@runpod/mcp-server 4.2.0 → 4.4.0

package/dist/http.mjs

@@ -1,5 +1,10 @@
 // src/http.ts
 import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";
+import {
+  CallToolRequestSchema as CallToolRequestSchema2,
+  InitializeRequestSchema,
+  JSONRPCMessageSchema
+} from "@modelcontextprotocol/sdk/types.js";
 
 // src/specgen/server.ts
 import { Server } from "@modelcontextprotocol/sdk/server/index.js";
@@ -87,11 +92,11 @@ function readParams(params) {
   }
   return { values, unreadable };
 }
-function parseExhaustedWindows(header) {
-  if (!header) return { windows: [], dropped: false };
+function parseExhaustedWindows(header2) {
+  if (!header2) return { windows: [], dropped: false };
   const windows = [];
   let dropped = false;
-  for (const item of splitOutsideStrings(header, ",")) {
+  for (const item of splitOutsideStrings(header2, ",")) {
     if (item.trim() === "") continue;
     const [rawName, ...params] = splitOutsideStrings(item, ";");
     if (!params.every(isWellFormedParam)) {
@@ -209,10 +214,13 @@ function validateArguments(inputSchema, args) {
 async function dispatchGeneratedTool(client, tool, args) {
   const rejected = validateArguments(tool.inputSchema, args);
   if (rejected) return rejected;
+  const required = new Set(tool.inputSchema.required ?? []);
   const pathParams = {};
   const query = {};
   for (const param of tool.params) {
     if (args[param.name] === void 0) continue;
+    if (param.location === "query" && !required.has(param.name) && isEmptyValue(args[param.name]))
+      continue;
     if (param.location === "path") {
       pathParams[param.name] = args[param.name];
       continue;
@@ -234,6 +242,11 @@ async function dispatchGeneratedTool(client, tool, args) {
       };
     }
   }
+  if (tool.hasBody && requestBody !== void 0) {
+    const properties = tool.inputSchema.properties ?? {};
+    const defs = tool.inputSchema.$defs ?? {};
+    requestBody = coerceScalars(requestBody, properties.body, defs);
+  }
   const verbs = client;
   const { data, error, response } = await verbs[tool.method](tool.path, {
     params: { path: pathParams, query },
@@ -246,6 +259,79 @@ async function dispatchGeneratedTool(client, tool, args) {
     payload: data ?? { status: response.status }
   };
 }
+function isEmptyValue(value) {
+  return value === null || value === "" || Array.isArray(value) && !value.length;
+}
+function asSchema(value) {
+  return value !== null && typeof value === "object" && !Array.isArray(value) ? value : void 0;
+}
+function resolveSchema(schema, defs, depth = 0) {
+  const node = asSchema(schema);
+  if (!node || depth > 16) return node;
+  const ref = node.$ref;
+  if (typeof ref === "string" && ref.startsWith("#/$defs/"))
+    return resolveSchema(defs[ref.slice("#/$defs/".length)], defs, depth + 1);
+  return node;
+}
+function branches(schema, defs, depth = 0) {
+  const node = resolveSchema(schema, defs);
+  if (!node || depth > 16) return [];
+  const out = [node];
+  for (const key of ["allOf", "anyOf", "oneOf"]) {
+    const list = node[key];
+    if (Array.isArray(list))
+      for (const sub of list) out.push(...branches(sub, defs, depth + 1));
+  }
+  return out;
+}
+function types(schemas) {
+  const out = /* @__PURE__ */ new Set();
+  for (const s of schemas) {
+    if (typeof s.type === "string") out.add(s.type);
+    else if (Array.isArray(s.type)) {
+      for (const t of s.type) if (typeof t === "string") out.add(t);
+    }
+  }
+  return out;
+}
+var NUMERIC = /^-?(?:0|[1-9]\d*)(?:\.\d+)?$/;
+function coerceScalars(value, schema, defs, depth = 0) {
+  if (depth > 32) return value;
+  const applicable = branches(schema, defs);
+  if (!applicable.length) return value;
+  const allowed = types(applicable);
+  if (typeof value === "string") {
+    if (allowed.has("string")) return value;
+    if ((allowed.has("integer") || allowed.has("number")) && NUMERIC.test(value)) {
+      const n = Number(value);
+      if (allowed.has("number") || Number.isInteger(n)) return n;
+    }
+    if (allowed.has("boolean") && (value === "true" || value === "false"))
+      return value === "true";
+    return value;
+  }
+  if (Array.isArray(value)) {
+    const items = applicable.map((s) => s.items).filter(Boolean);
+    if (!items.length) return value;
+    return value.map(
+      (item) => coerceScalars(item, { anyOf: items }, defs, depth + 1)
+    );
+  }
+  const object = asSchema(value);
+  if (!object) return value;
+  const out = {};
+  for (const [key, child] of Object.entries(object)) {
+    const childSchemas = [];
+    for (const s of applicable) {
+      const props = asSchema(s.properties);
+      if (props && key in props) childSchemas.push(props[key]);
+      else if (asSchema(s.additionalProperties))
+        childSchemas.push(s.additionalProperties);
+    }
+    out[key] = childSchemas.length ? coerceScalars(child, { anyOf: childSchemas }, defs, depth + 1) : child;
+  }
+  return out;
+}
 
 // src/specgen/generated/tools.gen.ts
 var generatedTools = [
@@ -263,6 +349,7 @@ var generatedTools = [
       idempotentHint: false,
       openWorldHint: true
     },
+    schemaOnDemand: true,
     inputSchema: {
       type: "object",
       properties: {
@@ -274,11 +361,10 @@ var generatedTools = [
       $defs: {
         BaseContainerConfig: {
           type: "object",
-          description: "Container configuration universal to every containerized resource. Compose ContainerConfig instead unless the resource cannot support a direct registry credential (clusters \u2014 there the registry credential arrives via a pod template, see CreateClusterRequest.templateId).\n",
           properties: {
             args: {
               type: "string",
-              description: 'The container\'s command, as a single raw string. This is the field `entrypoint` and `cmd` encode into, exposed in its stored form.\n\nTwo shapes are accepted. A bare shell string is treated as CMD and split into arguments, which is what the console\'s "Container start command" field writes. A JSON object of the form `{"entrypoint":[...],"cmd":[...]}` sets either or both explicitly.\n\nResponses always return both representations: `args` exactly as stored, plus the deconstructed `entrypoint` and `cmd`. Supplying `args` together with `entrypoint` or `cmd` is allowed only when they describe the same command, so a read-modify-write client can send back everything it received. Send `""` to clear, omit to leave unchanged.\n',
+              description: 'The container\'s command as one raw string: a bare shell string is treated as CMD and split into arguments; a JSON object `{"entrypoint":[...],"cmd":[...]}` sets either or both. Send `""` to clear; omit to leave unchanged.',
               examples: [
                 "--model meta-llama/Llama-3-8B --max-model-len 8192",
                 '{"entrypoint":["/bin/bash","-c"],"cmd":["python -u main.py"]}',
@@ -290,7 +376,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container CMD in exec form. When the image defines an ENTRYPOINT, this is the argument list passed to it. Encoded into the `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container CMD in exec form: when the image defines an ENTRYPOINT, the argument list passed to it. Encoded into `args`; supply both only when they describe the same command. Send `[]` to clear; omit to leave unchanged.",
               examples: [
                 ["--model", "meta-llama/Llama-3-8B", "--max-model-len", "8192"]
               ]
@@ -300,7 +386,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.",
               examples: [["/bin/bash", "-c"]]
             },
             disk: {
@@ -336,6 +422,21 @@ var generatedTools = [
             }
           }
         },
+        BaseMounts: {
+          type: "object",
+          properties: {
+            persistent: {
+              $ref: "#/$defs/PersistentMount"
+            },
+            network: {
+              type: "array",
+              maxItems: 1,
+              items: {
+                $ref: "#/$defs/NetworkMount"
+              }
+            }
+          }
+        },
         ClusterCompute: {
           type: "object",
           additionalProperties: false,
@@ -363,6 +464,15 @@ var generatedTools = [
             }
           }
         },
+        ClusterMounts: {
+          description: "Storage mounts for a cluster create: `persistent` or `network`, not both. Clusters cannot mount global volumes, so `global` is rejected.",
+          allOf: [
+            {
+              $ref: "#/$defs/BaseMounts"
+            }
+          ],
+          unevaluatedProperties: false
+        },
         ClusterType: {
           type: "string",
           description: "Cluster type. TRAINING is the generic distributed-training cluster; SLURM provisions a managed Slurm controller/compute topology; RAY provisions a managed Ray head/worker topology; APPLICATION is a general multi-node application cluster.",
@@ -377,7 +487,7 @@ var generatedTools = [
             {
               type: "object",
               required: ["name", "type", "compute"],
-              description: "Request body for creating a cluster. `compute` defines the\nhomogeneous pod shape; the container configuration (image, env, ports,\n\u2026) applies to every pod. Pass `templateId` to provision every member\npod from a pod template instead of inline container fields \u2014 the\ntemplate also supplies the container registry credential, the only\nprivate-image path for clusters (a bare `registry` property on this\nbody is rejected).\n",
+              description: "Request body for creating a cluster. `compute` defines the homogeneous pod shape; the container configuration (image, env, ports, \u2026) applies to every pod.",
               properties: {
                 compute: {
                   $ref: "#/$defs/ClusterCompute"
@@ -393,7 +503,7 @@ var generatedTools = [
                 templateId: {
                   type: "string",
                   minLength: 1,
-                  description: "ID of a pod template to provision every member pod from. The\ntemplate supplies the container settings (image, args, disk,\nenv, ports) and the container registry credential for private\nimages \u2014 the only private-image path for clusters. Mutually\nexclusive with `image`, `args`, `entrypoint`, `cmd`, `disk`,\n`env`, `ports`, and `mounts` (rejected with 400). The cluster\nretains the link: the `template` response field is set. Must be\na non-serverless pod template accessible to the caller.\n",
+                  description: "ID of a pod template to provision every member pod from: it supplies the container settings (image, args, disk, env, ports) and the registry credential for private images \u2014 the only private-image path for clusters. Mutually exclusive with the inline container fields (`image`, `args`, `entrypoint`, `cmd`, `disk`, `env`, `ports`, `mounts`). Must be a non-serverless template you can access.",
                   examples: ["30zmvf89kd"]
                 },
                 dataCenterIds: {
@@ -401,11 +511,11 @@ var generatedTools = [
                   items: {
                     type: "string"
                   },
-                  description: "Preferred data centers for placement. Omit or pass an empty\narray to let the scheduler choose. A cluster is always placed\nwithin a single data center.\n",
+                  description: "Preferred data centers for placement. Omit or pass an empty array to let the scheduler choose. A cluster is always placed within a single data center.",
                   example: ["US-TX-3"]
                 },
                 mounts: {
-                  $ref: "#/$defs/Mounts"
+                  $ref: "#/$defs/ClusterMounts"
                 },
                 startJupyter: {
                   type: "boolean",
@@ -422,28 +532,11 @@ var generatedTools = [
           ],
           unevaluatedProperties: false
         },
-        Mounts: {
-          type: "object",
-          additionalProperties: false,
-          description: "Storage mounts attached to a pod. At-most-one of `persistent` or\n`network` may be set today (mutually exclusive, enforced at the\nhandler with 400 if both are present). The `network` field is an\narray for forward compatibility with eventual multi-network-volume\nsupport, but `maxItems` is 1 today.\n\nPATCH semantics:\n- Omitting `mounts` or sending `{}` leaves the existing mount\n  unchanged.\n- An explicit `network: []` is rejected with 400 (clearing mounts\n  is not supported).\n- Mount kind is fixed at create \u2014 a PATCH that introduces a kind\n  not present at create (persistent on a network pod, network on\n  a persistent pod, or any mount on a previously-mountless pod)\n  is rejected with 400.\n- The `volumeId` of a network mount is immutable; a PATCH that\n  names a different `volumeId` is rejected with 400.\n- Partial mounts are not supported \u2014 every mount entry must\n  include the full schema (`size` + `path` for persistent,\n  `volumeId` + `path` for network). Missing required fields \u2192 422.\n",
-          properties: {
-            persistent: {
-              $ref: "#/$defs/PersistentMount"
-            },
-            network: {
-              type: "array",
-              maxItems: 1,
-              items: {
-                $ref: "#/$defs/NetworkMount"
-              }
-            }
-          }
-        },
         NetworkMount: {
           type: "object",
           required: ["volumeId", "path"],
           additionalProperties: false,
-          description: "Reference to a NetworkVolume. Custom paths are honored at runtime on\nboth GPU and CPU pods. The underlying `volumeId` is immutable\npost-create; the mount `path` may be changed via PATCH.\n",
+          description: "Reference to a NetworkVolume. Custom paths are honored at runtime on both GPU and CPU pods. The underlying `volumeId` is immutable post-create; the mount `path` may be changed via PATCH.",
           properties: {
             volumeId: {
               type: "string",
@@ -461,7 +554,7 @@ var generatedTools = [
           type: "object",
           required: ["size", "path"],
           additionalProperties: false,
-          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data\ndoes not survive a host failure. Disallowed on CPU pods. Mutually\nexclusive with NetworkMount. Deprecated: prefer NetworkMount for any\ndata you cannot recreate.\n",
+          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data does not survive a host failure. Disallowed on CPU pods. Mutually exclusive with NetworkMount. Deprecated: prefer NetworkMount for any data you cannot recreate.",
           properties: {
             size: {
               type: "integer",
@@ -538,6 +631,7 @@ var generatedTools = [
       idempotentHint: false,
       openWorldHint: true
     },
+    schemaOnDemand: true,
     inputSchema: {
       type: "object",
       properties: {
@@ -549,11 +643,10 @@ var generatedTools = [
       $defs: {
         BaseContainerConfig: {
           type: "object",
-          description: "Container configuration universal to every containerized resource. Compose ContainerConfig instead unless the resource cannot support a direct registry credential (clusters \u2014 there the registry credential arrives via a pod template, see CreateClusterRequest.templateId).\n",
           properties: {
             args: {
               type: "string",
-              description: 'The container\'s command, as a single raw string. This is the field `entrypoint` and `cmd` encode into, exposed in its stored form.\n\nTwo shapes are accepted. A bare shell string is treated as CMD and split into arguments, which is what the console\'s "Container start command" field writes. A JSON object of the form `{"entrypoint":[...],"cmd":[...]}` sets either or both explicitly.\n\nResponses always return both representations: `args` exactly as stored, plus the deconstructed `entrypoint` and `cmd`. Supplying `args` together with `entrypoint` or `cmd` is allowed only when they describe the same command, so a read-modify-write client can send back everything it received. Send `""` to clear, omit to leave unchanged.\n',
+              description: 'The container\'s command as one raw string: a bare shell string is treated as CMD and split into arguments; a JSON object `{"entrypoint":[...],"cmd":[...]}` sets either or both. Send `""` to clear; omit to leave unchanged.',
               examples: [
                 "--model meta-llama/Llama-3-8B --max-model-len 8192",
                 '{"entrypoint":["/bin/bash","-c"],"cmd":["python -u main.py"]}',
@@ -565,7 +658,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container CMD in exec form. When the image defines an ENTRYPOINT, this is the argument list passed to it. Encoded into the `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container CMD in exec form: when the image defines an ENTRYPOINT, the argument list passed to it. Encoded into `args`; supply both only when they describe the same command. Send `[]` to clear; omit to leave unchanged.",
               examples: [
                 ["--model", "meta-llama/Llama-3-8B", "--max-model-len", "8192"]
               ]
@@ -575,7 +668,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.",
               examples: [["/bin/bash", "-c"]]
             },
             disk: {
@@ -635,7 +728,7 @@ var generatedTools = [
             pools: {
               type: "array",
               minItems: 1,
-              description: 'Serverless GPU pool IDs (as returned by `GET /v2/catalog/gpus` in\n`pool`). Workers are placed on whichever listed pool has capacity.\nNarrow a pool down to specific cards with `excludedTypes`.\n\nOn `PATCH`, `pools` and `excludedTypes` are one selection and are\nreplaced together, so sending `pools` by itself **clears the\nexclusions**. Two cases:\n\n- **Changing pools, keeping exclusions** \u2014 send both fields in one\n  request: `{"gpu": {"pools": ["ADA_24"], "excludedTypes":\n  ["NVIDIA L40"]}}`. `GET` the endpoint first to read the current\n  `excludedTypes` and resend the ones that still apply to the new\n  pools; an exclusion naming a type outside `pools` is a 400.\n- **Changing only `count` or a CUDA constraint** \u2014 omit `pools`:\n  `{"gpu": {"minCudaVersion": "12.4"}}`. The pool list and the\n  exclusions are both left exactly as they are.\n\n`excludedTypes` documents the full rule.\n',
+              description: "Serverless GPU pool IDs (the `pool` field of `list-gpu-types`). Workers are placed on whichever listed pool has capacity; narrow to specific cards with `excludedTypes`. On update, `pools` and `excludedTypes` are replaced together, so sending `pools` alone clears the exclusions.",
               items: {
                 type: "string"
               },
@@ -644,7 +737,7 @@ var generatedTools = [
             excludedTypes: {
               type: "array",
               uniqueItems: true,
-              description: 'GPU **type** IDs to subtract from the selected pools \u2014 the `id`\nfield of `GET /v2/catalog/gpus`, the same identifiers pods take in\n`gpu.id`. Workers run on every type in `pools` except these. Omit to\nuse the whole pool.\n\nPools stay the unit of selection; types are the unit of\nsubtraction. There is no inclusive allowlist: a card later added to\none of your pools becomes eligible, which is the honest reading of\n"this pool, minus these".\n\nTied to `pools`, because the two together are one selection:\nsupplying `pools` replaces that selection wholesale, so a `PATCH`\nsending `pools` **without `excludedTypes`** **clears** them \u2014\nrestate them to keep them. A `PATCH` that omits `pools` leaves both\nthe pools and the exclusions untouched, so changing only a CUDA\nconstraint cannot widen a pinned endpoint.\n\nRejected with 400 if a value is not a GPU type in one of `pools`;\nupstream accepts unrecognized exclusions silently, so a typo would\notherwise produce a filter that does nothing. Surrounding whitespace\nis trimmed, so `" NVIDIA L40"` and `"NVIDIA L40"` mean the same card.\n',
+              description: "GPU type IDs to subtract from the selected pools \u2014 the `id` field of `list-gpu-types`, the same identifiers pods take in `gpu.id`. Workers run on every type in `pools` except these; omit to use the whole pool. A value outside `pools` is rejected. On update, restate them whenever you send `pools`, or they are cleared.",
               items: {
                 type: "string",
                 pattern: "^\\s*[^-\\s]"
@@ -661,7 +754,6 @@ var generatedTools = [
           }
         },
         ContainerConfig: {
-          description: "Reusable container configuration shared across templates, pods, and serverless endpoints. Adding a field here automatically propagates to all three resources.\n",
           allOf: [
             {
               $ref: "#/$defs/BaseContainerConfig"
@@ -687,7 +779,7 @@ var generatedTools = [
           unevaluatedProperties: false
         },
         CreateEndpointGpuConfig: {
-          description: "GPU request for an endpoint create. Carries the CUDA constraints, which\nlive here rather than at the body's top level so they are\nunrepresentable on a CPU endpoint.\n",
+          description: "GPU request for an endpoint create. Carries the CUDA constraints, which live here rather than at the body's top level so they are unrepresentable on a CPU endpoint.",
           allOf: [
             {
               $ref: "#/$defs/BaseEndpointGpuConfig"
@@ -702,13 +794,13 @@ var generatedTools = [
                     type: "string",
                     pattern: "^\\d+\\.\\d+$"
                   },
-                  description: "Acceptable CUDA versions for worker placement, as\n`major.minor`. Omit to accept any version (or inherit the\ntemplate's constraint when creating from `templateId`).\nMatching is exact \u2014 discover valid values per GPU type via\n`GET /v2/catalog/gpus?include=AVAILABILITY&product=SERVERLESS`\n(`cudaVersions`).\n\nA non-empty set is mutually exclusive with minCudaVersion (400\nif both are sent). An explicit `[]` states no constraint, so it\nmay accompany a floor.\n",
+                  description: "Acceptable CUDA versions for worker placement, as `major.minor`. Omit to accept any version (or inherit the template's constraint when creating from `templateId`).",
                   examples: [["12.8", "12.6"]]
                 },
                 minCudaVersion: {
                   type: "string",
                   pattern: "^\\d+\\.\\d+$",
-                  description: "Lowest acceptable CUDA version for worker placement, as\n`major.minor`, compared numerically rather than as a decimal \u2014\nso 12.11 is above 12.2. Use this for an open-ended floor and\nallowedCudaVersions for an exact set.\n\nMutually exclusive with a non-empty allowedCudaVersions (400 if\nboth are sent); an explicit `[]` there states no constraint and\nmay accompany this floor.\n",
+                  description: "Lowest acceptable CUDA version for worker placement, as `major.minor`, compared numerically rather than as a decimal \u2014 so 12.11 is above 12.2. Use this for an open-ended floor and allowedCudaVersions for an exact set.",
                   examples: ["12.1"]
                 }
               }
@@ -752,13 +844,13 @@ var generatedTools = [
                       $ref: "#/$defs/EndpointType"
                     }
                   ],
-                  description: "Request-routing model. Required \u2014 it determines the valid scaler\nand request URLs, so it must be chosen explicitly on every create.\n"
+                  description: "Request-routing model. Required \u2014 it determines the valid scaler and request URLs, so it must be chosen explicitly on every create."
                 },
                 cpu: {
                   type: "array",
                   minItems: 1,
                   uniqueItems: true,
-                  description: "Eligible CPU configurations for each worker. Memory is derived from the\nselected flavor's catalog RAM multiplier. Exact duplicate configurations\nare rejected; the same flavor may be listed at different vCPU counts.\n",
+                  description: "Eligible CPU configurations for each worker. Memory is derived from the selected flavor's catalog RAM multiplier. Exact duplicate configurations are rejected; the same flavor may be listed at different vCPU counts.",
                   items: {
                     $ref: "#/$defs/CreateCpuConfig"
                   }
@@ -782,12 +874,17 @@ var generatedTools = [
                   type: "array",
                   items: {
                     type: "string"
-                  }
+                  },
+                  description: "Network volume IDs.",
+                  deprecated: true
+                },
+                mounts: {
+                  $ref: "#/$defs/EndpointMountsInput"
                 },
                 templateId: {
                   type: "string",
                   minLength: 1,
-                  description: "ID of a serverless template to base this endpoint on. The\ntemplate is resolved at create time into the same container\nsettings you could otherwise spread into this body (image,\nargs, disk, ports, env, registry); explicit body fields\noverride the template's, except `env`, which is merged per\nkey with body values winning. The template's\nallowedCudaVersions seeds `gpu.allowedCudaVersions` when the\nbody omits it \u2014 but only for a GPU create, since a CPU endpoint\nhas no gpu block to seed into, and not when the body sets\n`gpu.minCudaVersion`, since seeding a set beside a floor would\nmanufacture the mutual-exclusion 400 from a valid request. Its\npod-specific startSsh/startJupyter flags are\nignored. Later template edits do not affect the endpoint.\nThe template may be one of your own or a public catalog\ntemplate \u2014 see `GET /v2/catalog/templates` (unknown or\ninaccessible ID \u2192 404) \u2014 and must be a serverless template\n(\u2192 422).\n",
+                  description: "ID of a serverless template to base this endpoint on (must be serverless). Its container settings (image, args, disk, ports, env, registry) fill in what this body omits; body fields win, and `env` merges per key with body winning. Its allowedCudaVersions seeds `gpu.allowedCudaVersions` when the body omits it; pod-only flags are ignored. Copied once: later edits do not apply.",
                   examples: ["30zmvf89kd"]
                 },
                 timeout: {
@@ -835,8 +932,42 @@ var generatedTools = [
             }
           }
         },
+        EndpointMount: {
+          type: "object",
+          required: ["volumeId"],
+          additionalProperties: false,
+          description: "A volume mounted on every worker, always at `/runpod-volume`. Custom paths are not supported yet.",
+          properties: {
+            volumeId: {
+              type: "string",
+              description: "ID of a NetworkVolume (under `network`) or GlobalVolume (under `global`).",
+              examples: ["vol_abc"]
+            }
+          }
+        },
+        EndpointMountsInput: {
+          type: "object",
+          additionalProperties: false,
+          description: "Volumes to mount on the endpoint's workers. A `network` or `global` list that is present replaces the endpoint's current list of that kind; an empty list detaches it, and an omitted list is left unchanged.",
+          properties: {
+            network: {
+              type: "array",
+              description: "Network volumes, at most one per data center.",
+              items: {
+                $ref: "#/$defs/EndpointMount"
+              }
+            },
+            global: {
+              type: "array",
+              maxItems: 1,
+              items: {
+                $ref: "#/$defs/EndpointMount"
+              }
+            }
+          }
+        },
         EndpointScaling: {
-          description: "Autoscaling signal \u2014 a discriminated union on `type`: `QUEUE_DELAY`\n(queue-based endpoints only) or `REQUEST_COUNT`. The scaler is chosen\nindependently of the endpoint's routing `type` and can be switched on\nupdate.\n",
+          description: "Autoscaling signal \u2014 a discriminated union on `type`: `QUEUE_DELAY` (queue-based endpoints only) or `REQUEST_COUNT`. The scaler is chosen independently of the endpoint's routing `type` and can be switched on update.",
           oneOf: [
             {
               $ref: "#/$defs/QueueDelayScaling"
@@ -855,7 +986,7 @@ var generatedTools = [
         },
         EndpointType: {
           type: "string",
-          description: "Request-routing semantics for a modern serverless endpoint.\n- `QUEUE` \u2014 submit asynchronous or synchronous jobs through the managed queue.\n- `LOAD_BALANCER` \u2014 send requests directly to worker-defined HTTP paths.\n  Configure via `env`: `PORT` (server port, default 80), `PORT_HEALTH`\n  (health-check port, default 80), and `HEALTH_CHECK_PATH` (path the\n  load balancer polls for worker health, default `/ping`).\n",
+          description: "Request-routing semantics for a modern serverless endpoint. `QUEUE` \u2014 submit asynchronous or synchronous jobs through the managed queue. `LOAD_BALANCER` \u2014 send requests directly to worker-defined HTTP paths.",
           "x-enum-varnames": ["EndpointTypeQueue", "EndpointTypeLoadBalancer"],
           enum: ["QUEUE", "LOAD_BALANCER"]
         },
@@ -879,14 +1010,14 @@ var generatedTools = [
               type: "integer",
               minimum: 1,
               maximum: 3600,
-              description: "Seconds before idle workers scale down. Not applicable to queue-based\nendpoints scaling on `requestCount` \u2014 rejected on create/update and\nomitted from responses for that combination.\n",
+              description: "Seconds before idle workers scale down. Not applicable to queue-based endpoints scaling on `requestCount` \u2014 rejected on create/update and omitted from responses for that combination.",
               examples: [5]
             }
           }
         },
         FlashBoot: {
           type: "string",
-          description: "FlashBoot cold-start acceleration mode.\n- `OFF`                \u2014 disabled\n- `FLASHBOOT`          \u2014 enabled\n- `PRIORITY_FLASHBOOT` \u2014 enabled with priority capacity\n",
+          description: "FlashBoot cold-start acceleration mode. `OFF` \u2014 disabled. `FLASHBOOT` \u2014 enabled. `PRIORITY_FLASHBOOT` \u2014 enabled with priority capacity",
           enum: ["OFF", "FLASHBOOT", "PRIORITY_FLASHBOOT"]
         },
         QueueDelayScaling: {
@@ -912,7 +1043,7 @@ var generatedTools = [
         RequestCountScaling: {
           type: "object",
           additionalProperties: false,
-          description: "Scale on concurrent in-flight requests per worker. Required for\nload-balancing endpoints; also selectable for queue-based.\n",
+          description: "Scale on concurrent in-flight requests per worker. Required for load-balancing endpoints; also selectable for queue-based.",
           required: ["type", "requestCount"],
           properties: {
             type: {
@@ -931,6 +1062,47 @@ var generatedTools = [
       }
     }
   },
+  {
+    name: "create-global-volume",
+    operationId: "createGlobalVolume",
+    description: "Create a global volume. Provisions a new global volume \u2014 storage that mounts into workloads in any data center. Only a `name` is required; storage is provisioned for you. Accounts have a cap on how many global volumes they can hold; creating one past it fails with `400`. Returns `201` with the created global volume, including its assigned `id`.",
+    method: "POST",
+    path: "/v2/global-volumes",
+    params: [],
+    hasBody: true,
+    annotations: {
+      readOnlyHint: false,
+      destructiveHint: false,
+      idempotentHint: false,
+      openWorldHint: true
+    },
+    inputSchema: {
+      type: "object",
+      properties: {
+        body: {
+          $ref: "#/$defs/CreateGlobalVolumeRequest"
+        }
+      },
+      required: ["body"],
+      $defs: {
+        CreateGlobalVolumeRequest: {
+          type: "object",
+          additionalProperties: false,
+          required: ["name"],
+          properties: {
+            name: {
+              type: "string",
+              minLength: 1,
+              maxLength: 191,
+              pattern: "[^\\p{Z}\\t-\\r\x85]",
+              description: "Human-readable name, 1\u2013191 characters with at least one non-whitespace character. Leading and trailing whitespace is trimmed.",
+              examples: ["training-data"]
+            }
+          }
+        }
+      }
+    }
+  },
   {
     name: "create-network-volume",
     operationId: "createNetworkVolume",
@@ -984,7 +1156,7 @@ var generatedTools = [
                   $ref: "#/$defs/VolumeType"
                 }
               ],
-              description: "Storage tier for the volume. Optional. When omitted, the volume is\nprovisioned using the requested data center's default (primary)\nstorage tier. HIGH_PERFORMANCE provisions a high-performance (HPS)\nvolume; STANDARD provisions a standard volume. A volume's tier is\nimmutable after creation.\n"
+              description: "Storage tier for the volume. Optional. When omitted, the volume is provisioned using the requested data center's default (primary) storage tier."
             }
           }
         },
@@ -1010,6 +1182,7 @@ var generatedTools = [
       idempotentHint: false,
       openWorldHint: true
     },
+    schemaOnDemand: true,
     inputSchema: {
       type: "object",
       properties: {
@@ -1021,11 +1194,10 @@ var generatedTools = [
       $defs: {
         BaseContainerConfig: {
           type: "object",
-          description: "Container configuration universal to every containerized resource. Compose ContainerConfig instead unless the resource cannot support a direct registry credential (clusters \u2014 there the registry credential arrives via a pod template, see CreateClusterRequest.templateId).\n",
           properties: {
             args: {
               type: "string",
-              description: 'The container\'s command, as a single raw string. This is the field `entrypoint` and `cmd` encode into, exposed in its stored form.\n\nTwo shapes are accepted. A bare shell string is treated as CMD and split into arguments, which is what the console\'s "Container start command" field writes. A JSON object of the form `{"entrypoint":[...],"cmd":[...]}` sets either or both explicitly.\n\nResponses always return both representations: `args` exactly as stored, plus the deconstructed `entrypoint` and `cmd`. Supplying `args` together with `entrypoint` or `cmd` is allowed only when they describe the same command, so a read-modify-write client can send back everything it received. Send `""` to clear, omit to leave unchanged.\n',
+              description: 'The container\'s command as one raw string: a bare shell string is treated as CMD and split into arguments; a JSON object `{"entrypoint":[...],"cmd":[...]}` sets either or both. Send `""` to clear; omit to leave unchanged.',
               examples: [
                 "--model meta-llama/Llama-3-8B --max-model-len 8192",
                 '{"entrypoint":["/bin/bash","-c"],"cmd":["python -u main.py"]}',
@@ -1037,7 +1209,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container CMD in exec form. When the image defines an ENTRYPOINT, this is the argument list passed to it. Encoded into the `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container CMD in exec form: when the image defines an ENTRYPOINT, the argument list passed to it. Encoded into `args`; supply both only when they describe the same command. Send `[]` to clear; omit to leave unchanged.",
               examples: [
                 ["--model", "meta-llama/Llama-3-8B", "--max-model-len", "8192"]
               ]
@@ -1047,7 +1219,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.",
               examples: [["/bin/bash", "-c"]]
             },
             disk: {
@@ -1119,13 +1291,27 @@ var generatedTools = [
             }
           }
         },
+        BaseMounts: {
+          type: "object",
+          properties: {
+            persistent: {
+              $ref: "#/$defs/PersistentMount"
+            },
+            network: {
+              type: "array",
+              maxItems: 1,
+              items: {
+                $ref: "#/$defs/NetworkMount"
+              }
+            }
+          }
+        },
         Cloud: {
           type: "string",
-          description: "Cloud tier.\n- `SECURE`    \u2014 Runpod-owned datacenter hardware\n- `COMMUNITY` \u2014 community-hosted hardware\n",
+          description: "Cloud tier. `SECURE` \u2014 Runpod-owned datacenter hardware. `COMMUNITY` \u2014 community-hosted hardware",
           enum: ["SECURE", "COMMUNITY"]
         },
         ContainerConfig: {
-          description: "Reusable container configuration shared across templates, pods, and serverless endpoints. Adding a field here automatically propagates to all three resources.\n",
           allOf: [
             {
               $ref: "#/$defs/BaseContainerConfig"
@@ -1151,7 +1337,7 @@ var generatedTools = [
           unevaluatedProperties: false
         },
         CreateGpuConfig: {
-          description: "GPU request for a pod create. Carries the CUDA host constraints, which\nlive here rather than at the body's top level so they are\nunrepresentable on a CPU pod.\n",
+          description: "GPU request for a pod create. Carries the CUDA host constraints, which live here rather than at the body's top level so they are unrepresentable on a CPU pod.",
           allOf: [
             {
               $ref: "#/$defs/BaseGpuConfig"
@@ -1165,13 +1351,13 @@ var generatedTools = [
                     type: "string",
                     pattern: "^\\d+\\.\\d+$"
                   },
-                  description: "Acceptable CUDA versions for the host machine, as `major.minor`.\nOmit to accept any version. Matching is exact, so a version no\nmachine reports yields a capacity error rather than a fallback \u2014\ndiscover valid values per GPU type via\n`GET /v2/catalog/gpus?include=AVAILABILITY&product=POD`\n(`cudaVersions`).\n\nA non-empty set is mutually exclusive with minCudaVersion (400\nif both are sent). An explicit `[]` states no constraint, so it\nmay accompany a floor.\n",
+                  description: "Acceptable CUDA versions for the host machine, as `major.minor`. Omit to accept any version.",
                   examples: [["12.8", "12.6"]]
                 },
                 minCudaVersion: {
                   type: "string",
                   pattern: "^\\d+\\.\\d+$",
-                  description: "Lowest acceptable CUDA version for the host machine, as\n`major.minor`, compared numerically rather than as a decimal \u2014\nso 12.11 is above 12.2. Use this for an open-ended floor and\nallowedCudaVersions for an exact set.\n\nMutually exclusive with a non-empty allowedCudaVersions (400 if\nboth are sent); an explicit `[]` there states no constraint and\nmay accompany this floor.\n",
+                  description: "Lowest acceptable CUDA version for the host machine, as `major.minor`, compared numerically rather than as a decimal \u2014 so 12.11 is above 12.2. Use this for an open-ended floor and allowedCudaVersions for an exact set.",
                   examples: ["12.1"]
                 },
                 minRamPerGpu: {
@@ -1199,7 +1385,7 @@ var generatedTools = [
             {
               type: "object",
               required: ["name"],
-              description: "Request body for creating a pod. Exactly one of `gpu` or `cpu`\nmust be set \u2014 enforced at the handler layer. For CPU pods, memory\nis derived by the API from the selected flavor's RAM multiplier;\nclients provide only CPU flavor and vCPU count. CPU pods support\ncontainer disk and network volumes only; `mounts.persistent` is\ninvalid when `cpu` is set.\n\n`image` is required unless `templateId` is set.\n",
+              description: "Request body for creating a pod. Exactly one of `gpu` or `cpu` must be set \u2014 enforced at the handler layer.",
               properties: {
                 name: {
                   type: "string",
@@ -1223,13 +1409,13 @@ var generatedTools = [
                   items: {
                     type: "string"
                   },
-                  description: "Preferred data centers for placement. Omit or pass an empty\narray to let the scheduler choose.\n",
+                  description: "Preferred data centers for placement. Omit or pass an empty array to let the scheduler choose.",
                   examples: [["US-TX-3"]]
                 },
                 globalNetworking: {
                   type: "boolean",
                   default: false,
-                  description: "Enable global networking, giving the pod a private IP reachable across data centers. Requires an NVIDIA GPU and a global-networking-enabled data center (both enforced upstream). See `GET /v2/catalog/datacenters` (`globalNetwork`) for eligible data centers.",
+                  description: "Enable global networking, giving the pod a private IP reachable across data centers. Requires an NVIDIA GPU and a global-networking-enabled data center (both enforced upstream).",
                   examples: [false]
                 },
                 gpu: {
@@ -1241,19 +1427,19 @@ var generatedTools = [
                 startJupyter: {
                   type: "boolean",
                   default: false,
-                  description: "Create-time flag telling the provisioner to start JupyterLab:\ninjects a generated `JUPYTER_PASSWORD` environment variable,\nunless the request already sets one. Only images that honor\nthe convention start Jupyter from it (Runpod official images\ndo); expose `8888/http` in `ports` to reach it.\n\nNot part of the pod's readable config \u2014 never returned by\nGET and not changeable by PATCH.\n",
+                  description: "Create-time flag: starts JupyterLab by injecting a generated `JUPYTER_PASSWORD` environment variable, unless the request sets one. Expose `8888/http` in `ports` to reach it.",
                   examples: [true]
                 },
                 startSsh: {
                   type: "boolean",
                   default: false,
-                  description: "Create-time flag telling the provisioner to set up SSH\naccess: injects a `PUBLIC_KEY` environment variable carrying\nyour account's registered SSH public keys, unless the request\nalready sets one. **Requires registered keys** (`PUT\n/v2/account/ssh-keys`) \u2014 with none registered the flag does\nnothing and the pod has no SSH access. Only images that honor\nthe convention start sshd from it (all Runpod official images\ndo). Connect using the pod's `ssh` block; the `ssh.direct`\nvariant additionally needs a `22/tcp` entry in `ports`.\n\nNot part of the pod's readable config \u2014 never returned by\nGET and not changeable by PATCH.\n",
+                  description: "Create-time flag: sets up SSH access by injecting a `PUBLIC_KEY` environment variable with your account's registered SSH keys (`update-ssh-keys`), unless the request sets one. With no registered keys it does nothing and the pod has no SSH access. Direct SSH also needs a `22/tcp` entry in `ports`.",
                   examples: [true]
                 },
                 templateId: {
                   type: "string",
                   minLength: 1,
-                  description: "ID of a pod template to base this pod on. The template is\nresolved at create time into the same container settings you\ncould otherwise spread into this body (image, args, disk,\nports, env, registry, persistent mount, startSsh,\nstartJupyter, allowedCudaVersions); explicit body fields\noverride the template's, except `env`, which is merged per\nkey with body values winning. Sending either CUDA field\n(`gpu.allowedCudaVersions` or `gpu.minCudaVersion`) replaces\nthe template's CUDA constraint entirely, and CPU pods ignore\nit (like the persistent mount). The template is a one-time\nsource of settings: later template edits do not affect the\npod, and the created pod does not retain a link to the\ntemplate (`template` stays null). The template may be one\nof your own or a public catalog template \u2014 see\n`GET /v2/catalog/templates` (unknown or inaccessible ID \u2192\n404) \u2014 and must not be a serverless template (\u2192 422). CPU\npods do not inherit a template's persistent mount.\n",
+                  description: "ID of a pod template to base this pod on (not a serverless one). Its container settings (image, args, disk, ports, env, registry, persistent mount) fill in what this body omits; body fields win, and `env` merges per key with body winning. A CUDA field in `gpu` replaces the template's CUDA constraint; CPU pods ignore its mount. Copied once: later template edits do not affect the pod.",
                   examples: ["30zmvf89kd"]
                 }
               }
@@ -1269,28 +1455,51 @@ var generatedTools = [
             required: ["image"]
           }
         },
-        Mounts: {
+        GlobalMount: {
           type: "object",
+          required: ["volumeId"],
           additionalProperties: false,
-          description: "Storage mounts attached to a pod. At-most-one of `persistent` or\n`network` may be set today (mutually exclusive, enforced at the\nhandler with 400 if both are present). The `network` field is an\narray for forward compatibility with eventual multi-network-volume\nsupport, but `maxItems` is 1 today.\n\nPATCH semantics:\n- Omitting `mounts` or sending `{}` leaves the existing mount\n  unchanged.\n- An explicit `network: []` is rejected with 400 (clearing mounts\n  is not supported).\n- Mount kind is fixed at create \u2014 a PATCH that introduces a kind\n  not present at create (persistent on a network pod, network on\n  a persistent pod, or any mount on a previously-mountless pod)\n  is rejected with 400.\n- The `volumeId` of a network mount is immutable; a PATCH that\n  names a different `volumeId` is rejected with 400.\n- Partial mounts are not supported \u2014 every mount entry must\n  include the full schema (`size` + `path` for persistent,\n  `volumeId` + `path` for network). Missing required fields \u2192 422.\n",
+          description: "Reference to a GlobalVolume. Usage-billed, so it has no size. The underlying `volumeId` is immutable post-create; the mount `path` may be changed via PATCH.",
           properties: {
-            persistent: {
-              $ref: "#/$defs/PersistentMount"
+            volumeId: {
+              type: "string",
+              description: "ID of an existing GlobalVolume.",
+              examples: ["gsb_abc"]
             },
-            network: {
-              type: "array",
-              maxItems: 1,
-              items: {
-                $ref: "#/$defs/NetworkMount"
-              }
+            path: {
+              type: "string",
+              description: "Mount path inside the container. Optional on create, where a default is chosen for you; a PATCH that omits it leaves the path unchanged.",
+              examples: ["/data"]
             }
           }
         },
+        Mounts: {
+          description: "Storage mounts for the pod: `persistent` (size + path) excludes both `network` and `global`; `network` (volumeId + path) and `global` (volumeId, optional path) take one entry each and can be combined. Every entry needs its full shape. On update, omit the field to leave the mounts as they are; the mount kind and a volume's volumeId cannot change, and mounts cannot be cleared.",
+          allOf: [
+            {
+              $ref: "#/$defs/BaseMounts"
+            },
+            {
+              type: "object",
+              properties: {
+                global: {
+                  type: "array",
+                  maxItems: 1,
+                  description: "At most one global volume, on GPU and CPU pods, with no data center constraint. Combinable with `network` but not with `persistent` (rejected): upstream drops a persistent disk whenever a volume is attached.",
+                  items: {
+                    $ref: "#/$defs/GlobalMount"
+                  }
+                }
+              }
+            }
+          ],
+          unevaluatedProperties: false
+        },
         NetworkMount: {
           type: "object",
           required: ["volumeId", "path"],
           additionalProperties: false,
-          description: "Reference to a NetworkVolume. Custom paths are honored at runtime on\nboth GPU and CPU pods. The underlying `volumeId` is immutable\npost-create; the mount `path` may be changed via PATCH.\n",
+          description: "Reference to a NetworkVolume. Custom paths are honored at runtime on both GPU and CPU pods. The underlying `volumeId` is immutable post-create; the mount `path` may be changed via PATCH.",
           properties: {
             volumeId: {
               type: "string",
@@ -1308,7 +1517,7 @@ var generatedTools = [
           type: "object",
           required: ["size", "path"],
           additionalProperties: false,
-          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data\ndoes not survive a host failure. Disallowed on CPU pods. Mutually\nexclusive with NetworkMount. Deprecated: prefer NetworkMount for any\ndata you cannot recreate.\n",
+          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data does not survive a host failure. Disallowed on CPU pods. Mutually exclusive with NetworkMount. Deprecated: prefer NetworkMount for any data you cannot recreate.",
           properties: {
             size: {
               type: "integer",
@@ -1391,6 +1600,7 @@ var generatedTools = [
       idempotentHint: false,
       openWorldHint: true
     },
+    schemaOnDemand: true,
     inputSchema: {
       type: "object",
       properties: {
@@ -1410,7 +1620,7 @@ var generatedTools = [
               minLength: 1,
               maxLength: 191,
               pattern: "^[a-zA-Z_][a-zA-Z0-9_.\\-/]*$",
-              description: "Unique name for the secret \u2014 referenced from environment variables as `{{ RUNPOD_SECRET_<name> }}`; immutable after creation. Maximum 191 characters, must start with a letter or underscore, and may contain letters, digits, and `_.-/`. Names beginning with the reserved prefix `RUNPOD` are rejected (case-insensitive).",
+              description: "Unique name for the secret \u2014 referenced from environment variables as `{{ RUNPOD_SECRET_<name> }}`; immutable after creation. Maximum 191 characters, must start with a letter or underscore, and may contain letters, digits, and `_.-/`.",
               examples: ["hf-token"]
             },
             value: {
@@ -1444,6 +1654,7 @@ var generatedTools = [
       idempotentHint: false,
       openWorldHint: true
     },
+    schemaOnDemand: true,
     inputSchema: {
       type: "object",
       properties: {
@@ -1455,11 +1666,10 @@ var generatedTools = [
       $defs: {
         BaseContainerConfig: {
           type: "object",
-          description: "Container configuration universal to every containerized resource. Compose ContainerConfig instead unless the resource cannot support a direct registry credential (clusters \u2014 there the registry credential arrives via a pod template, see CreateClusterRequest.templateId).\n",
           properties: {
             args: {
               type: "string",
-              description: 'The container\'s command, as a single raw string. This is the field `entrypoint` and `cmd` encode into, exposed in its stored form.\n\nTwo shapes are accepted. A bare shell string is treated as CMD and split into arguments, which is what the console\'s "Container start command" field writes. A JSON object of the form `{"entrypoint":[...],"cmd":[...]}` sets either or both explicitly.\n\nResponses always return both representations: `args` exactly as stored, plus the deconstructed `entrypoint` and `cmd`. Supplying `args` together with `entrypoint` or `cmd` is allowed only when they describe the same command, so a read-modify-write client can send back everything it received. Send `""` to clear, omit to leave unchanged.\n',
+              description: 'The container\'s command as one raw string: a bare shell string is treated as CMD and split into arguments; a JSON object `{"entrypoint":[...],"cmd":[...]}` sets either or both. Send `""` to clear; omit to leave unchanged.',
               examples: [
                 "--model meta-llama/Llama-3-8B --max-model-len 8192",
                 '{"entrypoint":["/bin/bash","-c"],"cmd":["python -u main.py"]}',
@@ -1471,7 +1681,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container CMD in exec form. When the image defines an ENTRYPOINT, this is the argument list passed to it. Encoded into the `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container CMD in exec form: when the image defines an ENTRYPOINT, the argument list passed to it. Encoded into `args`; supply both only when they describe the same command. Send `[]` to clear; omit to leave unchanged.",
               examples: [
                 ["--model", "meta-llama/Llama-3-8B", "--max-model-len", "8192"]
               ]
@@ -1481,7 +1691,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.",
               examples: [["/bin/bash", "-c"]]
             },
             disk: {
@@ -1518,7 +1728,6 @@ var generatedTools = [
           }
         },
         ContainerConfig: {
-          description: "Reusable container configuration shared across templates, pods, and serverless endpoints. Adding a field here automatically propagates to all three resources.\n",
           allOf: [
             {
               $ref: "#/$defs/BaseContainerConfig"
@@ -1555,7 +1764,7 @@ var generatedTools = [
                     type: "string",
                     pattern: "^\\d+\\.\\d+$"
                   },
-                  description: "Acceptable CUDA versions for containers created from this\ntemplate, as `major.minor`. Omit to accept any version \u2014 see\nthe same field on `createPod` for matching semantics.\nExpanded into GPU pod and serverless endpoint creates; CPU\npods ignore it.\n",
+                  description: "Acceptable CUDA versions for containers created from this template, as `major.minor`. Omit to accept any version \u2014 see the same field on `createPod` for matching semantics.",
                   examples: [["12.8", "12.6"]]
                 },
                 category: {
@@ -1581,13 +1790,13 @@ var generatedTools = [
                 startJupyter: {
                   type: "boolean",
                   default: true,
-                  description: "Start JupyterLab in containers created from this template:\ninjects a generated `JUPYTER_PASSWORD` environment variable,\nunless `env` already sets one. Only images that honor the\nconvention start Jupyter from it (Runpod official images do);\nexpose `8888/http` in `ports` to reach it. Defaults to `true`\nwhen omitted, matching console-created templates.\n",
+                  description: "Start JupyterLab in containers created from this template: injects a generated `JUPYTER_PASSWORD` environment variable, unless `env` already sets one.",
                   examples: [false]
                 },
                 startSsh: {
                   type: "boolean",
                   default: true,
-                  description: "Provision SSH access in containers created from this template:\ninjects a `PUBLIC_KEY` environment variable carrying the\ndeployer's registered SSH public keys (`PUT\n/v2/account/ssh-keys` \u2014 with none registered the flag does\nnothing), unless `env` already sets one. Only images that\nhonor the convention start sshd from it (all Runpod official\nimages do); direct SSH also needs a `22/tcp` entry in\n`ports`. Defaults to `true` when omitted, matching\nconsole-created templates.\n",
+                  description: "Provision SSH access in containers created from this template: injects a `PUBLIC_KEY` environment variable carrying the deployer's registered SSH public keys (`update-ssh-keys` \u2014 with none registered the flag does nothing), unless `env` already sets one.",
                   examples: [true]
                 }
               }
@@ -1599,7 +1808,7 @@ var generatedTools = [
           type: "object",
           required: ["size", "path"],
           additionalProperties: false,
-          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data\ndoes not survive a host failure. Disallowed on CPU pods. Mutually\nexclusive with NetworkMount. Deprecated: prefer NetworkMount for any\ndata you cannot recreate.\n",
+          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data does not survive a host failure. Disallowed on CPU pods. Mutually exclusive with NetworkMount. Deprecated: prefer NetworkMount for any data you cannot recreate.",
           properties: {
             size: {
               type: "integer",
@@ -1616,13 +1825,13 @@ var generatedTools = [
         },
         TemplateCategory: {
           type: "string",
-          description: "Controls how the template is grouped and filtered in the Runpod console.\nIt does not affect hardware selection, scheduling, or billing.\n- `CPU`    \u2014 CPU-only workloads\n- `NVIDIA` \u2014 NVIDIA GPU workloads\n- `AMD`    \u2014 AMD GPU workloads\n",
+          description: "Controls how the template is grouped and filtered in the Runpod console. It does not affect hardware selection, scheduling, or billing. `CPU` \u2014 CPU-only workloads. `NVIDIA` \u2014 NVIDIA GPU workloads. `AMD` \u2014 AMD GPU workloads",
           enum: ["CPU", "NVIDIA", "AMD"]
         },
         TemplateMounts: {
           type: "object",
           additionalProperties: false,
-          description: "Storage mounts attached to a template. Templates support only a\nsingle persistent mount today; any `network` property is rejected\nwith 422 by the schema validator.\n\nPATCH semantics: omitting `mounts` or sending `{}` leaves the\nexisting mount unchanged.\n",
+          description: "Storage mounts for containers from this template: a single `persistent` mount only (`network` is rejected). On update, omitting `mounts` or sending `{}` leaves the existing mount unchanged.",
           properties: {
             persistent: {
               $ref: "#/$defs/PersistentMount"
@@ -1692,6 +1901,36 @@ var generatedTools = [
       required: ["id"]
     }
   },
+  {
+    name: "delete-global-volume",
+    operationId: "deleteGlobalVolume",
+    description: "Delete a global volume. Permanently deletes a global volume and its stored data. Fails with `409` while the volume is still mounted: terminate every pod that mounts it (stopped pods count) and detach it from every serverless endpoint first. Nothing is detached or terminated on your behalf.",
+    method: "DELETE",
+    path: "/v2/global-volumes/{id}",
+    params: [
+      {
+        name: "id",
+        location: "path"
+      }
+    ],
+    hasBody: false,
+    annotations: {
+      readOnlyHint: false,
+      destructiveHint: true,
+      idempotentHint: true,
+      openWorldHint: true
+    },
+    inputSchema: {
+      type: "object",
+      properties: {
+        id: {
+          type: "string",
+          description: "Global volume identifier"
+        }
+      },
+      required: ["id"]
+    }
+  },
   {
     name: "delete-network-volume",
     operationId: "deleteNetworkVolume",
@@ -2057,12 +2296,42 @@ var generatedTools = [
         },
         buildId: {
           type: "string",
-          description: "GitHub build identifier (from GET /v2/serverless/{id}/builds or a release's `buildId`)"
+          description: "GitHub build identifier (from `list-endpoint-builds` or a release's `buildId`)"
         }
       },
       required: ["id", "buildId"]
     }
   },
+  {
+    name: "get-global-volume",
+    operationId: "getGlobalVolume",
+    description: "Get a global volume. Returns a single global volume by ID.",
+    method: "GET",
+    path: "/v2/global-volumes/{id}",
+    params: [
+      {
+        name: "id",
+        location: "path"
+      }
+    ],
+    hasBody: false,
+    annotations: {
+      readOnlyHint: true,
+      destructiveHint: false,
+      idempotentHint: true,
+      openWorldHint: true
+    },
+    inputSchema: {
+      type: "object",
+      properties: {
+        id: {
+          type: "string",
+          description: "Global volume identifier"
+        }
+      },
+      required: ["id"]
+    }
+  },
   {
     name: "get-gpu-type",
     operationId: "getGpuType",
@@ -2133,7 +2402,7 @@ var generatedTools = [
           items: {
             $ref: "#/$defs/Product"
           },
-          description: "Comma-separated availability product contexts. Supported values: POD, CLUSTER, SERVERLESS. Required with include=AVAILABILITY, and valid only with it (400 either way). There is no default: the same GPU type can be scarce for pods and plentiful for serverless, so the context has to be stated rather than assumed."
+          description: "Comma-separated availability product contexts. Supported values: POD, CLUSTER, SERVERLESS. Required with include=AVAILABILITY, and valid only with it (400 either way)."
         },
         count: {
           type: "integer",
@@ -2151,7 +2420,7 @@ var generatedTools = [
             type: "string",
             pattern: "^[A-Z]{2}$"
           },
-          description: "Comma-separated ISO 3166-1 alpha-2 country codes, uppercase, to constrain availability to \u2014 e.g. FR or FR,DE. Values within this filter use OR semantics. Valid only with include=AVAILABILITY (400 otherwise); a malformed entry is a 422. Scopes availability, lowest-price calculations and the dataCenters array to those countries, so a listed data center outside them is omitted rather than returned with availability NONE. On the list endpoint a GPU type with no data center in those countries drops out entirely; the single-GPU endpoint still returns the requested type, with availability NONE and dataCenters omitted, so a 404 keeps meaning the GPU type does not exist. Read the NONE on availability rather than the absence of dataCenters, which is also absent when availability was not requested."
+          description: "Comma-separated ISO 3166-1 alpha-2 country codes, uppercase, to constrain availability to \u2014 e.g. FR or FR,DE. Values within this filter use OR semantics. Valid only with include=AVAILABILITY; a malformed entry is a 422."
         },
         cudaVersions: {
           type: "array",
@@ -2159,12 +2428,12 @@ var generatedTools = [
             type: "string",
             pattern: "^\\d+\\.\\d+$"
           },
-          description: "Comma-separated CUDA versions to scope availability and lowest-price calculations to, matched exactly. Format: major.minor, e.g. 12.8 \u2014 a bare major is rejected here because it identifies no version. Valid only with include=AVAILABILITY (400 otherwise) and mutually exclusive with minCudaVersion (400 if both are sent); a malformed entry is a 422. Also narrows the returned cudaVersions array; omit it to enumerate every version offered."
+          description: "Comma-separated CUDA versions to scope availability and lowest-price calculations to, matched exactly. Format: major.minor, e.g. 12.8 \u2014 a bare major is rejected here because it identifies no version."
         },
         minCudaVersion: {
           type: "string",
           pattern: "^\\d+(\\.\\d+)?$",
-          description: "Lowest acceptable CUDA version to scope availability and lowest-price calculations to, compared numerically. Format: integer major or major.minor, e.g. 12 or 12.1 \u2014 unlike the `gpu.minCudaVersion` body field on pod and endpoint create, a bare major is accepted here and means any release of that major, because this filter only widens a read. Valid only with include=AVAILABILITY (400 otherwise) and mutually exclusive with cudaVersions (400 if both are sent); a malformed value is a 422. Use this for an open-ended floor and cudaVersions for an exact set."
+          description: "Lowest acceptable CUDA version to scope availability and lowest-price calculations to, compared numerically."
         }
       },
       required: ["id"],
@@ -2217,6 +2486,25 @@ var generatedTools = [
       required: ["id"]
     }
   },
+  {
+    name: "get-organization",
+    operationId: "getOrganization",
+    description: "Get your current organization. Returns the authenticated caller's current organization. Organization context comes from the caller's membership; no organization selector is required. Returns 404 when the caller does not belong to an organization, or 403 when organizations are unavailable to the caller.",
+    method: "GET",
+    path: "/v2/organization",
+    params: [],
+    hasBody: false,
+    annotations: {
+      readOnlyHint: true,
+      destructiveHint: false,
+      idempotentHint: true,
+      openWorldHint: true
+    },
+    inputSchema: {
+      type: "object",
+      properties: {}
+    }
+  },
   {
     name: "get-pod",
     operationId: "getPod",
@@ -2391,12 +2679,12 @@ var generatedTools = [
         startTime: {
           type: "string",
           format: "date-time",
-          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago. Snapped down to the start of its bucketSize bucket so the window aligns with the returned records; provide a boundary-aligned value (e.g. midnight for bucketSize=day) to avoid widening.\n"
+          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago."
         },
         endTime: {
           type: "string",
           format: "date-time",
-          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records.\n"
+          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records."
         },
         bucketSize: {
           $ref: "#/$defs/BillingBucketSize",
@@ -2405,7 +2693,7 @@ var generatedTools = [
         lastN: {
           type: "integer",
           minimum: 1,
-          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days". The resolved window is aligned to bucket boundaries: startTime is the start of the earliest bucket (e.g. midnight of the earliest day) and endTime is the end of the current bucket. Mutually exclusive with startTime/endTime; provide one or the other, not both.\n'
+          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days".'
         }
       },
       $defs: {
@@ -2467,12 +2755,12 @@ var generatedTools = [
         startTime: {
           type: "string",
           format: "date-time",
-          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago. Snapped down to the start of its bucketSize bucket so the window aligns with the returned records; provide a boundary-aligned value (e.g. midnight for bucketSize=day) to avoid widening.\n"
+          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago."
         },
         endTime: {
           type: "string",
           format: "date-time",
-          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records.\n"
+          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records."
         },
         bucketSize: {
           $ref: "#/$defs/BillingBucketSize",
@@ -2481,7 +2769,7 @@ var generatedTools = [
         lastN: {
           type: "integer",
           minimum: 1,
-          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days". The resolved window is aligned to bucket boundaries: startTime is the start of the earliest bucket (e.g. midnight of the earliest day) and endTime is the end of the current bucket. Mutually exclusive with startTime/endTime; provide one or the other, not both.\n'
+          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days".'
         },
         clusterId: {
           type: "string",
@@ -2539,10 +2827,19 @@ var generatedTools = [
   {
     name: "list-clusters",
     operationId: "listClusters",
-    description: "List clusters. Returns all clusters owned by the authenticated user.",
+    description: "List clusters. Returns clusters owned by the authenticated user, cursor-paginated newest-first; an omitted `limit` defaults to 1000. Follow `pagination.nextCursor` until `hasNextPage` is false.",
     method: "GET",
     path: "/v2/clusters",
-    params: [],
+    params: [
+      {
+        name: "cursor",
+        location: "query"
+      },
+      {
+        name: "limit",
+        location: "query"
+      }
+    ],
     hasBody: false,
     annotations: {
       readOnlyHint: true,
@@ -2552,7 +2849,19 @@ var generatedTools = [
     },
     inputSchema: {
       type: "object",
-      properties: {}
+      properties: {
+        cursor: {
+          type: "string",
+          minLength: 1,
+          description: "Opaque resume cursor \u2014 pass the previous response's `pagination.nextCursor` through verbatim; omit for the first page. A cursor is only valid for the operation and parameters that issued it; a malformed or foreign cursor is rejected."
+        },
+        limit: {
+          type: "integer",
+          minimum: 1,
+          maximum: 1e3,
+          description: "Page size, 1\u20131000. Defaults to 1000 when omitted."
+        }
+      }
     }
   },
   {
@@ -2806,12 +3115,12 @@ var generatedTools = [
         startTime: {
           type: "string",
           format: "date-time",
-          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago. Snapped down to the start of its bucketSize bucket so the window aligns with the returned records; provide a boundary-aligned value (e.g. midnight for bucketSize=day) to avoid widening.\n"
+          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago."
         },
         endTime: {
           type: "string",
           format: "date-time",
-          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records.\n"
+          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records."
         },
         bucketSize: {
           $ref: "#/$defs/BillingBucketSize",
@@ -2820,7 +3129,7 @@ var generatedTools = [
         lastN: {
           type: "integer",
           minimum: 1,
-          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days". The resolved window is aligned to bucket boundaries: startTime is the start of the earliest bucket (e.g. midnight of the earliest day) and endTime is the end of the current bucket. Mutually exclusive with startTime/endTime; provide one or the other, not both.\n'
+          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days".'
         }
       },
       $defs: {
@@ -2878,7 +3187,7 @@ var generatedTools = [
         cursor: {
           type: "string",
           minLength: 1,
-          description: "Opaque resume cursor \u2014 pass the previous response's `pagination.nextCursor` through verbatim; omit for the first page. A cursor is only valid for the operation and parameters that issued it; a malformed or foreign cursor is rejected with 422.\n"
+          description: "Opaque resume cursor \u2014 pass the previous response's `pagination.nextCursor` through verbatim; omit for the first page. A cursor is only valid for the operation and parameters that issued it; a malformed or foreign cursor is rejected."
         },
         limit: {
           type: "integer",
@@ -2927,7 +3236,7 @@ var generatedTools = [
         cursor: {
           type: "string",
           minLength: 1,
-          description: "Opaque resume cursor \u2014 pass the previous response's `pagination.nextCursor` through verbatim; omit for the first page. A cursor is only valid for the operation and parameters that issued it; a malformed or foreign cursor is rejected with 422.\n"
+          description: "Opaque resume cursor \u2014 pass the previous response's `pagination.nextCursor` through verbatim; omit for the first page. A cursor is only valid for the operation and parameters that issued it; a malformed or foreign cursor is rejected."
         },
         limit: {
           type: "integer",
@@ -2969,6 +3278,46 @@ var generatedTools = [
       required: ["id"]
     }
   },
+  {
+    name: "list-global-volumes",
+    operationId: "listGlobalVolumes",
+    description: "List global volumes. Returns global volumes owned by the authenticated user, cursor-paginated newest-first; an omitted `limit` defaults to 1000. Follow `pagination.nextCursor` until `hasNextPage` is false.",
+    method: "GET",
+    path: "/v2/global-volumes",
+    params: [
+      {
+        name: "cursor",
+        location: "query"
+      },
+      {
+        name: "limit",
+        location: "query"
+      }
+    ],
+    hasBody: false,
+    annotations: {
+      readOnlyHint: true,
+      destructiveHint: false,
+      idempotentHint: true,
+      openWorldHint: true
+    },
+    inputSchema: {
+      type: "object",
+      properties: {
+        cursor: {
+          type: "string",
+          minLength: 1,
+          description: "Opaque resume cursor \u2014 pass the previous response's `pagination.nextCursor` through verbatim; omit for the first page. A cursor is only valid for the operation and parameters that issued it; a malformed or foreign cursor is rejected."
+        },
+        limit: {
+          type: "integer",
+          minimum: 1,
+          maximum: 1e3,
+          description: "Page size, 1\u20131000. Defaults to 1000 when omitted."
+        }
+      }
+    }
+  },
   {
     name: "list-gpu-types",
     operationId: "listGpuTypes",
@@ -3032,7 +3381,7 @@ var generatedTools = [
           items: {
             $ref: "#/$defs/Product"
           },
-          description: "Comma-separated availability product contexts. Supported values: POD, CLUSTER, SERVERLESS. Required with include=AVAILABILITY, and valid only with it (400 either way). There is no default: the same GPU type can be scarce for pods and plentiful for serverless, so the context has to be stated rather than assumed."
+          description: "Comma-separated availability product contexts. Supported values: POD, CLUSTER, SERVERLESS. Required with include=AVAILABILITY, and valid only with it (400 either way)."
         },
         count: {
           type: "integer",
@@ -3050,7 +3399,7 @@ var generatedTools = [
             type: "string",
             pattern: "^[A-Z]{2}$"
           },
-          description: "Comma-separated ISO 3166-1 alpha-2 country codes, uppercase, to constrain availability to \u2014 e.g. FR or FR,DE. Values within this filter use OR semantics. Valid only with include=AVAILABILITY (400 otherwise); a malformed entry is a 422. Scopes availability, lowest-price calculations and the dataCenters array to those countries, so a listed data center outside them is omitted rather than returned with availability NONE. On the list endpoint a GPU type with no data center in those countries drops out entirely; the single-GPU endpoint still returns the requested type, with availability NONE and dataCenters omitted, so a 404 keeps meaning the GPU type does not exist. Read the NONE on availability rather than the absence of dataCenters, which is also absent when availability was not requested."
+          description: "Comma-separated ISO 3166-1 alpha-2 country codes, uppercase, to constrain availability to \u2014 e.g. FR or FR,DE. Values within this filter use OR semantics. Valid only with include=AVAILABILITY; a malformed entry is a 422."
         },
         cudaVersions: {
           type: "array",
@@ -3058,12 +3407,12 @@ var generatedTools = [
             type: "string",
             pattern: "^\\d+\\.\\d+$"
           },
-          description: "Comma-separated CUDA versions to scope availability and lowest-price calculations to, matched exactly. Format: major.minor, e.g. 12.8 \u2014 a bare major is rejected here because it identifies no version. Valid only with include=AVAILABILITY (400 otherwise) and mutually exclusive with minCudaVersion (400 if both are sent); a malformed entry is a 422. Also narrows the returned cudaVersions array; omit it to enumerate every version offered."
+          description: "Comma-separated CUDA versions to scope availability and lowest-price calculations to, matched exactly. Format: major.minor, e.g. 12.8 \u2014 a bare major is rejected here because it identifies no version."
         },
         minCudaVersion: {
           type: "string",
           pattern: "^\\d+(\\.\\d+)?$",
-          description: "Lowest acceptable CUDA version to scope availability and lowest-price calculations to, compared numerically. Format: integer major or major.minor, e.g. 12 or 12.1 \u2014 unlike the `gpu.minCudaVersion` body field on pod and endpoint create, a bare major is accepted here and means any release of that major, because this filter only widens a read. Valid only with include=AVAILABILITY (400 otherwise) and mutually exclusive with cudaVersions (400 if both are sent); a malformed value is a 422. Use this for an open-ended floor and cudaVersions for an exact set."
+          description: "Lowest acceptable CUDA version to scope availability and lowest-price calculations to, compared numerically."
         }
       },
       $defs: {
@@ -3126,12 +3475,12 @@ var generatedTools = [
         startTime: {
           type: "string",
           format: "date-time",
-          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago. Snapped down to the start of its bucketSize bucket so the window aligns with the returned records; provide a boundary-aligned value (e.g. midnight for bucketSize=day) to avoid widening.\n"
+          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago."
         },
         endTime: {
           type: "string",
           format: "date-time",
-          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records.\n"
+          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records."
         },
         bucketSize: {
           $ref: "#/$defs/BillingBucketSize",
@@ -3140,7 +3489,7 @@ var generatedTools = [
         lastN: {
           type: "integer",
           minimum: 1,
-          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days". The resolved window is aligned to bucket boundaries: startTime is the start of the earliest bucket (e.g. midnight of the earliest day) and endTime is the end of the current bucket. Mutually exclusive with startTime/endTime; provide one or the other, not both.\n'
+          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days".'
         },
         networkVolumeId: {
           type: "string",
@@ -3168,10 +3517,19 @@ var generatedTools = [
   {
     name: "list-network-volumes",
     operationId: "listNetworkVolumes",
-    description: "List network volumes. Returns all network volumes owned by the authenticated user.",
+    description: "List network volumes. Returns network volumes owned by the authenticated user, cursor-paginated newest-first; an omitted `limit` defaults to 1000. Follow `pagination.nextCursor` until `hasNextPage` is false.",
     method: "GET",
     path: "/v2/network-volumes",
-    params: [],
+    params: [
+      {
+        name: "cursor",
+        location: "query"
+      },
+      {
+        name: "limit",
+        location: "query"
+      }
+    ],
     hasBody: false,
     annotations: {
       readOnlyHint: true,
@@ -3181,7 +3539,19 @@ var generatedTools = [
     },
     inputSchema: {
       type: "object",
-      properties: {}
+      properties: {
+        cursor: {
+          type: "string",
+          minLength: 1,
+          description: "Opaque resume cursor \u2014 pass the previous response's `pagination.nextCursor` through verbatim; omit for the first page. A cursor is only valid for the operation and parameters that issued it; a malformed or foreign cursor is rejected."
+        },
+        limit: {
+          type: "integer",
+          minimum: 1,
+          maximum: 1e3,
+          description: "Page size, 1\u20131000. Defaults to 1000 when omitted."
+        }
+      }
     }
   },
   {
@@ -3225,12 +3595,12 @@ var generatedTools = [
         startTime: {
           type: "string",
           format: "date-time",
-          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago. Snapped down to the start of its bucketSize bucket so the window aligns with the returned records; provide a boundary-aligned value (e.g. midnight for bucketSize=day) to avoid widening.\n"
+          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago."
         },
         endTime: {
           type: "string",
           format: "date-time",
-          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records.\n"
+          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records."
         },
         bucketSize: {
           $ref: "#/$defs/BillingBucketSize",
@@ -3239,7 +3609,7 @@ var generatedTools = [
         lastN: {
           type: "integer",
           minimum: 1,
-          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days". The resolved window is aligned to bucket boundaries: startTime is the start of the earliest bucket (e.g. midnight of the earliest day) and endTime is the end of the current bucket. Mutually exclusive with startTime/endTime; provide one or the other, not both.\n'
+          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days".'
         },
         podId: {
           type: "string",
@@ -3267,7 +3637,7 @@ var generatedTools = [
   {
     name: "list-pods",
     operationId: "listPods",
-    description: "List pods. Returns pods owned by the authenticated user. Cluster member pods are excluded by default; set `includeClusterPods=true` to include them (each carries a non-null `cluster` membership block). Results are cursor-paginated newest-first; an omitted `limit` defaults to 1000. When cluster member pods are excluded, the exclusion applies to each page after it is cut, so a page may hold fewer than `limit` pods \u2014 follow `pagination.nextCursor` until `hasNextPage` is false rather than counting items.",
+    description: "List pods. Returns pods owned by the authenticated user. Cluster member pods are excluded by default; set `includeClusterPods=true` to include them (each carries a non-null `cluster` membership block). Results are cursor-paginated newest-first; an omitted `limit` defaults to 1000. The cluster-member exclusion is applied before the page is cut, so a page holds `limit` pods whenever that many remain. Follow `pagination.nextCursor` until `hasNextPage` is false.",
     method: "GET",
     path: "/v2/pods",
     params: [
@@ -3302,7 +3672,7 @@ var generatedTools = [
         cursor: {
           type: "string",
           minLength: 1,
-          description: "Opaque resume cursor \u2014 pass the previous response's `pagination.nextCursor` through verbatim; omit for the first page. A cursor is only valid for the operation and parameters that issued it; a malformed or foreign cursor is rejected with 422.\n"
+          description: "Opaque resume cursor \u2014 pass the previous response's `pagination.nextCursor` through verbatim; omit for the first page. A cursor is only valid for the operation and parameters that issued it; a malformed or foreign cursor is rejected."
         },
         limit: {
           type: "integer",
@@ -3339,7 +3709,7 @@ var generatedTools = [
           type: "string",
           enum: ["official", "verified", "community"],
           default: "official",
-          description: "Which slice of the catalog to return: `official` for\nRunpod-curated templates (default), `verified` for\nRunpod-verified community templates, or `community` for all other\npublicly shared templates.\n"
+          description: "Which slice of the catalog to return: `official` for Runpod-curated templates (default), `verified` for Runpod-verified community templates, or `community` for all other publicly shared templates."
         }
       }
     }
@@ -3433,12 +3803,12 @@ var generatedTools = [
         startTime: {
           type: "string",
           format: "date-time",
-          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago. Snapped down to the start of its bucketSize bucket so the window aligns with the returned records; provide a boundary-aligned value (e.g. midnight for bucketSize=day) to avoid widening.\n"
+          description: "Start of the billing period (RFC 3339). Defaults to 30 days ago."
         },
         endTime: {
           type: "string",
           format: "date-time",
-          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records.\n"
+          description: "End of the billing period (RFC 3339), exclusive. Defaults to now. Snapped up to the end of the bucketSize bucket it lands in (unless already on a boundary) so the window aligns with the returned records."
         },
         bucketSize: {
           $ref: "#/$defs/BillingBucketSize",
@@ -3447,7 +3817,7 @@ var generatedTools = [
         lastN: {
           type: "integer",
           minimum: 1,
-          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days". The resolved window is aligned to bucket boundaries: startTime is the start of the earliest bucket (e.g. midnight of the earliest day) and endTime is the end of the current bucket. Mutually exclusive with startTime/endTime; provide one or the other, not both.\n'
+          description: 'Return the last N buckets of bucketSize, ending with the current (in-progress) bucket \u2014 e.g. lastN=100 with bucketSize=day is "last 100 days".'
         },
         serverlessId: {
           type: "string",
@@ -3586,7 +3956,7 @@ var generatedTools = [
           type: "object",
           additionalProperties: false,
           minProperties: 1,
-          description: "Request body for updating a cluster. Only the cluster name can be\nchanged \u2014 this endpoint is a rename. Compute shape, type, and container\nconfiguration are fixed at creation.\n",
+          description: "Request body for updating a cluster. Only the cluster name can be changed \u2014 this endpoint is a rename. Compute shape, type, and container configuration are fixed at creation.",
           properties: {
             name: {
               type: "string",
@@ -3617,6 +3987,7 @@ var generatedTools = [
       idempotentHint: false,
       openWorldHint: true
     },
+    schemaOnDemand: true,
     inputSchema: {
       type: "object",
       properties: {
@@ -3632,11 +4003,10 @@ var generatedTools = [
       $defs: {
         BaseContainerConfig: {
           type: "object",
-          description: "Container configuration universal to every containerized resource. Compose ContainerConfig instead unless the resource cannot support a direct registry credential (clusters \u2014 there the registry credential arrives via a pod template, see CreateClusterRequest.templateId).\n",
           properties: {
             args: {
               type: "string",
-              description: 'The container\'s command, as a single raw string. This is the field `entrypoint` and `cmd` encode into, exposed in its stored form.\n\nTwo shapes are accepted. A bare shell string is treated as CMD and split into arguments, which is what the console\'s "Container start command" field writes. A JSON object of the form `{"entrypoint":[...],"cmd":[...]}` sets either or both explicitly.\n\nResponses always return both representations: `args` exactly as stored, plus the deconstructed `entrypoint` and `cmd`. Supplying `args` together with `entrypoint` or `cmd` is allowed only when they describe the same command, so a read-modify-write client can send back everything it received. Send `""` to clear, omit to leave unchanged.\n',
+              description: 'The container\'s command as one raw string: a bare shell string is treated as CMD and split into arguments; a JSON object `{"entrypoint":[...],"cmd":[...]}` sets either or both. Send `""` to clear; omit to leave unchanged.',
               examples: [
                 "--model meta-llama/Llama-3-8B --max-model-len 8192",
                 '{"entrypoint":["/bin/bash","-c"],"cmd":["python -u main.py"]}',
@@ -3648,7 +4018,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container CMD in exec form. When the image defines an ENTRYPOINT, this is the argument list passed to it. Encoded into the `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container CMD in exec form: when the image defines an ENTRYPOINT, the argument list passed to it. Encoded into `args`; supply both only when they describe the same command. Send `[]` to clear; omit to leave unchanged.",
               examples: [
                 ["--model", "meta-llama/Llama-3-8B", "--max-model-len", "8192"]
               ]
@@ -3658,7 +4028,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.",
               examples: [["/bin/bash", "-c"]]
             },
             disk: {
@@ -3718,7 +4088,7 @@ var generatedTools = [
             pools: {
               type: "array",
               minItems: 1,
-              description: 'Serverless GPU pool IDs (as returned by `GET /v2/catalog/gpus` in\n`pool`). Workers are placed on whichever listed pool has capacity.\nNarrow a pool down to specific cards with `excludedTypes`.\n\nOn `PATCH`, `pools` and `excludedTypes` are one selection and are\nreplaced together, so sending `pools` by itself **clears the\nexclusions**. Two cases:\n\n- **Changing pools, keeping exclusions** \u2014 send both fields in one\n  request: `{"gpu": {"pools": ["ADA_24"], "excludedTypes":\n  ["NVIDIA L40"]}}`. `GET` the endpoint first to read the current\n  `excludedTypes` and resend the ones that still apply to the new\n  pools; an exclusion naming a type outside `pools` is a 400.\n- **Changing only `count` or a CUDA constraint** \u2014 omit `pools`:\n  `{"gpu": {"minCudaVersion": "12.4"}}`. The pool list and the\n  exclusions are both left exactly as they are.\n\n`excludedTypes` documents the full rule.\n',
+              description: "Serverless GPU pool IDs (the `pool` field of `list-gpu-types`). Workers are placed on whichever listed pool has capacity; narrow to specific cards with `excludedTypes`. On update, `pools` and `excludedTypes` are replaced together, so sending `pools` alone clears the exclusions.",
               items: {
                 type: "string"
               },
@@ -3727,7 +4097,7 @@ var generatedTools = [
             excludedTypes: {
               type: "array",
               uniqueItems: true,
-              description: 'GPU **type** IDs to subtract from the selected pools \u2014 the `id`\nfield of `GET /v2/catalog/gpus`, the same identifiers pods take in\n`gpu.id`. Workers run on every type in `pools` except these. Omit to\nuse the whole pool.\n\nPools stay the unit of selection; types are the unit of\nsubtraction. There is no inclusive allowlist: a card later added to\none of your pools becomes eligible, which is the honest reading of\n"this pool, minus these".\n\nTied to `pools`, because the two together are one selection:\nsupplying `pools` replaces that selection wholesale, so a `PATCH`\nsending `pools` **without `excludedTypes`** **clears** them \u2014\nrestate them to keep them. A `PATCH` that omits `pools` leaves both\nthe pools and the exclusions untouched, so changing only a CUDA\nconstraint cannot widen a pinned endpoint.\n\nRejected with 400 if a value is not a GPU type in one of `pools`;\nupstream accepts unrecognized exclusions silently, so a typo would\notherwise produce a filter that does nothing. Surrounding whitespace\nis trimmed, so `" NVIDIA L40"` and `"NVIDIA L40"` mean the same card.\n',
+              description: "GPU type IDs to subtract from the selected pools \u2014 the `id` field of `list-gpu-types`, the same identifiers pods take in `gpu.id`. Workers run on every type in `pools` except these; omit to use the whole pool. A value outside `pools` is rejected. On update, restate them whenever you send `pools`, or they are cleared.",
               items: {
                 type: "string",
                 pattern: "^\\s*[^-\\s]"
@@ -3744,7 +4114,6 @@ var generatedTools = [
           }
         },
         ContainerConfig: {
-          description: "Reusable container configuration shared across templates, pods, and serverless endpoints. Adding a field here automatically propagates to all three resources.\n",
           allOf: [
             {
               $ref: "#/$defs/BaseContainerConfig"
@@ -3769,8 +4138,42 @@ var generatedTools = [
           ],
           unevaluatedProperties: false
         },
+        EndpointMount: {
+          type: "object",
+          required: ["volumeId"],
+          additionalProperties: false,
+          description: "A volume mounted on every worker, always at `/runpod-volume`. Custom paths are not supported yet.",
+          properties: {
+            volumeId: {
+              type: "string",
+              description: "ID of a NetworkVolume (under `network`) or GlobalVolume (under `global`).",
+              examples: ["vol_abc"]
+            }
+          }
+        },
+        EndpointMountsInput: {
+          type: "object",
+          additionalProperties: false,
+          description: "Volumes to mount on the endpoint's workers. A `network` or `global` list that is present replaces the endpoint's current list of that kind; an empty list detaches it, and an omitted list is left unchanged.",
+          properties: {
+            network: {
+              type: "array",
+              description: "Network volumes, at most one per data center.",
+              items: {
+                $ref: "#/$defs/EndpointMount"
+              }
+            },
+            global: {
+              type: "array",
+              maxItems: 1,
+              items: {
+                $ref: "#/$defs/EndpointMount"
+              }
+            }
+          }
+        },
         EndpointScaling: {
-          description: "Autoscaling signal \u2014 a discriminated union on `type`: `QUEUE_DELAY`\n(queue-based endpoints only) or `REQUEST_COUNT`. The scaler is chosen\nindependently of the endpoint's routing `type` and can be switched on\nupdate.\n",
+          description: "Autoscaling signal \u2014 a discriminated union on `type`: `QUEUE_DELAY` (queue-based endpoints only) or `REQUEST_COUNT`. The scaler is chosen independently of the endpoint's routing `type` and can be switched on update.",
           oneOf: [
             {
               $ref: "#/$defs/QueueDelayScaling"
@@ -3807,14 +4210,14 @@ var generatedTools = [
               type: "integer",
               minimum: 1,
               maximum: 3600,
-              description: "Seconds before idle workers scale down. Not applicable to queue-based\nendpoints scaling on `requestCount` \u2014 rejected on create/update and\nomitted from responses for that combination.\n",
+              description: "Seconds before idle workers scale down. Not applicable to queue-based endpoints scaling on `requestCount` \u2014 rejected on create/update and omitted from responses for that combination.",
               examples: [5]
             }
           }
         },
         FlashBoot: {
           type: "string",
-          description: "FlashBoot cold-start acceleration mode.\n- `OFF`                \u2014 disabled\n- `FLASHBOOT`          \u2014 enabled\n- `PRIORITY_FLASHBOOT` \u2014 enabled with priority capacity\n",
+          description: "FlashBoot cold-start acceleration mode. `OFF` \u2014 disabled. `FLASHBOOT` \u2014 enabled. `PRIORITY_FLASHBOOT` \u2014 enabled with priority capacity",
           enum: ["OFF", "FLASHBOOT", "PRIORITY_FLASHBOOT"]
         },
         QueueDelayScaling: {
@@ -3840,7 +4243,7 @@ var generatedTools = [
         RequestCountScaling: {
           type: "object",
           additionalProperties: false,
-          description: "Scale on concurrent in-flight requests per worker. Required for\nload-balancing endpoints; also selectable for queue-based.\n",
+          description: "Scale on concurrent in-flight requests per worker. Required for load-balancing endpoints; also selectable for queue-based.",
           required: ["type", "requestCount"],
           properties: {
             type: {
@@ -3857,7 +4260,7 @@ var generatedTools = [
           }
         },
         UpdateEndpointGpuConfig: {
-          description: "Partial GPU update \u2014 every field is optional and an omitted one is left\nunchanged. Unlike create, `pools` is optional, so changing only a CUDA\nconstraint does not require resending the pool list.\n\n`excludedTypes` requires `pools`, because the two are one selection and\nonly a supplied `pools` replaces it \u2014 an exclusion on its own would\notherwise be silently dropped.\n",
+          description: "Partial GPU update \u2014 every field is optional and an omitted one is left unchanged. `pools` is optional here, so a CUDA-only change does not resend the pool list; `excludedTypes` needs `pools` in the same request, and `pools` sent alone clears the exclusions.",
           allOf: [
             {
               $ref: "#/$defs/BaseEndpointGpuConfig"
@@ -3874,12 +4277,12 @@ var generatedTools = [
                     type: "string",
                     pattern: "^\\d+\\.\\d+$"
                   },
-                  description: "Acceptable CUDA versions for worker placement, as\n`major.minor`. An explicit `[]` clears the constraint;\nomitting the field leaves it unchanged. Takes effect as\nworkers are replaced.\n\nA non-empty set is mutually exclusive with minCudaVersion (400\nif both are sent). An explicit `[]` states no constraint, so it\nmay accompany a floor.\nSetting one does not clear the other \u2014 clear it explicitly in\nthe same patch if the endpoint already carries it.\n"
+                  description: "Acceptable CUDA versions for worker placement, as `major.minor`. An explicit `[]` clears the constraint; omitting the field leaves it unchanged. A non-empty set cannot be sent with minCudaVersion, and setting one does not clear the other: clear the other explicitly in the same patch. Takes effect as workers are replaced."
                 },
                 minCudaVersion: {
                   type: "string",
                   pattern: "^$|^\\d+\\.\\d+$",
-                  description: 'Lowest acceptable CUDA version for worker placement, as\n`major.minor`. An explicit `""` clears the floor; omitting the\nfield leaves it unchanged. Takes effect as workers are\nreplaced.\n\nMutually exclusive with a non-empty allowedCudaVersions (400 if\nboth are sent); an explicit `[]` there states no constraint and\nmay accompany this floor.\n',
+                  description: 'Lowest acceptable CUDA version for worker placement, as `major.minor`. An explicit `""` clears the floor; omitting the field leaves it unchanged. Cannot be sent with a non-empty allowedCudaVersions (an explicit `[]` there is fine); clear that explicitly in the same patch if the endpoint carries one. Takes effect as workers are replaced.',
                   examples: ["12.1"]
                 }
               }
@@ -3900,7 +4303,7 @@ var generatedTools = [
                   type: "array",
                   minItems: 1,
                   uniqueItems: true,
-                  description: "Complete replacement CPU selection. Valid only for an existing CPU\nendpoint; endpoint compute family cannot be changed.\n",
+                  description: "Complete replacement CPU selection. Valid only for an existing CPU endpoint; endpoint compute family cannot be changed.",
                   items: {
                     $ref: "#/$defs/CreateCpuConfig"
                   }
@@ -3926,14 +4329,19 @@ var generatedTools = [
                   type: "array",
                   items: {
                     type: "string"
-                  }
+                  },
+                  description: "Network volume IDs. Pass `[]` to detach.",
+                  deprecated: true
+                },
+                mounts: {
+                  $ref: "#/$defs/EndpointMountsInput"
                 },
                 scaling: {
                   $ref: "#/$defs/EndpointScaling"
                 },
                 templateId: {
                   type: "string",
-                  description: "ID of a serverless template whose container settings are\napplied as if they were provided in this PATCH body (image,\nargs, disk, ports, env, registry). Explicit body fields\noverride the template's; `env` merges template and body per\nkey (body wins) and, per PATCH semantics, replaces the\nendpoint's env. One-time application \u2014 no link to the\ntemplate is retained. Must be one of your templates or a\npublic template (unknown or inaccessible ID \u2192 404); must be\na serverless template (\u2192 422).\n",
+                  description: "ID of a serverless template whose container settings (image, args, disk, ports, env, registry) are applied as if sent in this PATCH body; explicit body fields win. `env` merges template and body per key (body wins) and replaces the endpoint's existing env. Applied once; must be a serverless template you can access.",
                   examples: ["30zmvf89kd"]
                 },
                 timeout: {
@@ -3944,8 +4352,59 @@ var generatedTools = [
                 }
               }
             }
-          ],
-          unevaluatedProperties: false
+          ],
+          unevaluatedProperties: false
+        }
+      }
+    }
+  },
+  {
+    name: "update-global-volume",
+    operationId: "updateGlobalVolume",
+    description: "Update a global volume. Updates mutable fields on a global volume. Only provided fields are changed.",
+    method: "PATCH",
+    path: "/v2/global-volumes/{id}",
+    params: [
+      {
+        name: "id",
+        location: "path"
+      }
+    ],
+    hasBody: true,
+    annotations: {
+      readOnlyHint: false,
+      destructiveHint: false,
+      idempotentHint: false,
+      openWorldHint: true
+    },
+    inputSchema: {
+      type: "object",
+      properties: {
+        id: {
+          type: "string",
+          description: "Global volume identifier"
+        },
+        body: {
+          $ref: "#/$defs/UpdateGlobalVolumeRequest"
+        }
+      },
+      required: ["id", "body"],
+      $defs: {
+        UpdateGlobalVolumeRequest: {
+          type: "object",
+          additionalProperties: false,
+          minProperties: 1,
+          description: "Only the provided fields are updated. At least one field must be present; an empty body is rejected.",
+          properties: {
+            name: {
+              type: "string",
+              minLength: 1,
+              maxLength: 191,
+              pattern: "[^\\p{Z}\\t-\\r\x85]",
+              description: "New human-readable name, 1\u2013191 characters with at least one non-whitespace character. Leading and trailing whitespace is trimmed.",
+              examples: ["renamed-volume"]
+            }
+          }
         }
       }
     }
@@ -3986,7 +4445,7 @@ var generatedTools = [
           type: "object",
           additionalProperties: false,
           minProperties: 1,
-          description: "Only the provided fields are updated. At least one field must be\npresent; an empty body is rejected.\n",
+          description: "Only the provided fields are updated. At least one field must be present; an empty body is rejected.",
           properties: {
             name: {
               type: "string",
@@ -3998,7 +4457,7 @@ var generatedTools = [
               type: "integer",
               minimum: 10,
               maximum: 4096,
-              description: "New size in GB. Must be greater than or equal to the current size \u2014\nnetwork volume storage cannot be reduced.\n",
+              description: "New size in GB. Must be greater than or equal to the current size \u2014 network volume storage cannot be reduced.",
               examples: [100]
             }
           }
@@ -4025,6 +4484,7 @@ var generatedTools = [
       idempotentHint: false,
       openWorldHint: true
     },
+    schemaOnDemand: true,
     inputSchema: {
       type: "object",
       properties: {
@@ -4040,11 +4500,10 @@ var generatedTools = [
       $defs: {
         BaseContainerConfig: {
           type: "object",
-          description: "Container configuration universal to every containerized resource. Compose ContainerConfig instead unless the resource cannot support a direct registry credential (clusters \u2014 there the registry credential arrives via a pod template, see CreateClusterRequest.templateId).\n",
           properties: {
             args: {
               type: "string",
-              description: 'The container\'s command, as a single raw string. This is the field `entrypoint` and `cmd` encode into, exposed in its stored form.\n\nTwo shapes are accepted. A bare shell string is treated as CMD and split into arguments, which is what the console\'s "Container start command" field writes. A JSON object of the form `{"entrypoint":[...],"cmd":[...]}` sets either or both explicitly.\n\nResponses always return both representations: `args` exactly as stored, plus the deconstructed `entrypoint` and `cmd`. Supplying `args` together with `entrypoint` or `cmd` is allowed only when they describe the same command, so a read-modify-write client can send back everything it received. Send `""` to clear, omit to leave unchanged.\n',
+              description: 'The container\'s command as one raw string: a bare shell string is treated as CMD and split into arguments; a JSON object `{"entrypoint":[...],"cmd":[...]}` sets either or both. Send `""` to clear; omit to leave unchanged.',
               examples: [
                 "--model meta-llama/Llama-3-8B --max-model-len 8192",
                 '{"entrypoint":["/bin/bash","-c"],"cmd":["python -u main.py"]}',
@@ -4056,7 +4515,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container CMD in exec form. When the image defines an ENTRYPOINT, this is the argument list passed to it. Encoded into the `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container CMD in exec form: when the image defines an ENTRYPOINT, the argument list passed to it. Encoded into `args`; supply both only when they describe the same command. Send `[]` to clear; omit to leave unchanged.",
               examples: [
                 ["--model", "meta-llama/Llama-3-8B", "--max-model-len", "8192"]
               ]
@@ -4066,7 +4525,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.",
               examples: [["/bin/bash", "-c"]]
             },
             disk: {
@@ -4102,8 +4561,22 @@ var generatedTools = [
             }
           }
         },
+        BaseMounts: {
+          type: "object",
+          properties: {
+            persistent: {
+              $ref: "#/$defs/PersistentMount"
+            },
+            network: {
+              type: "array",
+              maxItems: 1,
+              items: {
+                $ref: "#/$defs/NetworkMount"
+              }
+            }
+          }
+        },
         ContainerConfig: {
-          description: "Reusable container configuration shared across templates, pods, and serverless endpoints. Adding a field here automatically propagates to all three resources.\n",
           allOf: [
             {
               $ref: "#/$defs/BaseContainerConfig"
@@ -4120,28 +4593,51 @@ var generatedTools = [
             }
           ]
         },
-        Mounts: {
+        GlobalMount: {
           type: "object",
+          required: ["volumeId"],
           additionalProperties: false,
-          description: "Storage mounts attached to a pod. At-most-one of `persistent` or\n`network` may be set today (mutually exclusive, enforced at the\nhandler with 400 if both are present). The `network` field is an\narray for forward compatibility with eventual multi-network-volume\nsupport, but `maxItems` is 1 today.\n\nPATCH semantics:\n- Omitting `mounts` or sending `{}` leaves the existing mount\n  unchanged.\n- An explicit `network: []` is rejected with 400 (clearing mounts\n  is not supported).\n- Mount kind is fixed at create \u2014 a PATCH that introduces a kind\n  not present at create (persistent on a network pod, network on\n  a persistent pod, or any mount on a previously-mountless pod)\n  is rejected with 400.\n- The `volumeId` of a network mount is immutable; a PATCH that\n  names a different `volumeId` is rejected with 400.\n- Partial mounts are not supported \u2014 every mount entry must\n  include the full schema (`size` + `path` for persistent,\n  `volumeId` + `path` for network). Missing required fields \u2192 422.\n",
+          description: "Reference to a GlobalVolume. Usage-billed, so it has no size. The underlying `volumeId` is immutable post-create; the mount `path` may be changed via PATCH.",
           properties: {
-            persistent: {
-              $ref: "#/$defs/PersistentMount"
+            volumeId: {
+              type: "string",
+              description: "ID of an existing GlobalVolume.",
+              examples: ["gsb_abc"]
             },
-            network: {
-              type: "array",
-              maxItems: 1,
-              items: {
-                $ref: "#/$defs/NetworkMount"
-              }
+            path: {
+              type: "string",
+              description: "Mount path inside the container. Optional on create, where a default is chosen for you; a PATCH that omits it leaves the path unchanged.",
+              examples: ["/data"]
             }
           }
         },
+        Mounts: {
+          description: "Storage mounts for the pod: `persistent` (size + path) excludes both `network` and `global`; `network` (volumeId + path) and `global` (volumeId, optional path) take one entry each and can be combined. Every entry needs its full shape. On update, omit the field to leave the mounts as they are; the mount kind and a volume's volumeId cannot change, and mounts cannot be cleared.",
+          allOf: [
+            {
+              $ref: "#/$defs/BaseMounts"
+            },
+            {
+              type: "object",
+              properties: {
+                global: {
+                  type: "array",
+                  maxItems: 1,
+                  description: "At most one global volume, on GPU and CPU pods, with no data center constraint. Combinable with `network` but not with `persistent` (rejected): upstream drops a persistent disk whenever a volume is attached.",
+                  items: {
+                    $ref: "#/$defs/GlobalMount"
+                  }
+                }
+              }
+            }
+          ],
+          unevaluatedProperties: false
+        },
         NetworkMount: {
           type: "object",
           required: ["volumeId", "path"],
           additionalProperties: false,
-          description: "Reference to a NetworkVolume. Custom paths are honored at runtime on\nboth GPU and CPU pods. The underlying `volumeId` is immutable\npost-create; the mount `path` may be changed via PATCH.\n",
+          description: "Reference to a NetworkVolume. Custom paths are honored at runtime on both GPU and CPU pods. The underlying `volumeId` is immutable post-create; the mount `path` may be changed via PATCH.",
           properties: {
             volumeId: {
               type: "string",
@@ -4159,7 +4655,7 @@ var generatedTools = [
           type: "object",
           required: ["size", "path"],
           additionalProperties: false,
-          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data\ndoes not survive a host failure. Disallowed on CPU pods. Mutually\nexclusive with NetworkMount. Deprecated: prefer NetworkMount for any\ndata you cannot recreate.\n",
+          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data does not survive a host failure. Disallowed on CPU pods. Mutually exclusive with NetworkMount. Deprecated: prefer NetworkMount for any data you cannot recreate.",
           properties: {
             size: {
               type: "integer",
@@ -4184,7 +4680,7 @@ var generatedTools = [
               properties: {
                 globalNetworking: {
                   type: "boolean",
-                  description: "Enable (true) or disable (false) global networking. Takes effect on the next pod start/restart, not live. Requires an NVIDIA GPU and a global-networking-enabled data center (both enforced upstream). See `GET /v2/catalog/datacenters` (`globalNetwork`) for eligible data centers."
+                  description: "Enable (true) or disable (false) global networking. Takes effect on the next pod start/restart, not live. Requires an NVIDIA GPU and a global-networking-enabled data center (both enforced upstream)."
                 },
                 locked: {
                   type: "boolean",
@@ -4199,7 +4695,7 @@ var generatedTools = [
                 },
                 templateId: {
                   type: "string",
-                  description: "ID of a pod template whose container settings are applied as\nif they were provided in this PATCH body (image, args, disk,\nports, env, registry \u2014 mounts are not applied on update).\nExplicit body fields override the template's; `env` merges\ntemplate and body per key (body wins) and, per PATCH\nsemantics, replaces the pod's env. One-time application \u2014\nno link to the template is retained. Must be one of your\ntemplates or a public template (unknown or inaccessible ID\n\u2192 404); must not be a serverless template (\u2192 422).\n",
+                  description: "ID of a pod template whose container settings (image, args, disk, ports, env, registry; not mounts) are applied as if sent in this PATCH body; explicit body fields win. `env` merges template and body per key (body wins) and replaces the pod's existing env. Applied once; must be a non-serverless template you can access.",
                   examples: ["30zmvf89kd"]
                 }
               }
@@ -4229,6 +4725,7 @@ var generatedTools = [
       idempotentHint: false,
       openWorldHint: true
     },
+    schemaOnDemand: true,
     inputSchema: {
       type: "object",
       properties: {
@@ -4246,7 +4743,7 @@ var generatedTools = [
           type: "object",
           additionalProperties: false,
           minProperties: 1,
-          description: "Only the provided fields are updated; at least one field is required.\nThe secret's `name` is immutable.\n",
+          description: "Only the provided fields are updated; at least one field is required. The secret's `name` is immutable.",
           properties: {
             value: {
               type: "string",
@@ -4298,7 +4795,7 @@ var generatedTools = [
                 type: "string",
                 pattern: "^(ssh|ecdsa|sk)-[^\\s]+ [^\\s]+([ \\t][^\\n\\r]*)?$"
               },
-              description: "The full set of SSH public keys to register \u2014 this is a complete\nreplacement, not a merge. Each entry is an authorized_keys-style\nline: `<type> <base64-key> [comment]`, e.g. from\n`~/.ssh/id_ed25519.pub`. Send `[]` to remove all keys. These keys\nare provisioned into pods created with `startSsh` and\nauthenticate both SSH paths reported in the pod's `ssh` block.\n",
+              description: "The full set of SSH public keys to register \u2014 this is a complete replacement, not a merge. Each entry is an authorized_keys-style line: `<type> <base64-key> [comment]`, e.g. from `~/.ssh/id_ed25519.pub`. Send `[]` to remove all keys.",
               examples: [
                 [
                   "ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAILXGDN/SclOozk1xsDztpmhGiKkkrfQB9SKoO8dSIQQZ me@example.com"
@@ -4329,6 +4826,7 @@ var generatedTools = [
       idempotentHint: false,
       openWorldHint: true
     },
+    schemaOnDemand: true,
     inputSchema: {
       type: "object",
       properties: {
@@ -4343,11 +4841,10 @@ var generatedTools = [
       $defs: {
         BaseContainerConfig: {
           type: "object",
-          description: "Container configuration universal to every containerized resource. Compose ContainerConfig instead unless the resource cannot support a direct registry credential (clusters \u2014 there the registry credential arrives via a pod template, see CreateClusterRequest.templateId).\n",
           properties: {
             args: {
               type: "string",
-              description: 'The container\'s command, as a single raw string. This is the field `entrypoint` and `cmd` encode into, exposed in its stored form.\n\nTwo shapes are accepted. A bare shell string is treated as CMD and split into arguments, which is what the console\'s "Container start command" field writes. A JSON object of the form `{"entrypoint":[...],"cmd":[...]}` sets either or both explicitly.\n\nResponses always return both representations: `args` exactly as stored, plus the deconstructed `entrypoint` and `cmd`. Supplying `args` together with `entrypoint` or `cmd` is allowed only when they describe the same command, so a read-modify-write client can send back everything it received. Send `""` to clear, omit to leave unchanged.\n',
+              description: 'The container\'s command as one raw string: a bare shell string is treated as CMD and split into arguments; a JSON object `{"entrypoint":[...],"cmd":[...]}` sets either or both. Send `""` to clear; omit to leave unchanged.',
               examples: [
                 "--model meta-llama/Llama-3-8B --max-model-len 8192",
                 '{"entrypoint":["/bin/bash","-c"],"cmd":["python -u main.py"]}',
@@ -4359,7 +4856,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container CMD in exec form. When the image defines an ENTRYPOINT, this is the argument list passed to it. Encoded into the `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container CMD in exec form: when the image defines an ENTRYPOINT, the argument list passed to it. Encoded into `args`; supply both only when they describe the same command. Send `[]` to clear; omit to leave unchanged.",
               examples: [
                 ["--model", "meta-llama/Llama-3-8B", "--max-model-len", "8192"]
               ]
@@ -4369,7 +4866,7 @@ var generatedTools = [
               items: {
                 type: "string"
               },
-              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.\n",
+              description: "Container ENTRYPOINT in exec form, overriding the image's own. Encoded into `args` field; supplying both is allowed only when they describe the same command. Send `[]` to clear, omit to leave unchanged.",
               examples: [["/bin/bash", "-c"]]
             },
             disk: {
@@ -4406,7 +4903,6 @@ var generatedTools = [
           }
         },
         ContainerConfig: {
-          description: "Reusable container configuration shared across templates, pods, and serverless endpoints. Adding a field here automatically propagates to all three resources.\n",
           allOf: [
             {
               $ref: "#/$defs/BaseContainerConfig"
@@ -4427,7 +4923,7 @@ var generatedTools = [
           type: "object",
           required: ["size", "path"],
           additionalProperties: false,
-          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data\ndoes not survive a host failure. Disallowed on CPU pods. Mutually\nexclusive with NetworkMount. Deprecated: prefer NetworkMount for any\ndata you cannot recreate.\n",
+          description: "Host-local persistent storage. Pinned to the pod's host machine \u2014 data does not survive a host failure. Disallowed on CPU pods. Mutually exclusive with NetworkMount. Deprecated: prefer NetworkMount for any data you cannot recreate.",
           properties: {
             size: {
               type: "integer",
@@ -4444,13 +4940,13 @@ var generatedTools = [
         },
         TemplateCategory: {
           type: "string",
-          description: "Controls how the template is grouped and filtered in the Runpod console.\nIt does not affect hardware selection, scheduling, or billing.\n- `CPU`    \u2014 CPU-only workloads\n- `NVIDIA` \u2014 NVIDIA GPU workloads\n- `AMD`    \u2014 AMD GPU workloads\n",
+          description: "Controls how the template is grouped and filtered in the Runpod console. It does not affect hardware selection, scheduling, or billing. `CPU` \u2014 CPU-only workloads. `NVIDIA` \u2014 NVIDIA GPU workloads. `AMD` \u2014 AMD GPU workloads",
           enum: ["CPU", "NVIDIA", "AMD"]
         },
         TemplateMounts: {
           type: "object",
           additionalProperties: false,
-          description: "Storage mounts attached to a template. Templates support only a\nsingle persistent mount today; any `network` property is rejected\nwith 422 by the schema validator.\n\nPATCH semantics: omitting `mounts` or sending `{}` leaves the\nexisting mount unchanged.\n",
+          description: "Storage mounts for containers from this template: a single `persistent` mount only (`network` is rejected). On update, omitting `mounts` or sending `{}` leaves the existing mount unchanged.",
           properties: {
             persistent: {
               $ref: "#/$defs/PersistentMount"
@@ -5199,6 +5695,8 @@ var listHubRepos = {
 };
 var deployHubRepo = {
   name: "deploy-hub-repo",
+  // Two sequential upstream calls: together they can outlast a client's request timeout.
+  holdsResponse: true,
   annotations: write,
   description: "Deploy a Runpod Hub repo's listed release as a new Serverless endpoint (the same as clicking Deploy on the Hub). Identify the repo by `repo` (\"owner/name\" from list-hub-repos) or by `hubReleaseId`. The release supplies the prebuilt image, container disk, CUDA constraints, and env-var defaults; pass `env` to override or fill in values (required keys without a default must be provided \u2014 check list-hub-repos with includeConfig:true for the schema). GPU selection comes from the release config when it specifies one; otherwise pass gpuIds (GPU pool names, e.g. 'ADA_24' or 'ADA_80_PRO,AMPERE_80'). Uses the authenticated GraphQL API (no REST home for Hub deploys yet).",
   inputSchema: {
@@ -5687,6 +6185,7 @@ var runEndpoint = {
 };
 var runsyncEndpoint = {
   name: "runsync-endpoint",
+  holdsResponse: true,
   annotations: write,
   description: "Submit a synchronous job to a Serverless endpoint and wait for the result. Best for fast tasks: if the job outlives the wait, the response returns a job ID and a non-terminal status (IN_QUEUE or IN_PROGRESS) to poll with get-job-status. Max payload 20 MB; results expire after 1 minute. The wait parameter extends the server-side wait" + (HOSTED ? ` \u2014 this server runs behind a 60-second gateway deadline, so waits clamp to ${HTTP_LONG_POLL_BUDGET_MS} ms, which is also the effective default. For longer jobs use run-endpoint + get-job-status.` : " to up to 5 minutes (300000 ms); it defaults to 90000 (90 seconds)."),
   inputSchema: {
@@ -5731,6 +6230,7 @@ var runsyncEndpoint = {
 };
 var getJobStatus = {
   name: "get-job-status",
+  holdsResponse: true,
   annotations: readOnly,
   description: "Check the status of a Serverless job. Returns the current status and output when complete. Job statuses: IN_QUEUE, IN_PROGRESS, COMPLETED, FAILED, CANCELLED, TIMED_OUT. Pass `wait` (milliseconds, up to " + STATUS_WAIT_MAX_MS + ") to BLOCK server-side, polling until the job reaches a terminal status or the budget expires \u2014 use it to ride out a cold start (a first job on a fresh worker pulls the image and loads the model, commonly 1\u20135+ minutes) in a single call instead of returning on the first IN_QUEUE and tight-looping the tool. Without `wait` it returns the current status immediately (a single check). If the budget expires before a terminal status, it returns the latest non-terminal status with pollingTimedOut:true \u2014 call again (with `wait`) to keep blocking. IMPORTANT: a job stuck IN_QUEUE does not necessarily mean a capacity shortage \u2014 a worker may have spun up and crash-looped. When the job is IN_QUEUE this tool attaches a workerHealth summary and a hint. To dig deeper call list-endpoint-workers, then read stream-worker-logs regardless of the status a worker reports: a container failing to start loops while its worker still shows RUNNING or THROTTLED and unhealthy stays 0, so UNHEALTHY is one symptom of a crash loop and not a test for it. The logs are the authority.",
   inputSchema: {
@@ -5779,6 +6279,7 @@ var getJobStatus = {
 };
 var streamJob = {
   name: "stream-job",
+  holdsResponse: true,
   annotations: readOnly,
   description: `Retrieve streaming output from a Serverless job. The worker must support streaming output. Polls /stream/{jobId} and collects chunks until the status is COMPLETED, FAILED, CANCELLED, or TIMED_OUT, for up to ${formatBudget(STREAM_BUDGET_MS)}. If the budget expires first, returns the chunks collected so far with pollingTimedOut: true \u2014 call stream-job again to resume where it left off, or get-job-status to check without streaming.`,
   inputSchema: {
@@ -6011,8 +6512,11 @@ function publicGraphqlBase(env) {
 function authedGraphqlBase(env) {
   return pick(env.RUNPOD_AUTHED_GRAPHQL_URL, "https://api.runpod.io/graphql");
 }
+function knowledgeSearchBase(env) {
+  return pick(env.KNOWLEDGE_SEARCH_URL, "") || void 0;
+}
 
-// node_modules/.pnpm/@runpod+typescript-api-sdk@0.2.0/node_modules/@runpod/typescript-api-sdk/dist/index.js
+// node_modules/.pnpm/@runpod+typescript-api-sdk@0.3.0/node_modules/@runpod/typescript-api-sdk/dist/index.js
 var PATH_PARAM_RE = /\{[^{}]+\}/g;
 var supportsRequestInitExt = () => {
   return typeof process === "object" && Number.parseInt(process?.versions?.node?.substring(0, 2)) >= 18 && process.versions.undici;
@@ -6909,9 +7413,9 @@ function isDigit(char) {
     return false;
   return isDigitRegex.test(char);
 }
-function parseRetryAfter(header, now = Date.now()) {
-  if (header === null) return void 0;
-  const value = header.trim();
+function parseRetryAfter(header2, now = Date.now()) {
+  if (header2 === null) return void 0;
+  const value = header2.trim();
   if (/^\d+$/.test(value)) {
     const ms = Number(value) * 1e3;
     return Number.isSafeInteger(ms) ? ms : void 0;
@@ -7416,6 +7920,7 @@ function logParams(args) {
 }
 var streamPodLogs = {
   name: "stream-pod-logs",
+  holdsResponse: true,
   annotations: readOnly,
   description: "Read a bounded snapshot of a Pod's logs (container and/or system source). Holds the live stream open for maxWaitMs and returns the parsed log lines; `truncated: true` means the byte cap cut the output. Use `since` to resume from a timestamp.",
   inputSchema: {
@@ -7438,6 +7943,7 @@ var streamPodLogs = {
 };
 var streamWorkerLogs = {
   name: "stream-worker-logs",
+  holdsResponse: true,
   annotations: readOnly,
   description: "Read a bounded snapshot of a Serverless worker's logs (container and/or system source). Holds the live stream open for maxWaitMs and returns the parsed log lines; `truncated: true` means the byte cap cut the output. Get worker IDs from list-endpoint-workers. Use this whenever a job stays IN_QUEUE while workers exist, whatever status they report \u2014 a crash-looping container often still shows RUNNING or THROTTLED with unhealthy 0, so these logs, not the worker status, are what identify a crash loop.",
   inputSchema: {
@@ -7702,6 +8208,7 @@ var source = {
 var listGuides = {
   name: "list-guides",
   annotations: bundledRead,
+  recordedArguments: ["kind", "query", "concept"],
   description: "List the official Runpod guides: skills (how to work with each tool), reference docs (how Runpod works), and golden paths (live-verified end-to-end walkthroughs such as a dev pod, a Whisper endpoint or a fine-tune-to-serverless pipeline). Read one with read-guide. For a golden path, mcp says whether an agent with only these MCP tools can finish it: full (follow it with tool calls), partial (use it for the plan and facts; name the step the user must run in a terminal), none (reference only). Skills and reference docs have mcp null. concepts lists the concept ids it works with (see lookup-concept). Filter with kind, query or concept.",
   inputSchema: {
     type: "object",
@@ -7737,6 +8244,8 @@ var listGuides = {
 var readGuide = {
   name: "read-guide",
   annotations: bundledRead,
+  recordedArguments: ["id"],
+  recordedGuideKind: (args) => knowledge.guides.find((guide) => guide.id === String(args.id).trim())?.kind,
   description: 'Read one official Runpod guide in full, by the id list-guides returns (for example "runpod-usage/storage" or "golden-path/06-dev-pod").',
   inputSchema: {
     type: "object",
@@ -7774,6 +8283,7 @@ var readGuide = {
 var lookupConcept = {
   name: "lookup-concept",
   annotations: bundledRead,
+  recordedArguments: ["concept"],
   description: 'Look up one Runpod concept (pod, network volume, machine, worker, template, \u2026) by id, name or alias; case, plurals and small typos are tolerated, and `matched_as` shows the key used when the match was not exact. Returns its summary, fields, states, relations and rules, each rule with the public source it comes from, the concepts that point at it, and the linked guides: examples (golden paths that use it, with their mcp level) and docs (skills and reference docs that explain it). Use it for exact facts, for example "is a volume disk tied to one machine" or "what happens to a stopped pod\'s GPU".',
   inputSchema: {
     type: "object",
@@ -7826,6 +8336,7 @@ var lookupConcept = {
 var searchConcepts = {
   name: "search-concepts",
   annotations: bundledRead,
+  recordedArguments: ["query", "limit"],
   description: 'Search the rules of the Runpod concept graph by words, for example "restart stopped pod no gpu" or "network volume data center". Returns the best-matching rules with their concept and public evidence. Follow up with lookup-concept for the full concept.',
   inputSchema: {
     type: "object",
@@ -7890,9 +8401,143 @@ var guideResources = knowledge.guides.map((guide) => ({
   text: guide.body
 }));
 
+// src/specgen/prose.ts
+var ABBREVIATIONS = ["e.g.", "i.e.", "etc.", "vs."];
+var DOT = "\0";
+function splitSentences(text) {
+  let guarded = text;
+  for (const abbreviation of ABBREVIATIONS) {
+    guarded = guarded.replaceAll(
+      abbreviation,
+      abbreviation.replaceAll(".", DOT)
+    );
+  }
+  return guarded.split(/(?<=[.!?])\s+(?=[A-Z`"'([])/).map((sentence) => sentence.replaceAll(DOT, ".").trim()).filter(Boolean);
+}
+function leadSentence(text) {
+  return splitSentences(text)[0] ?? text;
+}
+
+// src/specgen/on-demand.ts
+var DESCRIBE_TOOL_NAME = "describe-tool";
+var ON_DEMAND_POINTER = ` Call ${DESCRIBE_TOOL_NAME} for the full argument schema before the first use.`;
+function listedAnnotations(annotations) {
+  const { idempotentHint, ...listed } = annotations;
+  return idempotentHint ? annotations : listed;
+}
+function listTool(tool) {
+  const { name, description, inputSchema } = tool;
+  const annotations = listedAnnotations(tool.annotations);
+  if (!tool.schemaOnDemand)
+    return { name, description, inputSchema, annotations };
+  return {
+    name,
+    description: description + ON_DEMAND_POINTER,
+    inputSchema: callShape(inputSchema),
+    annotations
+  };
+}
+function callShape(schema) {
+  const defs = schema.$defs ?? {};
+  const resolve = (node) => {
+    if (typeof node.$ref !== "string") return node;
+    const name = node.$ref.split("/").pop() ?? "";
+    return defs[name] ?? {};
+  };
+  const properties = Object.fromEntries(
+    Object.entries(schema.properties ?? {}).map(
+      ([key, value]) => [key, shallowProperty(resolve(value), resolve)]
+    )
+  );
+  return {
+    type: "object",
+    properties,
+    ...schema.required ? { required: schema.required } : {}
+  };
+}
+function shallowProperty(node, resolve) {
+  const out = {};
+  const type = schemaType(node, resolve);
+  if (type !== void 0) out.type = type;
+  if (type === "array") {
+    const items = resolve(node.items ?? {});
+    const itemType = schemaType(items, resolve);
+    if (itemType !== void 0) out.items = { type: itemType };
+  }
+  if (Array.isArray(node.enum)) out.enum = node.enum;
+  if (typeof node.description === "string")
+    out.description = leadSentence(node.description);
+  return out;
+}
+function schemaType(node, resolve) {
+  if (node.type !== void 0) return node.type;
+  if (node.properties || node.allOf) return "object";
+  const branches2 = node.anyOf ?? node.oneOf;
+  const allObjects = branches2?.every(
+    (branch) => schemaType(resolve(branch), resolve) === "object"
+  ) ?? false;
+  return allObjects ? "object" : void 0;
+}
+
+// src/specgen/tools/describe.ts
+function createDescribeTool(served) {
+  return {
+    name: DESCRIBE_TOOL_NAME,
+    description: "Return one tool's full definition: its description, complete argument schema (every field, nested objects, constraints) and annotations. Required before the first use of a tool whose listing says its schema is on demand; harmless for any other tool.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        name: {
+          type: "string",
+          description: "A tool name exactly as this server's tool list spells it."
+        }
+      },
+      required: ["name"]
+    },
+    annotations: {
+      readOnlyHint: true,
+      destructiveHint: false,
+      idempotentHint: true,
+      openWorldHint: false
+    },
+    handler: async (_ctx, args) => {
+      const name = String(args.name ?? "");
+      const tools = served();
+      const tool = tools.find((candidate) => candidate.name === name);
+      if (!tool) {
+        return {
+          ok: false,
+          status: 404,
+          payload: {
+            error: `No tool named ${name}.`,
+            suggestions: similarNames(
+              name,
+              tools.map((candidate) => candidate.name)
+            ),
+            hint: "Tool names are exactly as this server lists them; the suggestions share a word with the name you sent."
+          }
+        };
+      }
+      const { description, inputSchema, annotations } = tool;
+      return ok({ name: tool.name, description, inputSchema, annotations });
+    }
+  };
+}
+function similarNames(query, names) {
+  const wanted = new Set(words2(query));
+  return names.map((name) => {
+    const own = new Set(words2(name));
+    const shared = [...own].filter((word) => wanted.has(word)).length;
+    return { name, score: shared / (/* @__PURE__ */ new Set([...own, ...wanted])).size };
+  }).filter((entry) => entry.score > 0).sort((a, b) => b.score - a.score || a.name.localeCompare(b.name)).slice(0, 8).map((entry) => entry.name);
+}
+function words2(name) {
+  return name.toLowerCase().split(/[-_\s]+/).filter(Boolean).map((word) => word.replace(/s$/, ""));
+}
+
 // src/server.ts
 var SERVER_NAME = "Runpod API Server";
-var SERVER_VERSION = true ? "4.2.0" : "dev";
+var SERVER_VERSION = true ? "4.4.0" : "dev";
 
 // src/specgen/analytics.ts
 import { createHmac } from "crypto";
@@ -7908,7 +8553,22 @@ function analyticsCallerId(apiKey, accountId) {
   const prefix = accountId ? "a:" : "k:";
   return prefix + createHmac("sha256", salt).update(identity).digest("hex").slice(0, 16);
 }
-function captureToolCall(apiKey, event) {
+var MAX_RECORDED_VALUE_LENGTH = 200;
+function bounded(value) {
+  return value.slice(0, MAX_RECORDED_VALUE_LENGTH);
+}
+function recordedArgumentValues(names, args) {
+  if (!names?.length) return void 0;
+  const values = {};
+  for (const name of names) {
+    const value = args[name];
+    if (typeof value === "string" || typeof value === "number") {
+      values[name] = bounded(String(value).trim());
+    }
+  }
+  return Object.keys(values).length ? values : void 0;
+}
+function capture(event, distinctId, properties) {
   const posthogKey = process.env.POSTHOG_API_KEY;
   if (!posthogKey) return;
   void fetch(`${posthogHost()}/capture/`, {
@@ -7916,18 +8576,12 @@ function captureToolCall(apiKey, event) {
     headers: { "Content-Type": "application/json" },
     body: JSON.stringify({
       api_key: posthogKey,
-      event: "runpod_mcp_tool_call",
-      distinct_id: analyticsCallerId(apiKey, event.accountId),
+      event,
+      distinct_id: distinctId,
       timestamp: (/* @__PURE__ */ new Date()).toISOString(),
       properties: {
-        tool: event.tool,
-        ok: event.ok,
-        status: event.status,
-        duration_ms: event.durationMs,
-        transport: event.transport,
-        server_version: event.serverVersion,
+        ...properties,
         surface_version: SURFACE_VERSION,
-        ...event.clientName ? { client_name: event.clientName } : {},
         $process_person_profile: false
       }
     }),
@@ -7935,6 +8589,63 @@ function captureToolCall(apiKey, event) {
   }).catch(() => {
   });
 }
+function captureToolCall(apiKey, event) {
+  const args = Object.fromEntries(
+    Object.entries(event.args ?? {}).map(([name, value]) => [
+      `arg_${name}`,
+      value
+    ])
+  );
+  capture("runpod_mcp_tool_call", analyticsCallerId(apiKey, event.accountId), {
+    tool: event.tool,
+    ok: event.ok,
+    status: event.status,
+    duration_ms: event.durationMs,
+    transport: event.transport,
+    server_version: event.serverVersion,
+    ...event.clientName ? { client_name: event.clientName } : {},
+    ...event.guideKind ? { guide_kind: event.guideKind } : {},
+    ...event.blockedBy ? { blocked_by: event.blockedBy } : {},
+    ...args
+  });
+}
+function captureLoopBlocked(apiKey, event) {
+  capture(
+    "runpod_mcp_loop_blocked",
+    analyticsCallerId(apiKey, event.accountId),
+    {
+      tool: event.tool,
+      blocked_for_s: event.blockedForS,
+      transport: event.transport,
+      server_version: event.serverVersion,
+      ...event.clientName ? { client_name: event.clientName } : {}
+    }
+  );
+}
+function captureResourceCall(apiKey, event) {
+  capture(
+    "runpod_mcp_resource_call",
+    analyticsCallerId(apiKey, event.accountId),
+    {
+      method: event.method,
+      uri: bounded(event.uri),
+      ok: event.ok,
+      transport: event.transport,
+      server_version: event.serverVersion,
+      ...event.clientName ? { client_name: event.clientName } : {}
+    }
+  );
+}
+
+// src/alp/ingest.ts
+function isSinkUrl(value) {
+  try {
+    const url = new URL(value);
+    return url.protocol === "https:" && url.hostname.endsWith(".convex.site") && url.pathname === "/alp/submit";
+  } catch {
+    return false;
+  }
+}
 
 // src/alp/read.ts
 var DEFAULT_JOURNAL_LIMIT = 20;
@@ -8156,6 +8867,177 @@ function createAlpTools(opts) {
   return [reportFeedback, saveToJournal, askQuestion, readJournalTool];
 }
 
+// src/search/knowledge.ts
+var SEARCH_NAMESPACES = ["docs", "knowledge", "blog", "site"];
+var MAX_QUERY_CHARS = 2e3;
+var MAX_SEARCH_LIMIT = 20;
+var ATTEMPTS = 7;
+var BASE_DELAY_MS = 250;
+var ATTEMPT_TIMEOUT_MS = 1e4;
+var RETRY_DEADLINE_MS = 3e4;
+var PLATFORM_LIMIT = /TooManyConcurrentRequests|Too many concurrent requests/;
+var AT_CAPACITY = {
+  status: 503,
+  body: {
+    error: "Runpod knowledge search is at capacity.",
+    hint: "Retry the same call in a few seconds."
+  }
+};
+var UNREACHABLE = {
+  status: 504,
+  body: {
+    error: "Runpod knowledge search did not respond.",
+    hint: "Retry the same call once."
+  }
+};
+function searchConfigured(env = process.env) {
+  return Boolean(knowledgeSearchBase(env) && env.RUNPOD_ASSISTANT_API_KEY);
+}
+function parseSearchRequest(raw) {
+  if (typeof raw.query !== "string" || !raw.query.trim()) {
+    return { error: "query must be a non-empty string." };
+  }
+  if (raw.query.length > MAX_QUERY_CHARS) {
+    return { error: `query exceeds ${MAX_QUERY_CHARS} characters.` };
+  }
+  const request = { query: raw.query };
+  if (raw.namespaces !== void 0) {
+    if (!Array.isArray(raw.namespaces) || raw.namespaces.length === 0 || !raw.namespaces.every(
+      (ns) => SEARCH_NAMESPACES.includes(ns)
+    )) {
+      return {
+        error: `namespaces must be a non-empty list of: ${SEARCH_NAMESPACES.join(", ")}.`
+      };
+    }
+    request.namespaces = raw.namespaces;
+  }
+  if (raw.limit !== void 0) {
+    if (typeof raw.limit !== "number" || !Number.isInteger(raw.limit) || raw.limit < 1 || raw.limit > MAX_SEARCH_LIMIT) {
+      return {
+        error: `limit must be an integer from 1 to ${MAX_SEARCH_LIMIT}.`
+      };
+    }
+    request.limit = raw.limit;
+  }
+  return { request };
+}
+async function searchKnowledge(request, opts = {}) {
+  const env = opts.env ?? process.env;
+  const doFetch = opts.fetch ?? fetch;
+  const sleep2 = opts.sleep ?? ((ms) => new Promise((resolve) => setTimeout(resolve, ms)));
+  const now = opts.now ?? Date.now;
+  if (!searchConfigured(env)) {
+    return {
+      status: 503,
+      body: { error: "Knowledge search is not configured on this server." }
+    };
+  }
+  const url = `${knowledgeSearchBase(env)}/v1/search`;
+  const init = {
+    method: "POST",
+    headers: {
+      Authorization: `Bearer ${env.RUNPOD_ASSISTANT_API_KEY}`,
+      "Content-Type": "application/json"
+    },
+    body: JSON.stringify({ ...request, source: "runpod-mcp" })
+  };
+  const startedAt = now();
+  for (let attempt = 1; ; attempt++) {
+    try {
+      const res = await doFetch(url, {
+        ...init,
+        signal: AbortSignal.timeout(ATTEMPT_TIMEOUT_MS)
+      });
+      if (res.ok) {
+        const data = await res.json().catch(() => null);
+        if (!data) {
+          console.warn("knowledge_search_failed status=200 invalid_json");
+          return {
+            status: 502,
+            body: { error: "Runpod knowledge search failed." }
+          };
+        }
+        return {
+          status: 200,
+          body: {
+            text: String(data.text ?? ""),
+            totalResults: Number(data.totalResults ?? 0),
+            failedNamespaces: Array.isArray(data.failedNamespaces) ? data.failedNamespaces : []
+          }
+        };
+      }
+      const text = await res.text();
+      const retryable = res.status === 503 || res.status === 429 || PLATFORM_LIMIT.test(text);
+      if (!retryable) {
+        console.warn(`knowledge_search_failed status=${res.status}`);
+        return res.status === 400 ? { status: 400, body: { error: errorField(text) } } : { status: 502, body: { error: "Runpod knowledge search failed." } };
+      }
+    } catch (error) {
+      console.warn(
+        `knowledge_search_unreachable ${error instanceof Error ? error.name : "error"}`
+      );
+      return UNREACHABLE;
+    }
+    const delay = Math.random() * BASE_DELAY_MS * 2 ** (attempt - 1);
+    if (attempt >= ATTEMPTS || now() - startedAt + delay > RETRY_DEADLINE_MS) {
+      return AT_CAPACITY;
+    }
+    await sleep2(delay);
+  }
+}
+function errorField(text) {
+  try {
+    const parsed = JSON.parse(text);
+    if (typeof parsed.error === "string") return parsed.error;
+  } catch {
+  }
+  return "Invalid search request.";
+}
+
+// src/specgen/tools/search.ts
+function createSearchDocsTool(search) {
+  return {
+    name: "search-docs",
+    annotations: readOnly,
+    recordedArguments: ["limit"],
+    description: 'Search Runpod documentation, curated support knowledge, the Runpod blog and runpod.io for passages that answer a question, for example "how do I attach a network volume to a Serverless endpoint". Returns ranked passages with their source URLs. Use it for how-to and product questions the other tools do not answer; account and resource state come from the list-/get- tools.',
+    inputSchema: {
+      type: "object",
+      properties: {
+        query: {
+          type: "string",
+          maxLength: MAX_QUERY_CHARS,
+          description: "The question, in natural language."
+        },
+        namespaces: {
+          type: "array",
+          items: { type: "string", enum: [...SEARCH_NAMESPACES] },
+          minItems: 1,
+          description: "Sources to search: docs (docs.runpod.io), knowledge (curated support answers), blog, site (runpod.io pages). Default: docs, knowledge, blog."
+        },
+        limit: {
+          type: "integer",
+          minimum: 1,
+          maximum: MAX_SEARCH_LIMIT,
+          description: "Most passages to return. Default 8."
+        }
+      },
+      required: ["query"],
+      additionalProperties: false
+    },
+    handler: async (ctx, args) => {
+      const parsed = parseSearchRequest(args);
+      if ("error" in parsed) return badRequest(parsed.error);
+      const outcome = await search(parsed.request, ctx.apiKey);
+      return {
+        ok: outcome.status === 200,
+        status: outcome.status,
+        payload: outcome.body
+      };
+    }
+  };
+}
+
 // src/specgen/ops.ts
 import { createHash, randomBytes } from "crypto";
 var SALT = randomBytes(16).toString("hex");
@@ -8180,6 +9062,9 @@ var curatedTools = [
   listPublicEndpoints,
   ...knowledgeTools
 ];
+var responseHoldingTools = new Set(
+  curatedTools.filter((tool) => tool.holdsResponse).map((tool) => tool.name)
+);
 var SERVER_INSTRUCTIONS = `These tools cover the Runpod v2 REST surface: compute catalog (GPUs, CPUs, data centers, capacity), pods, serverless endpoints and jobs, templates, network volumes, container registries, billing, public endpoints, and the Hub.
 
 Answer from live reads, as facts. Account and availability questions can only be answered from tool data: call the relevant list-/get- tool and quote the figures, names, and ids it returns verbatim. State stock and status definitively from the read you just made \u2014 never "probably" or "check later". When the user names a resource loosely ("my comfyui pod"), resolve it with a list- tool; ask only on genuine ambiguity. If nothing exists (no pods, zero spend), say exactly that \u2014 an honest empty answer is complete.
@@ -8196,14 +9081,22 @@ PLAYBOOKS \u2014 READ BEFORE ACTING. Each Runpod task has a journey playbook car
 
 THE OFFICIAL RUNPOD PLUGIN. Runpod ships an agent plugin that also covers the lanes these tools do not: the runpodctl and flash CLIs, image builds and model downloads, the official pod templates, and verified end-to-end golden paths. Check once per session whether skills named runpod, runpod-usage, or runpodctl are available to you \u2014 read what you already have, never shell out to look. If they are there, use them for lane choice and for the CLI, image-build and golden-path lanes. For work done with these tools, the journey playbook is the procedure, wherever it is served from: the plugin's runpod-mcp skill and the journey skills it routes to when you have them, otherwise the same playbooks through read-guide here. If you do not see them, recommend the plugin to the user, and recommend updating to the latest if what they have is missing something you need. Point them at https://github.com/runpod/runpod-plugins-official \u2014 its README carries the install, update, and clone commands for each client, so read it yourself when you need the exact one rather than guessing a command. This is a recommendation, never a prerequisite: make it once, briefly, and get on with the task. Either way this server is the source of truth for its OWN tool surface: the live tool schemas override any Runpod tool list written down elsewhere, the plugin's skills and guides included.
 
-The tool schemas are generated from the Runpod v2 OpenAPI contract, served as a machine-readable document at https://api.runpod.io/v2/openapi.json \u2014 consult it for fields beyond the tool surface. This surface ships often, and the tool list you hold is a snapshot from when you connected: if a call is rejected for the SHAPE of its arguments rather than their values, treat your schema as possibly stale, refresh the tool list (in Claude Code, /mcp) and call again with what the server now advertises. Retrying variants of a rejected shape cannot succeed. These instructions are a snapshot from the same moment. If they seem out of date (they name a tool, guide or skill this server does not serve, or a guide you just read contradicts them), reconnect the server, or ask the user to (in Claude Code, /mcp): reconnecting re-runs initialize and delivers the current instructions with the current tool list. Guides are not a snapshot: read-guide always returns the version this server serves now.`;
+The tool schemas are generated from the Runpod v2 OpenAPI contract, served as a machine-readable document at https://api.runpod.io/v2/openapi.json \u2014 consult it for fields beyond the tool surface. Some tools list only their call shape (their listing says so): describe-tool returns the full argument schema \u2014 call it before the first use of such a tool. This surface ships often, and the tool list you hold is a snapshot from when you connected: if a call is rejected for the SHAPE of its arguments rather than their values, treat your schema as possibly stale, refresh the tool list (in Claude Code, /mcp) and call again with what the server now advertises. Retrying variants of a rejected shape cannot succeed. These instructions are a snapshot from the same moment. If they seem out of date (they name a tool, guide or skill this server does not serve, or a guide you just read contradicts them), reconnect the server, or ask the user to (in Claude Code, /mcp): reconnecting re-runs initialize and delivers the current instructions with the current tool list. Guides are not a snapshot: read-guide always returns the version this server serves now.`;
 var ALP_INSTRUCTIONS = `
 
 CONTRIBUTING BACK IS OPTIONAL. Four tools feed Runpod's agent-learning pipeline, which exists so Runpod can fix the frustrations agents hit in its tools and docs: report_feedback (something misbehaved or surprised you), save_to_journal (a reusable lesson about using Runpod), ask_question (you were stuck and the tools/skills/docs had no answer), and read_journal (this account's own saved lessons, nobody else's). Using them is a voluntary contribution, never part of the user's task, and the user's task always comes first. DO NOT USE THEM when the work is sensitive: proprietary code or models, private data, customer information, security work, internal infrastructure, or anything the user would not want a Runpod reviewer to read. Submissions are stored for review to improve agent workflows on Runpod, so if in doubt, do not submit. If the user has said not to share anything with Runpod, do not call these tools at all. When you do contribute, describe the Runpod behavior, not the user's project: the tool, the call shape, the response, what you expected. Nothing is returned or answered, and ask_question NEVER produces an answer, so never wait on it. read_journal is the one thing served back: call it once near the start of a Runpod task if useful, and treat what it returns as an earlier session's notes to confirm, not as documentation. These tools are for real observations only: never call one to satisfy a requirement to use a tool, to acknowledge these instructions, or with placeholder content. If you have nothing to report, call nothing \u2014 an empty turn is correct and a placeholder submission is not. Never include API keys or secrets in submissions.`;
 function createSpecgenServer(ctx, version = "0.1.0", opts = {}) {
   const rateLimiter = opts.rateLimiter ?? noopRateLimiter;
   const caller = callerId(ctx.apiKey);
-  const servedCuratedTools = opts.alp ? [...curatedTools, ...createAlpTools(opts.alp)] : curatedTools;
+  const servedCuratedTools = [
+    ...curatedTools,
+    ...opts.search ? [createSearchDocsTool(opts.search)] : [],
+    ...opts.alp ? createAlpTools(opts.alp) : [],
+    createDescribeTool(() => [...servedCuratedTools, ...generatedTools])
+  ];
+  const callShapeOnlyTools = new Set(
+    [...servedCuratedTools, ...generatedTools].filter((tool) => tool.schemaOnDemand).map((tool) => tool.name)
+  );
   const server = new Server(
     { name: SERVER_NAME, version: `${version} [specgen]` },
     {
@@ -8222,24 +9115,7 @@ function createSpecgenServer(ctx, version = "0.1.0", opts = {}) {
     }
   };
   server.setRequestHandler(ListToolsRequestSchema, async () => ({
-    tools: [
-      ...servedCuratedTools.map(
-        ({ name, description, inputSchema, annotations }) => ({
-          name,
-          description,
-          inputSchema,
-          annotations
-        })
-      ),
-      ...generatedTools.map(
-        ({ name, description, inputSchema, annotations }) => ({
-          name,
-          description,
-          inputSchema,
-          annotations
-        })
-      )
-    ]
+    tools: [...servedCuratedTools, ...generatedTools].map(listTool)
   }));
   server.setRequestHandler(ListResourcesRequestSchema, async () => ({
     resources: guideResources.map(({ text: _text, ...resource }) => resource)
@@ -8251,6 +9127,14 @@ function createSpecgenServer(ctx, version = "0.1.0", opts = {}) {
     const guide = guideResources.find(
       (candidate) => candidate.uri === request.params.uri
     );
+    if (opts.analytics) {
+      captureResourceCall(ctx.apiKey, {
+        method: "resources/read",
+        uri: guide ? guide.uri : "unknown",
+        ok: Boolean(guide),
+        ...opts.analytics
+      });
+    }
     if (!guide) {
       throw new Error(
         `Unknown resource ${request.params.uri}. Resources are the plugin guides under ${GUIDE_URI_PREFIX}; call list-guides for the guide ids.`
@@ -8273,6 +9157,16 @@ function createSpecgenServer(ctx, version = "0.1.0", opts = {}) {
         status: 429,
         durationMs: 0
       });
+      if (opts.analytics) {
+        captureToolCall(ctx.apiKey, {
+          tool: request.params.name,
+          ok: false,
+          status: 429,
+          durationMs: 0,
+          ...verdict.reason ? { blockedBy: verdict.reason } : {},
+          ...opts.analytics
+        });
+      }
       return {
         content: [
           {
@@ -8310,11 +9204,14 @@ function createSpecgenServer(ctx, version = "0.1.0", opts = {}) {
         ok: result.ok,
         status: result.status,
         durationMs: Date.now() - startedAt,
+        args: recordedArgumentValues(curated?.recordedArguments, args),
+        guideKind: curated?.recordedGuideKind?.(args),
         ...opts.analytics
       });
     }
+    opts.onToolResult?.(request.params.name, result.status);
     if (!result.ok && result.status === 401) opts.onUnauthorized?.();
-    const payload = result.ok ? result.payload : withRecoveryHint(result);
+    const payload = result.ok ? result.payload : withRecoveryHint(result, callShapeOnlyTools.has(request.params.name));
     return {
       content: [{ type: "text", text: JSON.stringify(payload, null, 2) }],
       isError: !result.ok
@@ -8332,9 +9229,29 @@ function createSpecgenServer(ctx, version = "0.1.0", opts = {}) {
   }
   return server;
 }
-function errorHint(status) {
-  if (status === 400 || status === 422)
-    return "Input shape mismatch: check this tool's schema for required fields. For job payloads, the worker's release config defines the expected input object \u2014 fix the payload and retry; don't abandon the task.";
+var SHAPE_HINT = "Input shape mismatch: check this tool's schema for required fields. For job payloads, the worker's release config defines the expected input object \u2014 fix the payload and retry; don't abandon the task.";
+var DETAIL_HINTS = [
+  [
+    /no longer any instances available|not enough free GPUs/i,
+    "Capacity, not input: the request is valid but no machine can take it right now. Don't change the payload shape. Retry with another GPU type, data center, cloud, CUDA version or count, or try again later. For a stopped Pod, its host's GPUs are taken; that Pod can only start once they free up."
+  ],
+  [
+    /at least \$\d|insufficient (?:funds|balance)|\bbalance\b/i,
+    "Account balance, not input: this operation needs funds on the account. Ask the user to add credit in the Runpod console; list-billing shows recent spend."
+  ],
+  [
+    /mutually exclusive|cannot be combined with/i,
+    "Two arguments conflict: send only one of them and leave the other out entirely. Do not change anything else."
+  ]
+];
+function errorHint(status, detail) {
+  if (status === 400 && detail) {
+    for (const [pattern, hint] of DETAIL_HINTS)
+      if (pattern.test(detail)) return hint;
+    if (/^unauthorized$/i.test(detail.trim()))
+      return "Upstream refused this write. With a key that can read, this usually means the account has no balance: ask the user to check their balance in the Runpod console before retrying.";
+  }
+  if (status === 400 || status === 422) return SHAPE_HINT;
   if (status === 401) return "The Runpod API key is missing or invalid.";
   if (status === 402)
     return "Check your account balance in the Runpod console before retrying billable operations; list-billing shows the recent spend behind it.";
@@ -8348,13 +9265,23 @@ function errorHint(status) {
     return "Upstream Runpod error: retry once; if it persists, report it as an API-side failure.";
   return void 0;
 }
-function withRecoveryHint(result) {
+var CALL_SHAPE_ONLY_HINT = `This tool lists its call shape only; ${DESCRIBE_TOOL_NAME} serves the full argument schema.`;
+function withRecoveryHint(result, listsCallShapeOnly = false) {
   const payload = asRecord(result.payload) ?? { error: result.payload };
+  const hint = payload.hint ?? asRecord(payload.detail)?.hint ?? errorHint(result.status, detailText(payload));
+  const shapeRejected = result.status === 400 || result.status === 422;
   return {
     ...payload,
-    hint: payload.hint ?? asRecord(payload.detail)?.hint ?? errorHint(result.status)
+    hint: listsCallShapeOnly && shapeRejected ? [hint, CALL_SHAPE_ONLY_HINT].filter(Boolean).join(" ") : hint
   };
 }
+function detailText(payload) {
+  if (Array.isArray(payload.errors) && payload.errors.length) return void 0;
+  const detail = payload.detail ?? payload.error;
+  if (typeof detail === "string") return detail;
+  const nested = asRecord(detail);
+  return typeof nested?.detail === "string" ? nested.detail : void 0;
+}
 function asRecord(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value) ? value : void 0;
 }
@@ -8520,6 +9447,58 @@ function createToolContext(options = {}) {
   };
 }
 
+// src/specgen/not-found-limiter.ts
+var SINK_TIMEOUT_MS = 3e3;
+var NOT_FOUND_LOOP = "not_found_loop";
+var blocked = /* @__PURE__ */ new Map();
+function notFoundSinkUrl(submitUrl) {
+  if (!isSinkUrl(submitUrl)) return null;
+  const url = new URL(submitUrl);
+  url.pathname = "/ratelimit/not-found";
+  return url.toString();
+}
+function createNotFoundLimiter(opts) {
+  const fetchImpl = opts.fetchImpl ?? fetch;
+  const now = opts.now ?? Date.now;
+  return {
+    rateLimiter: async (_caller, tool) => {
+      const pair = `${opts.key}|${tool}`;
+      const until = blocked.get(pair);
+      if (until === void 0) return { allowed: true };
+      if (until <= now()) {
+        blocked.delete(pair);
+        return { allowed: true };
+      }
+      return {
+        allowed: false,
+        retryAfterS: Math.ceil((until - now()) / 1e3),
+        reason: NOT_FOUND_LOOP
+      };
+    },
+    onToolResult: (tool, status) => {
+      if (status !== 404) return;
+      const pair = `${opts.key}|${tool}`;
+      void fetchImpl(opts.sinkUrl, {
+        method: "POST",
+        headers: {
+          "Content-Type": "application/json",
+          "X-ALP-Secret": opts.secret
+        },
+        body: JSON.stringify({ key: pair }),
+        signal: AbortSignal.timeout(SINK_TIMEOUT_MS)
+      }).then((response) => response.ok ? response.json() : null).then((body) => {
+        const reply = body;
+        const until = reply?.blockedUntil;
+        if (typeof until === "number" && until > now()) {
+          blocked.set(pair, until);
+          if (reply?.started === true) opts.onBlockStarted?.(tool, until);
+        }
+      }).catch(() => {
+      });
+    }
+  };
+}
+
 // src/_shared/credential-check.ts
 import { createHash as createHash2 } from "crypto";
 var VALID_TTL_MS = 6e4;
@@ -8631,8 +9610,7 @@ function extractBearerToken(req) {
   return parts[1];
 }
 function analyticsOptedOut(req) {
-  const raw = req.headers["x-runpod-analytics"];
-  const value = (Array.isArray(raw) ? raw[0] : raw)?.trim().toLowerCase();
+  const value = header(req, "x-runpod-analytics")?.trim().toLowerCase();
   return value === "off" || value === "0" || value === "false";
 }
 function getBaseUrl(req) {
@@ -8691,14 +9669,16 @@ function credentialCheckMayBeWrongEnvironment(env = process.env) {
   );
   return restMoved !== graphqlMoved;
 }
-function looksLikeMcpMessage(value) {
-  if (typeof value !== "object" || value === null) return false;
-  const msg = value;
-  return msg.jsonrpc === "2.0" && typeof msg.method === "string";
+function parseMcpMessages(body) {
+  const elements = Array.isArray(body) ? body : [body];
+  return elements.flatMap((element) => {
+    const parsed = JSONRPCMessageSchema.safeParse(element);
+    return parsed.success ? [parsed.data] : [];
+  });
 }
+var invokesMethod = (message) => "method" in message;
 function bodyIsCheckable(body) {
-  const messages = Array.isArray(body) ? body : [body];
-  return messages.some(looksLikeMcpMessage);
+  return parseMcpMessages(body).some(invokesMethod);
 }
 function isJsonContentType(req) {
   const ct = req.headers["content-type"];
@@ -8736,17 +9716,79 @@ function earlyCredentialCheckSkip(req) {
   if (credentialCheckMayBeWrongEnvironment()) return "skip_env_mismatch";
   return null;
 }
+var callsResponseHoldingTool = (message) => {
+  const call = CallToolRequestSchema2.safeParse(message);
+  return call.success && responseHoldingTools.has(call.data.params.name);
+};
+function prefersJsonResponse(messages) {
+  return messages.length > 0 && !messages.some(callsResponseHoldingTool);
+}
+var DEFAULT_REQUEST_LOG_SAMPLE_RATE = 0.1;
+function requestLogSampleRate(env = process.env) {
+  if (env.MCP_VERBOSE_LOGS === "true") return 1;
+  const raw = env.MCP_REQUEST_LOG_SAMPLE_RATE;
+  if (raw === void 0 || raw === "") return DEFAULT_REQUEST_LOG_SAMPLE_RATE;
+  const parsed = Number(raw);
+  if (!Number.isFinite(parsed)) return DEFAULT_REQUEST_LOG_SAMPLE_RATE;
+  return Math.min(1, Math.max(0, parsed));
+}
+function summarizeRpc(messages) {
+  const rpc = messages.map((message) => {
+    const call = CallToolRequestSchema2.safeParse(message);
+    if (call.success) return `tools/call:${call.data.params.name}`;
+    return "method" in message ? message.method : "response";
+  });
+  const init = messages.map((message) => InitializeRequestSchema.safeParse(message)).find((result) => result.success);
+  const client = init?.success ? `${init.data.params.clientInfo.name}/${init.data.params.clientInfo.version}` : void 0;
+  return { rpc: rpc.length > 0 ? rpc.join(",") : void 0, client };
+}
+function header(req, name) {
+  const value = req.headers[name];
+  return Array.isArray(value) ? value[0] : value;
+}
+function clientAddress(req) {
+  return header(req, "x-real-ip") || header(req, "x-forwarded-for")?.split(",")[0].trim() || req.socket?.remoteAddress || void 0;
+}
+function startRequestLog(req, res, bearerToken) {
+  const sampleRate = requestLogSampleRate();
+  if (Math.random() >= sampleRate) return { note() {
+  } };
+  const startedAt = Date.now();
+  let summary = {};
+  const address = clientAddress(req);
+  const userAgent = header(req, "user-agent");
+  let emitted = false;
+  const emit = () => {
+    if (emitted) return;
+    emitted = true;
+    console.log("mcp_request", {
+      method: req.method,
+      url: req.url,
+      status: res.writableFinished ? res.statusCode : void 0,
+      durationMs: Date.now() - startedAt,
+      rpc: summary.rpc,
+      client: summary.client,
+      ua: userAgent?.slice(0, 160),
+      caller: callerId(bearerToken ?? void 0),
+      ip: address ? callerId(address) : void 0,
+      sampleRate
+    });
+  };
+  res.on("finish", emit);
+  res.on("close", emit);
+  return {
+    note(next) {
+      summary = next;
+    }
+  };
+}
 function logCredentialCheck(outcome) {
   if (outcome === "pass" && process.env.MCP_VERBOSE_LOGS !== "true") return;
   console.log("credential_check", { outcome });
 }
 async function handleMcpRequest(req, res, opts = {}) {
   const bearerToken = extractBearerToken(req);
-  console.log("mcp_request", {
-    method: req.method,
-    url: req.url,
-    hasBearerToken: !!bearerToken
-  });
+  const requestLog = startRequestLog(req, res, bearerToken);
   if (req.method === "GET") {
     res.writeHead(405, {
       Allow: "POST, DELETE",
@@ -8775,9 +9817,7 @@ async function handleMcpRequest(req, res, opts = {}) {
   let body = req.body;
   let accountId;
   const early = earlyCredentialCheckSkip(req);
-  if (early) {
-    logCredentialCheck(early);
-  } else {
+  if (!early) {
     const read = await readJsonBody(req);
     if (read === BODY_TOO_LARGE) {
       res.writeHead(413, { "Content-Type": "application/json" });
@@ -8785,26 +9825,47 @@ async function handleMcpRequest(req, res, opts = {}) {
       return;
     }
     body = read;
-    if (!bodyIsCheckable(body)) {
-      logCredentialCheck("skip_not_mcp_body");
-    } else {
-      const verify = opts.verifyCredential ?? defaultCredentialChecker.verify;
-      const verdict = await verify(bearerToken);
-      if (verdict.status === "valid") accountId = verdict.accountId;
-      logCredentialCheck(
-        verdict.status === "invalid" ? "reject" : verdict.status === "valid" ? "pass" : "fail_open"
+  }
+  const messages = parseMcpMessages(body);
+  requestLog.note(summarizeRpc(messages));
+  if (early) {
+    logCredentialCheck(early);
+  } else if (!messages.some(invokesMethod)) {
+    logCredentialCheck("skip_not_mcp_body");
+  } else {
+    const verify = opts.verifyCredential ?? defaultCredentialChecker.verify;
+    const verdict = await verify(bearerToken);
+    if (verdict.status === "valid") accountId = verdict.accountId;
+    logCredentialCheck(
+      verdict.status === "invalid" ? "reject" : verdict.status === "valid" ? "pass" : "fail_open"
+    );
+    if (verdict.status === "invalid") {
+      writeUnauthorized(
+        req,
+        res,
+        `${verdict.reason} Re-authenticate to continue.`,
+        true
       );
-      if (verdict.status === "invalid") {
-        writeUnauthorized(
-          req,
-          res,
-          `${verdict.reason} Re-authenticate to continue.`,
-          true
-        );
-        return;
-      }
+      return;
     }
   }
+  const notFoundUrl = process.env.ALP_SINK_URL ? notFoundSinkUrl(process.env.ALP_SINK_URL) : null;
+  const notFoundLimiter = bearerToken && notFoundUrl && process.env.ALP_SINK_SECRET ? createNotFoundLimiter({
+    key: analyticsCallerId(bearerToken, accountId),
+    sinkUrl: notFoundUrl,
+    secret: process.env.ALP_SINK_SECRET,
+    // One usage event per block, unless this client opted out.
+    onBlockStarted: analyticsOptedOut(req) ? void 0 : (tool, blockedUntil) => captureLoopBlocked(bearerToken, {
+      tool,
+      blockedForS: Math.round((blockedUntil - Date.now()) / 1e3),
+      transport: "http",
+      serverVersion: opts.serverVersion ?? SERVER_VERSION,
+      clientName: sanitizeUaToken(
+        String(req.headers["user-agent"] ?? "unknown")
+      ),
+      accountId
+    })
+  }) : null;
   const server = createSpecgenServer(
     createToolContext({
       apiKey: bearerToken,
@@ -8845,6 +9906,13 @@ async function handleMcpRequest(req, res, opts = {}) {
           harnessSource: "user_agent"
         }
       } : {},
+      // search-docs: listed only where the deployment holds the assistant's
+      // service key; the search runs in-process with it.
+      ...searchConfigured() ? { search: (request) => searchKnowledge(request) } : {},
+      ...notFoundLimiter ? {
+        rateLimiter: notFoundLimiter.rateLimiter,
+        onToolResult: notFoundLimiter.onToolResult
+      } : {},
       // A tool call that 401s proves a cached "valid" verdict wrong before
       // its TTL — drop it so the next request re-checks and re-auths.
       onUnauthorized: () => (opts.invalidateCredential ?? defaultCredentialChecker.invalidate)(
@@ -8853,8 +9921,9 @@ async function handleMcpRequest(req, res, opts = {}) {
     }
   );
   const transport = new StreamableHTTPServerTransport({
-    sessionIdGenerator: void 0
+    sessionIdGenerator: void 0,
     // stateless — no session persistence
+    enableJsonResponse: prefersJsonResponse(messages)
   });
   res.on("close", () => {
     transport.close().catch(() => {
@@ -8868,7 +9937,12 @@ async function handleMcpRequest(req, res, opts = {}) {
 export {
   bodyIsCheckable,
   defaultCredentialChecker,
+  extractBearerToken,
   getBaseUrl,
-  handleMcpRequest
+  handleMcpRequest,
+  parseMcpMessages,
+  prefersJsonResponse,
+  requestLogSampleRate,
+  summarizeRpc
 };
 //# sourceMappingURL=http.mjs.map