(input) {
   }
   const { result, abortAndDrain, drained } = startAgentExecution(agent);
   return {
-    result: result.then(normalizeResult),
+    result: result.then((value) => normalizeResult(value, input.findingPolicy)),
     abortAndDrain,
     drained
   };

package/build/lib/{chunk-xhdpgpkm.js → chunk-mbsde9d0.js}

@@ -1,7 +1,7 @@
 import {
   createLogger,
   scopedLogger
-} from "./chunk-ba41jh41.js";
+} from "./chunk-kcw4cpr3.js";
 import {
   GOOGLE_SIGNIN_PROMPT_GUIDANCE,
   browserEngineForGoogleSignIn
@@ -12,7 +12,7 @@ import {
 import {
   AgentRuntime,
   defineAgent
-} from "./chunk-acgbyf5h.js";
+} from "./chunk-ydxhkgjw.js";
 
 // src/core/agents/specialized/authenticationAgent/agent.ts
 import { existsSync, readFileSync } from "node:fs";

package/build/lib/{chunk-9yksqgzr.js → chunk-mpvjfvz0.js}

@@ -1,7 +1,7 @@
 import {
   getCurrentVersion,
   config2
-} from "./chunk-w66ce82a.js";
+} from "./chunk-8njjwwcw.js";
 import {
   newSessionId2,
   isSessionId2
@@ -34,12 +34,13 @@ import {
   getPensarConsoleUrl2,
   buildAuthConfig2,
   normalizeStepUsage2
-} from "./chunk-ba41jh41.js";
+} from "./chunk-kcw4cpr3.js";
 import {
   buildSessionWorkspaceSection2,
   buildBaseSystemPrompt2,
-  OffensiveSecurityAgent2
-} from "./chunk-r4j8ybn9.js";
+  OffensiveSecurityAgent2,
+  EXTERNAL_FINDING_POLICY_CAPABILITY2
+} from "./chunk-hys7nemr.js";
 import {
   ASK_USER_QUESTIONS_TOOL_NAME2,
   apiRequest,
@@ -62,32 +63,32 @@ import {
   createDomain2,
   ALL_TOOL_NAMES2,
   SKILL_TOOL_NAMES2
-} from "./chunk-2qsq76xq.js";
+} from "./chunk-qkz5sm0d.js";
 import {
   BlackboxAttackSurfaceAgent2
-} from "./chunk-1fy8azee.js";
+} from "./chunk-fasqtjq1.js";
 import {
   runWhiteboxAttackSurfaceWorkflow2
-} from "./chunk-z137x08b.js";
+} from "./chunk-1rqbhx4b.js";
 import {
   runAuthenticationAgent3
-} from "./chunk-xhdpgpkm.js";
+} from "./chunk-mbsde9d0.js";
 import {
   readExecutionMetrics,
   writeExecutionMetrics,
   runPentestWorkflow2
-} from "./chunk-v3dk9j7k.js";
+} from "./chunk-aq9eyxas.js";
 import {
   PatchingAgent2
-} from "./chunk-1yct6cra.js";
+} from "./chunk-3hjvdf67.js";
 import {
   startAgentExecution,
   runTargetedPentestAgent2
-} from "./chunk-z96pbjhd.js";
+} from "./chunk-m4swxy5r.js";
 import {
   AgentRuntime,
   defineAgent
-} from "./chunk-acgbyf5h.js";
+} from "./chunk-ydxhkgjw.js";
 import {
   __export
 } from "./chunk-n0ba7khr.js";
@@ -98,6 +99,7 @@ __export(exports_api, {
   CLOSED_DISPOSITIONS: () => CLOSED_DISPOSITIONS2,
   DEFAULT_NATIVE_ROLLOUT_CAPTURE_LIMITS: () => DEFAULT_NATIVE_ROLLOUT_CAPTURE_LIMITS2,
   ENDPOINT_TRANSPORTS: () => ENDPOINT_TRANSPORTS2,
+  EXTERNAL_FINDING_POLICY_CAPABILITY: () => EXTERNAL_FINDING_POLICY_CAPABILITY2,
   KNOWLEDGE_CATEGORIES: () => KNOWLEDGE_CATEGORIES2,
   KNOWLEDGE_STATUSES: () => KNOWLEDGE_STATUSES2,
   PENSAR_API_BASE_URL: () => PENSAR_API_BASE_URL2,
@@ -2237,7 +2239,12 @@ async function runPentestAgent2(input) {
       tokenUsage
     });
   }
-  const { findings, findingsPath, pocsPath, reportPath } = result;
+  const { findings, findingsPath, pocsPath, reportPath, reportedFindings } = result;
+  if ((input.hooks?.findingPolicy ?? input.findingPolicy) === "external") {
+    console.log(`
+Reported ${reportedFindings?.length ?? 0} finding(s) for external evaluation`);
+    return result;
+  }
   console.log(`
 Found ${findings.length} vulnerabilities`);
   console.log(`Findings: ${findingsPath}`);

package/build/lib/{chunk-2qsq76xq.js → chunk-qkz5sm0d.js}

@@ -44,7 +44,7 @@ import {
 import {
   getCurrentVersion,
   config2
-} from "./chunk-w66ce82a.js";
+} from "./chunk-8njjwwcw.js";
 import {
   CredentialManager,
   createLogger,
@@ -53,7 +53,7 @@ import {
   ensureValidToken,
   signGatewayRequest,
   generateObjectResponse2
-} from "./chunk-ba41jh41.js";
+} from "./chunk-kcw4cpr3.js";
 import {
   EndpointTransportEnum2,
   GrpcEndpointMetadataSchema2,
@@ -74,7 +74,7 @@ import {
 } from "./chunk-zrbqn62x.js";
 import {
   FindingsRegistry2
-} from "./chunk-1azdancq.js";
+} from "./chunk-0dnx79b1.js";
 
 // src/core/agents/offSecAgent/tools/addMemory.ts
 import { tool } from "ai";
@@ -3651,6 +3651,7 @@ IMPORTANT: Pass protectedEndpoints in authHints when you've discovered 401/403 e
         const result = await spawner.spawn({
           spec: { type: "authentication", target, authHints },
           runtime: {
+            findingPolicy: ctx.findingPolicy,
             session: ctx.session,
             model: ctx.model,
             authConfig: ctx.authConfig,
@@ -4286,6 +4287,7 @@ async function generateThreatModelForEndpoint(ctx, input) {
           excludeTools: ["document_endpoint", "document_app"]
         },
         runtime: {
+          findingPolicy: ctx.findingPolicy,
           session: ctx.session,
           model,
           authConfig: ctx.authConfig,
@@ -7710,34 +7712,86 @@ Each directory entry is suffixed with "/" for easy identification.`,
 // src/core/agents/offSecAgent/tools/listMemories.ts
 import { tool as tool35 } from "ai";
 import { z as z37 } from "zod";
+var PAGE_SIZE = 20;
+var PAGE_CHAR_BUDGET = 12000;
 var listMemoriesInputSchema = z37.object({
   category: z37.enum(MEMORY_CATEGORIES2).optional().describe('Filter by category: "app", "framework", or "general". ' + "Omit to list memories across all categories."),
   tag: z37.string().optional().describe("Optional tag to filter memories by (exact match)"),
+  cursor: z37.object({
+    createdAt: z37.string(),
+    category: z37.enum(MEMORY_CATEGORIES2),
+    id: z37.string()
+  }).optional().describe("Pass the previous nextCursor to continue a listing"),
   toolCallDescription: z37.string().describe("A concise, human-readable description of what this tool call is doing (e.g., 'Listing stored recon notes')")
 });
+function compareMemories(a, b) {
+  const byCreatedAt = Date.parse(b.createdAt) - Date.parse(a.createdAt);
+  if (byCreatedAt > 0 || byCreatedAt < 0)
+    return byCreatedAt;
+  return a.category.localeCompare(b.category) || a.id.localeCompare(b.id);
+}
+function previewSummary(memory) {
+  const title = memory.title.slice(0, 160);
+  const tags = memory.tags.slice(0, 8).map((tag) => tag.slice(0, 48));
+  const truncated = title.length < memory.title.length || tags.length < memory.tags.length || tags.some((tag, index) => tag.length < (memory.tags[index]?.length ?? tag.length));
+  return {
+    id: memory.id,
+    category: memory.category,
+    title,
+    tags,
+    createdAt: memory.createdAt,
+    ...truncated ? { truncated: true } : {}
+  };
+}
 function listMemories2(ctx) {
   return tool35({
     description: `List saved memories, optionally filtered by category and/or tag.
 
-Returns lightweight summaries (id, category, title, tags, createdAt) sorted by
-most recent first. Use the returned category + id with get_memory to retrieve
-full content.`,
+Returns up to 20 lightweight summaries (id, category, title, tags, createdAt)
+sorted by most recent first. Use nextCursor to continue until it is null;
+count is the current total matching memories. A fresh listing includes newer
+memories added ahead of a cursor. Titles and tags may be abbreviated in the
+listing. Use category + id with get_memory to retrieve full content.`,
     inputSchema: listMemoriesInputSchema,
-    execute: async ({ category, tag }, { toolCallId }) => {
+    execute: async ({ category, tag, cursor }, { toolCallId }) => {
       try {
         const memories = await listMemories3({ category, tag }, { sessionId: ctx.session.id, toolCallId });
+        const ordered = [...memories].sort(compareMemories);
+        let start = 0;
+        if (cursor) {
+          start = ordered.findIndex((memory) => compareMemories(memory, cursor) > 0);
+          if (start < 0)
+            start = ordered.length;
+        }
+        const page = [];
+        let pageChars = 0;
+        for (const memory of ordered.slice(start, start + PAGE_SIZE)) {
+          const summary = previewSummary(memory);
+          const chars = JSON.stringify(summary).length + 1;
+          if (page.length > 0 && pageChars + chars > PAGE_CHAR_BUDGET)
+            break;
+          page.push(summary);
+          pageChars += chars;
+        }
+        const last = page.at(-1);
         return {
           success: true,
           error: "",
-          memories,
-          count: memories.length
+          memories: page,
+          count: memories.length,
+          nextCursor: start + page.length < ordered.length && last ? {
+            createdAt: last.createdAt,
+            category: last.category,
+            id: last.id
+          } : null
         };
       } catch (err) {
         return {
           success: false,
           error: err instanceof Error ? err.message : String(err),
           memories: [],
-          count: 0
+          count: 0,
+          nextCursor: null
         };
       }
     }
@@ -8381,7 +8435,8 @@ var ReportErrorReasonSchema2 = z46.enum([
   "blocked",
   "other"
 ]);
-function createReportErrorTool2(onError) {
+function createReportErrorTool2(onError, findingPolicy = "native") {
+  const vulnerabilityGuidance = findingPolicy === "external" ? "Do NOT use this for vulnerabilities (include them in your final response)" : "Do NOT use this for vulnerabilities (use document_vulnerability)";
   return tool44({
     description: `Report a blocking error that prevents you from completing the pentest, and END your run.
 
@@ -8390,7 +8445,7 @@ Call this ONLY when you genuinely cannot continue, for example:
 - The target is unreachable or responds in a way that makes testing impossible.
 - Some other runtime condition blocks you from carrying out the objectives.
 
-Do NOT use this for vulnerabilities (use document_vulnerability) or for normal completion (use response). Provide a clear, specific message describing exactly what blocked you — include the failing URL, status codes, or the auth step that failed where relevant — so it can be surfaced to the operator.`,
+${vulnerabilityGuidance} or for normal completion (use response). Provide a clear, specific message describing exactly what blocked you — include the failing URL, status codes, or the auth step that failed where relevant — so it can be surfaced to the operator.`,
     inputSchema: z46.object({
       reason: ReportErrorReasonSchema2.describe("Coarse category of the blocking error."),
       message: z46.string().min(1).describe("Clear, specific explanation of what blocked you (the failing URL, status codes, or auth step where relevant).")
@@ -8472,6 +8527,7 @@ should be passed directly to spawn_pentest_swarm for deep testing.`,
               attackSurfaceRegistry: ctx.attackSurfaceRegistry
             },
             runtime: {
+              findingPolicy: ctx.findingPolicy,
               session: ctx.session,
               model: ctx.model,
               authConfig: ctx.authConfig,
@@ -8524,6 +8580,7 @@ should be passed directly to spawn_pentest_swarm for deep testing.`,
             attackSurfaceRegistry: ctx.attackSurfaceRegistry
           },
           runtime: {
+            findingPolicy: ctx.findingPolicy,
             session: ctx.session,
             model: ctx.model,
             authConfig: ctx.authConfig,
@@ -8721,13 +8778,14 @@ ${output.stderr}` : "");
 import { tool as tool48 } from "ai";
 import { z as z50 } from "zod";
 function runPentestWorkflow3(ctx) {
+  const resultStep = ctx.findingPolicy === "external" ? "3. Return evidence-backed findings for the caller's evaluation" : "3. Aggregate findings and generate the final report";
   return tool48({
     description: `Run the full autonomous penetration test pipeline against a target.
 
 This single tool call executes the entire pentest workflow:
 1. Attack surface discovery (blackbox or whitebox)
 2. Spawn a swarm of targeted pentest agents
-3. Aggregate findings and generate the final report
+${resultStep}
 
 Use this when you want to run a complete, end-to-end pentest. The workflow
 handles session creation, concurrency, persistence, and reporting automatically.
@@ -8750,6 +8808,7 @@ Omit \`cwd\` for blackbox mode (live target probing only).`,
       const startedAt = Date.now();
       try {
         const result = await workflow({
+          findingPolicy: ctx.findingPolicy,
           target,
           cwd,
           model: ctx.model,
@@ -8762,6 +8821,7 @@ Omit \`cwd\` for blackbox mode (live target probing only).`,
           openAIReasoningEffort: ctx.openAIReasoningEffort,
           surfaceIntegrationEnabled: ctx.surfaceIntegrationEnabled,
           hooks: {
+            findingPolicy: ctx.findingPolicy,
             backends: ctx.backends,
             subagentSpawner: ctx.subagentSpawner,
             languageModelMiddleware: ctx.languageModelMiddleware,
@@ -8775,6 +8835,17 @@ Omit \`cwd\` for blackbox mode (live target probing only).`,
         const elapsedSeconds = Math.round((Date.now() - startedAt) / 1000);
         const minutes = Math.floor(elapsedSeconds / 60);
         const seconds = elapsedSeconds % 60;
+        if (ctx.findingPolicy === "external") {
+          const reportedFindings = result.reportedFindings ?? [];
+          return {
+            success: true,
+            totalFindings: reportedFindings.length,
+            reportedFindings,
+            objectiveResults: result.objectiveResults ?? [],
+            duration: `${minutes}m ${seconds}s`,
+            message: `Pentest complete. ${reportedFindings.length} finding(s) returned for external evaluation.`
+          };
+        }
         return {
           success: true,
           totalFindings: result.findings.length,
@@ -8821,13 +8892,14 @@ var ScanKindSchema = z51.enum([
   "fuzzing"
 ]);
 function runWhiteboxScan2(ctx) {
+  const verificationGuidance = ctx.findingPolicy === "external" ? "as recoverable conditions. Static results are candidates only; assess them against the caller's evidence requirements." : `as recoverable conditions. Static results are candidates only; verify before
+calling document_vulnerability.`;
   return tool49({
     description: `Run installed whitebox security scanners with bounded output.
 
 The tool detects applicable scanner adapters for the repository and normalizes
 results into concise summaries plus artifact paths. Missing tools are reported
-as recoverable conditions. Static results are candidates only; verify before
-calling document_vulnerability.`,
+${verificationGuidance}`,
     inputSchema: z51.object({
       path: z51.string().optional().describe("Repository root to scan; must stay under the configured codebase root."),
       kind: ScanKindSchema.optional().describe("Restrict scan family"),
@@ -9113,6 +9185,7 @@ async function runSingleCodingAgent(ctx, spawner, codebasePath, objective, _agen
   const result = await spawner.spawn({
     spec: { type: "code", codebasePath, objective },
     runtime: {
+      findingPolicy: ctx.findingPolicy,
       session: ctx.session,
       model: ctx.model,
       authConfig: ctx.authConfig,
@@ -9154,6 +9227,14 @@ import { z as z54 } from "zod";
 var WORKER_DRAIN_GRACE_MS = 30000;
 var ORIGIN_AUTH_FAILURE_GUIDANCE = `A 401/403 from one resource does not by itself prove authentication failure. Verify the session against a known protected or session endpoint and re-authenticate on the assigned origin only if needed. Call report_error with reason "authentication_failed" only if required authentication cannot be established and this blocks the assigned objective or all further testing; otherwise continue any reachable testing.`;
 function spawnPentestAgent2(ctx) {
+  const externalFindings = ctx.findingPolicy === "external";
+  const findingSteps = externalFindings ? `2. Reports concise evidence-backed findings in its response
+3. Leaves acceptance to the caller's external evaluator` : `2. Confirms vulnerabilities with POC scripts
+3. Documents findings to the shared findings registry only when the exploit path and material security impact are proven`;
+  const findingGuidance = externalFindings ? "False-positive discipline: report a security weakness only when observations support a material abuse path. Include the affected target, evidence, and impact; do not add a separate in-run finding review." : "False-positive discipline: workers should not call document_vulnerability for public/unauthenticated behavior, missing best-practice controls, generic errors, local HTTP deployment context, or other low-signal observations unless they prove a material abuse path such as non-public data access, credentialed cross-origin abuse, state change, denial of service, account takeover, or another concrete exploit chain.";
+  const returnGuidance = externalFindings ? "Returns the worker's summary, objective results, and reported findings so the parent can investigate chains." : "Returns the worker's summary, objective results, and finding count — but NOT the findings themselves (those land in the shared registry).";
+  const exploratoryGuidance = externalFindings ? "Spawn a follow-up exploratory worker only when the assigned objective requires broader coverage or a credible untested lead remains. For detection-only work, finish once the requested evidence is sufficient." : 'After fanning out one worker per objective, call this tool ONE MORE TIME for a final "exploratory" worker. Pass it the URLs of any suspicious behaviors discovered by earlier workers and ask it to chain findings together and probe for additional, unspecified vulnerabilities.';
+  const objectiveGuidance = externalFindings ? "Pass one focused objective per worker. A follow-up objective may cover a concrete chain or untested lead when the assignment calls for it." : "For per-objective fan-out, pass exactly one objective. For the final exploratory pass, pass a single composite objective describing chaining + exploration.";
   const targetScopeGuidance = ctx.systemScope && ctx.systemScope.memberHosts.length > 0 ? `For a cross-service follow-up, you may instead use a full URL on one of these authorized System Scope members: ${ctx.systemScope.memberHosts.join(", ")}.` : "No structured System Scope is present, so keep the assigned host unchanged.";
   return tool52({
     description: `Spawn a single focused pentest sub-agent (worker) to test ONE objective at a time.
@@ -9166,22 +9247,21 @@ Authentication state is origin-specific. Before spawning a cross-service worker
 
 Each spawn produces an independent TargetedPentestAgent that:
 1. Tests the assigned objective(s) deeply (execute_command, http_request, browser tools)
-2. Confirms vulnerabilities with POC scripts
-3. Documents findings to the shared findings registry only when the exploit path and material security impact are proven
+${findingSteps}
 
 Cloned browser session — the worker gets its OWN isolated Chromium instance, but it is seeded at spawn time with a snapshot of your current cookies and localStorage. If you've already authenticated in your browser (or recon set up useful state), the worker starts already authenticated for the same origins. Critically, the worker's browser is a COPY: its navigations, form fills, \`localStorage.setItem\`, alerts, and any other browser-side mutations stay inside the worker's own browser and do NOT leak back to you or to sibling workers. If you want state to be visible to the next worker, set it up in YOUR browser before spawning.
 
 The worker has no further spawn capability — it cannot recursively spawn other pentest agents.
 
-After fanning out one worker per objective, call this tool ONE MORE TIME for a final "exploratory" worker. Pass it the URLs of any suspicious behaviors discovered by earlier workers and ask it to chain findings together and probe for additional, unspecified vulnerabilities.
+${exploratoryGuidance}
 
-False-positive discipline: workers should not call document_vulnerability for public/unauthenticated behavior, missing best-practice controls, generic errors, local HTTP deployment context, or other low-signal observations unless they prove a material abuse path such as non-public data access, credentialed cross-origin abuse, state change, denial of service, account takeover, or another concrete exploit chain.
+${findingGuidance}
 
-Returns the worker's summary, objective results, and finding count — but NOT the findings themselves (those land in the shared registry).`,
+${returnGuidance}`,
     inputSchema: z54.object({
       name: z54.string().describe("Short human-readable label for this sub-agent shown in the UI (e.g. 'SQLi /api/users', 'IDOR orders', 'Final chain & explore')."),
       target: z54.string().describe(`Full target URL INCLUDING the endpoint path you want this worker to focus on (e.g. 'https://example.com/api/users/{id}', 'https://example.com/login'). Use the orchestrator's assigned URL by default and never strip it to a bare domain. A recon-supported sibling endpoint is allowed when it belongs to the objective. ${targetScopeGuidance} Never invent a host or endpoint.`),
-      objectives: z54.array(z54.string()).min(1).describe("Testing objectives for this worker. For per-objective fan-out, pass exactly one objective. For the final exploratory pass, pass a single composite objective describing chaining + exploration."),
+      objectives: z54.array(z54.string()).min(1).describe(`Testing objectives for this worker. ${objectiveGuidance}`),
       context: z54.string().optional().describe("Additional grounding context for the worker — e.g. summaries of what earlier workers found, suspicious endpoints to chain, or application-specific guidance. Treated as guidance, not as code to read."),
       toolCallDescription: z54.string().describe("Concise, human-readable description of what this spawn is doing (shown in the UI).")
     }),
@@ -9217,7 +9297,7 @@ Returns the worker's summary, objective results, and finding count — but NOT t
       const spawner = ctx.subagentSpawner;
       let subagentId = "";
       try {
-        const { findings, objectiveResults } = await spawner.spawn({
+        const { findings, reportedFindings, objectiveResults } = await spawner.spawn({
           spec: {
             type: "pentest",
             target,
@@ -9230,6 +9310,7 @@ Returns the worker's summary, objective results, and finding count — but NOT t
             browserSession: sharedBrowserSession
           },
           runtime: {
+            findingPolicy: ctx.findingPolicy,
             session: ctx.session,
             model: ctx.model,
             authConfig: ctx.authConfig,
@@ -9262,9 +9343,10 @@ Returns the worker's summary, objective results, and finding count — but NOT t
         return {
           success: true,
           subagentId,
-          findingsCount: findings.length,
+          findingsCount: externalFindings ? reportedFindings?.length ?? 0 : findings.length,
+          ...externalFindings ? { reportedFindings: reportedFindings ?? [] } : {},
           objectiveResults,
-          message: `Worker '${name}' complete. Documented ${findings.length} finding(s).`
+          message: externalFindings ? `Worker '${name}' complete. Reported ${reportedFindings?.length ?? 0} finding(s) for external evaluation.` : `Worker '${name}' complete. Documented ${findings.length} finding(s).`
         };
       } catch (error) {
         const message = error instanceof Error ? error.message : String(error);
@@ -9302,15 +9384,19 @@ import { tool as tool53 } from "ai";
 import { z as z55 } from "zod";
 var log7 = scopedLogger(() => createLogger("spawn-pentest-swarm"));
 function spawnPentestSwarm2(ctx) {
+  const externalFindings = ctx.findingPolicy === "external";
+  const findingSteps = externalFindings ? `2. Surface concise evidence-backed findings in its response
+3. Leave finding acceptance to the caller's evaluator` : `2. Create POC scripts for confirmed vulnerabilities
+3. Document findings with evidence and remediation`;
+  const findingDestination = externalFindings ? "Reported findings are returned to this caller; no native finding review runs." : "All findings are written to the shared session directory.";
   return tool53({
     description: `Launch a swarm of targeted pentest sub-agents to test multiple targets in parallel.
 
 Call this AFTER completing attack surface discovery. Each entry in 'targets' spawns an independent TargetedPentestAgent that will:
 1. Test the specified target for the given objectives
-2. Create POC scripts for confirmed vulnerabilities
-3. Document findings with evidence and remediation
+${findingSteps}
 
-All findings are written to the shared session directory.
+${findingDestination}
 
 Pass every target you want tested — the swarm handles concurrency automatically.`,
     inputSchema: z55.object({
@@ -9339,6 +9425,7 @@ Pass every target you want tested — the swarm handles concurrency automaticall
       const total = targets.length;
       log7.info(`Spawning pentest swarm: ${total} agents (concurrency: ${DEFAULT_CONCURRENCY})`);
       const swarmResults = await runPentestSwarm({
+        findingPolicy: ctx.findingPolicy,
         targets,
         model: ctx.model,
         session: ctx.session,
@@ -9354,6 +9441,7 @@ Pass every target you want tested — the swarm handles concurrency automaticall
         openAIReasoningEffort: ctx.openAIReasoningEffort,
         display: ctx.display,
         hooks: {
+          findingPolicy: ctx.findingPolicy,
           backends: ctx.backends,
           subagentSpawner: ctx.subagentSpawner,
           languageModelMiddleware: ctx.languageModelMiddleware,
@@ -9364,7 +9452,7 @@ Pass every target you want tested — the swarm handles concurrency automaticall
           sandbox: ctx.sandbox
         }
       });
-      const totalFindings = swarmResults.reduce((sum, r) => sum + (r?.findings.length ?? 0), 0);
+      const totalFindings = swarmResults.reduce((sum, r) => sum + (externalFindings ? r?.reportedFindings?.length ?? 0 : r?.findings.length ?? 0), 0);
       const failedCount = swarmResults.filter((r) => r == null).length;
       return {
         success: true,
@@ -9373,7 +9461,11 @@ Pass every target you want tested — the swarm handles concurrency automaticall
         failedTargets: failedCount,
         results: targets.map((t, i) => ({
           target: t.target,
-          findingsCount: swarmResults[i]?.findings.length ?? 0,
+          findingsCount: externalFindings ? swarmResults[i]?.reportedFindings?.length ?? 0 : swarmResults[i]?.findings.length ?? 0,
+          ...externalFindings ? {
+            reportedFindings: swarmResults[i]?.reportedFindings ?? [],
+            objectiveResults: swarmResults[i]?.objectiveResults ?? []
+          } : {},
           error: swarmResults[i] == null ? "Agent failed" : undefined
         })),
         message: `Swarm complete. ${totalFindings} findings across ${total} targets. ${failedCount > 0 ? `${failedCount} targets failed.` : "All targets tested successfully."}`
@@ -9623,12 +9715,13 @@ var updateTaskInputSchema = z59.object({
   observation: z59.string().optional().describe("Intermediate observations or notes from testing")
 });
 function updateTask2(ctx) {
+  const completedGuidance = ctx.findingPolicy === "external" ? "vulnerability identified with supporting evidence, OR conclusively determined the endpoint is not vulnerable / materiality is not met for this technique." : "vulnerability found and documented via document_vulnerability, OR conclusively determined the endpoint is not vulnerable / materiality is not met for this technique. Documentation is NOT required to complete a non-vulnerable task.";
   return tool57({
     description: `Update a task's status and record results.
 
 Status transitions:
 - "in_progress" — Set when you START testing a task. Pick one pending task at a time.
-- "completed" — Set when testing is done: vulnerability found and documented via document_vulnerability, OR conclusively determined the endpoint is not vulnerable / materiality is not met for this technique. Documentation is NOT required to complete a non-vulnerable task.
+- "completed" — Set when testing is done: ${completedGuidance}
 - "failed" — Set when the technique didn't work: rate limited, blocked by WAF, dead end, or approach not applicable. Then create a NEW task with an alternative technique.
 
 Always include a result or observation explaining the outcome.`,
@@ -9977,12 +10070,13 @@ var CandidateStateSchema = z62.enum([
   "deferred"
 ]);
 function createWhiteboxCandidate2(ctx) {
+  const candidateGuidance = ctx.findingPolicy === "external" ? "for the caller's final evaluation. Candidates keep hypotheses, source-to-sink traces, scanner artifacts, fuzzing logs, and verification status separate from reported findings." : `for document_vulnerability. Candidates keep hypotheses, source-to-sink traces,
+scanner artifacts, fuzzing logs, and verification status out of official findings.`;
   return tool60({
     description: `Create a whitebox vulnerability candidate.
 
 Use this for plausible source-code findings that are not yet confirmed enough
-for document_vulnerability. Candidates keep hypotheses, source-to-sink traces,
-scanner artifacts, fuzzing logs, and verification status out of official findings.`,
+${candidateGuidance}`,
     inputSchema: z62.object({
       title: z62.string(),
       vulnerabilityClass: z62.string(),
@@ -10009,7 +10103,7 @@ scanner artifacts, fuzzing logs, and verification status out of official finding
         artifactPaths: candidate.artifacts.map((artifact) => artifact.path),
         nextActions: [
           "Move candidate to investigating after reading source context.",
-          "Attempt a reproducer or dynamic verification before document_vulnerability."
+          ctx.findingPolicy === "external" ? "Gather the evidence required by the caller before reporting this candidate." : "Attempt a reproducer or dynamic verification before document_vulnerability."
         ],
         truncated: false
       };
@@ -10053,7 +10147,7 @@ State machine: hypothesis -> investigating -> repro_attempted -> confirmed (requ
           data: { candidate },
           artifactPaths: candidate.artifacts.map((artifact) => artifact.path),
           nextActions: candidate.state === "confirmed" ? [
-            "Use document_vulnerability only if the proof demonstrates exploitability."
+            ctx.findingPolicy === "external" ? "Report only the evidence-backed conclusion to the caller." : "Use document_vulnerability only if the proof demonstrates exploitability."
           ] : [
             "Continue tracing or verification based on the candidate state."
           ],

package/build/lib/{chunk-nhfax4gb.js → chunk-rfyjqybp.js}

@@ -1,7 +1,7 @@
 import {
   createLogger,
   scopedLogger
-} from "./chunk-ba41jh41.js";
+} from "./chunk-kcw4cpr3.js";
 import {
   __export
 } from "./chunk-n0ba7khr.js";

package/build/lib/{chunk-mg03jdmt.js → chunk-vsnwzne5.js}

@@ -7,7 +7,7 @@ import {
   PATCHING_ACTIVE_TOOLS2,
   readAgentsMd2,
   PatchingAgent2
-} from "./chunk-1yct6cra.js";
+} from "./chunk-3hjvdf67.js";
 import {
   __export
 } from "./chunk-n0ba7khr.js";

package/build/lib/{chunk-acgbyf5h.js → chunk-ydxhkgjw.js}

@@ -1,10 +1,11 @@
 import {
   OffensiveSecurityAgent2
-} from "./chunk-r4j8ybn9.js";
+} from "./chunk-hys7nemr.js";
 
 // src/core/agents/agentRuntime.ts
 function assembleAgentHooks(opts) {
   return {
+    findingPolicy: opts.findingPolicy,
     backends: opts.backends,
     subagentSpawner: opts.subagentSpawner,
     languageModelMiddleware: opts.languageModelMiddleware,
@@ -39,7 +40,7 @@ class AgentRuntime extends OffensiveSecurityAgent2 {
       responseSchema: definition.responseSchema?.(opts, state),
       stopWhen: definition.stopWhen?.(opts, state) ?? opts.stopWhen,
       extraTools: mergeExtraTools(opts.extraTools, definition.extraTools?.(opts, state)),
-      resolveResult: resolveResult ? (streamResult) => resolveResult(opts, state, streamResult) : undefined,
+      resolveResult: resolveResult ? (streamResult, validatedResponse) => resolveResult(opts, state, streamResult, validatedResponse) : undefined,
       target: definition.target?.(opts, state) ?? opts.target,
       grpc: definition.grpc?.(opts, state) ?? opts.grpc,
       browserEngine: definition.browserEngine?.(opts, state) ?? opts.browserEngine,