@kirrosh/zond 0.23.0 → 0.26.1

package/src/core/probe/security/orchestrator.ts

@@ -0,0 +1,398 @@
+/**
+ * `zond probe-security <classes>` — live SSRF / CRLF / open-redirect probes.
+ *
+ * Mirrors the `probe-mass-assignment` shape: live runner, optional regression
+ * YAML emission, idempotent cleanup. Where mass-assignment injects extra
+ * suspect fields, this probe replaces a single benign field with a security
+ * payload (SSRF / CRLF / open-redirect) and classifies the response.
+ */
+import type { EndpointInfo, SecuritySchemeInfo } from "../../generator/types.ts";
+import { executeRequest } from "../../runner/http-client.ts";
+import {
+  classifyPostSemantics,
+  findDeleteCounterpart,
+  pathTouchesSeededVar,
+} from "../shared.ts";
+import {
+  buildBaselineFromSpec,
+  buildBodyAuthHeaders,
+  buildProbeUrl,
+  hasProbeBody,
+  serializeProbeBody,
+} from "../probe-harness.ts";
+import { detectFields, PAYLOADS } from "./detectors.ts";
+import {
+  restoreOriginal,
+  sendBaseline,
+  snapshotOriginal,
+} from "./baseline.ts";
+import { classify } from "./classify.ts";
+import { tryCleanup } from "./cleanup.ts";
+import type {
+  CleanupFeasibility,
+  ProbeStepOpts,
+  SecurityFieldHit,
+  SecurityProbeOptions,
+  SecurityProbeResult,
+  SecuritySeverity,
+  SecurityVerdict,
+} from "./types.ts";
+
+export async function runSecurityProbes(
+  opts: SecurityProbeOptions,
+): Promise<SecurityProbeResult> {
+  const verdicts: SecurityVerdict[] = [];
+  const warnings: string[] = [];
+  let totalEndpoints = 0;
+
+  // ARV-140: pre-flight cleanup-feasibility scan. For each POST target, look
+  // up the DELETE counterpart in the spec once. Without --allow-leaks any
+  // attack against a POST-without-DELETE is dropped — orphan tracker can't
+  // clean it (no DELETE path to retry) so it would linger in the user's
+  // tenant indefinitely (feedback round-01/02 Sentry: 18 manual cleanups).
+  const feasibility: CleanupFeasibility = {
+    status: {},
+    skippedNoCleanup: 0,
+    forcedNoCleanup: 0,
+    actionNoCleanupNeeded: 0,
+  };
+  for (const ep of opts.endpoints) {
+    if (ep.deprecated) continue;
+    if (ep.method.toUpperCase() !== "POST") continue;
+    const key = `POST ${ep.path}`;
+    // ARV-153: action POSTs (`/capture`, `/verify`, `/cancel`, …) don't
+    // allocate a new resource — there is nothing to DELETE. Attacking them
+    // without `--allow-leaks` is safe; classifying them up front prevents
+    // the feasibility pre-flight from masking 18/22 Stripe action endpoints.
+    const semantics = classifyPostSemantics(ep);
+    if (semantics === "action") {
+      feasibility.status[key] = "action";
+      feasibility.actionNoCleanupNeeded += 1;
+      continue;
+    }
+    const hasDelete = findDeleteCounterpart(ep, opts.endpoints) !== undefined;
+    feasibility.status[key] = hasDelete ? "has-delete" : "no-delete-counterpart";
+    if (!hasDelete) {
+      if (opts.allowLeaks) feasibility.forcedNoCleanup += 1;
+      else feasibility.skippedNoCleanup += 1;
+    }
+  }
+
+  for (const ep of opts.endpoints) {
+    if (ep.deprecated) continue;
+    const m = ep.method.toUpperCase();
+    if (m !== "POST" && m !== "PUT" && m !== "PATCH") continue;
+    totalEndpoints++;
+
+    // ARV-140: cleanup-feasibility gate. POST without a DELETE counterpart
+    // (and without --allow-leaks) is dropped before any live request fires.
+    // PUT/PATCH have snapshot/restore so they're unaffected here.
+    if (m === "POST" && !opts.allowLeaks) {
+      const status = feasibility.status[`POST ${ep.path}`];
+      if (status === "no-delete-counterpart") {
+        verdicts.push(skipped(ep, "skipped: no DELETE counterpart in spec (cleanup-feasibility pre-flight; pass --allow-leaks to override)"));
+        continue;
+      }
+    }
+
+    // TASK-264: --isolated guard. Mutation on a seeded fixture would corrupt
+    // user data the next `zond run` depends on; skip the endpoint instead.
+    if (opts.isolated && (m === "PUT" || m === "PATCH") && pathTouchesSeededVar(ep.path, opts.vars)) {
+      verdicts.push(skipped(ep, "skipped: --isolated mode protects seeded fixtures (PUT/PATCH on seeded path-params)"));
+      continue;
+    }
+
+    // ARV-161 (round-08 F18): parity with mass-assignment — accept
+    // application/x-www-form-urlencoded endpoints too. Stripe v1 declares
+    // user-controlled URL fields (webhook url, return_url, ...) only on
+    // form-encoded bodies; the previous JSON-only gate hid 78+ POSTs from
+    // SSRF/CRLF/open-redirect probing.
+    if (!hasProbeBody(ep)) {
+      verdicts.push(skipped(ep, "no JSON or form-urlencoded request body"));
+      continue;
+    }
+
+    const detected = detectFields(ep, opts.classes);
+    if (detected.length === 0) {
+      verdicts.push(skipped(ep, `no fields matched classes: ${opts.classes.join(",")}`));
+      continue;
+    }
+
+    if (opts.dryRun) {
+      verdicts.push({
+        method: m,
+        path: ep.path,
+        severity: "skipped",
+        summary: "dry-run: would attack " + detected.map(d => `${d.field}/${d.class}`).join(", "),
+        detectedFields: detected,
+        findings: [],
+        skipReason: "dry-run",
+      });
+      continue;
+    }
+
+    const verdict = await probeOneEndpoint(
+      ep,
+      opts.endpoints,
+      opts.securitySchemes,
+      opts.vars,
+      detected,
+      {
+        noCleanup: opts.noCleanup === true,
+        timeoutMs: opts.timeoutMs,
+        cleanupRetryDelaysMs: opts.cleanupRetryDelaysMs,
+        seedBody: opts.seedBodies?.get(`${ep.method.toUpperCase()} ${ep.path}`),
+      },
+    );
+    verdicts.push(verdict);
+  }
+
+  return {
+    classes: opts.classes,
+    totalEndpoints,
+    specProbed: verdicts.length,
+    verdicts,
+    warnings,
+    cleanupFeasibility: feasibility,
+  };
+}
+
+async function probeOneEndpoint(
+  ep: EndpointInfo,
+  allEndpoints: EndpointInfo[],
+  schemes: SecuritySchemeInfo[],
+  vars: Record<string, string>,
+  detected: SecurityFieldHit[],
+  opts: ProbeStepOpts,
+): Promise<SecurityVerdict> {
+  const m = ep.method.toUpperCase();
+  const verdict: SecurityVerdict = {
+    method: m,
+    path: ep.path,
+    severity: "ok",
+    summary: "",
+    detectedFields: detected,
+    findings: [],
+  };
+
+  // Build baseline body. Same recipe as mass-assignment: spec → generators → vars.
+  // ARV-269: agent-authored `seed_body` overlay wins when present.
+  const baseline = buildBaselineFromSpec(ep, vars, opts.seedBody);
+  if (baseline === null) {
+    return skipped(ep, "request body not a JSON object");
+  }
+
+  const { url, unresolved } = buildProbeUrl(ep, vars);
+  if (unresolved.length > 0) {
+    return skipped(ep, `cannot resolve path placeholders: ${unresolved.join(", ")}`);
+  }
+
+  // ARV-161: Content-Type follows the spec — form-urlencoded for Stripe v1,
+  // JSON otherwise. All outbound payloads in this function (baseline, per-
+  // attack, restore-PUT) flow through serializeProbeBody for matching wire
+  // encoding.
+  const headers = buildBodyAuthHeaders(ep, schemes, vars);
+
+  // ── Snapshot original state (TASK-151) ────────────────────────────────
+  // For PUT/PATCH we MUST capture original state before any mutation. The
+  // old DELETE-cleanup is wrong for rename'ы — it can't undo a renamed
+  // DSN-key / team-name / webhook URL. Snapshot first, restore after each
+  // 2xx. POST falls back to DELETE-cleanup (correct semantics there).
+  const isUpdate = m === "PUT" || m === "PATCH";
+  const snapshot = isUpdate && !opts.noCleanup
+    ? await snapshotOriginal(ep, allEndpoints, schemes, vars, opts)
+    : null;
+
+  // ── Baseline-OK gate ────────────────────────────────────────────────────
+  // Eliminates the "5 × 404" output the markdown template produced in the
+  // audit. If baseline isn't 2xx, attacks would just hit the same 4xx
+  // wall and tell us nothing.
+  const fullBaseline = await sendBaseline(ep, m, url, headers, baseline, opts);
+  if (fullBaseline.kind === "network") {
+    verdict.severity = "high";
+    verdict.summary = `baseline network error: ${fullBaseline.reason}`;
+    return verdict;
+  }
+  verdict.baseline = { status: fullBaseline.status };
+
+  // ── Partial-body fallback (TASK-152) ──────────────────────────────────
+  // common SaaS-style APIs accept partial PUT — full bodies
+  // generated from spec get rejected (422 / 400). Walking each detected
+  // field with a single-key body recovers the proven-HIGH cases that
+  // otherwise fall into INCONCLUSIVE-BASELINE.
+  let fullOk = fullBaseline.kind === "ok" && fullBaseline.status >= 200 && fullBaseline.status < 300;
+  const perFieldBaseline = new Map<string, Record<string, unknown>>();
+  if (!fullOk && isUpdate && fullBaseline.kind === "ok") {
+    for (const hit of detected) {
+      // Reuse spec value when present; otherwise fall back to the substituted
+      // generator output for the field. Either way the partial body has
+      // exactly one key, which is what partial-PUT APIs accept.
+      const partial: Record<string, unknown> = {};
+      if (hit.field in baseline) partial[hit.field] = baseline[hit.field];
+      else partial[hit.field] = "";
+      const partResp = await sendBaseline(ep, m, url, headers, partial, opts);
+      if (partResp.kind === "ok" && partResp.status >= 200 && partResp.status < 300) {
+        perFieldBaseline.set(hit.field, partial);
+      }
+    }
+    if (perFieldBaseline.size > 0 && snapshot) {
+      // Each successful partial baseline mutated only its single key; restore
+      // exactly those before attacks start.
+      await restoreOriginal(
+        ep, snapshot, headers, schemes, vars, opts, verdict,
+        perFieldBaseline.keys(),
+      );
+    }
+  }
+
+  if (!fullOk && perFieldBaseline.size === 0) {
+    // fullBaseline.kind === "network" was already returned above; here it
+    // must be "ok" with non-2xx status.
+    const status = fullBaseline.kind === "ok" ? fullBaseline.status : 0;
+    verdict.severity = "inconclusive-baseline";
+    verdict.summary = isUpdate
+      ? `baseline ${status} on full body; partial-body per-field also rejected — fixture/scope issue`
+      : `baseline ${status} — endpoint unreachable or fixture invalid; skipping attacks`;
+    return verdict;
+  }
+
+  // Cleanup state mutated by the (full) baseline, before issuing attacks.
+  // With a snapshot → restore PUT (full baseline mutated every key).
+  // Without snapshot → DELETE-counterpart (POST flow).
+  if (fullOk && fullBaseline.kind === "ok" && !opts.noCleanup) {
+    if (snapshot) {
+      await restoreOriginal(
+        ep, snapshot, headers, schemes, vars, opts, verdict,
+        Object.keys(baseline),
+      );
+    } else {
+      await tryCleanup(
+        ep, allEndpoints, schemes, vars,
+        fullBaseline.body, verdict, opts,
+      );
+    }
+  }
+
+  // ── Attacks ──────────────────────────────────────────────────────────────
+  for (const hit of detected) {
+    // Pick the body shape that this endpoint actually accepts.
+    let baseBody: Record<string, unknown> | undefined;
+    let mode: "full" | "partial" | "none" = "none";
+    if (fullOk) {
+      baseBody = baseline;
+      mode = "full";
+    } else if (perFieldBaseline.has(hit.field)) {
+      baseBody = perFieldBaseline.get(hit.field)!;
+      mode = "partial";
+    }
+
+    if (mode === "none" || !baseBody) {
+      // Field doesn't have a usable baseline body shape — record one
+      // INCONCLUSIVE per payload so the digest still exposes the field.
+      for (const payload of PAYLOADS[hit.class]) {
+        verdict.findings.push({
+          field: hit.field,
+          class: hit.class,
+          payload,
+          status: 0,
+          echoed: false,
+          severity: "inconclusive",
+          reason: "no baseline body shape accepted (full+partial both rejected)",
+        });
+      }
+      continue;
+    }
+
+    for (const payload of PAYLOADS[hit.class]) {
+      const body = { ...baseBody, [hit.field]: payload };
+      let resp;
+      try {
+        resp = await executeRequest(
+          { method: m, url, headers, body: serializeProbeBody(ep, body).content },
+          { timeout: opts.timeoutMs ?? 30000, retries: 0 },
+        );
+      } catch (err) {
+        verdict.findings.push({
+          field: hit.field,
+          class: hit.class,
+          payload,
+          status: 0,
+          echoed: false,
+          severity: "inconclusive",
+          reason: `network error: ${err instanceof Error ? err.message : String(err)}`,
+        });
+        continue;
+      }
+      const finding = classify(hit, payload, resp, { endpoint: ep });
+      // Annotate which body shape was used for this attack — useful for
+      // case-studies and emit-tests.
+      finding.reason = mode === "partial"
+        ? `${finding.reason} [partial-body]`
+        : finding.reason;
+      verdict.findings.push(finding);
+
+      // Per-finding cleanup. Snapshot path takes precedence — DELETE on a
+      // PUT-rename'd resource would wipe a live entity, restore-PUT puts
+      // it back to the captured original. Only restore the single field
+      // this attack mutated — sending a multi-key body trips
+      // `422 use partial PUT` on common SaaS-shaped APIs.
+      if (resp.status >= 200 && resp.status < 300 && !opts.noCleanup) {
+        if (snapshot) {
+          await restoreOriginal(
+            ep, snapshot, headers, schemes, vars, opts, verdict,
+            [hit.field],
+          );
+        } else {
+          await tryCleanup(
+            ep, allEndpoints, schemes, vars,
+            resp.body_parsed ?? resp.body, verdict, opts,
+          );
+        }
+      }
+    }
+  }
+
+  verdict.severity = rollupSecuritySeverity(verdict.findings);
+  verdict.summary = summaryLine(verdict);
+  return verdict;
+}
+
+/**
+ * Roll up per-finding severities to the worst verdict-level severity.
+ * ARV-253: "info" sits below "low" (single_signal sanitization-only).
+ * ARV-254: "medium" sits between "high" and "low" (SSRF accept on an
+ * endpoint declaring delivery). Exported so callers can recompute the
+ * verdict rollup after adjusting individual finding severities.
+ */
+export function rollupSecuritySeverity(
+  findings: readonly { severity: SecuritySeverity }[],
+): SecuritySeverity {
+  const severities = findings.map((f) => f.severity);
+  if (severities.includes("high")) return "high";
+  if (severities.includes("inconclusive")) return "inconclusive";
+  if (severities.includes("medium")) return "medium";
+  if (severities.includes("low")) return "low";
+  if (severities.includes("info")) return "info";
+  return "ok";
+}
+
+function summaryLine(v: SecurityVerdict): string {
+  const counts: Record<SecuritySeverity, number> = {
+    high: 0, medium: 0, low: 0, info: 0, inconclusive: 0, "inconclusive-baseline": 0, ok: 0, skipped: 0,
+  };
+  for (const f of v.findings) counts[f.severity]++;
+  const fields = Array.from(new Set(v.detectedFields.map(d => d.field))).join(", ");
+  return `fields=[${fields}] · HIGH=${counts.high} MED=${counts.medium} LOW=${counts.low} INFO=${counts.info} INCONCLUSIVE=${counts.inconclusive} OK=${counts.ok}`;
+}
+
+function skipped(ep: EndpointInfo, reason: string): SecurityVerdict {
+  return {
+    method: ep.method.toUpperCase(),
+    path: ep.path,
+    severity: "skipped",
+    summary: `skipped: ${reason}`,
+    detectedFields: [],
+    findings: [],
+    skipReason: reason,
+  };
+}

package/src/core/probe/security/regression.ts

@@ -0,0 +1,103 @@
+import type { EndpointInfo, SecuritySchemeInfo } from "../../generator/types.ts";
+import { generateFromSchema } from "../../generator/data-factory.ts";
+import type { RawStep, RawSuite } from "../../generator/serializer.ts";
+import { flattenToFormFields } from "../../runner/form-encode.ts";
+import {
+  captureFieldFor,
+  convertPath,
+  endpointStem,
+  findDeleteCounterpart,
+  getAuthHeaders,
+} from "../shared.ts";
+import type { SecurityProbeResult } from "./types.ts";
+
+const ATTACK_EXPECTED_STATUS = [400, 403, 404, 405, 409, 415, 422];
+
+function shortPayload(s: string): string {
+  return s.length > 40 ? s.slice(0, 37) + "…" : s;
+}
+
+export function emitSecurityRegressionSuites(
+  result: SecurityProbeResult,
+  endpoints: EndpointInfo[],
+  schemes: SecuritySchemeInfo[],
+): RawSuite[] {
+  const suites: RawSuite[] = [];
+  for (const v of result.verdicts) {
+    // ARV-247 (R-04/F18): `high` findings are 2xx-with-echoed-payload — the
+    // strongest regression signal we have. Skipping them meant CI had nothing
+    // to gate on for confirmed stored injections. Treat them like `ok` here:
+    // the suite locks in the *expected* state (attack rejected) once the API
+    // owner ships a fix, and fails loud while it's still broken.
+    if (v.severity !== "ok" && v.severity !== "low" && v.severity !== "high") continue;
+    const ep = endpoints.find(
+      e => e.path === v.path && e.method.toUpperCase() === v.method,
+    );
+    if (!ep) continue;
+    const suiteHeaders = getAuthHeaders(ep, schemes);
+    const tests: RawStep[] = [];
+    for (const f of v.findings) {
+      // `ok` = attack already rejected; `high` = attack accepted+echoed (the
+      // regression target is rejection, same expected set as `ok`). `low` =
+      // attack accepted but no echo — lock in the 2xx-without-echo shape.
+      const expected = (f.severity === "ok" || f.severity === "high") ? ATTACK_EXPECTED_STATUS : [200, 201, 202, 204];
+      const body = ep.requestBodySchema ? generateFromSchema(ep.requestBodySchema) : {};
+      if (typeof body === "object" && body !== null && !Array.isArray(body)) {
+        (body as Record<string, unknown>)[f.field] = f.payload;
+      }
+      // ARV-161: mirror mass-assignment-probe — emit `form:` instead of
+      // `json:` for form-urlencoded endpoints (Stripe v1), otherwise the
+      // regression suite gets rejected with 400 "check that your POST
+      // content type is application/x-www-form-urlencoded" and never
+      // exercises the actual attack vector.
+      const bodyField =
+        ep.requestBodyContentType === "application/x-www-form-urlencoded"
+          ? { form: flattenToFormFields(body) }
+          : { json: body };
+      const step: RawStep = {
+        name: `${f.class}: ${f.field}=${shortPayload(f.payload)} must ${f.severity === "ok" ? "be rejected" : "not echo"}`,
+        source: {
+          generator: "probe-security",
+          endpoint: `${v.method} ${v.path}`,
+          response_branch: expected.map(String).join("|"),
+        },
+        [v.method]: convertPath(ep.path),
+        ...bodyField,
+        expect: { status: expected },
+      };
+      tests.push(step);
+    }
+    if (tests.length === 0) continue;
+    // Attach a generic cleanup step keyed off `created_id` (only fires when
+    // a previous step captured one — same `always:true` semantics other
+    // probes use).
+    const delEp = findDeleteCounterpart(ep, endpoints);
+    if (delEp) {
+      const idField = captureFieldFor(ep);
+      tests[0]!.expect.body = { ...(tests[0]!.expect.body ?? {}), [idField]: { capture: "created_id" } };
+      const idParam = (delEp.path.match(/\{([^}]+)\}/) ?? [])[1] ?? "id";
+      const delStep: RawStep = {
+        name: "cleanup",
+        source: { generator: "probe-security-cleanup", endpoint: `DELETE ${delEp.path}` },
+        always: true,
+        DELETE: convertPath(delEp.path).replace(`{{${idParam}}}`, "{{created_id}}"),
+        expect: { status: [200, 202, 204, 404] },
+      } as RawStep & { always: boolean };
+      tests.push(delStep);
+    }
+    suites.push({
+      name: `probe-security ${v.method} ${v.path}`,
+      tags: ["probe-security", ...result.classes],
+      source: {
+        type: "probe-suite",
+        generator: "probe-security",
+        endpoint: `${v.method} ${v.path}`,
+      },
+      fileStem: `probe-security-${endpointStem(ep)}`,
+      base_url: "{{base_url}}",
+      ...(suiteHeaders ? { headers: suiteHeaders } : {}),
+      tests,
+    });
+  }
+  return suites;
+}

package/src/core/probe/security/types.ts

@@ -0,0 +1,151 @@
+import type { EndpointInfo, SecuritySchemeInfo } from "../../generator/types.ts";
+import type { SeedBodyConfig } from "../../generator/resources-builder.ts";
+import type { RecommendedAction } from "../../diagnostics/failure-hints.ts";
+
+export type SecurityClass = "ssrf" | "crlf" | "open-redirect";
+
+export const SECURITY_CLASSES: SecurityClass[] = ["ssrf", "crlf", "open-redirect"];
+
+/**
+ * Security-probe severity ladder. Includes 'info' (ARV-253) for
+ * sanitization-only signals (CRLF accept-without-reflection) and
+ * 'medium' (ARV-254) for SSRF accept on endpoints declaring delivery
+ * semantics. The full m-21 matrix governs the cap: HIGH requires
+ * evidence_chain proof, OOB-backed SSRF lands here only when ARV-177
+ * lifts.
+ */
+export type SecuritySeverity =
+  | "high"
+  | "medium"
+  | "low"
+  | "info"
+  | "inconclusive"
+  | "inconclusive-baseline"
+  | "ok"
+  | "skipped";
+
+export interface SecurityFieldHit {
+  /** Field name in the request body. */
+  field: string;
+  /** Class that triggered (a field can hit multiple — we record all). */
+  class: SecurityClass;
+}
+
+export interface SecurityFinding {
+  field: string;
+  class: SecurityClass;
+  payload: string;
+  /** Raw HTTP status of the attack request. */
+  status: number;
+  /** Whether the response body echoes the payload (suggesting stored injection). */
+  echoed: boolean;
+  /** PASS / FAIL classification per finding. */
+  severity: SecuritySeverity;
+  reason: string;
+  /** TASK-294: agent-routable action. FAIL/WARN → `report_backend_bug`;
+   *  PASS → undefined (no action needed). */
+  recommended_action?: RecommendedAction;
+  /** Reserved suppression trace (audit-trail). Currently unset — the
+   *  agent triages severity from the raw finding. */
+  suppressed_by?: { source: string; rule_index: number; reason: string };
+}
+
+export interface SecurityVerdict {
+  method: string;
+  path: string;
+  /** Most-severe finding wins. */
+  severity: SecuritySeverity;
+  summary: string;
+  /** Field hits detected on this endpoint (some may have produced no findings). */
+  detectedFields: SecurityFieldHit[];
+  /** All attempted attacks. Empty for SKIPPED endpoints. */
+  findings: SecurityFinding[];
+  baseline?: { status: number };
+  cleanup?: {
+    attempted: boolean;
+    status?: number;
+    error?: string;
+    /** TASK-278: created resource id (slug/uuid/...) so `zond cleanup --orphans`
+     *  can retry DELETE without re-running the probe. */
+    id?: string | number;
+    /** TASK-278: concrete DELETE URL path with the id substituted. */
+    deletePath?: string;
+  };
+  skipReason?: string;
+}
+
+export interface SecurityProbeOptions {
+  endpoints: EndpointInfo[];
+  securitySchemes: SecuritySchemeInfo[];
+  vars: Record<string, string>;
+  classes: SecurityClass[];
+  noCleanup?: boolean;
+  timeoutMs?: number;
+  /** When true, only print which endpoints/fields would be attacked. */
+  dryRun?: boolean;
+  /**
+   * DELETE-cleanup retry delays in ms (round-5: handles eventual
+   * consistency between write replica and read replica). Default
+   * `[200, 1000]` — two retries on 404, total worst-case ~1.2s. Tests
+   * pass `[]` to disable; ops can pass longer for laggier replicas.
+   */
+  cleanupRetryDelaysMs?: number[];
+  /** TASK-264: when true, refuse to attack PUT/PATCH/DELETE endpoints whose
+   *  path-params are filled from `.env.yaml` (a.k.a. seeded fixtures). The
+   *  trade-off: lower coverage (those endpoints get SKIPPED), but a
+   *  guaranteed «probe doesn't mutate fixtures the user spent time
+   *  bootstrapping» property. POST endpoints still run — they create their
+   *  own resources, so isolation is automatic, with cleanup falling back to
+   *  the existing DELETE-counterpart + orphan-tracker flow (TASK-278). */
+  isolated?: boolean;
+  /** ARV-140: opt-in to attacks that have no cleanup path (POSTs without a
+   *  DELETE counterpart). By default we now skip them — round-01/02 Sentry
+   *  runs left ~18 manually-cleanable orphans in prod because the probe
+   *  happily POSTed to `/teams/`, `/symbol-sources/`, etc., where the spec
+   *  has no DELETE. The pre-flight feasibility map drops these unless the
+   *  caller explicitly accepts the leak. */
+  allowLeaks?: boolean;
+  /** ARV-269: agent-authored `seed_body` overlays from `.api-resources.local.yaml`,
+   *  keyed by `"METHOD /path"`. Wins over `generateFromSchema` when the
+   *  endpoint matches — see `MassAssignmentOptions.seedBodies` for the
+   *  rationale (strict APIs reject random-scalar baselines). */
+  seedBodies?: Map<string, SeedBodyConfig>;
+}
+
+/** ARV-140: cleanup-feasibility map. Built once before the live loop so
+ *  every POST verdict can see whether the spec has a DELETE counterpart;
+ *  the summary digest also reports counts for skipped/forced endpoints.
+ *
+ *  ARV-153 extends the status enum with "action": POSTs whose last path
+ *  segment is a known action verb (`/capture`, `/verify`, `/cancel`, …)
+ *  operate on an existing resource and never allocate a new one, so a
+ *  DELETE counterpart isn't meaningful. These are attacked the same way
+ *  as POSTs with a real DELETE — without `--allow-leaks` — because there
+ *  is no resource to leak. */
+export interface CleanupFeasibility {
+  status: Record<string, "has-delete" | "no-delete-counterpart" | "action">;
+  skippedNoCleanup: number;
+  forcedNoCleanup: number;
+  /** ARV-153: POSTs we attacked even though no DELETE counterpart exists,
+   *  because the operation is semantically an action (no resource created). */
+  actionNoCleanupNeeded: number;
+}
+
+export interface SecurityProbeResult {
+  classes: SecurityClass[];
+  totalEndpoints: number;
+  specProbed: number;
+  verdicts: SecurityVerdict[];
+  warnings: string[];
+  /** ARV-140: cleanup-feasibility digest (POSTs without DELETE counterpart). */
+  cleanupFeasibility?: CleanupFeasibility;
+}
+
+/** Internal: per-step options shared between orchestrator/baseline/cleanup. */
+export interface ProbeStepOpts {
+  noCleanup: boolean;
+  timeoutMs?: number;
+  cleanupRetryDelaysMs?: number[];
+  /** ARV-269: optional `seed_body` overlay for this endpoint. */
+  seedBody?: SeedBodyConfig;
+}

package/src/core/probe/security-probe-class.ts

@@ -18,7 +18,8 @@ import {
   type SecurityProbeResult,
   type SecurityVerdict,
 } from "./security-probe.ts";
-import { hasJsonBody, pathTouchesSeededVar } from "./shared.ts";
+import { pathTouchesSeededVar } from "./shared.ts";
+import { hasProbeBody } from "./probe-harness.ts";
 
 const FLAGS: ProbeFlags = {
   api: true,
@@ -55,7 +56,12 @@ function planForSecurity(
       continue;
     }
 
-    if (!hasJsonBody(ep)) {
+    // ARV-313: parity with the live orchestrator (which uses hasProbeBody)
+    // and the mass-assignment planner — accept application/x-www-form-urlencoded
+    // bodies, not just JSON. The old hasJsonBody gate made the dry-run
+    // inventory a no-op on form-encoded APIs (Stripe v1), so `probe security
+    // --dry-run` planned 0/291 while mass-assignment planned 290/291.
+    if (!hasProbeBody(ep)) {
       out.push({
         path: ep.path,
         method: m,