@octocodeai/freellama-mcp-server 0.1.0 → 0.6.0

package/dist/index.js

@@ -3109,6 +3109,7 @@ var require_utils = __commonJS({
     "use strict";
     var isUUID = RegExp.prototype.test.bind(/^[\da-f]{8}-[\da-f]{4}-[\da-f]{4}-[\da-f]{4}-[\da-f]{12}$/iu);
     var isIPv4 = RegExp.prototype.test.bind(/^(?:(?:25[0-5]|2[0-4]\d|1\d{2}|[1-9]\d|\d)\.){3}(?:25[0-5]|2[0-4]\d|1\d{2}|[1-9]\d|\d)$/u);
+    var isPort = RegExp.prototype.test.bind(/^\d*$/u);
     var isHexPair = RegExp.prototype.test.bind(/^[\da-f]{2}$/iu);
     var isUnreserved = RegExp.prototype.test.bind(/^[\da-z\-._~]$/iu);
     var isPathCharacter = RegExp.prototype.test.bind(/^[A-Za-z0-9\-._~!$&'()*+,;=:@/]$/u);
@@ -3261,8 +3262,8 @@ var require_utils = __commonJS({
       }
       return ind;
     }
-    function removeDotSegments(path5) {
-      let input = path5;
+    function removeDotSegments(path6) {
+      let input = path6;
       const output = [];
       let nextSlash = -1;
       let len = 0;
@@ -3574,8 +3575,12 @@ var require_utils = __commonJS({
         uriTokens.push(host);
       }
       if (typeof component.port === "number" || typeof component.port === "string") {
+        const port = String(component.port);
+        if (!isPort(port)) {
+          throw new TypeError("URI port is malformed.");
+        }
         uriTokens.push(":");
-        uriTokens.push(String(component.port));
+        uriTokens.push(port);
       }
       return uriTokens.length ? uriTokens.join("") : void 0;
     }
@@ -3667,8 +3672,8 @@ var require_schemes = __commonJS({
       }
       if (wsComponent.resourceName) {
         const queryIndex = wsComponent.resourceName.indexOf("?");
-        const path5 = queryIndex === -1 ? wsComponent.resourceName : wsComponent.resourceName.slice(0, queryIndex);
-        wsComponent.path = path5 && path5 !== "/" ? path5 : void 0;
+        const path6 = queryIndex === -1 ? wsComponent.resourceName : wsComponent.resourceName.slice(0, queryIndex);
+        wsComponent.path = path6 && path6 !== "/" ? path6 : void 0;
         wsComponent.query = queryIndex === -1 ? void 0 : wsComponent.resourceName.slice(queryIndex + 1);
         wsComponent.resourceName = void 0;
       }
@@ -3832,7 +3837,7 @@ var require_fast_uri = __commonJS({
         normalizeString(uri, options);
       } else if (typeof uri === "object") {
         uri = /** @type {T} */
-        parse3(serialize2(uri, options), options);
+        parse3(serialize(uri, options), options);
       }
       return uri;
     }
@@ -3867,13 +3872,13 @@ var require_fast_uri = __commonJS({
         throw new Error(resolved.error);
       }
       schemelessOptions.skipEscape = true;
-      return serialize2(resolved, schemelessOptions);
+      return serialize(resolved, schemelessOptions);
     }
     function resolveComponent(base, relative, options, skipNormalization) {
       const target = {};
       if (!skipNormalization) {
-        base = parse3(serialize2(base, options), options);
-        relative = parse3(serialize2(relative, options), options);
+        base = parse3(serialize(base, options), options);
+        relative = parse3(serialize(relative, options), options);
       }
       options = options || {};
       if (!options.tolerant && relative.scheme) {
@@ -3927,7 +3932,7 @@ var require_fast_uri = __commonJS({
       const normalizedB = normalizeComparableURI(uriB, options);
       return normalizedA !== void 0 && normalizedB !== void 0 && normalizedA === normalizedB;
     }
-    function serialize2(cmpts, opts) {
+    function serialize(cmpts, opts) {
       const component = {
         host: cmpts.host,
         scheme: cmpts.scheme,
@@ -4018,12 +4023,15 @@ var require_fast_uri = __commonJS({
       }
       return false;
     }
+    function isIPLiteral(host) {
+      return host[0] === "[" && host[host.length - 1] === "]";
+    }
     function hasMalformedComponentPercentEncoding(matches) {
       const host = matches[4];
-      return hasMalformedPercentEncoding(matches[3]) || host !== void 0 && !(host[0] === "[" && host[host.length - 1] === "]") && hasMalformedPercentEncoding(host) || hasMalformedPercentEncoding(matches[6]) || hasMalformedPercentEncoding(matches[7]) || hasMalformedPercentEncoding(matches[8]);
+      return hasMalformedPercentEncoding(matches[3]) || host !== void 0 && !isIPLiteral(host) && hasMalformedPercentEncoding(host) || hasMalformedPercentEncoding(matches[6]) || hasMalformedPercentEncoding(matches[7]) || hasMalformedPercentEncoding(matches[8]);
     }
     function canonicalizeHost(parsed, options, schemeHandler, isIP) {
-      if (!options.unicodeSupport && (!schemeHandler || !schemeHandler.unicodeSupport) && parsed.host && parsed.host[0] !== "[" && (options.domainHost || schemeHandler && schemeHandler.domainHost) && isIP === false && nonSimpleDomain(parsed.host)) {
+      if (!options.unicodeSupport && (!schemeHandler || !schemeHandler.unicodeSupport) && parsed.host && !isIPLiteral(parsed.host) && (options.domainHost || schemeHandler && schemeHandler.domainHost) && isIP === false && nonSimpleDomain(parsed.host)) {
         try {
           parsed.host = new URL("http://" + parsed.host).hostname;
         } catch (e) {
@@ -4110,10 +4118,11 @@ var require_fast_uri = __commonJS({
         if (parsed.host) {
           const ipv4result = isIPv4(parsed.host);
           if (ipv4result === false) {
-            const bracketedIPLiteral = parsed.host[0] === "[" && parsed.host[parsed.host.length - 1] === "]";
+            const bracketedIPLiteral = isIPLiteral(parsed.host);
+            const hasIPLiteralBracket = parsed.host.indexOf("[") !== -1 || parsed.host.indexOf("]") !== -1;
             const ipv6result = normalizeIPv6(parsed.host);
             isIP = ipv6result.isIPV6 || ipv6result.isIPVFuture === true;
-            malformedIPLiteral = bracketedIPLiteral && ipv6result.error === true;
+            malformedIPLiteral = hasIPLiteralBracket && (!bracketedIPLiteral || ipv6result.error === true);
             parsed.host = isIP ? ipv6result.host : ipv6result.host.toLowerCase();
             if (malformedIPLiteral) {
               parsed.error = parsed.error || "URI host is malformed.";
@@ -4136,14 +4145,17 @@ var require_fast_uri = __commonJS({
           parsed.error = parsed.error || "URI is not a " + options.reference + " reference.";
         }
         const schemeHandler = getSchemeHandler(options.scheme || parsed.scheme);
-        malformedHost = canonicalizeHost(parsed, options, schemeHandler, isIP);
-        if (!schemeHandler || schemeHandler && !schemeHandler.skipNormalize) {
-          if (uri.indexOf("%") !== -1) {
-            if (parsed.host !== void 0 && !malformedIPLiteral) {
-              const host = isIP ? parsed.host : normalizePercentEncoding(parsed.host, true);
-              parsed.host = reescapeHostDelimiters(host, isIP);
-            }
+        if (!malformedIPLiteral) {
+          malformedHost = canonicalizeHost(parsed, options, schemeHandler, isIP);
+        }
+        if (uri.indexOf("%") !== -1 && parsed.host !== void 0 && !malformedIPLiteral) {
+          let host = isIP ? parsed.host : normalizePercentEncoding(parsed.host, true);
+          if (!isIP) {
+            host = normalizePercentEncoding(host.toLowerCase());
           }
+          parsed.host = reescapeHostDelimiters(host, isIP);
+        }
+        if (!schemeHandler || schemeHandler && !schemeHandler.skipNormalize) {
           if (parsed.path) {
             parsed.path = normalizePathEncoding(parsed.path);
           }
@@ -4174,7 +4186,7 @@ var require_fast_uri = __commonJS({
     function normalizeStringWithStatus(uri, opts) {
       const { parsed, malformedAuthorityOrPort, malformedPercentEncoding, malformedSchemeSpecific, malformedHost, malformedScheme } = parseWithStatus(uri, opts);
       return {
-        normalized: malformedAuthorityOrPort || malformedPercentEncoding || malformedSchemeSpecific || malformedHost || malformedScheme ? uri : serialize2(parsed, opts),
+        normalized: malformedAuthorityOrPort || malformedPercentEncoding || malformedSchemeSpecific || malformedHost || malformedScheme ? uri : serialize(parsed, opts),
         malformedAuthorityOrPort,
         malformedPercentEncoding,
         malformedSchemeSpecific,
@@ -4188,7 +4200,7 @@ var require_fast_uri = __commonJS({
       }
       let value;
       try {
-        value = typeof uri === "string" ? uri : serialize2(uri, opts);
+        value = typeof uri === "string" ? uri : serialize(uri, opts);
       } catch {
         return void 0;
       }
@@ -4201,7 +4213,7 @@ var require_fast_uri = __commonJS({
       resolve,
       resolveComponent,
       equal,
-      serialize: serialize2,
+      serialize,
       parse: parse3
     };
     module.exports = fastUri;
@@ -7187,16 +7199,6 @@ var require_dist = __commonJS({
   }
 });
 
-// src/index.ts
-import { execFile } from "node:child_process";
-import { existsSync as existsSync4, readdirSync } from "node:fs";
-import { mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
-import { createHash } from "node:crypto";
-import { tmpdir } from "node:os";
-import path4 from "node:path";
-import { fileURLToPath as fileURLToPath2 } from "node:url";
-import { promisify } from "node:util";
-
 // ../../node_modules/zod/v3/external.js
 var external_exports = {};
 __export(external_exports, {
@@ -7675,8 +7677,8 @@ function getErrorMap() {
 
 // ../../node_modules/zod/v3/helpers/parseUtil.js
 var makeIssue = (params) => {
-  const { data, path: path5, errorMaps, issueData } = params;
-  const fullPath = [...path5, ...issueData.path || []];
+  const { data, path: path6, errorMaps, issueData } = params;
+  const fullPath = [...path6, ...issueData.path || []];
   const fullIssue = {
     ...issueData,
     path: fullPath
@@ -7688,15 +7690,15 @@ var makeIssue = (params) => {
       message: issueData.message
     };
   }
-  let errorMessage = "";
+  let errorMessage2 = "";
   const maps = errorMaps.filter((m) => !!m).slice().reverse();
   for (const map of maps) {
-    errorMessage = map(fullIssue, { data, defaultError: errorMessage }).message;
+    errorMessage2 = map(fullIssue, { data, defaultError: errorMessage2 }).message;
   }
   return {
     ...issueData,
     path: fullPath,
-    message: errorMessage
+    message: errorMessage2
   };
 };
 var EMPTY_PATH = [];
@@ -7731,18 +7733,18 @@ var ParseStatus = class _ParseStatus {
     if (this.value !== "aborted")
       this.value = "aborted";
   }
-  static mergeArray(status, results) {
+  static mergeArray(status2, results) {
     const arrayValue = [];
     for (const s of results) {
       if (s.status === "aborted")
         return INVALID;
       if (s.status === "dirty")
-        status.dirty();
+        status2.dirty();
       arrayValue.push(s.value);
     }
-    return { status: status.value, value: arrayValue };
+    return { status: status2.value, value: arrayValue };
   }
-  static async mergeObjectAsync(status, pairs) {
+  static async mergeObjectAsync(status2, pairs) {
     const syncPairs = [];
     for (const pair of pairs) {
       const key = await pair.key;
@@ -7752,9 +7754,9 @@ var ParseStatus = class _ParseStatus {
         value
       });
     }
-    return _ParseStatus.mergeObjectSync(status, syncPairs);
+    return _ParseStatus.mergeObjectSync(status2, syncPairs);
   }
-  static mergeObjectSync(status, pairs) {
+  static mergeObjectSync(status2, pairs) {
     const finalObject = {};
     for (const pair of pairs) {
       const { key, value } = pair;
@@ -7763,14 +7765,14 @@ var ParseStatus = class _ParseStatus {
       if (value.status === "aborted")
         return INVALID;
       if (key.status === "dirty")
-        status.dirty();
+        status2.dirty();
       if (value.status === "dirty")
-        status.dirty();
+        status2.dirty();
       if (key.value !== "__proto__" && (typeof value.value !== "undefined" || pair.alwaysSet)) {
         finalObject[key.value] = value.value;
       }
     }
-    return { status: status.value, value: finalObject };
+    return { status: status2.value, value: finalObject };
   }
 };
 var INVALID = Object.freeze({
@@ -7792,11 +7794,11 @@ var errorUtil;
 
 // ../../node_modules/zod/v3/types.js
 var ParseInputLazyPath = class {
-  constructor(parent, value, path5, key) {
+  constructor(parent, value, path6, key) {
     this._cachedPath = [];
     this.parent = parent;
     this.data = value;
-    this._path = path5;
+    this._path = path6;
     this._key = key;
   }
   get path() {
@@ -8230,7 +8232,7 @@ var ZodString = class _ZodString2 extends ZodType {
       });
       return INVALID;
     }
-    const status = new ParseStatus();
+    const status2 = new ParseStatus();
     let ctx = void 0;
     for (const check2 of this._def.checks) {
       if (check2.kind === "min") {
@@ -8244,7 +8246,7 @@ var ZodString = class _ZodString2 extends ZodType {
             exact: false,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "max") {
         if (input.data.length > check2.value) {
@@ -8257,7 +8259,7 @@ var ZodString = class _ZodString2 extends ZodType {
             exact: false,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "length") {
         const tooBig = input.data.length > check2.value;
@@ -8283,7 +8285,7 @@ var ZodString = class _ZodString2 extends ZodType {
               message: check2.message
             });
           }
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "email") {
         if (!emailRegex.test(input.data)) {
@@ -8293,7 +8295,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "emoji") {
         if (!emojiRegex) {
@@ -8306,7 +8308,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "uuid") {
         if (!uuidRegex.test(input.data)) {
@@ -8316,7 +8318,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "nanoid") {
         if (!nanoidRegex.test(input.data)) {
@@ -8326,7 +8328,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "cuid") {
         if (!cuidRegex.test(input.data)) {
@@ -8336,7 +8338,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "cuid2") {
         if (!cuid2Regex.test(input.data)) {
@@ -8346,7 +8348,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "ulid") {
         if (!ulidRegex.test(input.data)) {
@@ -8356,7 +8358,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "url") {
         try {
@@ -8368,7 +8370,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "regex") {
         check2.regex.lastIndex = 0;
@@ -8380,7 +8382,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "trim") {
         input.data = input.data.trim();
@@ -8392,7 +8394,7 @@ var ZodString = class _ZodString2 extends ZodType {
             validation: { includes: check2.value, position: check2.position },
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "toLowerCase") {
         input.data = input.data.toLowerCase();
@@ -8406,7 +8408,7 @@ var ZodString = class _ZodString2 extends ZodType {
             validation: { startsWith: check2.value },
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "endsWith") {
         if (!input.data.endsWith(check2.value)) {
@@ -8416,7 +8418,7 @@ var ZodString = class _ZodString2 extends ZodType {
             validation: { endsWith: check2.value },
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "datetime") {
         const regex = datetimeRegex(check2);
@@ -8427,7 +8429,7 @@ var ZodString = class _ZodString2 extends ZodType {
             validation: "datetime",
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "date") {
         const regex = dateRegex;
@@ -8438,7 +8440,7 @@ var ZodString = class _ZodString2 extends ZodType {
             validation: "date",
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "time") {
         const regex = timeRegex(check2);
@@ -8449,7 +8451,7 @@ var ZodString = class _ZodString2 extends ZodType {
             validation: "time",
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "duration") {
         if (!durationRegex.test(input.data)) {
@@ -8459,7 +8461,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "ip") {
         if (!isValidIP(input.data, check2.version)) {
@@ -8469,7 +8471,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "jwt") {
         if (!isValidJWT(input.data, check2.alg)) {
@@ -8479,7 +8481,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "cidr") {
         if (!isValidCidr(input.data, check2.version)) {
@@ -8489,7 +8491,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "base64") {
         if (!base64Regex.test(input.data)) {
@@ -8499,7 +8501,7 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "base64url") {
         if (!base64urlRegex.test(input.data)) {
@@ -8509,13 +8511,13 @@ var ZodString = class _ZodString2 extends ZodType {
             code: ZodIssueCode.invalid_string,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else {
         util.assertNever(check2);
       }
     }
-    return { status: status.value, value: input.data };
+    return { status: status2.value, value: input.data };
   }
   _regex(regex, validation, message) {
     return this.refinement((data) => regex.test(data), {
@@ -8791,7 +8793,7 @@ var ZodNumber = class _ZodNumber extends ZodType {
       return INVALID;
     }
     let ctx = void 0;
-    const status = new ParseStatus();
+    const status2 = new ParseStatus();
     for (const check2 of this._def.checks) {
       if (check2.kind === "int") {
         if (!util.isInteger(input.data)) {
@@ -8802,7 +8804,7 @@ var ZodNumber = class _ZodNumber extends ZodType {
             received: "float",
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "min") {
         const tooSmall = check2.inclusive ? input.data < check2.value : input.data <= check2.value;
@@ -8816,7 +8818,7 @@ var ZodNumber = class _ZodNumber extends ZodType {
             exact: false,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "max") {
         const tooBig = check2.inclusive ? input.data > check2.value : input.data >= check2.value;
@@ -8830,7 +8832,7 @@ var ZodNumber = class _ZodNumber extends ZodType {
             exact: false,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "multipleOf") {
         if (floatSafeRemainder(input.data, check2.value) !== 0) {
@@ -8840,7 +8842,7 @@ var ZodNumber = class _ZodNumber extends ZodType {
             multipleOf: check2.value,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "finite") {
         if (!Number.isFinite(input.data)) {
@@ -8849,13 +8851,13 @@ var ZodNumber = class _ZodNumber extends ZodType {
             code: ZodIssueCode.not_finite,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else {
         util.assertNever(check2);
       }
     }
-    return { status: status.value, value: input.data };
+    return { status: status2.value, value: input.data };
   }
   gte(value, message) {
     return this.setLimit("min", value, true, errorUtil.toString(message));
@@ -9020,7 +9022,7 @@ var ZodBigInt = class _ZodBigInt extends ZodType {
       return this._getInvalidInput(input);
     }
     let ctx = void 0;
-    const status = new ParseStatus();
+    const status2 = new ParseStatus();
     for (const check2 of this._def.checks) {
       if (check2.kind === "min") {
         const tooSmall = check2.inclusive ? input.data < check2.value : input.data <= check2.value;
@@ -9033,7 +9035,7 @@ var ZodBigInt = class _ZodBigInt extends ZodType {
             inclusive: check2.inclusive,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "max") {
         const tooBig = check2.inclusive ? input.data > check2.value : input.data >= check2.value;
@@ -9046,7 +9048,7 @@ var ZodBigInt = class _ZodBigInt extends ZodType {
             inclusive: check2.inclusive,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "multipleOf") {
         if (input.data % check2.value !== BigInt(0)) {
@@ -9056,13 +9058,13 @@ var ZodBigInt = class _ZodBigInt extends ZodType {
             multipleOf: check2.value,
             message: check2.message
           });
-          status.dirty();
+          status2.dirty();
         }
       } else {
         util.assertNever(check2);
       }
     }
-    return { status: status.value, value: input.data };
+    return { status: status2.value, value: input.data };
   }
   _getInvalidInput(input) {
     const ctx = this._getOrReturnCtx(input);
@@ -9220,7 +9222,7 @@ var ZodDate = class _ZodDate extends ZodType {
       });
       return INVALID;
     }
-    const status = new ParseStatus();
+    const status2 = new ParseStatus();
     let ctx = void 0;
     for (const check2 of this._def.checks) {
       if (check2.kind === "min") {
@@ -9234,7 +9236,7 @@ var ZodDate = class _ZodDate extends ZodType {
             minimum: check2.value,
             type: "date"
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (check2.kind === "max") {
         if (input.data.getTime() > check2.value) {
@@ -9247,14 +9249,14 @@ var ZodDate = class _ZodDate extends ZodType {
             maximum: check2.value,
             type: "date"
           });
-          status.dirty();
+          status2.dirty();
         }
       } else {
         util.assertNever(check2);
       }
     }
     return {
-      status: status.value,
+      status: status2.value,
       value: new Date(input.data.getTime())
     };
   }
@@ -9440,7 +9442,7 @@ ZodVoid.create = (params) => {
 };
 var ZodArray = class _ZodArray extends ZodType {
   _parse(input) {
-    const { ctx, status } = this._processInputParams(input);
+    const { ctx, status: status2 } = this._processInputParams(input);
     const def = this._def;
     if (ctx.parsedType !== ZodParsedType.array) {
       addIssueToContext(ctx, {
@@ -9463,7 +9465,7 @@ var ZodArray = class _ZodArray extends ZodType {
           exact: true,
           message: def.exactLength.message
         });
-        status.dirty();
+        status2.dirty();
       }
     }
     if (def.minLength !== null) {
@@ -9476,7 +9478,7 @@ var ZodArray = class _ZodArray extends ZodType {
           exact: false,
           message: def.minLength.message
         });
-        status.dirty();
+        status2.dirty();
       }
     }
     if (def.maxLength !== null) {
@@ -9489,20 +9491,20 @@ var ZodArray = class _ZodArray extends ZodType {
           exact: false,
           message: def.maxLength.message
         });
-        status.dirty();
+        status2.dirty();
       }
     }
     if (ctx.common.async) {
       return Promise.all([...ctx.data].map((item, i) => {
         return def.type._parseAsync(new ParseInputLazyPath(ctx, item, ctx.path, i));
       })).then((result2) => {
-        return ParseStatus.mergeArray(status, result2);
+        return ParseStatus.mergeArray(status2, result2);
       });
     }
     const result = [...ctx.data].map((item, i) => {
       return def.type._parseSync(new ParseInputLazyPath(ctx, item, ctx.path, i));
     });
-    return ParseStatus.mergeArray(status, result);
+    return ParseStatus.mergeArray(status2, result);
   }
   get element() {
     return this._def.type;
@@ -9591,7 +9593,7 @@ var ZodObject = class _ZodObject extends ZodType {
       });
       return INVALID;
     }
-    const { status, ctx } = this._processInputParams(input);
+    const { status: status2, ctx } = this._processInputParams(input);
     const { shape, keys: shapeKeys } = this._getCached();
     const extraKeys = [];
     if (!(this._def.catchall instanceof ZodNever && this._def.unknownKeys === "strip")) {
@@ -9626,7 +9628,7 @@ var ZodObject = class _ZodObject extends ZodType {
             code: ZodIssueCode.unrecognized_keys,
             keys: extraKeys
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (unknownKeys === "strip") {
       } else {
@@ -9660,10 +9662,10 @@ var ZodObject = class _ZodObject extends ZodType {
         }
         return syncPairs;
       }).then((syncPairs) => {
-        return ParseStatus.mergeObjectSync(status, syncPairs);
+        return ParseStatus.mergeObjectSync(status2, syncPairs);
       });
     } else {
-      return ParseStatus.mergeObjectSync(status, pairs);
+      return ParseStatus.mergeObjectSync(status2, pairs);
     }
   }
   get shape() {
@@ -10141,7 +10143,7 @@ function mergeValues(a, b) {
 }
 var ZodIntersection = class extends ZodType {
   _parse(input) {
-    const { status, ctx } = this._processInputParams(input);
+    const { status: status2, ctx } = this._processInputParams(input);
     const handleParsed = (parsedLeft, parsedRight) => {
       if (isAborted(parsedLeft) || isAborted(parsedRight)) {
         return INVALID;
@@ -10154,9 +10156,9 @@ var ZodIntersection = class extends ZodType {
         return INVALID;
       }
       if (isDirty(parsedLeft) || isDirty(parsedRight)) {
-        status.dirty();
+        status2.dirty();
       }
-      return { status: status.value, value: merged.data };
+      return { status: status2.value, value: merged.data };
     };
     if (ctx.common.async) {
       return Promise.all([
@@ -10194,7 +10196,7 @@ ZodIntersection.create = (left, right, params) => {
 };
 var ZodTuple = class _ZodTuple extends ZodType {
   _parse(input) {
-    const { status, ctx } = this._processInputParams(input);
+    const { status: status2, ctx } = this._processInputParams(input);
     if (ctx.parsedType !== ZodParsedType.array) {
       addIssueToContext(ctx, {
         code: ZodIssueCode.invalid_type,
@@ -10222,7 +10224,7 @@ var ZodTuple = class _ZodTuple extends ZodType {
         exact: false,
         type: "array"
       });
-      status.dirty();
+      status2.dirty();
     }
     const items = [...ctx.data].map((item, itemIndex) => {
       const schema = this._def.items[itemIndex] || this._def.rest;
@@ -10232,10 +10234,10 @@ var ZodTuple = class _ZodTuple extends ZodType {
     }).filter((x) => !!x);
     if (ctx.common.async) {
       return Promise.all(items).then((results) => {
-        return ParseStatus.mergeArray(status, results);
+        return ParseStatus.mergeArray(status2, results);
       });
     } else {
-      return ParseStatus.mergeArray(status, items);
+      return ParseStatus.mergeArray(status2, items);
     }
   }
   get items() {
@@ -10267,7 +10269,7 @@ var ZodRecord = class _ZodRecord extends ZodType {
     return this._def.valueType;
   }
   _parse(input) {
-    const { status, ctx } = this._processInputParams(input);
+    const { status: status2, ctx } = this._processInputParams(input);
     if (ctx.parsedType !== ZodParsedType.object) {
       addIssueToContext(ctx, {
         code: ZodIssueCode.invalid_type,
@@ -10287,9 +10289,9 @@ var ZodRecord = class _ZodRecord extends ZodType {
       });
     }
     if (ctx.common.async) {
-      return ParseStatus.mergeObjectAsync(status, pairs);
+      return ParseStatus.mergeObjectAsync(status2, pairs);
     } else {
-      return ParseStatus.mergeObjectSync(status, pairs);
+      return ParseStatus.mergeObjectSync(status2, pairs);
     }
   }
   get element() {
@@ -10320,7 +10322,7 @@ var ZodMap = class extends ZodType {
     return this._def.valueType;
   }
   _parse(input) {
-    const { status, ctx } = this._processInputParams(input);
+    const { status: status2, ctx } = this._processInputParams(input);
     if (ctx.parsedType !== ZodParsedType.map) {
       addIssueToContext(ctx, {
         code: ZodIssueCode.invalid_type,
@@ -10347,11 +10349,11 @@ var ZodMap = class extends ZodType {
             return INVALID;
           }
           if (key.status === "dirty" || value.status === "dirty") {
-            status.dirty();
+            status2.dirty();
           }
           finalMap.set(key.value, value.value);
         }
-        return { status: status.value, value: finalMap };
+        return { status: status2.value, value: finalMap };
       });
     } else {
       const finalMap = /* @__PURE__ */ new Map();
@@ -10362,11 +10364,11 @@ var ZodMap = class extends ZodType {
           return INVALID;
         }
         if (key.status === "dirty" || value.status === "dirty") {
-          status.dirty();
+          status2.dirty();
         }
         finalMap.set(key.value, value.value);
       }
-      return { status: status.value, value: finalMap };
+      return { status: status2.value, value: finalMap };
     }
   }
 };
@@ -10380,7 +10382,7 @@ ZodMap.create = (keyType, valueType, params) => {
 };
 var ZodSet = class _ZodSet extends ZodType {
   _parse(input) {
-    const { status, ctx } = this._processInputParams(input);
+    const { status: status2, ctx } = this._processInputParams(input);
     if (ctx.parsedType !== ZodParsedType.set) {
       addIssueToContext(ctx, {
         code: ZodIssueCode.invalid_type,
@@ -10400,7 +10402,7 @@ var ZodSet = class _ZodSet extends ZodType {
           exact: false,
           message: def.minSize.message
         });
-        status.dirty();
+        status2.dirty();
       }
     }
     if (def.maxSize !== null) {
@@ -10413,7 +10415,7 @@ var ZodSet = class _ZodSet extends ZodType {
           exact: false,
           message: def.maxSize.message
         });
-        status.dirty();
+        status2.dirty();
       }
     }
     const valueType = this._def.valueType;
@@ -10423,10 +10425,10 @@ var ZodSet = class _ZodSet extends ZodType {
         if (element.status === "aborted")
           return INVALID;
         if (element.status === "dirty")
-          status.dirty();
+          status2.dirty();
         parsedSet.add(element.value);
       }
-      return { status: status.value, value: parsedSet };
+      return { status: status2.value, value: parsedSet };
     }
     const elements = [...ctx.data.values()].map((item, i) => valueType._parse(new ParseInputLazyPath(ctx, item, ctx.path, i)));
     if (ctx.common.async) {
@@ -10757,15 +10759,15 @@ var ZodEffects = class extends ZodType {
     return this._def.schema._def.typeName === ZodFirstPartyTypeKind.ZodEffects ? this._def.schema.sourceType() : this._def.schema;
   }
   _parse(input) {
-    const { status, ctx } = this._processInputParams(input);
+    const { status: status2, ctx } = this._processInputParams(input);
     const effect = this._def.effect || null;
     const checkCtx = {
       addIssue: (arg) => {
         addIssueToContext(ctx, arg);
         if (arg.fatal) {
-          status.abort();
+          status2.abort();
         } else {
-          status.dirty();
+          status2.dirty();
         }
       },
       get path() {
@@ -10777,7 +10779,7 @@ var ZodEffects = class extends ZodType {
       const processed = effect.transform(ctx.data, checkCtx);
       if (ctx.common.async) {
         return Promise.resolve(processed).then(async (processed2) => {
-          if (status.value === "aborted")
+          if (status2.value === "aborted")
             return INVALID;
           const result = await this._def.schema._parseAsync({
             data: processed2,
@@ -10788,12 +10790,12 @@ var ZodEffects = class extends ZodType {
             return INVALID;
           if (result.status === "dirty")
             return DIRTY(result.value);
-          if (status.value === "dirty")
+          if (status2.value === "dirty")
             return DIRTY(result.value);
           return result;
         });
       } else {
-        if (status.value === "aborted")
+        if (status2.value === "aborted")
           return INVALID;
         const result = this._def.schema._parseSync({
           data: processed,
@@ -10804,7 +10806,7 @@ var ZodEffects = class extends ZodType {
           return INVALID;
         if (result.status === "dirty")
           return DIRTY(result.value);
-        if (status.value === "dirty")
+        if (status2.value === "dirty")
           return DIRTY(result.value);
         return result;
       }
@@ -10829,17 +10831,17 @@ var ZodEffects = class extends ZodType {
         if (inner.status === "aborted")
           return INVALID;
         if (inner.status === "dirty")
-          status.dirty();
+          status2.dirty();
         executeRefinement(inner.value);
-        return { status: status.value, value: inner.value };
+        return { status: status2.value, value: inner.value };
       } else {
         return this._def.schema._parseAsync({ data: ctx.data, path: ctx.path, parent: ctx }).then((inner) => {
           if (inner.status === "aborted")
             return INVALID;
           if (inner.status === "dirty")
-            status.dirty();
+            status2.dirty();
           return executeRefinement(inner.value).then(() => {
-            return { status: status.value, value: inner.value };
+            return { status: status2.value, value: inner.value };
           });
         });
       }
@@ -10857,13 +10859,13 @@ var ZodEffects = class extends ZodType {
         if (result instanceof Promise) {
           throw new Error(`Asynchronous transform encountered during synchronous parse operation. Use .parseAsync instead.`);
         }
-        return { status: status.value, value: result };
+        return { status: status2.value, value: result };
       } else {
         return this._def.schema._parseAsync({ data: ctx.data, path: ctx.path, parent: ctx }).then((base) => {
           if (!isValid(base))
             return INVALID;
           return Promise.resolve(effect.transform(base.value, checkCtx)).then((result) => ({
-            status: status.value,
+            status: status2.value,
             value: result
           }));
         });
@@ -11042,7 +11044,7 @@ var ZodBranded = class extends ZodType {
 };
 var ZodPipeline = class _ZodPipeline extends ZodType {
   _parse(input) {
-    const { status, ctx } = this._processInputParams(input);
+    const { status: status2, ctx } = this._processInputParams(input);
     if (ctx.common.async) {
       const handleAsync = async () => {
         const inResult = await this._def.in._parseAsync({
@@ -11053,7 +11055,7 @@ var ZodPipeline = class _ZodPipeline extends ZodType {
         if (inResult.status === "aborted")
           return INVALID;
         if (inResult.status === "dirty") {
-          status.dirty();
+          status2.dirty();
           return DIRTY(inResult.value);
         } else {
           return this._def.out._parseAsync({
@@ -11073,7 +11075,7 @@ var ZodPipeline = class _ZodPipeline extends ZodType {
       if (inResult.status === "aborted")
         return INVALID;
       if (inResult.status === "dirty") {
-        status.dirty();
+        status2.dirty();
         return {
           status: "dirty",
           value: inResult.value
@@ -11433,10 +11435,10 @@ function assignProp(target, prop, value) {
     configurable: true
   });
 }
-function getElementAtPath(obj, path5) {
-  if (!path5)
+function getElementAtPath(obj, path6) {
+  if (!path6)
     return obj;
-  return path5.reduce((acc, key) => acc?.[key], obj);
+  return path6.reduce((acc, key) => acc?.[key], obj);
 }
 function promiseAllObject(promisesObj) {
   const keys = Object.keys(promisesObj);
@@ -11756,11 +11758,11 @@ function aborted(x, startIndex = 0) {
   }
   return false;
 }
-function prefixIssues(path5, issues) {
+function prefixIssues(path6, issues) {
   return issues.map((iss) => {
     var _a;
     (_a = iss).path ?? (_a.path = []);
-    iss.path.unshift(path5);
+    iss.path.unshift(path6);
     return iss;
   });
 }
@@ -15171,11 +15173,11 @@ function normalizeObjectSchema(schema) {
   }
   return void 0;
 }
-function getDotPath(path5) {
-  if (path5.length === 0) {
+function getDotPath(path6) {
+  if (path6.length === 0) {
     return "object root";
   }
-  return path5.reduce((acc, seg, index) => {
+  return path6.reduce((acc, seg, index) => {
     if (index === 0) {
       return String(seg);
     }
@@ -17474,8 +17476,8 @@ var UrlElicitationRequiredError = class extends McpError {
 };
 
 // ../../node_modules/@modelcontextprotocol/sdk/dist/esm/experimental/tasks/interfaces.js
-function isTerminal(status) {
-  return status === "completed" || status === "failed" || status === "cancelled";
+function isTerminal(status2) {
+  return status2 === "completed" || status2 === "failed" || status2 === "cancelled";
 }
 
 // ../../node_modules/zod-to-json-schema/dist/esm/Options.js
@@ -17534,19 +17536,19 @@ var getRefs = (options) => {
 };
 
 // ../../node_modules/zod-to-json-schema/dist/esm/errorMessages.js
-function addErrorMessage(res, key, errorMessage, refs) {
+function addErrorMessage(res, key, errorMessage2, refs) {
   if (!refs?.errorMessages)
     return;
-  if (errorMessage) {
+  if (errorMessage2) {
     res.errorMessage = {
       ...res.errorMessage,
-      [key]: errorMessage
+      [key]: errorMessage2
     };
   }
 }
-function setResponseValueAndErrors(res, key, value, errorMessage, refs) {
+function setResponseValueAndErrors(res, key, value, errorMessage2, refs) {
   res[key] = value;
-  addErrorMessage(res, key, errorMessage, refs);
+  addErrorMessage(res, key, errorMessage2, refs);
 }
 
 // ../../node_modules/zod-to-json-schema/dist/esm/getRelativePath.js
@@ -18857,8 +18859,8 @@ var Protocol = class {
                   if (queuedMessage.type === "response") {
                     resolver(message);
                   } else {
-                    const errorMessage = message;
-                    const error2 = new McpError(errorMessage.error.code, errorMessage.error.message, errorMessage.error.data);
+                    const errorMessage2 = message;
+                    const error2 = new McpError(errorMessage2.error.code, errorMessage2.error.message, errorMessage2.error.data);
                     resolver(error2);
                   }
                 } else {
@@ -19694,8 +19696,8 @@ var Protocol = class {
         }
         return task;
       },
-      storeTaskResult: async (taskId, status, result) => {
-        await taskStore.storeTaskResult(taskId, status, result, sessionId);
+      storeTaskResult: async (taskId, status2, result) => {
+        await taskStore.storeTaskResult(taskId, status2, result, sessionId);
         const task = await taskStore.getTask(taskId, sessionId);
         if (task) {
           const notification = TaskStatusNotificationSchema.parse({
@@ -19711,15 +19713,15 @@ var Protocol = class {
       getTaskResult: (taskId) => {
         return taskStore.getTaskResult(taskId, sessionId);
       },
-      updateTaskStatus: async (taskId, status, statusMessage) => {
+      updateTaskStatus: async (taskId, status2, statusMessage) => {
         const task = await taskStore.getTask(taskId, sessionId);
         if (!task) {
           throw new McpError(ErrorCode.InvalidParams, `Task "${taskId}" not found - it may have been cleaned up`);
         }
         if (isTerminal(task.status)) {
-          throw new McpError(ErrorCode.InvalidParams, `Cannot update task "${taskId}" from terminal status "${task.status}" to "${status}". Terminal states (completed, failed, cancelled) cannot transition to other states.`);
+          throw new McpError(ErrorCode.InvalidParams, `Cannot update task "${taskId}" from terminal status "${task.status}" to "${status2}". Terminal states (completed, failed, cancelled) cannot transition to other states.`);
         }
-        await taskStore.updateTaskStatus(taskId, status, statusMessage, sessionId);
+        await taskStore.updateTaskStatus(taskId, status2, statusMessage, sessionId);
         const updatedTask = await taskStore.getTask(taskId, sessionId);
         if (updatedTask) {
           const notification = TaskStatusNotificationSchema.parse({
@@ -20149,23 +20151,23 @@ var Server = class extends Protocol {
       const wrappedHandler = async (request, extra) => {
         const validatedRequest = safeParse2(CallToolRequestSchema, request);
         if (!validatedRequest.success) {
-          const errorMessage = validatedRequest.error instanceof Error ? validatedRequest.error.message : String(validatedRequest.error);
-          throw new McpError(ErrorCode.InvalidParams, `Invalid tools/call request: ${errorMessage}`);
+          const errorMessage2 = validatedRequest.error instanceof Error ? validatedRequest.error.message : String(validatedRequest.error);
+          throw new McpError(ErrorCode.InvalidParams, `Invalid tools/call request: ${errorMessage2}`);
         }
         const { params } = validatedRequest.data;
         const result = await Promise.resolve(handler(request, extra));
         if (params.task) {
           const taskValidationResult = safeParse2(CreateTaskResultSchema, result);
           if (!taskValidationResult.success) {
-            const errorMessage = taskValidationResult.error instanceof Error ? taskValidationResult.error.message : String(taskValidationResult.error);
-            throw new McpError(ErrorCode.InvalidParams, `Invalid task creation result: ${errorMessage}`);
+            const errorMessage2 = taskValidationResult.error instanceof Error ? taskValidationResult.error.message : String(taskValidationResult.error);
+            throw new McpError(ErrorCode.InvalidParams, `Invalid task creation result: ${errorMessage2}`);
           }
           return taskValidationResult.data;
         }
         const validationResult = safeParse2(CallToolResultSchema, result);
         if (!validationResult.success) {
-          const errorMessage = validationResult.error instanceof Error ? validationResult.error.message : String(validationResult.error);
-          throw new McpError(ErrorCode.InvalidParams, `Invalid tools/call result: ${errorMessage}`);
+          const errorMessage2 = validationResult.error instanceof Error ? validationResult.error.message : String(validationResult.error);
+          throw new McpError(ErrorCode.InvalidParams, `Invalid tools/call result: ${errorMessage2}`);
         }
         return validationResult.data;
       };
@@ -20659,12 +20661,12 @@ var McpServer = class {
    * @param errorMessage - The error message.
    * @returns The tool error result.
    */
-  createToolError(errorMessage) {
+  createToolError(errorMessage2) {
     return {
       content: [
         {
           type: "text",
-          text: errorMessage
+          text: errorMessage2
         }
       ],
       isError: true
@@ -20682,8 +20684,8 @@ var McpServer = class {
     const parseResult = await safeParseAsync2(schemaToParse, args);
     if (!parseResult.success) {
       const error2 = "error" in parseResult ? parseResult.error : "Unknown error";
-      const errorMessage = getParseErrorMessage(error2);
-      throw new McpError(ErrorCode.InvalidParams, `Input validation error: Invalid arguments for tool ${toolName}: ${errorMessage}`);
+      const errorMessage2 = getParseErrorMessage(error2);
+      throw new McpError(ErrorCode.InvalidParams, `Input validation error: Invalid arguments for tool ${toolName}: ${errorMessage2}`);
     }
     return parseResult.data;
   }
@@ -20707,8 +20709,8 @@ var McpServer = class {
     const parseResult = await safeParseAsync2(outputObj, result.structuredContent);
     if (!parseResult.success) {
       const error2 = "error" in parseResult ? parseResult.error : "Unknown error";
-      const errorMessage = getParseErrorMessage(error2);
-      throw new McpError(ErrorCode.InvalidParams, `Output validation error: Invalid structured content for tool ${toolName}: ${errorMessage}`);
+      const errorMessage2 = getParseErrorMessage(error2);
+      throw new McpError(ErrorCode.InvalidParams, `Output validation error: Invalid structured content for tool ${toolName}: ${errorMessage2}`);
     }
   }
   /**
@@ -20920,8 +20922,8 @@ var McpServer = class {
         const parseResult = await safeParseAsync2(argsObj, request.params.arguments);
         if (!parseResult.success) {
           const error2 = "error" in parseResult ? parseResult.error : "Unknown error";
-          const errorMessage = getParseErrorMessage(error2);
-          throw new McpError(ErrorCode.InvalidParams, `Invalid arguments for prompt ${request.params.name}: ${errorMessage}`);
+          const errorMessage2 = getParseErrorMessage(error2);
+          throw new McpError(ErrorCode.InvalidParams, `Invalid arguments for prompt ${request.params.name}: ${errorMessage2}`);
         }
         const args = parseResult.data;
         const cb = prompt.callback;
@@ -21431,32 +21433,47 @@ var StdioServerTransport = class {
   }
 };
 
+// src/native.ts
+import { createRequire } from "node:module";
+import native from "../native/index.js";
+var { createScope, getScope, forkScope, deleteScope, warmModelRequest, doctor, machine, health, status, usage, createSession, deleteSession, killSession, listModels, route, runTaskRequest, runTaskBatchRequest, listTaskJobs, getTaskJob, cancelTaskJob, removeTaskJob } = native;
+var { version: SERVER_VERSION } = createRequire(import.meta.url)("../package.json");
+
+// src/lifecycle.ts
+import { spawn as spawn2 } from "node:child_process";
+
+// src/serve.ts
+import { spawn } from "node:child_process";
+import { existsSync as existsSync2 } from "node:fs";
+import { createRequire as createRequire2 } from "node:module";
+import path2 from "node:path";
+import { fileURLToPath as fileURLToPath2 } from "node:url";
+
 // src/config.ts
 import { existsSync } from "node:fs";
 import { realpath } from "node:fs/promises";
 import path from "node:path";
 import { fileURLToPath } from "node:url";
 import { homedir } from "node:os";
+var PACKAGE_ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
 function findRepoRoot() {
   let dir = path.dirname(fileURLToPath(import.meta.url));
   for (let depth = 0; depth < 10; depth += 1) {
-    if (existsSync(path.join(dir, "Cargo.toml"))) return dir;
+    if (existsSync(path.join(dir, "Cargo.toml")) && path.join(dir, "packages", "mcp") === PACKAGE_ROOT) return dir;
     const parent = path.dirname(dir);
     if (parent === dir) break;
     dir = parent;
   }
   return void 0;
 }
-var REPO_ROOT = findRepoRoot() ?? path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
-var PACKAGE_ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
+var CHECKOUT_ROOT = findRepoRoot();
+var REPO_ROOT = CHECKOUT_ROOT ?? PACKAGE_ROOT;
 function adapterPath(file) {
-  const bundled = path.join(PACKAGE_ROOT, "adapters", file);
-  if (existsSync(bundled)) return bundled;
-  return path.join(REPO_ROOT, "benchmark/local/scripts", file);
+  return path.join(PACKAGE_ROOT, "dist", "research", file);
 }
 var RESEARCH_ADAPTERS = {
-  bash: adapterPath("bash_agent.py"),
-  octocode: adapterPath("octocode_agent.py")
+  bash: adapterPath("bash-agent.js"),
+  octocode: adapterPath("octocode-agent.js")
 };
 var DEFAULT_RESEARCH_ADAPTER = process.env.FREELLAMA_MCP_DEFAULT_ADAPTER === "octocode" ? "octocode" : "bash";
 function envInt(name, fallback) {
@@ -21467,7 +21484,7 @@ function envInt(name, fallback) {
 }
 var DEFAULT_OLLAMA_ENDPOINT = process.env.FREELLAMA_OLLAMA_ENDPOINT ?? "http://127.0.0.1:11434";
 var DEFAULT_SERVE_ENDPOINT = process.env.FREELLAMA_SERVE_ENDPOINT ?? "http://127.0.0.1:11435";
-var DEFAULT_DELEGATE_MODEL = process.env.FREELLAMA_MCP_DEFAULT_MODEL ?? "qwen3.8:27b-mlx";
+var DEFAULT_DELEGATE_MODEL = process.env.FREELLAMA_MCP_DEFAULT_MODEL?.trim() || void 0;
 var DEFAULT_DELEGATE_MAX_TURNS = envInt("FREELLAMA_MCP_MAX_TURNS", 8);
 var DEFAULT_DELEGATE_TIMEOUT_SECONDS = envInt("FREELLAMA_MCP_DELEGATE_TIMEOUT_SECONDS", 180);
 var DEFAULT_PULL_TIMEOUT_SECONDS = envInt("FREELLAMA_MCP_PULL_TIMEOUT_SECONDS", 1200);
@@ -21481,7 +21498,7 @@ function parseAllowedResearchRoots(raw, delimiter = path.delimiter) {
   return raw.split(delimiter).filter(Boolean).map((root) => path.resolve(root));
 }
 var ALLOWED_RESEARCH_ROOTS = parseAllowedResearchRoots(
-  process.env.FREELLAMA_MCP_ALLOWED_ROOTS ?? findRepoRoot() ?? ""
+  process.env.FREELLAMA_MCP_ALLOWED_ROOTS ?? CHECKOUT_ROOT ?? ""
 );
 var resolvedRootsPromise = null;
 function allowedResearchRoots() {
@@ -21521,59 +21538,315 @@ async function assertAllowedWorkspace(workspacePath) {
   }
   return resolved;
 }
+var DELEGATE_ENV_ALLOWLIST = /* @__PURE__ */ new Set([
+  "PATH",
+  "HOME",
+  "USER",
+  "LOGNAME",
+  "LANG",
+  "LC_ALL",
+  "LC_CTYPE",
+  "TZ",
+  "TMPDIR",
+  "TEMP",
+  "TMP",
+  "SYSTEMROOT",
+  "SYSTEMDRIVE",
+  "WINDIR",
+  "COMSPEC",
+  "PATHEXT",
+  "APPDATA",
+  "LOCALAPPDATA",
+  "USERPROFILE",
+  "HTTP_PROXY",
+  "HTTPS_PROXY",
+  "NO_PROXY",
+  "http_proxy",
+  "https_proxy",
+  "no_proxy",
+  "NODE_EXTRA_CA_CERTS",
+  "SSL_CERT_FILE",
+  "SSL_CERT_DIR",
+  "REQUESTS_CA_BUNDLE",
+  "XDG_CACHE_HOME",
+  "XDG_DATA_HOME",
+  "npm_config_cache"
+]);
+function delegateEnvironment(source = process.env) {
+  const env = {};
+  for (const [key, value] of Object.entries(source)) {
+    if (value !== void 0 && (DELEGATE_ENV_ALLOWLIST.has(key) || key.startsWith("FREELLAMA_"))) env[key] = value;
+  }
+  return env;
+}
 
-// src/native.ts
-import { createRequire } from "node:module";
-import native from "../native/index.js";
-var { doctor, machine, health, createSession, deleteSession, listModels, route, runTask, runTaskRequest, runTaskBatchRequest } = native;
-var { version: SERVER_VERSION } = createRequire(import.meta.url)("../package.json");
-
-// src/helpers.ts
-import { existsSync as existsSync2, realpathSync, statSync } from "node:fs";
-import path2 from "node:path";
-function extractExistingWorkspacePath(command, workspace) {
-  let root;
+// src/serve.ts
+var HEALTH_PATH = "/_freellama/v1/health";
+var STARTUP_TIMEOUT_MS = 15e3;
+var child = null;
+var starting = null;
+function withServe(call) {
+  return async (endpoint, ...args) => {
+    await ensureServe(endpoint ?? void 0);
+    return call(endpoint, ...args);
+  };
+}
+function autostartEnabled() {
+  return process.env.FREELLAMA_MCP_AUTOSTART_SERVE !== "0";
+}
+function loopbackListen(endpoint) {
   try {
-    root = realpathSync(workspace);
+    const url = new URL(endpoint);
+    if (!["127.0.0.1", "localhost", "[::1]"].includes(url.hostname)) return null;
+    return `${url.hostname === "localhost" ? "127.0.0.1" : url.hostname}:${url.port || "80"}`;
   } catch {
     return null;
   }
-  const tokens = command.match(/(?:[^\s"'\\]+|"(?:\\.|[^"])*"|'[^']*')+/g) ?? [];
-  const patternFlags = /* @__PURE__ */ new Set(["-name", "-iname", "-path", "-ipath", "-regex", "-iregex", "-wholename"]);
-  for (let index = 0; index < tokens.length; index += 1) {
-    const raw = tokens[index] ?? "";
-    if (index > 0 && patternFlags.has(tokens[index - 1] ?? "")) continue;
-    const token = raw.replace(/^["']|["']$/g, "").replace(/[|;&)]+$/g, "").replace(/:\d+(?::\d+)?$/, "");
-    if (!token || token.startsWith("-") || /[*?$`]/.test(token)) continue;
-    const candidate = path2.isAbsolute(token) ? token : path2.resolve(root, token);
-    if (!existsSync2(candidate)) continue;
+}
+async function answers(endpoint, timeoutMs) {
+  try {
+    await fetch(`${endpoint.replace(/\/$/, "")}${HEALTH_PATH}`, { signal: AbortSignal.timeout(timeoutMs) });
+    return true;
+  } catch {
+    return false;
+  }
+}
+function serveBinaryCandidates() {
+  const exe = process.platform === "win32" ? "freellama.exe" : "freellama";
+  const candidates = [
+    path2.join(REPO_ROOT, "target", "release", exe),
+    path2.join(REPO_ROOT, "target", "debug", exe)
+  ];
+  const requireFromHere = createRequire2(fileURLToPath2(import.meta.url));
+  const suffixes = process.platform === "linux" ? ["-gnu", "-musl"] : process.platform === "win32" ? ["-msvc"] : [""];
+  for (const suffix of suffixes) {
     try {
-      const resolved = realpathSync(candidate);
-      const relative = path2.relative(root, resolved);
-      const withinRoot = relative && relative !== ".." && !relative.startsWith(`..${path2.sep}`);
-      if (withinRoot && statSync(resolved).isFile()) return relative;
+      const manifest = requireFromHere.resolve(`@octocodeai/freellama-native-${process.platform}-${process.arch}${suffix}/package.json`);
+      candidates.push(path2.join(path2.dirname(manifest), exe));
     } catch {
     }
   }
-  return null;
+  if (process.env.FREELLAMA_SERVE_BINARY) candidates.unshift(process.env.FREELLAMA_SERVE_BINARY);
+  return candidates;
+}
+function stopAutostartedServe() {
+  if (child && child.exitCode === null) child.kill("SIGTERM");
+  child = null;
+}
+async function ensureServe(endpoint) {
+  const target = endpoint ?? DEFAULT_SERVE_ENDPOINT;
+  if (!autostartEnabled() || target !== DEFAULT_SERVE_ENDPOINT) return;
+  const listen = loopbackListen(target);
+  if (!listen) return;
+  if (child && child.exitCode === null && await answers(target, 1e3)) return;
+  starting ??= (async () => {
+    try {
+      if (await answers(target, 1e3)) return;
+      const binary = serveBinaryCandidates().find((candidate) => existsSync2(candidate));
+      if (!binary) return;
+      const spawned = spawn(binary, ["serve", "--listen", listen, "--upstream", DEFAULT_OLLAMA_ENDPOINT], {
+        stdio: ["ignore", "ignore", "pipe"]
+      });
+      child = spawned;
+      spawned.stderr?.on("data", (chunk) => process.stderr.write(`[freellama serve] ${chunk}`));
+      spawned.once("exit", () => {
+        if (child === spawned) child = null;
+      });
+      await new Promise((resolve, reject) => {
+        spawned.once("spawn", resolve);
+        spawned.once("error", (error2) => {
+          if (child === spawned) child = null;
+          reject(new Error(`Could not start freellama serve at ${target}: ${error2.message}`));
+        });
+      });
+      const deadline = Date.now() + STARTUP_TIMEOUT_MS;
+      while (Date.now() < deadline && child === spawned) {
+        if (await answers(target, 500)) return;
+        await new Promise((resolve) => setTimeout(resolve, 200));
+      }
+      if (child === spawned) {
+        stopAutostartedServe();
+        throw new Error(`freellama serve did not become ready at ${target} within ${STARTUP_TIMEOUT_MS}ms.`);
+      }
+      throw new Error(`freellama serve exited before becoming ready at ${target} (exit ${spawned.exitCode}, signal ${spawned.signalCode}).`);
+    } finally {
+      starting = null;
+    }
+  })();
+  await starting;
+}
+
+// src/lifecycle.ts
+var liveDelegates = /* @__PURE__ */ new Set();
+var DETACH_DELEGATES = process.platform !== "win32";
+function killDelegateTree(child2) {
+  if (DETACH_DELEGATES && child2.pid !== void 0) {
+    try {
+      process.kill(-child2.pid, "SIGKILL");
+      return;
+    } catch {
+    }
+  }
+  try {
+    child2.kill("SIGKILL");
+  } catch {
+  }
+}
+function runDelegate(file, args, options) {
+  const child2 = spawn2(file, args, { env: options.env, detached: DETACH_DELEGATES, stdio: ["ignore", "pipe", "pipe"] });
+  const done = new Promise((resolve, reject) => {
+    const out = [];
+    const err = [];
+    let outLen = 0;
+    let errLen = 0;
+    let failure;
+    let timedOut = false;
+    let settled = false;
+    const fail = (error2) => {
+      if (failure === void 0) failure = error2;
+      killDelegateTree(child2);
+    };
+    const timer = setTimeout(() => {
+      timedOut = true;
+      killDelegateTree(child2);
+    }, options.timeoutMs);
+    const onAbort = () => {
+      const abort = new Error("The operation was aborted", { cause: options.signal?.reason });
+      abort.name = "AbortError";
+      abort.code = "ABORT_ERR";
+      fail(abort);
+    };
+    if (options.signal?.aborted) onAbort();
+    else options.signal?.addEventListener("abort", onAbort, { once: true });
+    const collect = (chunks, name) => (chunk) => {
+      const length = name === "stdout" ? outLen += chunk.length : errLen += chunk.length;
+      if (length > options.maxBuffer) {
+        const overflow = new RangeError(`${name} maxBuffer length exceeded`);
+        overflow.code = "ERR_CHILD_PROCESS_STDIO_MAXBUFFER";
+        fail(overflow);
+        return;
+      }
+      chunks.push(chunk);
+    };
+    child2.stdout.on("data", collect(out, "stdout"));
+    child2.stderr.on("data", collect(err, "stderr"));
+    const finish = (code, signal, spawnError) => {
+      if (settled) return;
+      settled = true;
+      clearTimeout(timer);
+      options.signal?.removeEventListener("abort", onAbort);
+      const stdout = Buffer.concat(out).toString("utf8");
+      const stderr = Buffer.concat(err).toString("utf8");
+      if (spawnError ?? failure) {
+        reject(spawnError ?? failure);
+        return;
+      }
+      if (code === 0 && signal === null) {
+        resolve({ stdout, stderr });
+        return;
+      }
+      const error2 = new Error(`Command failed: ${[file, ...args].join(" ")}
+${stderr}`);
+      Object.assign(error2, { code, killed: timedOut, signal, stdout, stderr });
+      reject(error2);
+    };
+    child2.once("error", (error2) => finish(null, null, error2));
+    child2.once("close", (code, signal) => finish(code, signal));
+  });
+  return { child: child2, done };
+}
+function killLiveDelegates() {
+  for (const child2 of liveDelegates) killDelegateTree(child2);
+  liveDelegates.clear();
+}
+function installProcessLifecycle() {
+  for (const stream of [process.stdout, process.stderr]) {
+    stream.on("error", (error2) => {
+      if (error2.code === "EPIPE") process.exit(0);
+      throw error2;
+    });
+  }
+  process.on("exit", () => {
+    killLiveDelegates();
+    stopAutostartedServe();
+  });
+  process.stdin.on("close", () => {
+    killLiveDelegates();
+    stopAutostartedServe();
+    process.exit(0);
+  });
+  for (const signal of ["SIGINT", "SIGTERM", "SIGHUP"]) {
+    process.on(signal, () => {
+      killLiveDelegates();
+      process.exit(0);
+    });
+  }
+}
+
+// src/schema-dialect.ts
+function stripSchemaDialect(message) {
+  const tools = message.result?.tools;
+  if (!Array.isArray(tools)) return message;
+  for (const tool of tools) {
+    delete tool.inputSchema?.$schema;
+    delete tool.outputSchema?.$schema;
+  }
+  return message;
+}
+function withoutSchemaDialect(transport2) {
+  const send = transport2.send.bind(transport2);
+  transport2.send = (message, options) => send(stripSchemaDialect(message), options);
+  return transport2;
+}
+
+// src/docs-resources.ts
+import { existsSync as existsSync3, readdirSync } from "node:fs";
+import { readFile } from "node:fs/promises";
+import path3 from "node:path";
+import { fileURLToPath as fileURLToPath3 } from "node:url";
+function registerDocsResources(server2) {
+  const PACKAGED_DOCS_DIR = path3.resolve(path3.dirname(fileURLToPath3(import.meta.url)), "..", "docs");
+  if (!existsSync3(path3.join(PACKAGED_DOCS_DIR, "README.md"))) {
+    throw new Error(`FreeLlama MCP documentation index is missing at ${PACKAGED_DOCS_DIR}/README.md; run the package build.`);
+  }
+  const packagedDocs = readdirSync(PACKAGED_DOCS_DIR).filter((name) => name.endsWith(".md")).sort();
+  for (const name of packagedDocs) {
+    const isIndex = name === "README.md";
+    const slug = isIndex ? "index" : name.replace(/\.md$/, "");
+    const uri = `freellama://docs/${slug}`;
+    server2.registerResource(
+      `freellama-docs-${slug.toLowerCase()}`,
+      uri,
+      {
+        mimeType: "text/markdown",
+        description: isIndex ? "Index of packaged FreeLlama operator and agent documentation. Read this first, then fetch one relevant document." : `Packaged FreeLlama documentation: ${name}.`
+      },
+      async () => ({
+        contents: [{ uri, mimeType: "text/markdown", text: await readFile(path3.join(PACKAGED_DOCS_DIR, name), "utf8") }]
+      })
+    );
+  }
 }
-async function ollamaFetch(endpoint, path5, init = {}) {
+
+// src/helpers.ts
+async function ollamaFetch(endpoint, path6, init = {}) {
   const base = (endpoint ?? DEFAULT_OLLAMA_ENDPOINT).replace(/\/$/, "");
   const controller = new AbortController();
   const timeout = setTimeout(
     () => controller.abort(),
     init.timeoutMs ?? DEFAULT_OLLAMA_FETCH_TIMEOUT_SECONDS * 1e3
   );
+  const signal = init.signal ? AbortSignal.any([controller.signal, init.signal]) : controller.signal;
   try {
-    const response = await fetch(`${base}${path5}`, {
+    const response = await fetch(`${base}${path6}`, {
       method: init.method ?? "GET",
       headers: init.body ? { "content-type": "application/json" } : void 0,
       body: init.body ? JSON.stringify(init.body) : void 0,
-      signal: controller.signal
+      signal
     });
-    const text = await response.text();
+    const text = init.onLine && response.ok && response.body ? await readNdjson(response.body, init.onLine) : await response.text();
     if (!response.ok) {
-      throw new Error(`Ollama ${init.method ?? "GET"} ${path5} -> HTTP ${response.status}: ${text}`);
+      throw new Error(`Ollama ${init.method ?? "GET"} ${path6} -> HTTP ${response.status}: ${text}`);
     }
     if (!text) return {};
     if (init.parse === "ndjson") return text;
@@ -21582,62 +21855,28 @@ async function ollamaFetch(endpoint, path5, init = {}) {
     clearTimeout(timeout);
   }
 }
-function summarizeOllamaPullStream(text) {
-  const events = [];
-  for (const line of text.split(/\r?\n/)) {
-    const trimmed = line.trim();
-    if (!trimmed) continue;
-    try {
-      const parsed = JSON.parse(trimmed);
-      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-        events.push(parsed);
-      }
-    } catch {
-    }
-  }
-  if (events.length === 0) {
-    if (!text.trim()) return { status: "empty" };
-    try {
-      const parsed = JSON.parse(text);
-      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-        return parsed;
+async function readNdjson(body, onLine) {
+  const decoder = new TextDecoder();
+  let text = "";
+  let pending = "";
+  for await (const chunk of body) {
+    const piece = decoder.decode(chunk, { stream: true });
+    text += piece;
+    pending += piece;
+    const lines = pending.split("\n");
+    pending = lines.pop() ?? "";
+    for (const line of lines) {
+      try {
+        const parsed = JSON.parse(line);
+        if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) onLine(parsed);
+      } catch {
       }
-    } catch {
-      return { status: "unparsed", raw: clipText(text, 500) };
     }
-    return { status: "unparsed", raw: clipText(text, 500) };
   }
-  const last = events[events.length - 1] ?? {};
-  const withBytes = [...events].reverse().find((event) => typeof event.total === "number" && event.total > 0);
-  const completed = typeof withBytes?.completed === "number" ? withBytes.completed : null;
-  const total = typeof withBytes?.total === "number" ? withBytes.total : null;
-  const percent = completed != null && total != null && total > 0 ? Math.round(completed / total * 1e3) / 10 : null;
-  const status = typeof last.error === "string" && last.error ? "error" : last.status;
-  return {
-    ...last,
-    ...status === void 0 ? {} : { status },
-    progress: {
-      events: events.length,
-      completed,
-      total,
-      percent,
-      lastStatus: last.status ?? null
-    }
-  };
-}
-async function ollamaPull(endpoint, model, timeoutSeconds) {
-  const raw = await ollamaFetch(endpoint, "/api/pull", {
-    method: "POST",
-    body: { name: model, stream: true },
-    timeoutMs: (timeoutSeconds ?? DEFAULT_PULL_TIMEOUT_SECONDS) * 1e3,
-    parse: "ndjson"
-  });
-  if (typeof raw === "string") return summarizeOllamaPullStream(raw);
-  if (raw && typeof raw === "object" && !Array.isArray(raw)) return raw;
-  return { status: "empty" };
+  return text + decoder.decode();
 }
-var endpointParam = external_exports.string().min(1).optional().describe("serve endpoint, default :11435");
-var ollamaEndpointParam = external_exports.string().min(1).optional().describe("Ollama endpoint, default :11434");
+var endpointParam = external_exports.string().min(1).optional();
+var ollamaEndpointParam = external_exports.string().min(1).optional();
 var TASK_KINDS = [
   "completion",
   "coding",
@@ -21649,28 +21888,14 @@ var TASK_KINDS = [
   "long_context"
 ];
 var MCP_TASK_KINDS = [...TASK_KINDS, "code_review"];
-var taskParam = external_exports.enum(MCP_TASK_KINDS).describe("code_review=coding.");
+var taskParam = external_exports.enum(MCP_TASK_KINDS).default("completion").describe("caller owns prompts and output format.");
 function canonicalTaskKind(task) {
   return task === "code_review" ? "coding" : task;
 }
-var objectiveParam = external_exports.enum(["fastest", "balanced", "quality"]).optional().describe('"balanced"/"quality" need a policy; "fastest" does not');
-var executionPreferenceParam = external_exports.enum(["auto", "prefer_cpu", "prefer_gpu"]).optional().describe(
-  'Backend hint. "auto" uses guarded runtime feedback; prefer_* falls back when no eligible operator-assigned model exists.'
-);
-var minPlacementEvidenceParam = external_exports.enum(["configured", "observed"]).optional().describe('"observed" fails closed unless Ollama /api/ps confirms the selected processor');
-var CONFIDENCE_RANK = { low: 1, medium: 2 };
-var minConfidenceParam = external_exports.enum(["low", "medium"]).optional().describe('Fail closed below this. "medium" needs policy AND benchmark; "low" accepts capability metadata.');
-function belowConfidence(decision, minConfidence) {
-  if (!minConfidence) return null;
-  const actual = typeof decision.confidence === "string" ? decision.confidence : "low";
-  if ((CONFIDENCE_RANK[actual] ?? 0) >= (CONFIDENCE_RANK[minConfidence] ?? 0)) return null;
-  return errorResult(
-    new Error(
-      `Route refused: confidence is "${actual}" (evidence: ${decision.evidence}), below the requested minimum "${minConfidence}". Selected model would have been "${decision.selected_model}" for reasons [${decision.reasons?.join(", ")}].
-This is a fail-closed refusal, not a failure: the local router cannot justify this pick. Escalate to your own model, pass an explicit \`model\`, or run \`bench-all\` and configure a task policy to raise the evidence level.`
-    )
-  );
-}
+var objectiveParam = external_exports.enum(["fastest", "balanced", "quality"]).optional();
+var executionPreferenceParam = external_exports.enum(["auto", "prefer_cpu", "prefer_gpu"]).optional();
+var minPlacementEvidenceParam = external_exports.enum(["configured", "observed"]).optional();
+var minConfidenceParam = external_exports.enum(["low", "medium"]).optional();
 var REQUIRED_CAPABILITIES = [
   "completion",
   "tools",
@@ -21679,22 +21904,51 @@ var REQUIRED_CAPABILITIES = [
   "thinking",
   "embedding"
 ];
-var requiredCapabilitiesParam = external_exports.array(external_exports.enum(REQUIRED_CAPABILITIES)).max(REQUIRED_CAPABILITIES.length).optional().describe('Additional hard requirements, e.g. ["vision"] or ["tools"]. Fails closed if unmet.');
+var requiredCapabilitiesParam = external_exports.array(external_exports.enum(REQUIRED_CAPABILITIES)).max(REQUIRED_CAPABILITIES.length).optional();
+var systemPromptParam = external_exports.string().optional().describe("Local model instructions; verbatim.");
+var messagesParam = external_exports.array(external_exports.object({
+  role: external_exports.enum(["system", "user", "assistant", "tool"]),
+  content: external_exports.string().optional()
+}).passthrough()).min(1).optional();
+var localToolsParam = external_exports.array(external_exports.object({
+  type: external_exports.literal("function"),
+  function: external_exports.object({
+    name: external_exports.string().min(1),
+    description: external_exports.string().optional(),
+    parameters: external_exports.record(external_exports.unknown()).optional()
+  }).passthrough()
+}).passthrough()).min(1).optional();
+function taskMessages({ systemPrompt, messages, prompt, images }) {
+  if (systemPrompt === void 0) return messages ?? [];
+  const conversation = messages ?? (prompt === void 0 ? [] : [{
+    role: "user",
+    content: prompt,
+    ...images === void 0 ? {} : { images }
+  }]);
+  return [{ role: "system", content: systemPrompt }, ...conversation];
+}
 var batchTaskParam = external_exports.object({
-  task: taskParam,
+  // Nested `task` shares the top-level description; repeating it re-sends the text every request.
+  task: taskParam.removeDefault().optional(),
   objective: objectiveParam,
   model: external_exports.string().min(1).optional(),
   sessionId: external_exports.string().uuid().optional(),
+  scopeId: external_exports.string().uuid().optional(),
+  scopeRevision: external_exports.number().int().nonnegative().optional(),
   contextTokens: external_exports.number().int().positive().optional(),
   executionPreference: executionPreferenceParam,
   minPlacementEvidence: minPlacementEvidenceParam,
   requiredCapabilities: requiredCapabilitiesParam,
   priority: external_exports.enum(["interactive", "normal", "background"]).optional(),
+  maxWaitSeconds: external_exports.number().int().positive().optional(),
+  timeoutSeconds: external_exports.number().int().positive().optional(),
+  stopOnRepetition: external_exports.boolean().optional().describe("Stream and stop once output repeats itself; keeps all generated text"),
   prompt: external_exports.string().min(1).optional(),
-  messages: external_exports.array(external_exports.object({ role: external_exports.enum(["system", "user", "assistant", "tool"]), content: external_exports.string() }).passthrough()).min(1).optional(),
+  systemPrompt: systemPromptParam,
+  messages: messagesParam,
   images: external_exports.array(external_exports.string().min(1)).min(1).optional(),
   input: external_exports.union([external_exports.string().min(1), external_exports.array(external_exports.string().min(1)).min(1)]).optional(),
-  tools: external_exports.array(external_exports.record(external_exports.unknown())).min(1).optional(),
+  tools: localToolsParam,
   keepAlive: external_exports.string().min(1).optional(),
   format: external_exports.union([external_exports.literal("json"), external_exports.record(external_exports.unknown())]).optional(),
   think: external_exports.union([external_exports.boolean(), external_exports.enum(["low", "medium", "high"])]).optional(),
@@ -21703,52 +21957,27 @@ var batchTaskParam = external_exports.object({
   topLogprobs: external_exports.number().int().nonnegative().optional(),
   minConfidence: minConfidenceParam
 }).strict();
+function fresh(schema) {
+  const def = { ...schema._def };
+  for (const [key, value] of Object.entries(def)) {
+    if (value instanceof external_exports.ZodType) def[key] = fresh(value);
+    else if (Array.isArray(value) && value.length > 0 && value.every((v) => v instanceof external_exports.ZodType)) {
+      def[key] = value.map((v) => fresh(v));
+    }
+  }
+  if (schema instanceof external_exports.ZodObject) {
+    const shape = Object.fromEntries(Object.entries(schema.shape).map(([k, v]) => [k, fresh(v)]));
+    def.shape = () => shape;
+  }
+  return new schema.constructor(def);
+}
+var BATCH_TASK_FIELDS = new Set(Object.keys(batchTaskParam.shape));
 var batchItemParam = external_exports.object({
   id: external_exports.string().min(1),
   independent: external_exports.literal(true),
-  task: batchTaskParam
+  task: fresh(batchTaskParam)
 }).strict();
 var objectResultSchema = external_exports.object({}).passthrough();
-var doctorResultSchema = external_exports.object({
-  summary: external_exports.string().optional(),
-  endpoint: external_exports.string().optional(),
-  scheduler: external_exports.unknown().optional()
-}).passthrough();
-var sessionResultSchema = external_exports.object({
-  session_id: external_exports.string().optional(),
-  deleted: external_exports.boolean().optional()
-}).passthrough();
-var manageResultSchema = external_exports.object({
-  status: external_exports.string().optional(),
-  progress: external_exports.unknown().optional()
-}).passthrough();
-var deleteResultSchema = external_exports.object({
-  deleted: external_exports.string()
-}).passthrough();
-var modelsResultSchema = external_exports.object({
-  models: external_exports.array(external_exports.record(external_exports.unknown())).optional(),
-  tags: external_exports.array(external_exports.record(external_exports.unknown())).optional(),
-  page: external_exports.unknown().optional()
-}).passthrough();
-var taskResultSchema = external_exports.object({
-  selected_model: external_exports.string().optional(),
-  context_window_fit: external_exports.string().optional(),
-  execution: external_exports.unknown().optional(),
-  response: external_exports.unknown().optional(),
-  telemetry: external_exports.unknown().optional()
-}).passthrough();
-var batchResultSchema = external_exports.object({
-  results: external_exports.unknown().optional(),
-  telemetry: external_exports.unknown().optional()
-}).passthrough();
-var researchResultSchema = external_exports.object({
-  answer: external_exports.string(),
-  summary: external_exports.string(),
-  citations: external_exports.unknown().optional(),
-  verification: external_exports.unknown(),
-  usage: external_exports.unknown().optional(),
-  telemetry: external_exports.unknown().optional()
-}).passthrough();
 function configuredExternalCost(env = process.env) {
   const model = env.FREELLAMA_EXTERNAL_COST_MODEL;
   const inputRaw = env.FREELLAMA_EXTERNAL_COST_INPUT_USD_PER_M;
@@ -21763,22 +21992,22 @@ function configuredExternalCost(env = process.env) {
   }
   return { model, input, output };
 }
-function costTelemetry(usage, externalCost) {
+function costTelemetry(usage2, externalCost) {
   const local = {
-    inputTokens: usage.inputTokens,
-    outputTokens: usage.outputTokens,
-    totalTokens: usage.inputTokens === null || usage.outputTokens === null ? null : usage.inputTokens + usage.outputTokens,
-    totalDurationNs: usage.totalDurationNs ?? null
+    inputTokens: usage2.inputTokens,
+    outputTokens: usage2.outputTokens,
+    totalTokens: usage2.inputTokens === null || usage2.outputTokens === null ? null : usage2.inputTokens + usage2.outputTokens,
+    totalDurationNs: usage2.totalDurationNs ?? null
   };
-  if (!externalCost || usage.inputTokens === null || usage.outputTokens === null) {
+  if (!externalCost || usage2.inputTokens === null || usage2.outputTokens === null) {
     return {
       local,
       externalEquivalent: null,
       note: externalCost ? "External equivalent unavailable because Ollama did not report both input and output token counts." : "No external rate card supplied; local usage is observed, but avoided external cost is not estimated."
     };
   }
-  const inputUsd = usage.inputTokens * externalCost.input / 1e6;
-  const outputUsd = usage.outputTokens * externalCost.output / 1e6;
+  const inputUsd = usage2.inputTokens * externalCost.input / 1e6;
+  const outputUsd = usage2.outputTokens * externalCost.output / 1e6;
   return {
     local,
     externalEquivalent: {
@@ -21805,20 +22034,67 @@ function clipText(text, limit) {
   const head = Math.floor(usable * 2 / 3);
   return text.slice(0, head) + marker + text.slice(-(usable - head));
 }
-var PRETTY_PRINT_MAX_BYTES = 8 * 1024;
-function serialize(value) {
-  const compact = JSON.stringify(value);
-  return compact.length > PRETTY_PRINT_MAX_BYTES ? compact : JSON.stringify(value, null, 2);
-}
 function resultSummary(value) {
   const explicit = value.summary;
   if (typeof explicit === "string" && explicit.trim()) return clipText(explicit.trim(), 500);
   const keys = Object.keys(value);
   return `Structured result available (${keys.slice(0, 8).join(", ") || "empty object"}).`;
 }
+var ANSWER_TEXT_MAX_CHARS = 8 * 1024;
+function record2(value) {
+  return value && typeof value === "object" && !Array.isArray(value) ? value : void 0;
+}
+function taskAnswerText(payload) {
+  const response = record2(payload.response);
+  if (!response) return void 0;
+  const model = record2(payload.route)?.selected_model ?? payload.selected_model ?? response.model;
+  const header = typeof model === "string" && model ? `[${model}] ` : "";
+  const message = record2(response.message);
+  if (message) {
+    const parts = [];
+    if (typeof message.content === "string" && message.content.trim()) parts.push(answerContent(message.content, payload));
+    if (Array.isArray(message.tool_calls) && message.tool_calls.length > 0) {
+      parts.push(`tool_calls: ${JSON.stringify(message.tool_calls)}`);
+    }
+    if (parts.length === 0) parts.push("(empty response)");
+    return clipText(header + parts.join("\n"), ANSWER_TEXT_MAX_CHARS);
+  }
+  if (typeof response.response === "string") return clipText(header + response.response.trim(), ANSWER_TEXT_MAX_CHARS);
+  const omitted = record2(response.embeddings_omitted);
+  if (Array.isArray(response.embeddings) || omitted) {
+    const count = Array.isArray(response.embeddings) ? response.embeddings.length : omitted?.count;
+    const where = omitted ? "vectors withheld; pass returnEmbeddings:true" : "vectors are in structuredContent";
+    return `${header}embedding complete (${String(count ?? "?")} vectors); ${where}.`;
+  }
+  return void 0;
+}
+function answerContent(content, payload) {
+  const check2 = record2(payload.output_check);
+  const keep = check2?.unique_prefix_chars;
+  if (check2?.repetition_loop !== true || typeof keep !== "number") return content.trim();
+  const characters = Array.from(content);
+  const trimmed = characters.slice(0, keep).join("").trim();
+  return `${trimmed}
+[freellama: trimmed ${characters.length - keep} repeated characters; the model looped until num_predict \u2014 treat as low confidence]`;
+}
+function batchAnswerText(payload) {
+  if (!Array.isArray(payload.results)) return void 0;
+  const perItem = Math.max(512, Math.floor(ANSWER_TEXT_MAX_CHARS / Math.max(payload.results.length, 1)));
+  const lines = payload.results.map((row) => {
+    const item = record2(row) ?? {};
+    const id = String(item.id ?? "?");
+    if (item.ok !== true) return `## ${id} (failed)
+${clipText(JSON.stringify(item.error ?? item), perItem)}`;
+    const answer = taskAnswerText(record2(item.response) ?? {}) ?? "(no answer)";
+    return `## ${id}
+${clipText(answer, perItem)}`;
+  });
+  return lines.join("\n\n");
+}
 function structuredResult(value, options = {}) {
+  const text = options.text ?? resultSummary(value);
   return {
-    content: [{ type: "text", text: options.legacyJson ? serialize(value) : resultSummary(value) }],
+    content: [{ type: "text", text }],
     structuredContent: value
   };
 }
@@ -21839,44 +22115,44 @@ ${clipText(raw, 2e3)}`
   }
   return structuredResult(value);
 }
+function isStructuredSuccess(result) {
+  return !("isError" in result && result.isError) && "structuredContent" in result && result.structuredContent !== void 0;
+}
+var escapeRegExp = (text) => text.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+var SERVE_ADDRESS_PATTERN = (() => {
+  let url;
+  try {
+    url = new URL(DEFAULT_SERVE_ENDPOINT);
+  } catch {
+    return /_freellama/i;
+  }
+  const { hostname: hostname2, port } = url;
+  const hosts = ["127.0.0.1", "localhost"].includes(hostname2) ? ["127.0.0.1", "localhost"] : [hostname2];
+  const addresses = hosts.map((host) => escapeRegExp(port ? `${host}:${port}` : host));
+  return new RegExp(`(?:_freellama|${addresses.join("|")})`, "i");
+})();
 function errorResult(error2) {
   const message = error2 instanceof Error ? error2.message : String(error2);
-  const managedServeUnavailable = /(?:connection refused|connect error|failed to connect|error sending request)/i.test(message) && /(?:_freellama|127\.0\.0\.1:11435|localhost:11435)/i.test(message);
+  let receipt;
+  try {
+    const parsed = JSON.parse(message);
+    if (parsed !== null && typeof parsed === "object" && !Array.isArray(parsed) && typeof parsed.error === "string") {
+      receipt = parsed;
+    }
+  } catch {
+  }
+  const managedServeUnavailable = /(?:connection refused|connect error|failed to connect|error sending request)/i.test(message) && SERVE_ADDRESS_PATTERN.test(message);
+  const ollamaUnavailable = !managedServeUnavailable && /(?:connection refused|connect error|failed to connect|error sending request)/i.test(message) && /\/api\/(?:tags|ps|chat|generate|embed|show)/.test(message);
   const actionableMessage = managedServeUnavailable ? `${message}
 
-FreeLlama managed serve is unreachable. Start it with \`freellama serve\`, or set \`FREELLAMA_SERVE_ENDPOINT\` to a running managed endpoint; then retry. Run \`doctor\` to inspect the configured endpoint.` : message;
-  return { content: [{ type: "text", text: actionableMessage }], isError: true };
-}
-var adapterCallSchema = external_exports.object({
-  raw_name: external_exports.string().optional(),
-  status: external_exports.string().optional(),
-  arguments: external_exports.object({
-    tool: external_exports.string().optional(),
-    command: external_exports.string().optional(),
-    queries: external_exports.object({ path: external_exports.string().optional() }).passthrough().optional()
-  }).passthrough().optional()
-}).passthrough();
-var adapterResultSchema = external_exports.object({
-  final_answer: external_exports.string(),
-  tool_calls: external_exports.array(adapterCallSchema).default([]),
-  usage: external_exports.object({
-    input_tokens: external_exports.number().nullable().optional(),
-    output_tokens: external_exports.number().nullable().optional()
-  }).passthrough().default({}),
-  model_metadata: external_exports.object({
-    context_compactions: external_exports.number().int().nonnegative().optional(),
-    runtime_config: external_exports.record(external_exports.string(), external_exports.unknown()).optional(),
-    context_management: external_exports.object({
-      token_counting: external_exports.enum(["configured_estimate", "model_calibrated_estimate"]),
-      estimate_scale: external_exports.number().positive(),
-      calibration_samples: external_exports.number().int().nonnegative(),
-      pinned_overflow: external_exports.enum(["error", "clip"]),
-      compactions: external_exports.number().int().nonnegative()
-    }).passthrough().optional()
-  }).passthrough().optional()
-});
-function parseAdapterResult(raw) {
-  return adapterResultSchema.parse(JSON.parse(raw));
+FreeLlama managed serve is unreachable. Start it with \`freellama serve\`, or set \`FREELLAMA_SERVE_ENDPOINT\` to a running managed endpoint; then retry. Run \`doctor\` to inspect the configured endpoint.` : ollamaUnavailable ? `${message}
+
+Ollama is not running at that address. Ask the user to start Ollama (the app, or \`ollama serve\`); retrying before that cannot succeed.` : message;
+  return {
+    content: [{ type: "text", text: actionableMessage }],
+    isError: true,
+    ...receipt ? { structuredContent: receipt } : {}
+  };
 }
 function summarizeEmbeddings(payload) {
   const response = payload.response;
@@ -21899,13 +22175,211 @@ function summarizeEmbeddings(payload) {
   };
 }
 
+// src/tools/doctor.ts
+var doctorResultSchema = external_exports.object({
+  summary: external_exports.string().optional(),
+  endpoint: external_exports.string().optional()
+}).passthrough();
+function registerDoctorTool(server2) {
+  server2.registerTool(
+    "doctor",
+    {
+      description: "Use when: checking whether Ollama/FreeLlama are healthy, reading live queues/limits (view:status) or token usage (view:usage). Do not use when: choosing a model (models). Returns: a short summary; pick a view for detail.",
+      inputSchema: external_exports.object({
+        ollamaEndpoint: ollamaEndpointParam,
+        serveEndpoint: endpointParam,
+        view: external_exports.enum(["summary", "full", "status", "usage"]).optional(),
+        days: external_exports.number().int().min(1).max(366).optional()
+      }).strict(),
+      outputSchema: doctorResultSchema,
+      annotations: { readOnlyHint: true }
+    },
+    async ({ ollamaEndpoint, serveEndpoint, view, days }) => {
+      try {
+        if (days !== void 0 && view !== "usage") return errorResult(new Error("days is only valid with view: usage."));
+        if (view === "status") return parsedResult(await status(serveEndpoint));
+        if (view === "usage") return parsedResult(await usage(serveEndpoint, days));
+        const report = parsedResult(await doctor(ollamaEndpoint));
+        if (!isStructuredSuccess(report)) return report;
+        try {
+          report.structuredContent.machine = JSON.parse(await machine(serveEndpoint));
+          delete report.structuredContent.machine_unavailable;
+        } catch (error2) {
+          if (report.structuredContent.machine == null) {
+            report.structuredContent.machine_unavailable = `freellama serve unreachable, so no machine profile: ${error2 instanceof Error ? error2.message : String(error2)}`;
+          }
+        }
+        try {
+          report.structuredContent.platform_health = JSON.parse(await health(serveEndpoint));
+        } catch (error2) {
+          report.structuredContent.platform_health_unavailable = `freellama serve unreachable: ${error2 instanceof Error ? error2.message : String(error2)}`;
+        }
+        const full = { ...report.structuredContent };
+        delete full.ollama_env_config;
+        if ((view ?? "summary") === "full") return structuredResult(full);
+        const running = full.running?.models ?? [];
+        const compact = {
+          status: "ok",
+          summary: `Ollama ${String(full.version?.version ?? "unknown")}; ${running.length} resident model(s).`,
+          endpoint: full.endpoint,
+          ollama: { endpoint: full.endpoint, version: full.version, resident_model_count: running.length },
+          machine: full.machine,
+          local_conservative_config_posture: full.local_conservative_config_posture,
+          host_runtime_signals: full.host_runtime_signals,
+          evidence: { runtime: "ollama_api_version_and_ps", configuration: full.ollama_env_config_source },
+          next: 'Use models{view:"installed"} to select a local model; view:"full" has Ollama configuration and admission.'
+        };
+        return structuredResult(compact);
+      } catch (error2) {
+        return errorResult(error2);
+      }
+    }
+  );
+}
+
+// src/serve-client.ts
+var createSession2 = withServe(createSession);
+var deleteSession2 = withServe(deleteSession);
+var killSession2 = withServe(killSession);
+var listModels2 = withServe(listModels);
+var route2 = withServe(route);
+var runTaskRequest2 = withServe(runTaskRequest);
+var runTaskBatchRequest2 = withServe(runTaskBatchRequest);
+var warmModelRequest2 = withServe(warmModelRequest);
+var listTaskJobs2 = withServe(listTaskJobs);
+var getTaskJob2 = withServe(getTaskJob);
+var cancelTaskJob2 = withServe(cancelTaskJob);
+var removeTaskJob2 = withServe(removeTaskJob);
+var createScope2 = withServe(createScope);
+var getScope2 = withServe(getScope);
+var forkScope2 = withServe(forkScope);
+var deleteScope2 = withServe(deleteScope);
+
+// src/tools/session.ts
+var sessionResultSchema = external_exports.object({
+  session_id: external_exports.string().optional(),
+  deleted: external_exports.boolean().optional(),
+  killed: external_exports.boolean().optional(),
+  runner_stop_confirmed: external_exports.boolean().optional()
+}).passthrough();
+function registerSessionTool(server2) {
+  server2.registerTool(
+    "session",
+    {
+      description: "Use when: several related run_task calls should stay on one model (create, pass sessionId), or to kill a session's in-flight work. Do not use when: cancelling one deferred job (task_jobs), keeping history (scope), or unloading (ollama_manage). Returns: receipt.",
+      inputSchema: external_exports.object({
+        action: external_exports.enum(["create", "delete", "kill"]),
+        sessionId: external_exports.string().uuid().optional(),
+        endpoint: endpointParam
+      }).strict(),
+      outputSchema: sessionResultSchema,
+      annotations: { destructiveHint: true }
+    },
+    async ({ action, sessionId, endpoint }) => {
+      try {
+        if (action === "create") {
+          if (sessionId !== void 0) return errorResult(new Error("sessionId is only valid for action: delete or kill."));
+          return structuredResult(JSON.parse(await createSession2(endpoint)));
+        }
+        if (sessionId === void 0) return errorResult(new Error(`action: ${action} requires sessionId.`));
+        if (action === "kill") return parsedResult(await killSession2(endpoint, sessionId));
+        await deleteSession2(endpoint, sessionId);
+        return structuredResult({ session_id: sessionId, deleted: true });
+      } catch (error2) {
+        return errorResult(error2);
+      }
+    }
+  );
+}
+
+// src/paging.ts
+import { createHash } from "node:crypto";
+function pageLiveList(items, limit, cursor, identity) {
+  const pageSize = limit ?? 20;
+  const fingerprint = createHash("sha256").update(items.map(identity).join("\n")).digest("base64url").slice(0, 16);
+  let offset = 0;
+  if (cursor) {
+    try {
+      const parsed = JSON.parse(Buffer.from(cursor, "base64url").toString("utf8"));
+      if (!Number.isInteger(parsed.offset) || parsed.offset < 0 || parsed.fingerprint !== fingerprint) {
+        throw new Error("invalid or stale cursor");
+      }
+      offset = parsed.offset;
+    } catch {
+      throw new Error("`cursor` is invalid or the model list changed; restart without cursor.");
+    }
+  }
+  const page = items.slice(offset, offset + pageSize);
+  const nextOffset = offset + page.length;
+  return {
+    items: page,
+    returned: page.length,
+    total: items.length,
+    next_cursor: nextOffset < items.length ? Buffer.from(JSON.stringify({ offset: nextOffset, fingerprint }), "utf8").toString("base64url") : null
+  };
+}
+
 // src/model-search.ts
+function decodeHtml(text) {
+  const entities = { amp: "&", lt: "<", gt: ">", quot: '"', apos: "'", nbsp: " ", middot: "\xB7", bull: "\u2022" };
+  return text.replace(/&(#x[\da-f]+|#\d+|[a-z]+);/gi, (raw, entity) => {
+    if (!entity.startsWith("#")) return entities[entity.toLowerCase()] ?? raw;
+    const code = entity[1].toLowerCase() === "x" ? Number.parseInt(entity.slice(2), 16) : Number(entity.slice(1));
+    return code > 0 && code <= 1114111 && !(code >= 55296 && code <= 57343) ? String.fromCodePoint(code) : raw;
+  });
+}
+function attribute(tag, name) {
+  return tag.match(new RegExp(`(?:^|\\s)${name}\\s*=\\s*(["'])([\\s\\S]*?)\\1`, "i"))?.[2];
+}
+function htmlText(html) {
+  return decodeHtml(html.replace(/<(script|style|svg)\b[^>]*>[\s\S]*?<\/\1>/gi, "").replace(/<\/(?:p|li|div|h[1-6]|pre|tr)>|<br\s*\/?\s*>/gi, "\n").replace(/<[^>]*>/g, " ")).split("\n").map((line) => line.replace(/\s+/g, " ").trim()).filter(Boolean).join("\n");
+}
+function divContent(html, id) {
+  const tags = /<\/?div\b[^>]*>/gi;
+  let start;
+  let depth = 0;
+  for (const match of html.matchAll(tags)) {
+    if (start === void 0) {
+      if (attribute(match[0], "id") !== id) continue;
+      start = match.index + match[0].length;
+      depth = 1;
+    } else {
+      depth += /^<\//.test(match[0]) ? -1 : 1;
+      if (depth === 0) return html.slice(start, match.index);
+    }
+  }
+  return void 0;
+}
+function parseModelCard(html) {
+  const meta = [...html.matchAll(/<meta\b[^>]*>/gi)].find((m) => attribute(m[0], "name") === "description");
+  const description = decodeHtml(meta ? attribute(meta[0], "content") ?? "" : "").trim();
+  const editor = [...html.matchAll(/<textarea\b([^>]*)>([\s\S]*?)<\/textarea>/gi)].find((m) => attribute(m[1], "id") === "editor");
+  const rendered = divContent(html, "display");
+  const readme = editor ? decodeHtml(editor[2]).trim() : rendered === void 0 ? "" : htmlText(rendered);
+  if (!description && !editor && rendered === void 0) throw new Error("Ollama library markup was not recognized.");
+  const header = html.split(/<div[^>]*\bid=["']readme["']/i)[0];
+  const capabilities = [...new Set([...header.matchAll(/<span\b[^>]*class=["'][^"']*text-(?:indigo-600|cyan-500)[^"']*["'][^>]*>\s*([a-z]+)\s*<\/span>/g)].map((m) => m[1]))];
+  return { description, readme, readmeFormat: editor ? "markdown" : "text", capabilities };
+}
+function libraryTrialCandidate(tag, embedding) {
+  return {
+    tag,
+    evidence: "download_size_preflight_only",
+    taskQuality: "unmeasured",
+    why: embedding ? "Smallest listed download inside the conservative host-memory budget. This is a trial candidate; smaller size does not establish better embedding quality." : "Largest listed download inside the conservative host-memory budget. This is a trial candidate; download size establishes neither accelerator fit nor task quality.",
+    configure: 'Set contextTokens and output limits for the workload, then inspect preview and execution receipts. Ollama defaults depend on its configuration and hardware. Use keepAlive:"0" for one-off work; keep related repeated work warm when resources permit.',
+    caution: "Measure this exact tag on representative tasks before choosing it. Include resident runners, context/KV caches and OS headroom in memory fit. Discovery does not authorize a pull."
+  };
+}
 function parseModelSearch(html) {
   const results = [];
   for (const block of html.split(/<li\s/).slice(1)) {
-    const name = block.match(/href="\/library\/([^"]+)"/)?.[1];
+    const link = block.match(/href="\/(?:library\/([\w.-]+)|([\w.-]+\/[\w.-]+))"/);
+    const name = link?.[1] ?? link?.[2];
     if (!name) continue;
-    const description = block.match(/<p class="max-w-lg[^"]*">([\s\S]*?)<\/p>/)?.[1]?.replace(/<[^>]+>/g, "").replace(/&#39;/g, "'").replace(/&amp;/g, "&").replace(/&quot;/g, '"').replace(/\s+/g, " ").trim() ?? "";
+    if (!/<p\b[^>]*class="max-w-lg/.test(block)) continue;
+    const rawDescription = block.match(/<p class="max-w-lg[^"]*">([\s\S]*?)<\/p>/)?.[1];
+    const description = rawDescription === void 0 ? "" : decodeHtml(rawDescription.replace(/<[^>]+>/g, "")).replace(/\s+/g, " ").trim();
     const capabilities = [...block.matchAll(/text-(?:indigo-600|cyan-500)[^>]*>([a-z]+)<\/span>/g)].map(
       (m) => m[1]
     );
@@ -21918,7 +22392,7 @@ function parseModelSearch(html) {
       capabilities,
       pulls: stat("Pulls"),
       tags: stat("Tag"),
-      cloudOnly: capabilities.includes("cloud")
+      cloudAvailable: capabilities.includes("cloud")
     });
   }
   return results;
@@ -21926,1046 +22400,1545 @@ function parseModelSearch(html) {
 function parseModelTags(html, family) {
   const tags = [];
   const seen = /* @__PURE__ */ new Set();
-  const row = /<a href="\/library\/([^"]+)" class="sm:hidden[\s\S]*?<p class="flex text-neutral-500">([^<]*)<\/p>/g;
+  const row = /<a\b([^>]*)>([\s\S]*?)<\/a>/gi;
   for (const m of html.matchAll(row)) {
-    const tag = m[1];
+    let tag;
+    try {
+      tag = decodeURIComponent(attribute(m[1], "href") ?? "").replace(/^\/(?:library\/)?/, "");
+    } catch {
+      continue;
+    }
+    if (!tag.startsWith(`${family}:`)) continue;
+    const text = htmlText(m[2]).replace(/\n/g, " ");
+    if (!/context window|[\d.]\s*[MG]B\b|\bcloud\b/i.test(text)) continue;
     if (seen.has(tag)) continue;
     seen.add(tag);
-    const meta = m[2].replace(/&middot;/g, "\xB7").split("\xB7").map((x) => x.trim());
+    const meta = text.split(/[·•]/).map((x) => x.trim());
     const sizeText = meta.find((x) => /^[\d.]+\s?[MG]B$/i.test(x)) ?? null;
     const bytes = sizeText ? Number.parseFloat(sizeText) * (/GB/i.test(sizeText) ? 1e9 : 1e6) : null;
     tags.push({
       tag,
+      digest: text.replace(tag, "").match(/\b[\da-f]{12,64}\b/i)?.[0]?.toLowerCase() ?? null,
+      cloud: /-cloud$/i.test(tag) || meta.some((part) => /^cloud$/i.test(part)),
       size: sizeText,
       sizeBytes: bytes,
       context: meta.find((x) => /context window/i.test(x))?.replace(/\s*context window\s*/i, "") ?? null,
-      modalities: meta.find((x) => /^(Text|Image|Audio)/i.test(x)) ?? null,
+      modalities: meta.find((x) => /^(Text|Image|Audio)/i.test(x))?.replace(/\s+input.*$/i, "") ?? null,
       updated: meta.at(-1) ?? null
     });
   }
   return { family, tags };
 }
 
-// src/delegate.ts
-import { existsSync as existsSync3, readFileSync } from "node:fs";
-import path3 from "node:path";
-var MODEL_EVIDENCE = (() => {
-  const configured = process.env.FREELLAMA_MCP_MODEL_EVIDENCE;
-  const candidates = [
-    configured,
-    path3.join(REPO_ROOT, "benchmark/evidence/model-evidence.json")
-  ].filter((x) => Boolean(x));
-  for (const file of candidates) {
+// src/model-knowledge.ts
+var CACHE_TTL_MS = 60 * 60 * 1e3;
+var FAILURE_TTL_MS = 60 * 1e3;
+var MAX_PAGE_BYTES = 2 * 1024 * 1024;
+var MAX_README_CHARS = 64 * 1024;
+var errorMessage = (error2) => error2 instanceof Error ? error2.message : String(error2);
+function isRemoteModel(entry, name) {
+  return Boolean(entry.remote_host || entry.remote_model) || /-cloud$/i.test(name);
+}
+function libraryReference(model) {
+  const match = model.match(/^(?:library\/)?([a-zA-Z0-9][\w.-]*(?:\/[a-zA-Z0-9][\w.-]*)?)(?::([\w.-]+))?$/);
+  if (!match || match[1].split("/").some((part) => part === "." || part === "..")) return null;
+  const family = match[1];
+  const pathname = family.includes("/") ? family : `library/${family}`;
+  return { family, tag: `${family}:${match[2] ?? "latest"}`, url: `https://ollama.com/${pathname}` };
+}
+var LibraryFetchError = class extends Error {
+  constructor(status2, message) {
+    super(message);
+    this.status = status2;
+  }
+  status;
+};
+var OllamaLibraryClient = class {
+  constructor(options = {}) {
+    this.options = options;
+  }
+  options;
+  cache = /* @__PURE__ */ new Map();
+  pending = /* @__PURE__ */ new Map();
+  async search(params) {
+    const url = `https://ollama.com/search?${params.toString()}`;
+    const models = parseModelSearch(await this.page(url));
+    return { url, models, fetchedAt: new Date((this.options.now ?? Date.now)()).toISOString() };
+  }
+  async lookup(model) {
+    const ref = libraryReference(model);
+    if (!ref) return {
+      status: "unmatched",
+      sourceUrl: null,
+      tagsSourceUrl: null,
+      fetchedAt: null,
+      expiresAt: null,
+      cached: false,
+      message: "No exact Ollama public name could be resolved. Base-model provenance is unknown."
+    };
+    const now = this.options.now ?? Date.now;
+    const cached2 = this.cache.get(ref.family);
+    if (cached2 && Date.parse(cached2.expiresAt) > now()) {
+      this.cache.delete(ref.family);
+      this.cache.set(ref.family, cached2);
+      return { ...cached2, cached: true };
+    }
+    const pending = this.pending.get(ref.family);
+    if (pending) return { ...await pending, cached: true };
+    const work = this.load(ref).then((record3) => {
+      this.cache.delete(ref.family);
+      this.cache.set(ref.family, record3);
+      while (this.cache.size > (this.options.maxEntries ?? 128)) this.cache.delete(this.cache.keys().next().value);
+      return record3;
+    });
+    this.pending.set(ref.family, work);
     try {
-      if (!existsSync3(file)) continue;
-      const parsed = JSON.parse(readFileSync(file, "utf8"));
-      return parsed.models ?? {};
-    } catch {
+      return await work;
+    } finally {
+      this.pending.delete(ref.family);
     }
   }
-  return {};
-})();
-function assessDelegatedAnswer(question, evidenceCount, model) {
-  const grounded = evidenceCount > 0;
-  const evidence = MODEL_EVIDENCE[model];
-  if (evidence?.grade === "unusable") {
-    return {
-      recommendation: "escalate",
-      grounded,
-      why: `${model} is not viable for research on this machine (${evidence.note}). It answers fast and confidently while being wrong \u2014 a fast wrong answer is not a speed win. Re-run with a ~27B model (see README), or answer it yourself.`,
-      measuredBaseRate: `${model}: ${evidence.note}`
-    };
-  }
-  if (!evidence) {
-    return {
-      recommendation: "verify",
-      grounded,
-      why: `${model} has no measured accuracy in this repo's benchmarks, so no base rate applies to this answer. Treat it as unverified until it has been evaluated \u2014 accuracy fell off a cliff below ~12B in the models that were measured.`,
-      measuredBaseRate: "no measured base rate for this model"
-    };
-  }
-  if (evidence.grade === "weak" && grounded) {
-    return {
-      recommendation: "verify",
-      grounded,
-      why: `${model} holds up on simple single-file lookups but not beyond (${evidence.note}). Check the evidence trail, or re-run on a ~27B model (see README) if the answer matters.`,
-      measuredBaseRate: `${model}: ${evidence.note}`
-    };
-  }
-  if (!grounded) {
-    return {
-      recommendation: "escalate",
-      grounded: false,
-      why: "The model answered without reading a single file, so this is parametric recall, not research \u2014 the failure mode where `run_task` was verified inventing wrong facts about this project's own architecture. Re-ask with a narrower question, or answer it yourself.",
-      measuredBaseRate: "ungrounded answers have no measured accuracy \u2014 they were never the tested path"
-    };
-  }
-  const judgmentSignals = /\b(should (we|i|you)|better|best|worth|review|assess|evaluate|improve|opinion|recommend|why is|is it (good|safe|correct)|design|refactor)\b/i;
-  if (judgmentSignals.test(question)) {
-    return {
-      recommendation: "verify",
-      grounded: true,
-      why: "The question reads as a judgment call (keyword heuristic), which is the ~67%-accurate bucket rather than the 98.9% one \u2014 and the tone is identical either way. Check the evidence trail against the claim before acting on it.",
-      measuredBaseRate: `${model}: ~67% on judgment calls vs 98.9% on grounded lookups`
-    };
+  async page(url) {
+    const response = await (this.options.fetch ?? fetch)(url, {
+      method: "GET",
+      redirect: "error",
+      headers: { accept: "text/html" },
+      signal: AbortSignal.timeout(this.options.timeoutMs ?? Math.min(DEFAULT_OLLAMA_FETCH_TIMEOUT_SECONDS * 1e3, 1e4))
+    });
+    if (!response.ok) {
+      await response.body?.cancel();
+      throw new LibraryFetchError(response.status === 404 ? "not_found" : "unavailable", `Ollama library returned HTTP ${response.status}.`);
+    }
+    if (!(response.headers.get("content-type") ?? "").includes("text/html")) {
+      await response.body?.cancel();
+      throw new LibraryFetchError("parse_error", "Ollama library did not return HTML.");
+    }
+    if (Number(response.headers.get("content-length")) > MAX_PAGE_BYTES) {
+      await response.body?.cancel();
+      throw new LibraryFetchError("unavailable", "Ollama library page exceeds the 2 MiB limit.");
+    }
+    const reader = response.body?.getReader();
+    if (!reader) throw new LibraryFetchError("parse_error", "Ollama library returned an empty body.");
+    const chunks = [];
+    let bytes = 0;
+    try {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        bytes += value.byteLength;
+        if (bytes > MAX_PAGE_BYTES) throw new LibraryFetchError("unavailable", "Ollama library page exceeds the 2 MiB limit.");
+        chunks.push(value);
+      }
+      return Buffer.concat(chunks).toString("utf8");
+    } finally {
+      await reader.cancel();
+    }
   }
-  if (evidenceCount > 5) {
+  async load(ref) {
+    let result;
+    try {
+      const pages = await Promise.allSettled([this.page(ref.url), this.page(`${ref.url}/tags`)]);
+      if (pages[0].status === "rejected") throw pages[0].reason;
+      if (pages[1].status === "rejected") throw pages[1].reason;
+      let card;
+      try {
+        card = parseModelCard(pages[0].value);
+      } catch {
+        throw new LibraryFetchError("parse_error", "Ollama model-card markup was not recognized.");
+      }
+      const { tags } = parseModelTags(pages[1].value, ref.family);
+      if (tags.length === 0) throw new LibraryFetchError("parse_error", "No tag rows were recognized; check the source page before concluding a model is absent.");
+      result = {
+        status: "available",
+        tags,
+        readmeTruncated: card.readme.length > MAX_README_CHARS,
+        card: { ...card, readme: card.readme.slice(0, MAX_README_CHARS) }
+      };
+    } catch (error2) {
+      result = {
+        status: error2 instanceof LibraryFetchError ? error2.status : "unavailable",
+        message: error2 instanceof LibraryFetchError ? error2.message : `Ollama library lookup failed or timed out (${errorMessage(error2)}). Local metadata remains available.`
+      };
+    }
+    const now = (this.options.now ?? Date.now)();
     return {
-      recommendation: "verify",
-      grounded: true,
-      why: `${evidenceCount} tool calls is outside the 1-5 file envelope this tool was measured on. Wide searches are where it drifts; spot-check the evidence trail.`,
-      measuredBaseRate: `${model}: 98.9% on grounded lookups, measured within a 1-5 file scope`
+      ...result,
+      sourceUrl: ref.url,
+      tagsSourceUrl: `${ref.url}/tags`,
+      fetchedAt: new Date(now).toISOString(),
+      expiresAt: new Date(now + (result.status === "available" ? CACHE_TTL_MS : FAILURE_TTL_MS)).toISOString(),
+      cached: false
     };
   }
+};
+var supportedUses = {
+  completion: { use: "text_generation", reason: "Ollama advertises completion support; task accuracy still needs evaluation." },
+  tools: { use: "tool_calling", reason: "Ollama advertises tool-call support; this does not establish reliable agent execution." },
+  vision: { use: "image_understanding", reason: "Ollama advertises image-input support; OCR and visual accuracy are unmeasured here." },
+  audio: { use: "audio_input", reason: "Ollama advertises audio support; inspect the model instructions for the supported audio workflow." },
+  embedding: { use: "embeddings", reason: "Ollama advertises embedding support for retrieval and similarity; retrieval quality needs evaluation." },
+  thinking: { use: "thinking", reason: "Ollama advertises a thinking mode; it does not establish reasoning accuracy." },
+  insert: { use: "fill_in_middle", reason: "Ollama advertises suffix-based insertion support." },
+  image: { use: "image_generation", reason: "Ollama advertises image-generation support." }
+};
+var advertisedUses = [
+  { task: "coding", pattern: /\b(coding|code generation|software engineering)\b/i, requires: "completion" },
+  { task: "tool_calling", pattern: /\b(tool (?:use|calling)|function calling)\b/i, requires: "tools" },
+  { task: "image_understanding", pattern: /\b(vision|image understanding|processing text and images|multimodal|OCR|optical character recognition)\b/i, requires: "vision" },
+  { task: "reasoning", pattern: /\b(reasoning|mathematical|math tasks)\b/i, requires: "completion" },
+  { task: "embeddings", pattern: /\b(embedding|embeddings|text encoder)\b/i, requires: "embedding" },
+  { task: "summarization", pattern: /\b(summarization|summarisation|summarize|summarise)\b/i, requires: "completion" },
+  { task: "question_answering", pattern: /\b(question answering|question-answering)\b/i, requires: "completion" },
+  { task: "translation", pattern: /\b(translation|translate)\b/i, requires: "completion" },
+  { task: "long_context", pattern: /\b(long[- ]context|repository[- ]scale understanding)\b/i, requires: "completion" }
+];
+function prose(card) {
+  return `${card.description}
+${card.readme}`.replace(/```[\s\S]*?```|~~~[\s\S]*?~~~/g, "").replace(/<script\b[^>]*>[\s\S]*?<\/script>/gi, "").replace(/!\[[^\]]*\]\([^)]*\)/g, "").replace(/\[([^\]]+)\]\([^)]*\)/g, "$1").split(/\n|(?<=[.!?])\s+/).map((line) => line.replace(/^[\s#>*-]+/, "").trim()).filter(Boolean);
+}
+function publicModelGuidance(record3, capabilities, includeReadme = false) {
+  const card = record3.card;
+  if (!card) return { advertisedUseCases: [], usageNotes: [] };
+  const sentences = prose(card);
+  const advertisedUseCases = advertisedUses.flatMap(({ task, pattern, requires }) => {
+    if (capabilities && !capabilities.includes(requires)) return [];
+    const sentence = sentences.find((sentence2) => pattern.test(sentence2) && !/\b(not|never|cannot|unsupported|without)\b|doesn['’]t/i.test(sentence2));
+    const start = sentence ? Math.max(0, sentence.search(pattern) - 120) : 0;
+    return sentence ? [{
+      task,
+      basis: "advertised",
+      scope: "family",
+      quality: "unmeasured",
+      evidence: { excerpt: sentence.slice(start, start + 320), sourceUrl: record3.sourceUrl, fetchedAt: record3.fetchedAt }
+    }] : [];
+  });
+  const usageNotes = sentences.filter((sentence) => /\b(requires?|only|cannot|limitations?|must|memory|languages?|prefix)\b|not support|context (window|length)/i.test(sentence)).slice(0, 8).map((text) => ({ text: text.slice(0, 400), scope: "family", sourceUrl: record3.sourceUrl }));
   return {
-    recommendation: "accept",
-    grounded: true,
-    why: `Grounded in ${evidenceCount} tool call(s) within the measured 1-5 file envelope, the question is lookup-shaped, and ${model} is measured strong for this (${evidence.note}).`,
-    measuredBaseRate: `${model}: 98.9% on grounded lookups (100+ questions)`
+    description: card.description,
+    advertisedFamilyCapabilities: card.capabilities,
+    advertisedUseCases,
+    usageNotes,
+    readmeAvailable: card.readme.length > 0,
+    ...includeReadme ? { readme: { text: card.readme, format: card.readmeFormat, truncated: record3.readmeTruncated ?? false } } : {}
   };
 }
-
-// src/index.ts
-var execFileAsync = promisify(execFile);
-var EXTERNAL_COST = configuredExternalCost();
-function withTaskTelemetry(result) {
-  if (!("structuredContent" in result)) return result;
-  const payload = result.structuredContent;
-  const metrics = payload.metrics;
-  return structuredResult({
-    ...payload,
-    telemetry: costTelemetry({
-      inputTokens: typeof metrics?.prompt_tokens === "number" ? metrics.prompt_tokens : null,
-      outputTokens: typeof metrics?.output_tokens === "number" ? metrics.output_tokens : null,
-      totalDurationNs: typeof metrics?.total_duration_ns === "number" ? metrics.total_duration_ns : null
-    }, EXTERNAL_COST)
-  });
+function provenance(record3) {
+  const { card: _card, tags: _tags, readmeTruncated: _truncated, ...source } = record3;
+  return source;
 }
-function withBatchTelemetry(result) {
-  if (!("structuredContent" in result)) return result;
-  const payload = result.structuredContent;
-  const rows = Array.isArray(payload.results) ? payload.results : [];
-  let inputTokens = 0;
-  let outputTokens = 0;
-  let complete = 0;
-  const results = rows.map((row) => {
-    if (!row || typeof row !== "object") return row;
-    const item = row;
-    if (item.ok !== true || !item.response || typeof item.response !== "object") return item;
-    const response = item.response;
-    const metrics = response.metrics;
-    const input = typeof metrics?.prompt_tokens === "number" ? metrics.prompt_tokens : null;
-    const output = typeof metrics?.output_tokens === "number" ? metrics.output_tokens : null;
-    if (input !== null && output !== null) {
-      inputTokens += input;
-      outputTokens += output;
-      complete += 1;
+function modelKnowledge(model, data, record3, includeReadme = false, supportedFeaturesSource = "ollama_api") {
+  const capabilities = Array.isArray(data.capabilities) ? data.capabilities.filter((item) => typeof item === "string") : [];
+  const supportedUseCases = capabilities.flatMap((capability) => Object.hasOwn(supportedUses, capability) ? [{ capability, ...supportedUses[capability], basis: "supported_feature" }] : []);
+  const local = {
+    supportedFeatures: capabilities,
+    supportedFeaturesSource,
+    supportedUseCases,
+    quality: "metadata_only",
+    note: "Supported features and advertised uses do not establish task quality. Keep exact-model benchmark and policy evidence separate."
+  };
+  if (isRemoteModel(data, model)) {
+    return { ...local, localModel: false, library: { status: "remote_model", message: "This Ollama entry points to a remote model." } };
+  }
+  if (!record3) return { ...local, localModel: true, library: { status: "not_requested" } };
+  if (record3.status !== "available") return { ...local, localModel: true, library: provenance(record3) };
+  const ref = libraryReference(model);
+  const tag = record3.tags?.find((tag2) => tag2.tag === ref.tag);
+  const digest = typeof data.digest === "string" ? data.digest.replace(/^sha256:/, "").toLowerCase() : null;
+  const validDigest = digest && /^[\da-f]{64}$/.test(digest);
+  const matched = validDigest && tag?.digest && digest.startsWith(tag.digest);
+  const status2 = !tag ? "tag_not_found" : !validDigest || !tag.digest ? "unverified" : matched ? "matched" : "digest_mismatch";
+  if (!matched || tag?.cloud) return { ...local, localModel: true, library: {
+    ...provenance(record3),
+    status: tag?.cloud ? "remote_tag" : status2,
+    message: "Public guidance was withheld because an exact local tag and digest match could not be verified.",
+    tag: tag ?? null
+  } };
+  return { ...local, localModel: true, library: {
+    ...provenance(record3),
+    status: status2,
+    tag,
+    ...publicModelGuidance(record3, capabilities, includeReadme)
+  } };
+}
+async function enrichInstalledModels(models, client, includeLibrary, loadTags) {
+  const results = new Array(models.length);
+  const inventories = /* @__PURE__ */ new Map();
+  let next = 0;
+  await Promise.all(Array.from({ length: Math.min(4, models.length) }, async () => {
+    while (next < models.length) {
+      const index = next++;
+      const model = models[index];
+      const name = String(model.name ?? model.model ?? "");
+      let local = model;
+      let featureSource = "freellama_catalog";
+      let localError;
+      if (includeLibrary && loadTags) {
+        const upstream = model.execution?.upstream;
+        let inventory = inventories.get(upstream);
+        if (!inventory) {
+          inventory = loadTags(upstream);
+          inventories.set(upstream, inventory);
+        }
+        try {
+          const entry = (await inventory).find((entry2) => String(entry2.name ?? entry2.model) === name);
+          if (!entry || entry.digest !== model.digest) localError = "Local inventory changed; inspect models again before matching public guidance.";
+          else {
+            local = { ...model, ...entry };
+            if (Array.isArray(entry.capabilities)) featureSource = "ollama_api";
+          }
+        } catch (error2) {
+          localError = `The selected Ollama backend's inventory could not be verified (${errorMessage(error2)}); public guidance was withheld.`;
+        }
+      }
+      const record3 = includeLibrary && !localError && !isRemoteModel(local, name) ? await client.lookup(name) : void 0;
+      const knowledge = modelKnowledge(name, local, record3, false, featureSource);
+      results[index] = { ...model, knowledge: localError ? { ...knowledge, library: { status: "unverified", message: localError } } : knowledge };
     }
-    return { ...item, response: { ...response, telemetry: costTelemetry({ inputTokens: input, outputTokens: output }, EXTERNAL_COST) } };
-  });
-  return structuredResult({
-    ...payload,
-    results,
-    telemetry: complete === rows.filter((row) => row?.ok === true).length ? costTelemetry({ inputTokens, outputTokens }, EXTERNAL_COST) : { local: null, externalEquivalent: null, note: "Batch aggregate unavailable because one or more successful items omitted token counts." }
-  });
+  }));
+  return results;
 }
-function pageLiveList(items, limit, cursor, identity) {
-  const pageSize = limit ?? 20;
-  const fingerprint = createHash("sha256").update(items.map(identity).join("\n")).digest("base64url").slice(0, 16);
-  let offset = 0;
-  if (cursor) {
+
+// src/tools/models.ts
+var modelsResultSchema = external_exports.object({
+  models: external_exports.array(external_exports.record(external_exports.unknown())).optional(),
+  tags: external_exports.array(external_exports.record(external_exports.unknown())).optional(),
+  page: external_exports.unknown().optional(),
+  knowledge: external_exports.unknown().optional(),
+  library: external_exports.unknown().optional()
+}).passthrough();
+var ollamaLibrary = new OllamaLibraryClient();
+var HOST_MEMORY_FIT_FRACTION = 0.6;
+async function libraryLookup({
+  model,
+  query,
+  capabilities,
+  order,
+  limit,
+  cursor,
+  includeReadme,
+  endpoint,
+  ollamaEndpoint
+}) {
+  const localState = async () => {
+    let installed = null;
+    let installedUnavailable;
+    let memoryBytes = null;
     try {
-      const parsed = JSON.parse(Buffer.from(cursor, "base64url").toString("utf8"));
-      if (!Number.isInteger(parsed.offset) || parsed.offset < 0 || parsed.fingerprint !== fingerprint) {
-        throw new Error("invalid or stale cursor");
-      }
-      offset = parsed.offset;
+      const tags = await ollamaFetch(ollamaEndpoint, "/api/tags");
+      installed = new Set((tags.models ?? []).map((m) => m.name ?? ""));
+    } catch (error2) {
+      installedUnavailable = `Local inventory could not be read (${error2 instanceof Error ? error2.message : String(error2)}); installed is null.`;
+    }
+    try {
+      const profile = JSON.parse(await machine(endpoint));
+      memoryBytes = profile.memory_bytes ?? profile.unified_memory_bytes ?? null;
     } catch {
-      throw new Error("`cursor` is invalid or the model list changed; restart without cursor.");
     }
+    return { installed, installedUnavailable, memoryBytes };
+  };
+  if (!model && cursor !== void 0) {
+    return errorResult(new Error("Library step 1 has no `cursor`; it pages step-2 tags. Use `limit` for search, or pass `model`."));
   }
-  const page = items.slice(offset, offset + pageSize);
-  const nextOffset = offset + page.length;
-  return {
-    items: page,
-    returned: page.length,
-    total: items.length,
-    next_cursor: nextOffset < items.length ? Buffer.from(JSON.stringify({ offset: nextOffset, fingerprint }), "utf8").toString("base64url") : null
-  };
-}
-var liveDelegates = /* @__PURE__ */ new Set();
-function killLiveDelegates() {
-  for (const child of liveDelegates) {
-    try {
-      child.kill("SIGKILL");
-    } catch {
-    }
-  }
-  liveDelegates.clear();
-}
-for (const stream of [process.stdout, process.stderr]) {
-  stream.on("error", (error2) => {
-    if (error2.code === "EPIPE") process.exit(0);
-    throw error2;
-  });
-}
-process.on("exit", killLiveDelegates);
-for (const signal of ["SIGINT", "SIGTERM", "SIGHUP"]) {
-  process.on(signal, () => {
-    killLiveDelegates();
-    process.exit(0);
-  });
-}
-var INSTRUCTIONS = `caller owns task decomposition; operator owns endpoints, exact --cpu-model assignments, lifecycle.
-Ollama plus the OS/driver run physical CPU/GPU.
-Efficient loop: models{view:"installed"}, then models{view:"resident"}; doctor only for runtime diagnosis. Preview consequential work before executing.
-delegate_research is only for narrow allowed-workspace research. Keep full diagnostics, raw embeddings, and long evidence out of active context; read freellama://docs/index on demand.
-ask approval for one exact tag and reported size before ollama_manage; search or recommendation is never download permission.
-run_task preview never executes; code_review aliases coding.
-Use requiredCapabilities:["tools"] to preview tool eligibility; omit preview and supply the payload to execute.
-Docs: freellama://docs/index.`;
-var server = new McpServer(
-  { name: "freellama", version: SERVER_VERSION },
-  { instructions: INSTRUCTIONS }
-);
-var PACKAGED_DOCS_DIR = path4.resolve(path4.dirname(fileURLToPath2(import.meta.url)), "..", "docs");
-if (!existsSync4(PACKAGED_DOCS_DIR)) {
-  throw new Error(`FreeLlama MCP documentation is missing at ${PACKAGED_DOCS_DIR}; run the package build.`);
-}
-var packagedDocs = readdirSync(PACKAGED_DOCS_DIR).filter((name) => name.endsWith(".md")).sort();
-if (!packagedDocs.includes("INDEX.md")) {
-  throw new Error(`FreeLlama MCP documentation index is missing at ${PACKAGED_DOCS_DIR}/INDEX.md; run the package build.`);
-}
-for (const name of packagedDocs) {
-  const uri = `freellama://docs/${name === "INDEX.md" ? "index" : name.replace(/\.md$/, "")}`;
-  server.registerResource(
-    `freellama-docs-${name.toLowerCase().replace(/\.md$/, "")}`,
-    uri,
-    {
-      mimeType: "text/markdown",
-      description: name === "INDEX.md" ? "Index of packaged FreeLlama operator and agent documentation. Read this first, then fetch one relevant document." : `Packaged FreeLlama documentation: ${name}.`
-    },
-    async () => ({
-      contents: [{ uri, mimeType: "text/markdown", text: await readFile(path4.join(PACKAGED_DOCS_DIR, name), "utf8") }]
-    })
-  );
-}
-server.registerTool(
-  "doctor",
-  {
-    description: "Use when: runtime/config diagnosis. Do not use when: model selection. Returns: compact summary by default; config/full are opt-in.",
-    inputSchema: {
-      endpoint: ollamaEndpointParam,
-      serveEndpoint: endpointParam,
-      view: external_exports.enum(["summary", "scheduler", "config", "full"]).optional().describe("summary default; scheduler/config/full are verbose")
-    },
-    outputSchema: doctorResultSchema,
-    annotations: { readOnlyHint: true }
-  },
-  async ({ endpoint, serveEndpoint, view }) => {
-    try {
-      const report = parsedResult(await doctor(endpoint));
-      if (!("structuredContent" in report)) return report;
-      try {
-        report.structuredContent.machine = JSON.parse(await machine(serveEndpoint));
-        delete report.structuredContent.machine_unavailable;
-      } catch (error2) {
-        if (report.structuredContent.machine == null) {
-          report.structuredContent.machine_unavailable = `freellama serve unreachable, so no machine profile: ${error2 instanceof Error ? error2.message : String(error2)}`;
-        }
-      }
-      try {
-        report.structuredContent.platform_health = JSON.parse(await health(serveEndpoint));
-      } catch (error2) {
-        report.structuredContent.platform_health_unavailable = `freellama serve unreachable: ${error2 instanceof Error ? error2.message : String(error2)}`;
-      }
-      const full = { ...report.structuredContent };
-      const flatConfig = full.ollama_env_config;
-      delete full.ollama_env_config;
-      if (view === "config") {
-        return structuredResult({
-          status: "ok",
-          summary: "Categorized Ollama configuration; values are source-qualified, not endpoint/PID proof.",
-          endpoint: full.endpoint,
-          ollama_config: full.ollama_config,
-          ollama_env_config_source: full.ollama_env_config_source,
-          ollama_env_config_warning: full.ollama_env_config_warning,
-          host_runtime_signals: full.host_runtime_signals
-        });
-      }
-      if ((view ?? "summary") === "full") return structuredResult(full);
-      const running = full.running?.models ?? [];
-      const platformHealth = full.platform_health;
-      const compact = {
-        status: "ok",
-        summary: `Ollama ${String(full.version?.version ?? "unknown")}; ${running.length} resident model(s).`,
-        endpoint: full.endpoint,
-        ollama: { endpoint: full.endpoint, version: full.version, resident_model_count: running.length },
-        machine: full.machine,
-        local_conservative_config_posture: full.local_conservative_config_posture,
-        host_runtime_signals: full.host_runtime_signals,
-        ...view === "scheduler" ? {
-          scheduler: {
-            admission: platformHealth?.admission ?? null,
-            ollama_num_parallel: flatConfig?.OLLAMA_NUM_PARALLEL,
-            ollama_max_queue: flatConfig?.OLLAMA_MAX_QUEUE,
-            proof_level: "configured_and_snapshot_only; measure concurrent execution before claiming throughput"
-          }
-        } : {},
-        evidence: { runtime: "ollama_api_version_and_ps", configuration: full.ollama_env_config_source },
-        next: view === "scheduler" ? "Use run_task preview for a per-task advisory receipt." : 'Use models{view:"installed"} to select a local model.'
-      };
-      return structuredResult(compact);
-    } catch (error2) {
-      return errorResult(error2);
-    }
-  }
-);
-server.registerTool(
-  "session",
-  {
-    description: "Use when: retaining model affinity. Do not use when: storing history or KV. Inputs: action, sessionId for delete. Returns: affinity handle or confirmation.",
-    inputSchema: {
-      action: external_exports.enum(["create", "delete"]).describe("create | release"),
-      sessionId: external_exports.string().uuid().optional().describe("delete only"),
-      endpoint: endpointParam
-    },
-    outputSchema: sessionResultSchema,
-    annotations: { destructiveHint: false }
-  },
-  async ({ action, sessionId, endpoint }) => {
-    try {
-      if (action === "create") {
-        if (sessionId !== void 0) return errorResult(new Error("sessionId is only valid for action: delete."));
-        return structuredResult(JSON.parse(await createSession(endpoint)));
-      }
-      if (sessionId === void 0) return errorResult(new Error("action: delete requires sessionId."));
-      await deleteSession(endpoint, sessionId);
-      return structuredResult({ session_id: sessionId, deleted: true });
-    } catch (error2) {
-      return errorResult(error2);
-    }
-  }
-);
-server.registerTool(
-  "models",
-  {
-    description: 'Use when: inspecting models. Do not use when: executing or changing state; search never permits a pull. Inputs: one view and its fields. Returns: inventory/detail, placement, or candidates. Next: library family, then model:"<family>" for tags/fit.',
-    inputSchema: {
-      view: external_exports.enum(["installed", "resident", "detail", "raw", "library"]).optional().describe("installed (default, needs serve) | resident (needs serve) | detail | raw | library"),
-      model: external_exports.string().min(1).optional().describe('required for view "detail"; for "library", step 2 family name'),
-      includeVerbose: external_exports.boolean().optional().describe('"detail" only. Adds license/modelfile \u2014 the bulk of that payload, never routing-relevant'),
-      query: external_exports.string().min(1).optional().describe('"library" step 1: free text, e.g. "qwen", "embed"'),
-      capabilities: external_exports.array(external_exports.enum(["vision", "tools", "thinking", "embedding", "cloud"])).min(1).max(5).optional().describe('"library" step 1: filter chips; combined as AND by the site'),
-      order: external_exports.enum(["popular", "newest"]).optional().describe('"library" step 1. default "popular" \u2014 prefer it'),
-      limit: external_exports.number().int().positive().max(50).optional().describe('"library" search or raw/tag page size; raw default 20'),
-      cursor: external_exports.string().min(1).optional().describe("opaque continuation cursor for raw models or library tags"),
-      endpoint: endpointParam,
-      ollamaEndpoint: ollamaEndpointParam
-    },
-    outputSchema: modelsResultSchema,
-    annotations: { readOnlyHint: true }
-  },
-  async ({ view, model, includeVerbose, query, capabilities, order, limit, cursor, endpoint, ollamaEndpoint }) => {
-    try {
-      const selectedView = view ?? "installed";
-      if (selectedView !== "library" && [query, capabilities, order].some((value) => value !== void 0)) {
-        return errorResult(new Error(`view "${selectedView}" does not accept library search fields.`));
-      }
-      if (selectedView !== "library" && selectedView !== "raw" && [limit, cursor].some((value) => value !== void 0)) {
-        return errorResult(new Error(`view "${selectedView}" accepts neither pagination nor library search fields.`));
-      }
-      if (selectedView !== "detail" && includeVerbose !== void 0) {
-        return errorResult(new Error('`includeVerbose` is valid only for view "detail".'));
-      }
-      if (selectedView !== "detail" && selectedView !== "library" && model !== void 0) {
-        return errorResult(new Error('`model` is valid only for views "detail" and "library".'));
-      }
-      if (selectedView === "library" && model && [query, capabilities, order].some((value) => value !== void 0)) {
-        return errorResult(
-          new Error("Library step 2 accepts only `model` plus endpoint overrides; omit step-1 search fields.")
-        );
-      }
-      switch (selectedView) {
-        case "raw": {
-          const raw = await ollamaFetch(ollamaEndpoint, "/api/tags");
-          const models = Array.isArray(raw.models) ? raw.models : [];
-          const page = pageLiveList(models, limit, cursor, (model2) => String(model2.name ?? JSON.stringify(model2)));
-          return structuredResult({ ...raw, models: page.items, page: { returned: page.returned, total: page.total, next_cursor: page.next_cursor } });
-        }
-        case "resident": {
-          const managed = parsedResult(await listModels(endpoint));
-          if (!("structuredContent" in managed)) return managed;
-          const data = managed.structuredContent;
-          const models = (data.models ?? []).filter((entry) => entry.resident === true).map((entry) => {
-            const size = typeof entry.size === "number" ? entry.size : null;
-            const vram = typeof entry.resident_vram === "number" ? entry.resident_vram : null;
-            if (size === null || vram === null || size === 0) return entry;
-            const gpuPercent = Math.round(vram / size * 100);
-            const execution = entry.execution;
-            const assignedCpu = execution?.placement === "cpu";
-            const observedProcessor = execution?.observation?.processor;
-            const processor = observedProcessor === "cpu" || observedProcessor === "gpu" || observedProcessor === "mixed" ? observedProcessor : gpuPercent >= 100 ? "gpu" : gpuPercent <= 0 ? "cpu" : "mixed";
-            const mismatch = execution?.observation?.status === "mismatch";
-            return {
-              ...entry,
-              placement: {
-                gpu_percent: gpuPercent,
-                assigned: assignedCpu,
-                processor: processor === "gpu" ? "100% GPU" : processor === "cpu" ? "100% CPU" : `${gpuPercent}% GPU / ${100 - gpuPercent}% CPU`,
-                ...mismatch ? {
-                  warning: `Configured ${execution?.placement ?? "unknown"} backend disagrees with Ollama /api/ps: ${processor} was physically observed. This sample is excluded from adaptive routing feedback.`
-                } : processor === "mixed" ? {
-                  warning: 'Partially offloaded to CPU \u2014 expect a large slowdown. Free VRAM (`ollama_manage` action "stop") or lower the context length.'
-                } : {}
-              }
-            };
-          });
-          return structuredResult({ ...data, models });
-        }
-        case "detail": {
-          if (!model) {
-            return errorResult(new Error('view "detail" needs `model` set to an installed tag.'));
-          }
-          const data = await ollamaFetch(ollamaEndpoint, "/api/show", {
-            method: "POST",
-            body: { model }
-          });
-          const { license, modelfile, ...rest } = data;
-          const modelInfo = rest.model_info ?? {};
-          const contextEntry = Object.entries(modelInfo).find(
-            ([key, value]) => key.endsWith(".context_length") && typeof value === "number"
-          );
-          return structuredResult({
-            ...rest,
-            max_context_length: contextEntry?.[1] ?? null,
-            ...includeVerbose ? { license, modelfile } : {}
-          });
-        }
-        case "library":
-          return await libraryLookup({ model, query, capabilities, order, limit, cursor, endpoint, ollamaEndpoint });
-        default:
-          return parsedResult(await listModels(endpoint));
-      }
-    } catch (error2) {
-      return errorResult(error2);
-    }
-  }
-);
-async function libraryLookup({
-  model,
-  query,
-  capabilities,
-  order,
-  limit,
-  cursor,
-  endpoint,
-  ollamaEndpoint
-}) {
-  const fetchPage = async (url) => {
-    const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(), DEFAULT_OLLAMA_FETCH_TIMEOUT_SECONDS * 1e3);
-    try {
-      const response = await fetch(url, { signal: controller.signal });
-      if (!response.ok) throw new Error(`ollama.com returned HTTP ${response.status}`);
-      return await response.text();
-    } finally {
-      clearTimeout(timer);
-    }
-  };
-  const localState = async () => {
-    let installed = /* @__PURE__ */ new Set();
-    let memoryBytes = null;
-    try {
-      const tags = await ollamaFetch(ollamaEndpoint, "/api/tags");
-      installed = new Set((tags.models ?? []).map((m) => m.name ?? ""));
-    } catch {
-    }
-    try {
-      const profile = JSON.parse(await machine(endpoint));
-      memoryBytes = profile.memory_bytes ?? profile.unified_memory_bytes ?? null;
-    } catch {
-    }
-    return { installed, memoryBytes };
-  };
   try {
     if (model) {
-      const family = model.split(":")[0];
-      const familyPage = await fetchPage(`https://ollama.com/library/${encodeURIComponent(family)}`);
-      const { tags } = parseModelTags(familyPage, family);
-      const { installed: installed2, memoryBytes } = await localState();
-      const budget = memoryBytes ? memoryBytes * 0.6 : null;
+      const ref = libraryReference(model);
+      if (!ref) throw new Error("Use an exact Ollama family or namespace/model name, not a URL or path.");
+      const { family } = ref;
+      const library = await ollamaLibrary.lookup(model);
+      if (library.status !== "available") return structuredResult({
+        family,
+        sourcePage: ref.url,
+        tags: [],
+        tagsUnavailable: library.message,
+        library,
+        recommendation: null
+      });
+      const tags = library.tags;
+      const { installed: installed2, installedUnavailable: installedUnavailable2, memoryBytes } = await localState();
+      const budget = memoryBytes ? memoryBytes * HOST_MEMORY_FIT_FRACTION : null;
       const annotated = tags.map((t) => ({
         ...t,
-        installed: installed2.has(t.tag),
+        installed: installed2 ? installed2.has(t.tag) : null,
         fitsInMemory: budget && t.sizeBytes ? t.sizeBytes <= budget : null,
         fitScope: "host_memory_budget_only"
       }));
-      const runnable = annotated.filter((t) => t.fitsInMemory === true && !/cloud/.test(t.tag));
-      const embeddingMentions = (familyPage.match(/embedding/gi) ?? []).length;
-      const isEmbeddingFamily = embeddingMentions >= 3 || /embed/i.test(family);
+      const runnable = annotated.filter((t) => t.fitsInMemory === true && !t.cloud);
+      const isEmbeddingFamily = library.card.capabilities.includes("embedding");
       const sized = runnable.filter((t) => t.sizeBytes);
       const best = budget ? isEmbeddingFamily ? sized.sort((a, b) => (a.sizeBytes ?? 0) - (b.sizeBytes ?? 0))[0] : sized.sort((a, b) => (b.sizeBytes ?? 0) - (a.sizeBytes ?? 0))[0] : void 0;
       const tagPage = pageLiveList(annotated, limit, cursor, (tag) => tag.tag);
       return structuredResult({
         family,
-        sourcePage: `https://ollama.com/library/${encodeURIComponent(family)}`,
-        ...tags.length === 0 ? {
-          tagsUnavailable: `No pullable tags found for "${family}". Either it is cloud-only (no local download), the family name is wrong, or ollama.com changed its markup. Open the page above to check before concluding the model does not exist.`
-        } : {},
+        sourcePage: ref.url,
+        library: {
+          status: library.status,
+          sourceUrl: library.sourceUrl,
+          tagsSourceUrl: library.tagsSourceUrl,
+          fetchedAt: library.fetchedAt,
+          expiresAt: library.expiresAt,
+          cached: library.cached,
+          ...publicModelGuidance(library, void 0, includeReadme)
+        },
+        ...installedUnavailable2 ? { installedUnavailable: installedUnavailable2 } : {},
         machineMemoryBytes: memoryBytes,
         fitBudgetBytes: budget,
         tags: tagPage.items,
         page: { returned: tagPage.returned, total: tagPage.total, next_cursor: tagPage.next_cursor },
         recommendationUnavailable: budget ? void 0 : "No machine profile (freellama serve unreachable), so memory fit could not be checked and no tag is recommended. Start serve, or read the sizes yourself.",
-        recommendation: best ? isEmbeddingFamily ? {
-          tag: best.tag,
-          why: "Smallest tag that fits \u2014 this is an embedding family, where bigger is NOT better. Measured on real retrieval over this repo: nomic-embed-text (274MB) scored 5/6 recall@3 in 4.2s, while qwen3-embedding:0.6b (2.3x the size) scored 4/6 at 3.5x the indexing cost. Embeddings do no sampling, so there is no accuracy cliff for size to buy you past.",
-          configure: 'Batch your inputs in one `run_task` call and use keepAlive:"0" \u2014 an embedding model has no reason to hold memory after the vectors are computed. Leave returnEmbeddings off unless you are storing the vectors yourself.',
-          caution: "Index once, query many times, and own the staleness: FreeLlama stores no vectors, and a stale index fails silently by returning confidently wrong files. Also, if you know the keyword, grep beat embedding search here on accuracy, latency and cost at the same time \u2014 reach for embeddings when there is no keyword to search for."
-        } : {
-          tag: best.tag,
-          why: `Largest tag inside the conservative ~60% host-memory preflight${memoryBytes ? ` of ${Math.round(memoryBytes / 1e9)}GB` : ""}. This proves neither accelerator fit nor task quality; preview and benchmark before pulling.`,
-          configure: `Send an explicit num_ctx rather than inheriting Ollama's default, which is 4096 tokens by default unless the Ollama service or request sets num_ctx. Use keepAlive:"0" for one-off calls so it does not hold memory after.`,
-          caution: 'Do not co-resident large models until their combined runner sizes, contexts, KV caches, and OS headroom fit this host. Check `models{view:"resident"}` and accelerator telemetry first; benchmark packaging rather than inferring quality or speed from a model suffix.'
-        } : null
+        recommendation: best ? libraryTrialCandidate(best.tag, isEmbeddingFamily) : null
       });
     }
     const params = new URLSearchParams();
     for (const c of capabilities ?? []) params.append("c", c);
     if (query) params.set("q", query);
     if (order === "newest") params.set("o", "newest");
-    const url = `https://ollama.com/search?${params.toString()}`;
-    const parsed = parseModelSearch(await fetchPage(url)).slice(0, limit ?? 10);
-    const { installed } = await localState();
-    const installedFamilies = new Set([...installed].map((n) => n.split(":")[0]));
+    const search = await ollamaLibrary.search(params);
+    const url = search.url;
+    const parsed = search.models.slice(0, limit ?? 10);
+    const { installed, installedUnavailable } = await localState();
+    const installedFamilies = installed ? new Set([...installed].map((n) => n.split(":")[0])) : null;
     return structuredResult({
       query: url,
+      fetchedAt: search.fetchedAt,
+      ...parsed.length === 0 ? { searchUnavailable: "No result cards were recognized. The query may have no matches or Ollama's markup may have changed; inspect the source URL." } : {},
       order: order ?? "popular",
       count: parsed.length,
+      ...installedUnavailable ? { installedUnavailable } : {},
       nextStep: 'Not pullable yet. Call again with view:"library", model:"<name>" to get tags, sizes and memory fit.',
-      models: parsed.map((m) => ({ ...m, installed: installedFamilies.has(m.name) }))
+      models: parsed.map((m) => ({ ...m, installed: installedFamilies ? installedFamilies.has(m.name) : null }))
     });
   } catch (error2) {
     return errorResult(error2);
   }
 }
-server.registerTool(
-  "run_task",
-  {
-    description: "Use when: routing or executing supplied content. Do not use when: workspace files must be read; use delegate_research. Preview consequential work first; it never generates. Returns: decision or response with receipts. Next: inspect structured observation and verification.",
-    inputSchema: {
-      endpoint: endpointParam,
-      task: taskParam,
-      objective: objectiveParam,
-      model: external_exports.string().min(1).optional().describe("Force this exact installed model name."),
-      sessionId: external_exports.string().min(1).optional().describe("Session id for model affinity across calls."),
-      contextTokens: external_exports.number().int().positive().optional().describe("Total Ollama context window (num_ctx), including input and output."),
-      executionPreference: executionPreferenceParam,
-      minPlacementEvidence: minPlacementEvidenceParam,
-      requiredCapabilities: requiredCapabilitiesParam,
-      prompt: external_exports.string().min(1).optional().describe("Chat input when messages is omitted."),
-      images: external_exports.array(external_exports.string().min(1)).min(1).optional().describe("base64, no data-URI prefix; prompt mode only; requires an explicit tested vision model"),
-      messages: external_exports.array(
-        external_exports.object({ role: external_exports.enum(["system", "user", "assistant", "tool"]), content: external_exports.string() }).passthrough()
-      ).min(1).optional().describe(
-        "wins over prompt; preserves Ollama images, thinking, tool_calls, tool_name, and other message fields"
-      ),
-      input: external_exports.union([external_exports.string().min(1), external_exports.array(external_exports.string().min(1)).min(1)]).optional().describe("embedding only; batch strings because one call is far cheaper than one call per item"),
-      tools: external_exports.array(external_exports.record(external_exports.unknown())).min(1).optional().describe("Ollama function definitions for chat tasks."),
-      keepAlive: external_exports.string().min(1).optional().describe('"0" unloads now, "-1" pins, default 5m'),
-      format: external_exports.union([external_exports.literal("json"), external_exports.record(external_exports.unknown())]).optional().describe('Ollama structured output: "json" or a JSON schema object'),
-      think: external_exports.union([external_exports.boolean(), external_exports.enum(["low", "medium", "high"])]).optional().describe("Override the task profile for thinking-capable models"),
-      options: external_exports.record(external_exports.unknown()).optional().describe("Advanced Ollama options; num_ctx/contextTokens and num_gpu/placement are routing-owned. num_predict caps output."),
-      logprobs: external_exports.boolean().optional(),
-      topLogprobs: external_exports.number().int().nonnegative().optional().describe("Requires logprobs:true"),
-      minConfidence: minConfidenceParam,
-      priority: external_exports.enum(["interactive", "normal", "background"]).optional().describe("Admission class only; normal default. Fair scheduling prevents background starvation."),
-      returnEmbeddings: external_exports.boolean().optional().describe("false (default) withholds the raw vectors; they are large and unreadable to a model"),
-      preview: external_exports.boolean().optional().describe(
-        "true = routing fields only; rejects payloads and runtime controls; never executes"
-      )
+function registerModelsTool(server2) {
+  server2.registerTool(
+    "models",
+    {
+      description: "Use when: listing installed or loaded models with capabilities, policy rank and memory fit, inspecting one model, or searching ollama.com (view:library). Do not use when: running work (run_task) or pulling (ollama_manage). Returns: models; library lists families first, then one family's tags.",
+      inputSchema: external_exports.object({
+        view: external_exports.enum(["installed", "resident", "detail", "raw", "library"]).optional(),
+        model: external_exports.string().min(1).optional(),
+        includeVerbose: external_exports.boolean().optional(),
+        includeLibrary: external_exports.boolean().optional(),
+        includeReadme: external_exports.boolean().optional(),
+        query: external_exports.string().min(1).optional(),
+        capabilities: external_exports.array(external_exports.enum(["vision", "tools", "thinking", "embedding", "cloud"])).min(1).max(5).optional(),
+        order: external_exports.enum(["popular", "newest"]).optional(),
+        limit: external_exports.number().int().positive().max(50).optional(),
+        cursor: external_exports.string().min(1).optional(),
+        endpoint: endpointParam,
+        ollamaEndpoint: ollamaEndpointParam
+      }).strict(),
+      outputSchema: modelsResultSchema,
+      annotations: { readOnlyHint: true }
     },
-    outputSchema: taskResultSchema,
-    annotations: { destructiveHint: false }
-  },
-  async ({
-    endpoint,
-    task,
-    objective,
-    model,
-    sessionId,
-    contextTokens,
-    executionPreference,
-    minPlacementEvidence,
-    requiredCapabilities,
-    prompt,
-    images,
-    messages,
-    input,
-    tools,
-    keepAlive,
-    format,
-    think,
-    options,
-    logprobs,
-    topLogprobs,
-    minConfidence,
-    priority,
-    returnEmbeddings,
-    preview
-  }) => {
-    try {
-      const canonicalTask = canonicalTaskKind(task);
-      if (preview) {
-        const executionOnlyFields = [
-          ["prompt", prompt],
-          ["images", images],
-          ["messages", messages],
-          ["input", input],
-          ["tools", tools],
-          ["keepAlive", keepAlive],
-          ["format", format],
-          ["think", think],
-          ["options", options],
-          ["logprobs", logprobs],
-          ["topLogprobs", topLogprobs],
-          ["returnEmbeddings", returnEmbeddings]
-        ].filter(([, value]) => value !== void 0).map(([name]) => `\`${name}\``);
-        if (executionOnlyFields.length > 0) {
-          return errorResult(
-            new Error(
-              `\`preview:true\` accepts routing fields only; remove execution-only fields ${executionOnlyFields.join(", ")}. Use \`requiredCapabilities:["tools"]\` to preview tool capability, then make a separate execution call with the payload.`
-            )
-          );
+    async ({ view, model, includeVerbose, includeLibrary, includeReadme, query, capabilities, order, limit, cursor, endpoint, ollamaEndpoint }) => {
+      try {
+        const selectedView = view ?? "installed";
+        if (includeLibrary !== void 0 && selectedView !== "installed" && selectedView !== "detail") {
+          return errorResult(new Error('`includeLibrary` is valid only for views "installed" and "detail".'));
         }
-      }
-      const effectiveRequiredCapabilities = tools === void 0 ? requiredCapabilities : withRequiredCapability(requiredCapabilities, "tools");
-      if (topLogprobs !== void 0 && logprobs !== true) {
-        return errorResult(new Error("`topLogprobs` requires `logprobs:true`."));
-      }
-      if (!preview && task === "embedding") {
-        if (input === void 0) return errorResult(new Error('task "embedding" requires `input`.'));
-        if ([prompt, images, messages, tools, format, think, logprobs, topLogprobs].some(
-          (value) => value !== void 0
-        )) {
-          return errorResult(
-            new Error('task "embedding" accepts `input`, `options`, and routing fields; chat controls are not valid.')
-          );
+        if (includeReadme !== void 0 && !(selectedView === "library" && model || selectedView === "detail" && includeLibrary === true)) {
+          return errorResult(new Error('`includeReadme` needs library step 2, or view "detail" with `includeLibrary:true`.'));
         }
-      }
-      if (!preview && task !== "embedding") {
-        if (input !== void 0 || returnEmbeddings !== void 0) {
-          return errorResult(new Error('`input` and `returnEmbeddings` are valid only for task "embedding".'));
+        if (selectedView !== "library" && [query, capabilities, order].some((value) => value !== void 0)) {
+          return errorResult(new Error(`view "${selectedView}" does not accept library search fields.`));
         }
-        if (prompt === void 0 && (messages === void 0 || messages.length === 0)) {
-          return errorResult(new Error(`task "${task}" requires \`prompt\` or \`messages\`.`));
+        if (selectedView !== "library" && selectedView !== "raw" && [limit, cursor].some((value) => value !== void 0)) {
+          return errorResult(new Error(`view "${selectedView}" accepts neither pagination nor library search fields.`));
         }
-        if (images !== void 0 && messages !== void 0) {
-          return errorResult(
-            new Error("Top-level `images` is valid only with `prompt`; put images inside messages instead.")
-          );
+        if (selectedView !== "detail" && includeVerbose !== void 0) {
+          return errorResult(new Error('`includeVerbose` is valid only for view "detail".'));
         }
-        if (images !== void 0 && model === void 0) {
-          return errorResult(new Error("Images require an explicit tested vision-capable `model`."));
+        if (selectedView !== "detail" && selectedView !== "library" && model !== void 0) {
+          return errorResult(new Error('`model` is valid only for views "detail" and "library".'));
         }
-      }
-      if (preview) {
-        const result2 = parsedResult(
-          await route(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
-        );
-        if ("structuredContent" in result2) {
-          const refusal = belowConfidence(result2.structuredContent, minConfidence);
-          if (refusal) return refusal;
-        }
-        return result2;
-      }
-      if (minConfidence) {
-        const decision = parsedResult(
-          // minConfidence is forwarded so the CORE gate refuses, with its actionable message naming
-          // the two commands that raise the grade. The belowConfidence() check below stays only as a
-          // fallback for servers older than the core gate.
-          await route(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
-        );
-        if ("structuredContent" in decision) {
-          const refusal = belowConfidence(decision.structuredContent, minConfidence);
-          if (refusal) return refusal;
+        if (selectedView === "library" && model && [query, capabilities, order].some((value) => value !== void 0)) {
+          return errorResult(
+            new Error("Library step 2 accepts only `model` plus endpoint overrides; omit step-1 search fields.")
+          );
         }
-      }
-      const result = parsedResult(
-        await runTaskRequest(endpoint ?? DEFAULT_SERVE_ENDPOINT, {
-          task: canonicalTask,
-          objective: objective ?? "balanced",
-          model,
-          session_id: sessionId,
-          context_tokens: contextTokens,
-          required_capabilities: effectiveRequiredCapabilities ?? [],
-          prompt,
-          images,
-          messages: messages ?? [],
-          input,
-          tools,
-          keep_alive: keepAlive,
-          min_confidence: minConfidence,
-          priority: priority ?? "normal",
-          execution_preference: executionPreference ?? "auto",
-          min_placement_evidence: minPlacementEvidence ?? "configured",
-          request_options: {
-            format,
-            think,
-            options,
-            logprobs,
-            top_logprobs: topLogprobs
+        switch (selectedView) {
+          case "raw": {
+            const raw = await ollamaFetch(ollamaEndpoint, "/api/tags");
+            const models = Array.isArray(raw.models) ? raw.models : [];
+            const page = pageLiveList(models, limit, cursor, (model2) => String(model2.name ?? JSON.stringify(model2)));
+            return structuredResult({ ...raw, models: page.items, page: { returned: page.returned, total: page.total, next_cursor: page.next_cursor } });
           }
-        })
-      );
-      if (!returnEmbeddings && "structuredContent" in result) {
-        const trimmed = summarizeEmbeddings(result.structuredContent);
-        if (trimmed) return withTaskTelemetry(structuredResult(trimmed));
-      }
-      return withTaskTelemetry(result);
-    } catch (error2) {
-      return errorResult(error2);
-    }
-  }
-);
-server.registerTool(
-  "run_task_batch",
-  {
-    description: "Use when: independent work. Do not use when: dependencies. Inputs: [{id, independent:true, task}]; maxParallelism caps dispatch. Returns: ordered receipts.",
-    inputSchema: {
-      tasks: external_exports.array(batchItemParam).min(1).max(64),
-      maxParallelism: external_exports.number().int().positive().max(64).optional(),
-      endpoint: endpointParam
-    },
-    outputSchema: batchResultSchema,
-    annotations: { destructiveHint: false }
-  },
-  async ({ tasks, maxParallelism, endpoint }) => {
-    try {
-      for (const item of tasks) {
-        const task = item.task;
-        if (task.task === "embedding") {
-          if (task.input === void 0 || task.prompt !== void 0 || task.messages !== void 0 || task.tools !== void 0) {
-            return errorResult(new Error(`batch item ${item.id}: embedding requires input and accepts no chat payload.`));
+          case "resident": {
+            const managed = parsedResult(await listModels2(endpoint));
+            if (!isStructuredSuccess(managed)) return managed;
+            const data = managed.structuredContent;
+            const models = (data.models ?? []).filter((entry) => entry.resident === true).map((entry) => {
+              const { resident_view: view2, ...rest } = entry;
+              if (!view2) return rest;
+              return {
+                ...rest,
+                placement: {
+                  gpu_percent: view2.gpu_percent,
+                  assigned: view2.assigned,
+                  processor: view2.processor_label,
+                  ...view2.warning !== void 0 ? { warning: view2.warning } : {}
+                }
+              };
+            });
+            return structuredResult({ ...data, models });
           }
-        } else if (task.input !== void 0 || task.prompt === void 0 && task.messages === void 0) {
-          return errorResult(new Error(`batch item ${item.id}: chat work requires prompt or messages and accepts no input.`));
-        }
-        if (task.images !== void 0 && task.model === void 0) {
-          return errorResult(new Error(`batch item ${item.id}: images require an explicit tested vision model.`));
-        }
-      }
-      return withBatchTelemetry(parsedResult(await runTaskBatchRequest(endpoint ?? DEFAULT_SERVE_ENDPOINT, {
-        tasks: tasks.map((item) => {
-          const task = item.task;
-          return {
-            id: item.id,
-            independent: item.independent,
-            task: {
-              task: task.task === "code_review" ? "coding" : task.task,
-              objective: task.objective ?? "balanced",
-              model: task.model,
-              session_id: task.sessionId,
-              context_tokens: task.contextTokens,
-              execution_preference: task.executionPreference ?? "auto",
-              min_placement_evidence: task.minPlacementEvidence ?? "configured",
-              required_capabilities: task.requiredCapabilities ?? [],
-              priority: task.priority ?? "normal",
-              prompt: task.prompt,
-              images: task.images,
-              messages: task.messages ?? [],
-              input: task.input,
-              tools: task.tools,
-              keep_alive: task.keepAlive,
-              min_confidence: task.minConfidence,
-              request_options: {
-                format: task.format,
-                think: task.think,
-                options: task.options,
-                logprobs: task.logprobs,
-                top_logprobs: task.topLogprobs
+          case "detail": {
+            if (!model) {
+              return errorResult(new Error('view "detail" needs `model` set to an installed tag.'));
+            }
+            const data = await ollamaFetch(ollamaEndpoint, "/api/show", {
+              method: "POST",
+              body: { model }
+            });
+            let local = data;
+            if (includeLibrary && !isRemoteModel(data, model)) {
+              try {
+                const inventory = await ollamaFetch(ollamaEndpoint, "/api/tags");
+                const canonical = libraryReference(model)?.tag;
+                const entry = canonical && inventory.models?.find((entry2) => libraryReference(String(entry2.name ?? entry2.model))?.tag === canonical);
+                if (entry) local = { ...entry, ...data, digest: entry.digest };
+              } catch {
               }
             }
-          };
-        }),
-        max_parallelism: maxParallelism
-      })));
-    } catch (error2) {
-      return errorResult(error2);
-    }
-  }
-);
-server.registerTool(
-  "ollama_manage",
-  {
-    description: "Use when: pulling an approved tag or stopping a loaded model. Do not use when: deleting/recommending. Inputs: action/model; timeoutSeconds is pull-only. Returns: lifecycle result. Next: inspect models.",
-    inputSchema: {
-      action: external_exports.enum(["pull", "stop"]).describe('"pull" = disk, "stop" = memory'),
-      model: external_exports.string().min(1),
-      ollamaEndpoint: ollamaEndpointParam,
-      timeoutSeconds: external_exports.number().int().positive().optional().describe(`"pull" only. Defaults to ${DEFAULT_PULL_TIMEOUT_SECONDS}s.`)
-    },
-    outputSchema: manageResultSchema,
-    annotations: { destructiveHint: false }
-  },
-  async ({ action, model, ollamaEndpoint, timeoutSeconds }) => {
-    try {
-      if (action === "stop" && timeoutSeconds !== void 0) {
-        return errorResult(new Error('`timeoutSeconds` is valid only for action "pull".'));
-      }
-      const data = action === "pull" ? await ollamaPull(ollamaEndpoint, model, timeoutSeconds) : await ollamaFetch(ollamaEndpoint, "/api/generate", {
-        method: "POST",
-        body: { model, keep_alive: 0 }
-      });
-      const payload = data;
-      if (typeof payload.error === "string" && payload.error) {
-        return errorResult(new Error(`${action} ${model} failed: ${payload.error}`));
+            const library = includeLibrary && !isRemoteModel(local, model) ? await ollamaLibrary.lookup(model) : void 0;
+            const { license, modelfile, ...rest } = data;
+            const modelInfo = rest.model_info ?? {};
+            const contextEntry = Object.entries(modelInfo).find(
+              ([key, value]) => key.endsWith(".context_length") && typeof value === "number"
+            );
+            return structuredResult({
+              ...rest,
+              max_context_length: contextEntry?.[1] ?? null,
+              knowledge: modelKnowledge(model, local, library, includeReadme),
+              ...includeVerbose ? { license, modelfile } : {}
+            });
+          }
+          case "library":
+            return await libraryLookup({ model, query, capabilities, order, limit, cursor, includeReadme, endpoint, ollamaEndpoint });
+          default: {
+            const result = parsedResult(await listModels2(endpoint));
+            if (!isStructuredSuccess(result)) return result;
+            const data = result.structuredContent;
+            return structuredResult({ ...data, models: await enrichInstalledModels(
+              data.models ?? [],
+              ollamaLibrary,
+              includeLibrary ?? false,
+              async (upstream) => (await ollamaFetch(upstream ?? ollamaEndpoint, "/api/tags")).models ?? []
+            ) });
+          }
+        }
+      } catch (error2) {
+        return errorResult(error2);
       }
-      return structuredResult(payload);
-    } catch (error2) {
-      return errorResult(error2);
-    }
-  }
-);
-server.registerTool(
-  "ollama_delete",
-  {
-    description: "DESTRUCTIVE AND IRREVERSIBLE. Use when: a human requests one exact tag. Do not use when: freeing memory, cleaning by age, or inferring a tag. Inputs: exact tag. Returns: deleted tag. Next: refresh models.",
-    inputSchema: {
-      model: external_exports.string().min(1),
-      ollamaEndpoint: ollamaEndpointParam
-    },
-    outputSchema: deleteResultSchema,
-    annotations: { destructiveHint: true }
-  },
-  async ({ model, ollamaEndpoint }) => {
-    try {
-      await ollamaFetch(ollamaEndpoint, "/api/delete", {
-        method: "DELETE",
-        body: { name: model }
-      });
-      return structuredResult({ deleted: model });
-    } catch (error2) {
-      return errorResult(error2);
     }
-  }
-);
-server.registerTool(
-  "delegate_research",
-  {
-    description: "Use when: a narrow lookup needs an allowed workspace. Do not use when: mutation, broad judgment, or outside-workspace facts are needed. Pages tool evidence instead of flooding caller context. Returns: answer, citations, evidence, verdict. Next: verify or escalate yourself.",
-    inputSchema: {
-      question: external_exports.string().min(1).describe("narrow and self-contained, answerable by reading files"),
-      workspacePath: external_exports.string().min(1).describe("directory inside FREELLAMA_MCP_ALLOWED_ROOTS (absolute preferred; relative is resolved)"),
-      adapter: external_exports.enum(["bash", "octocode"]).optional().describe('"bash" (default) beat "octocode" on every model measured, and is faster'),
-      model: external_exports.string().min(1).optional().describe("omit to use FREELLAMA_MCP_DEFAULT_MODEL; research accuracy collapses below ~12B"),
+  );
+}
+
+// src/tools/scope.ts
+function registerScopeTool(server2) {
+  server2.registerTool("scope", {
+    description: "Use when: a multi-turn local conversation should keep its message history server-side (create/get/fork/delete), then continue with run_task {scopeId, scopeRevision}. Do not use when: you want model affinity (session) or runner KV control. Returns: process-local scope metadata; message history only with includeMessages:true.",
+    inputSchema: external_exports.object({
+      action: external_exports.enum(["create", "get", "fork", "delete"]),
       endpoint: endpointParam,
-      executionPreference: executionPreferenceParam,
-      minPlacementEvidence: minPlacementEvidenceParam,
-      legacyText: external_exports.boolean().optional().describe("false default: compact text cue; true: legacy serialized JSON text"),
-      agent: external_exports.object({
-        maxTurns: external_exports.number().int().positive().optional(),
-        contextTokens: external_exports.number().int().positive().optional().describe("Total agent context; input reserves outputTokens and safety margin."),
-        outputTokens: external_exports.number().int().positive().optional().describe("Maximum generated tokens per model call; reserved from context."),
-        temperature: external_exports.number().nonnegative().optional(),
-        seed: external_exports.number().int().nonnegative().optional(),
-        think: external_exports.boolean().optional(),
-        keepAlive: external_exports.string().min(1).optional().describe('Ollama duration such as "5m", "0", or "-1"'),
-        requestTimeoutSeconds: external_exports.number().positive().optional(),
-        toolTimeoutSeconds: external_exports.number().positive().optional(),
-        retryAttempts: external_exports.number().int().positive().optional(),
-        retryBackoffSeconds: external_exports.number().nonnegative().optional(),
-        maxParseRepairs: external_exports.number().int().nonnegative().optional(),
-        parseRepairEchoChars: external_exports.number().int().positive().optional(),
-        context: external_exports.object({
-          charsPerToken: external_exports.number().positive().optional(),
-          safetyMarginTokens: external_exports.number().int().nonnegative().optional(),
-          imageTokenEstimate: external_exports.number().int().nonnegative().optional(),
-          keepRecent: external_exports.number().int().nonnegative().optional(),
-          compactPreviewChars: external_exports.number().int().positive().optional(),
-          compactRetainRatio: external_exports.number().positive().lt(1).optional(),
-          clipHeadRatio: external_exports.number().positive().lt(1).optional(),
-          observationPageChars: external_exports.number().int().positive().optional(),
-          pinnedOverflow: external_exports.enum(["error", "clip"]).optional()
-        }).optional()
-      }).optional().describe("Per-call agent budget, recovery, and compaction controls.")
-    },
-    outputSchema: researchResultSchema,
-    annotations: { destructiveHint: false }
-  },
-  async ({ question, workspacePath, adapter, model, endpoint, executionPreference, minPlacementEvidence, legacyText, agent }) => {
-    const chosenAdapter = adapter ?? DEFAULT_RESEARCH_ADAPTER;
-    const chosenModel = model ?? DEFAULT_DELEGATE_MODEL;
-    let resolvedWorkspace;
+      scopeId: external_exports.string().uuid().optional(),
+      revision: external_exports.number().int().nonnegative().optional(),
+      messages: messagesParam.unwrap().min(0).optional(),
+      includeMessages: external_exports.boolean().optional()
+    }).strict(),
+    outputSchema: objectResultSchema,
+    annotations: { destructiveHint: true }
+  }, async ({ action, endpoint, scopeId, revision, messages, includeMessages }) => {
     try {
-      resolvedWorkspace = await assertAllowedWorkspace(workspacePath);
+      if (action === "create" !== (scopeId === void 0)) throw new Error("scopeId is required for get/fork/delete, forbidden for create.");
+      if (action === "fork" !== (revision !== void 0)) throw new Error("revision is required only for fork.");
+      if (includeMessages !== void 0 && action !== "get") throw new Error("includeMessages is valid only for get.");
+      if (messages !== void 0 && action !== "create") throw new Error("messages is valid only for create.");
+      if (action === "delete") {
+        await deleteScope2(endpoint, scopeId);
+        return structuredResult({ deleted: true, scope_id: scopeId });
+      }
+      return parsedResult(action === "create" ? await createScope2(endpoint, { messages }) : action === "fork" ? await forkScope2(endpoint, scopeId, { revision }) : await getScope2(endpoint, scopeId, includeMessages));
     } catch (error2) {
       return errorResult(error2);
     }
-    const known = MODEL_EVIDENCE[chosenModel];
-    if (known?.grade === "unusable") {
-      return structuredResult({
-        adapter: chosenAdapter,
-        verification: assessDelegatedAnswer(question, 0, chosenModel),
-        answer: "",
-        toolCallCount: 0,
-        usage: { inputTokens: null, outputTokens: null },
-        telemetry: costTelemetry({ inputTokens: null, outputTokens: null }, EXTERNAL_COST),
-        evidence: [],
-        summary: `Refused before running: ${chosenModel} is measured unusable for research here (${known.note}). Re-run with a ~27B model (see README), or answer it yourself.`
-      });
+  });
+}
+
+// src/telemetry.ts
+var EXTERNAL_COST = configuredExternalCost();
+function withTaskTelemetry(result) {
+  if (!isStructuredSuccess(result)) return result;
+  const payload = result.structuredContent;
+  const metrics = payload.metrics;
+  return structuredResult({
+    ...payload,
+    telemetry: costTelemetry({
+      inputTokens: typeof metrics?.prompt_tokens === "number" ? metrics.prompt_tokens : null,
+      outputTokens: typeof metrics?.output_tokens === "number" ? metrics.output_tokens : null,
+      totalDurationNs: typeof metrics?.total_duration_ns === "number" ? metrics.total_duration_ns : null
+    }, EXTERNAL_COST)
+  }, { text: taskAnswerText(payload) });
+}
+function withBatchTelemetry(result) {
+  if (!isStructuredSuccess(result)) return result;
+  const payload = result.structuredContent;
+  const rows = Array.isArray(payload.results) ? payload.results : [];
+  let inputTokens = 0;
+  let outputTokens = 0;
+  let complete = 0;
+  const results = rows.map((row) => {
+    if (!row || typeof row !== "object") return row;
+    const item = row;
+    if (item.ok !== true || !item.response || typeof item.response !== "object") return item;
+    const response = item.response;
+    const metrics = response.metrics;
+    const input = typeof metrics?.prompt_tokens === "number" ? metrics.prompt_tokens : null;
+    const output = typeof metrics?.output_tokens === "number" ? metrics.output_tokens : null;
+    if (input !== null && output !== null) {
+      inputTokens += input;
+      outputTokens += output;
+      complete += 1;
     }
-    const dir = await mkdtemp(path4.join(tmpdir(), "freellama-delegate-"));
-    const promptFile = path4.join(dir, "prompt.md");
-    const resultFile = path4.join(dir, "result.json");
-    try {
-      await writeFile(promptFile, `${question}
-`, "utf8");
-      const adapter2 = RESEARCH_ADAPTERS[chosenAdapter];
-      if (!existsSync4(adapter2)) {
-        return errorResult(
-          new Error(
-            `research adapter not found at ${adapter2}. In a published install it should be bundled under <package>/adapters; in-repo it comes from benchmark/local/scripts. Reinstall, or run \`npm run build\` from packages/mcp/ to re-copy it.`
-          )
-        );
-      }
-      const running = execFileAsync("python3", [adapter2], {
-        env: {
-          ...process.env,
-          FREELLAMA_TARGET_MODEL: chosenModel,
-          FREELLAMA_AGENT_MANAGED_ENDPOINT: endpoint ?? DEFAULT_SERVE_ENDPOINT,
-          FREELLAMA_AGENT_EXECUTION_PREFERENCE: executionPreference ?? "auto",
-          FREELLAMA_AGENT_MIN_PLACEMENT_EVIDENCE: minPlacementEvidence ?? "configured",
-          FREELLAMA_AGENT_TOKEN_CALIBRATION_DIR: DEFAULT_TOKEN_CALIBRATION_DIR,
-          FREELLAMA_BENCH_WORKSPACE: resolvedWorkspace,
-          FREELLAMA_BENCH_PROMPT: promptFile,
-          FREELLAMA_AGENT_RESULT: resultFile,
-          FREELLAMA_AGENT_MAX_TURNS: String(agent?.maxTurns ?? DEFAULT_DELEGATE_MAX_TURNS),
-          ...agent?.contextTokens !== void 0 ? { FREELLAMA_AGENT_NUM_CTX: String(agent.contextTokens) } : {},
-          ...agent?.outputTokens !== void 0 ? { FREELLAMA_AGENT_NUM_PREDICT: String(agent.outputTokens) } : {},
-          ...agent?.temperature !== void 0 ? { FREELLAMA_AGENT_TEMPERATURE: String(agent.temperature) } : {},
-          ...agent?.seed !== void 0 ? { FREELLAMA_AGENT_SEED: String(agent.seed) } : {},
-          ...agent?.think !== void 0 ? { FREELLAMA_AGENT_THINK: String(agent.think) } : {},
-          ...agent?.keepAlive !== void 0 ? { FREELLAMA_AGENT_KEEP_ALIVE: agent.keepAlive } : {},
-          ...agent?.requestTimeoutSeconds !== void 0 ? { FREELLAMA_AGENT_REQUEST_TIMEOUT_SECONDS: String(agent.requestTimeoutSeconds) } : {},
-          ...agent?.toolTimeoutSeconds !== void 0 ? { FREELLAMA_AGENT_TOOL_TIMEOUT_SECONDS: String(agent.toolTimeoutSeconds) } : {},
-          ...agent?.retryAttempts !== void 0 ? { FREELLAMA_AGENT_RETRY_ATTEMPTS: String(agent.retryAttempts) } : {},
-          ...agent?.retryBackoffSeconds !== void 0 ? { FREELLAMA_AGENT_RETRY_BACKOFF_SECONDS: String(agent.retryBackoffSeconds) } : {},
-          ...agent?.maxParseRepairs !== void 0 ? { FREELLAMA_AGENT_MAX_PARSE_REPAIRS: String(agent.maxParseRepairs) } : {},
-          ...agent?.parseRepairEchoChars !== void 0 ? { FREELLAMA_AGENT_PARSE_REPAIR_ECHO_CHARS: String(agent.parseRepairEchoChars) } : {},
-          ...agent?.context?.charsPerToken !== void 0 ? { FREELLAMA_AGENT_CHARS_PER_TOKEN: String(agent.context.charsPerToken) } : {},
-          ...agent?.context?.safetyMarginTokens !== void 0 ? { FREELLAMA_AGENT_SAFETY_MARGIN_TOKENS: String(agent.context.safetyMarginTokens) } : {},
-          ...agent?.context?.imageTokenEstimate !== void 0 ? { FREELLAMA_AGENT_IMAGE_TOKEN_ESTIMATE: String(agent.context.imageTokenEstimate) } : {},
-          ...agent?.context?.keepRecent !== void 0 ? { FREELLAMA_AGENT_KEEP_RECENT: String(agent.context.keepRecent) } : {},
-          ...agent?.context?.compactPreviewChars !== void 0 ? { FREELLAMA_AGENT_COMPACT_PREVIEW_CHARS: String(agent.context.compactPreviewChars) } : {},
-          ...agent?.context?.compactRetainRatio !== void 0 ? { FREELLAMA_AGENT_COMPACT_RETAIN_RATIO: String(agent.context.compactRetainRatio) } : {},
-          ...agent?.context?.clipHeadRatio !== void 0 ? { FREELLAMA_AGENT_CLIP_HEAD_RATIO: String(agent.context.clipHeadRatio) } : {},
-          ...agent?.context?.observationPageChars !== void 0 ? { FREELLAMA_AGENT_OBSERVATION_PAGE_CHARS: String(agent.context.observationPageChars) } : {},
-          ...agent?.context?.pinnedOverflow !== void 0 ? { FREELLAMA_AGENT_PINNED_OVERFLOW: agent.context.pinnedOverflow } : {}
-        },
-        timeout: DEFAULT_DELEGATE_TIMEOUT_SECONDS * 1e3,
-        // The answer is read from `resultFile`, never from stdout — but execFile's default 1 MB
-        // maxBuffer still applies to the agent's own progress logging, and overflowing it kills
-        // the subprocess and surfaces as a research failure with no explanation. Give the logs
-        // room; nothing here is proportional to the size of the answer.
-        maxBuffer: 32 * 1024 * 1024,
-        // On timeout, don't negotiate: SIGTERM can be swallowed by a python process blocked in a
-        // long model call, which would leave the child (and its VRAM) alive past the deadline the
-        // caller was promised.
-        killSignal: "SIGKILL"
-      });
-      liveDelegates.add(running.child);
-      let adapterError = null;
-      try {
-        await running;
-      } catch (error2) {
-        adapterError = error2;
-      } finally {
-        liveDelegates.delete(running.child);
-      }
-      let result;
-      try {
-        result = parseAdapterResult(await readFile(resultFile, "utf8"));
-      } catch {
-        return errorResult(
-          adapterError ?? new Error(
-            `research adapter exited without writing a readable result file \u2014 it was killed before it could report. Check that the model is loadable and that ${DEFAULT_DELEGATE_TIMEOUT_SECONDS}s is enough for this question.`
-          )
-        );
-      }
-      const evidence = result.tool_calls.map((call, index) => {
-        const target = call.arguments?.queries?.path;
-        const rawDetail = call.arguments?.command ?? null;
-        const detail = rawDetail ? clipText(rawDetail, 400) : null;
-        return {
-          step: index + 1,
-          tool: call.arguments?.tool ?? call.raw_name ?? "?",
-          // The adapters record "ok" | "error" | "repeat" per call. Carrying it through is what
-          // lets a reader tell a run that read three files from one that failed three commands —
-          // indistinguishable in the trail before, and graded identically.
-          status: call.status ?? "ok",
-          path: target ? path4.relative(resolvedWorkspace, target) : detail ? extractExistingWorkspacePath(detail, resolvedWorkspace) : null,
-          detail,
-          detail_truncated: rawDetail !== detail
-        };
-      });
-      const succeeded = evidence.filter((step) => step.status === "ok");
-      const failed = evidence.length - succeeded.length;
-      const evidenceText = evidence.map(
-        (step) => `  ${step.step}. ${step.tool}${step.status === "ok" ? "" : ` [${step.status}]`}${step.path ? ` -> ${step.path}` : ""}${step.detail ? `: ${clipText(step.detail, 400)}` : ""}`
-      ).join("\n");
-      const citations = succeeded.map((step) => ({
-        step: step.step,
-        tool: step.tool,
-        path: step.path,
-        command: step.detail,
-        command_truncated: step.detail_truncated
-      }));
-      if (adapterError) {
-        return errorResult(
-          new Error(
-            `research adapter failed: ${result.final_answer}` + (evidenceText ? `
-Evidence collected before the failure:
-${evidenceText}` : "")
-          )
+    return { ...item, response: { ...response, telemetry: costTelemetry({ inputTokens: input, outputTokens: output }, EXTERNAL_COST) } };
+  });
+  return structuredResult({
+    ...payload,
+    results,
+    telemetry: complete === rows.filter((row) => row?.ok === true).length ? costTelemetry({ inputTokens, outputTokens }, EXTERNAL_COST) : { local: null, externalEquivalent: null, note: "Batch aggregate unavailable because one or more successful items omitted token counts." }
+  }, { text: batchAnswerText(payload) });
+}
+
+// src/tools/run-task.ts
+var taskResultSchema = external_exports.object({
+  selected_model: external_exports.string().optional(),
+  context_window_fit: external_exports.string().optional(),
+  execution: external_exports.unknown().optional(),
+  response: external_exports.unknown().optional(),
+  telemetry: external_exports.unknown().optional()
+}).passthrough();
+var batchResultSchema = external_exports.object({
+  results: external_exports.unknown().optional(),
+  telemetry: external_exports.unknown().optional()
+}).passthrough();
+function validateTaskPayload(task, label) {
+  const fail = (message) => label ? `${label}: ${message}` : message;
+  if (task.scopeId === void 0 !== (task.scopeRevision === void 0)) {
+    return fail("scopeId and scopeRevision must be supplied together.");
+  }
+  if (task.topLogprobs !== void 0 && task.logprobs !== true) return fail("`topLogprobs` requires `logprobs:true`.");
+  if (task.task === "embedding") {
+    if (task.input === void 0) return fail('task "embedding" requires `input`.');
+    const chatControls = [
+      task.prompt,
+      task.systemPrompt,
+      task.images,
+      task.messages,
+      task.tools,
+      task.format,
+      task.think,
+      task.logprobs,
+      task.topLogprobs
+    ];
+    if (chatControls.some((value) => value !== void 0)) {
+      return fail('task "embedding" accepts `input`, `options`, and routing fields; chat controls are not valid.');
+    }
+    return null;
+  }
+  if (task.input !== void 0 || task.returnEmbeddings !== void 0) {
+    return fail('`input` and `returnEmbeddings` are valid only for task "embedding".');
+  }
+  if (task.prompt === void 0 && task.systemPrompt === void 0 && (task.messages === void 0 || task.messages.length === 0)) {
+    return fail(`task "${task.task ?? "completion"}" requires \`prompt\`, \`systemPrompt\`, or \`messages\`.`);
+  }
+  if (task.images !== void 0 && (task.messages !== void 0 || task.prompt === void 0)) {
+    return fail("Top-level `images` is valid only with `prompt`; put images inside messages instead.");
+  }
+  return null;
+}
+async function runBatch(tasks, maxParallelism, endpoint) {
+  for (const item of tasks) {
+    const problem = validateTaskPayload(item.task, `batch item ${item.id}`);
+    if (problem) return errorResult(new Error(problem));
+  }
+  return withBatchTelemetry(parsedResult(await runTaskBatchRequest2(endpoint ?? DEFAULT_SERVE_ENDPOINT, {
+    tasks: tasks.map(({ id, independent, task }) => ({
+      id,
+      independent,
+      task: {
+        task: task.task === void 0 ? task.scopeId === void 0 ? "completion" : void 0 : canonicalTaskKind(task.task),
+        objective: task.objective ?? (task.scopeId === void 0 ? "balanced" : void 0),
+        model: task.model,
+        session_id: task.sessionId,
+        scope_id: task.scopeId,
+        scope_revision: task.scopeRevision,
+        context_tokens: task.contextTokens,
+        execution_preference: task.executionPreference,
+        min_placement_evidence: task.minPlacementEvidence,
+        required_capabilities: task.requiredCapabilities,
+        priority: task.priority ?? "normal",
+        max_wait_seconds: task.maxWaitSeconds,
+        timeout_seconds: task.timeoutSeconds,
+        stop_on_repetition: task.stopOnRepetition,
+        prompt: task.prompt,
+        images: task.images,
+        messages: taskMessages(task),
+        input: task.input,
+        tools: task.tools,
+        keep_alive: task.keepAlive,
+        min_confidence: task.minConfidence,
+        request_options: {
+          format: task.format,
+          think: task.think,
+          options: task.options,
+          logprobs: task.logprobs,
+          top_logprobs: task.topLogprobs
+        }
+      }
+    })),
+    max_parallelism: maxParallelism
+  })));
+}
+function registerRunTaskTool(server2) {
+  server2.registerTool(
+    "run_task",
+    {
+      description: "Use when: offloading caller-supplied work (chat, small tasks, code, tool calls, OCR/vision, embeddings) to a local model. Modes: single (`prompt`/`messages`, or `input` for embedding); `tasks` = up to 64 independent items, ordered results, `maxParallelism` caps dispatch; `preview:true` decides only; `warm:true` preloads `model`. Images: base64 `images` + the model's own prompt; `stopOnRepetition` ends a looping answer early, keeping its text. Do not use when: it needs file access (pass content, or delegate_research) or high-stakes judgment. Returns: answer text; receipts in structuredContent.",
+      inputSchema: external_exports.object({
+        endpoint: endpointParam,
+        task: taskParam.removeDefault().optional().describe("caller owns prompts and output format."),
+        objective: objectiveParam,
+        model: external_exports.string().min(1).optional(),
+        sessionId: external_exports.string().uuid().optional(),
+        scopeId: external_exports.string().uuid().optional(),
+        scopeRevision: external_exports.number().int().nonnegative().optional(),
+        contextTokens: external_exports.number().int().positive().optional().describe("Total input + output window (num_ctx)."),
+        executionPreference: executionPreferenceParam,
+        minPlacementEvidence: minPlacementEvidenceParam,
+        requiredCapabilities: requiredCapabilitiesParam,
+        prompt: external_exports.string().min(1).optional(),
+        systemPrompt: systemPromptParam,
+        images: external_exports.array(external_exports.string().min(1)).min(1).optional().describe("base64, no data-URI prefix; prompt mode only; needs `model` or [policies.vision]"),
+        messages: messagesParam.describe(
+          "Caller system prompts; no injected task instructions. Overrides prompt; extra fields preserved."
+        ),
+        input: external_exports.union([external_exports.string().min(1), external_exports.array(external_exports.string().min(1)).min(1)]).optional(),
+        tools: localToolsParam,
+        keepAlive: external_exports.string().min(1).optional().describe('"0" unloads now, "-1" pins; omitted = adaptive finite TTL'),
+        format: external_exports.union([external_exports.literal("json"), external_exports.record(external_exports.unknown())]).optional(),
+        think: external_exports.union([external_exports.boolean(), external_exports.enum(["low", "medium", "high"])]).optional(),
+        options: external_exports.record(external_exports.unknown()).optional().describe("num_ctx/contextTokens and num_gpu/placement are routing-owned; num_predict caps output."),
+        logprobs: external_exports.boolean().optional(),
+        topLogprobs: external_exports.number().int().nonnegative().optional(),
+        minConfidence: minConfidenceParam,
+        priority: external_exports.enum(["interactive", "normal", "background"]).optional(),
+        maxWaitSeconds: external_exports.number().int().positive().optional(),
+        timeoutSeconds: external_exports.number().int().positive().optional(),
+        stopOnRepetition: external_exports.boolean().optional().describe("Stream and stop once output repeats itself (a loop to num_predict); keeps all generated text"),
+        defer: external_exports.boolean().optional(),
+        returnEmbeddings: external_exports.boolean().optional(),
+        preview: external_exports.boolean().optional().describe(
+          "true = routing fields only; rejects payloads and runtime controls; never executes"
+        ),
+        warm: external_exports.boolean().optional().describe("Load exact `model`; no prompt"),
+        tasks: external_exports.array(batchItemParam).min(1).max(64).optional().describe("Batch mode: independent items; per-item fields go inside each item's `task`"),
+        maxParallelism: external_exports.number().int().positive().max(64).optional().describe("Batch mode only")
+      }).strict(),
+      outputSchema: taskResultSchema.merge(batchResultSchema),
+      annotations: { destructiveHint: false }
+    },
+    async ({
+      endpoint,
+      task,
+      objective,
+      model,
+      sessionId,
+      scopeId,
+      scopeRevision,
+      contextTokens,
+      executionPreference,
+      minPlacementEvidence,
+      requiredCapabilities,
+      prompt,
+      systemPrompt,
+      images,
+      messages,
+      input,
+      tools,
+      keepAlive,
+      format,
+      think,
+      options,
+      logprobs,
+      topLogprobs,
+      minConfidence,
+      priority,
+      maxWaitSeconds,
+      timeoutSeconds,
+      stopOnRepetition,
+      defer,
+      returnEmbeddings,
+      preview,
+      warm,
+      tasks,
+      maxParallelism
+    }) => {
+      try {
+        const singleFields = {
+          task,
+          objective,
+          model,
+          sessionId,
+          scopeId,
+          scopeRevision,
+          contextTokens,
+          executionPreference,
+          minPlacementEvidence,
+          requiredCapabilities,
+          prompt,
+          systemPrompt,
+          images,
+          messages,
+          input,
+          tools,
+          keepAlive,
+          format,
+          think,
+          options,
+          logprobs,
+          topLogprobs,
+          minConfidence,
+          priority,
+          maxWaitSeconds,
+          timeoutSeconds,
+          stopOnRepetition,
+          defer,
+          returnEmbeddings,
+          preview,
+          warm
+        };
+        if (tasks !== void 0) {
+          const stray = Object.entries(singleFields).filter(([, value]) => value !== void 0).map(([name]) => name);
+          const unsupported = stray.filter((name) => !BATCH_TASK_FIELDS.has(name)).map((name) => `\`${name}\``);
+          if (unsupported.length > 0) {
+            return errorResult(new Error(`${unsupported.join(", ")} ${unsupported.length === 1 ? "is" : "are"} not supported with \`tasks\`.`));
+          }
+          const movable = stray.map((name) => `\`${name}\``);
+          if (movable.length > 0) {
+            return errorResult(new Error(`With \`tasks\`, put ${movable.join(", ")} inside each item's \`task\`.`));
+          }
+          return await runBatch(tasks, maxParallelism, endpoint);
+        }
+        if (maxParallelism !== void 0) return errorResult(new Error("`maxParallelism` is valid only with `tasks`."));
+        if (warm) {
+          const allowed = /* @__PURE__ */ new Set([
+            "warm",
+            "model",
+            "task",
+            "contextTokens",
+            "executionPreference",
+            "minPlacementEvidence",
+            "priority",
+            "keepAlive",
+            "maxWaitSeconds",
+            "timeoutSeconds",
+            "defer"
+          ]);
+          const stray = Object.entries(singleFields).filter(([name, value]) => value !== void 0 && !allowed.has(name)).map(([name]) => `\`${name}\``);
+          if (stray.length > 0) return errorResult(new Error(`\`warm:true\` loads a model without generation; remove ${stray.join(", ")}.`));
+          if (model === void 0) return errorResult(new Error("`warm:true` requires an exact installed `model`."));
+          if (task === "embedding") return errorResult(new Error("`warm:true` supports chat models; embedding warming is unsupported."));
+          return parsedResult(await warmModelRequest2(endpoint, {
+            model,
+            task: task === void 0 ? void 0 : canonicalTaskKind(task),
+            context_tokens: contextTokens,
+            execution_preference: executionPreference,
+            min_placement_evidence: minPlacementEvidence,
+            priority,
+            keep_alive: keepAlive,
+            max_wait_seconds: maxWaitSeconds,
+            timeout_seconds: timeoutSeconds,
+            defer
+          }));
+        }
+        const canonicalTask = canonicalTaskKind(task ?? "completion");
+        if (preview) {
+          const executionOnlyFields = [
+            ["scopeId", scopeId],
+            ["scopeRevision", scopeRevision],
+            ["prompt", prompt],
+            ["systemPrompt", systemPrompt],
+            ["images", images],
+            ["messages", messages],
+            ["input", input],
+            ["tools", tools],
+            ["keepAlive", keepAlive],
+            ["format", format],
+            ["think", think],
+            ["options", options],
+            ["logprobs", logprobs],
+            ["topLogprobs", topLogprobs],
+            ["returnEmbeddings", returnEmbeddings],
+            ["maxWaitSeconds", maxWaitSeconds],
+            ["priority", priority],
+            ["timeoutSeconds", timeoutSeconds],
+            ["stopOnRepetition", stopOnRepetition],
+            ["defer", defer]
+          ].filter(([, value]) => value !== void 0).map(([name]) => `\`${name}\``);
+          if (executionOnlyFields.length > 0) {
+            return errorResult(
+              new Error(
+                `\`preview:true\` accepts routing fields only; remove execution-only fields ${executionOnlyFields.join(", ")}. Use \`requiredCapabilities:["tools"]\` to preview tool capability, then make a separate execution call with the payload.`
+              )
+            );
+          }
+        }
+        const effectiveRequiredCapabilities = tools === void 0 ? requiredCapabilities : withRequiredCapability(requiredCapabilities, "tools");
+        if (preview) {
+          return parsedResult(
+            await route2(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
+          );
+        }
+        const problem = validateTaskPayload({
+          task,
+          scopeId,
+          scopeRevision,
+          prompt,
+          systemPrompt,
+          messages,
+          images,
+          input,
+          tools,
+          format,
+          think,
+          logprobs,
+          topLogprobs,
+          returnEmbeddings
+        });
+        if (problem) return errorResult(new Error(problem));
+        if (minConfidence && defer && scopeId === void 0) {
+          const decision = parsedResult(
+            await route2(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
+          );
+          if (!isStructuredSuccess(decision)) return decision;
+        }
+        const result = parsedResult(
+          await runTaskRequest2(endpoint ?? DEFAULT_SERVE_ENDPOINT, {
+            task: scopeId !== void 0 && task === void 0 ? void 0 : canonicalTask,
+            objective: objective ?? (scopeId === void 0 ? "balanced" : void 0),
+            model,
+            session_id: sessionId,
+            scope_id: scopeId,
+            scope_revision: scopeRevision,
+            context_tokens: contextTokens,
+            required_capabilities: effectiveRequiredCapabilities,
+            prompt,
+            images,
+            messages: taskMessages({ systemPrompt, messages, prompt, images }),
+            input,
+            tools,
+            keep_alive: keepAlive,
+            min_confidence: minConfidence,
+            priority: priority ?? "normal",
+            max_wait_seconds: maxWaitSeconds,
+            timeout_seconds: timeoutSeconds,
+            stop_on_repetition: stopOnRepetition,
+            defer: defer ?? false,
+            execution_preference: executionPreference,
+            min_placement_evidence: minPlacementEvidence,
+            request_options: {
+              format,
+              think,
+              options,
+              logprobs,
+              top_logprobs: topLogprobs
+            }
+          })
         );
+        if (isStructuredSuccess(result) && result.structuredContent.deferred === true) return result;
+        if (!returnEmbeddings && isStructuredSuccess(result)) {
+          const trimmed = summarizeEmbeddings(result.structuredContent);
+          if (trimmed) return withTaskTelemetry(structuredResult(trimmed));
+        }
+        return withTaskTelemetry(result);
+      } catch (error2) {
+        return errorResult(error2);
       }
-      const verification = assessDelegatedAnswer(question, succeeded.length, chosenModel);
-      const summary = `Delegated answer ready: ${result.tool_calls.length} tool call(s)` + (failed > 0 ? `, ${failed} of which did not succeed` : "") + `; ${result.usage.input_tokens ?? "?"} input / ${result.usage.output_tokens ?? "?"} output local tokens; verification=${verification.recommendation}. Read structuredContent.answer and citations.`;
-      const payload = {
-        adapter: chosenAdapter,
-        verification,
-        answer: result.final_answer,
-        // recommendation + why live under `verification`; `citations` completes the triple so a
-        // caller never has to read `summary` to act on the result.
-        citations,
-        toolCallCount: result.tool_calls.length,
-        successfulToolCallCount: succeeded.length,
-        usage: {
-          inputTokens: result.usage.input_tokens,
-          outputTokens: result.usage.output_tokens
-        },
-        telemetry: costTelemetry({
-          inputTokens: result.usage.input_tokens ?? null,
-          outputTokens: result.usage.output_tokens ?? null
-        }, EXTERNAL_COST),
-        contextManagement: result.model_metadata?.context_management ?? null,
-        execution: {
-          preference: executionPreference ?? "auto",
-          minPlacementEvidence: minPlacementEvidence ?? "configured",
-          receipts: result.model_metadata?.execution_receipts ?? []
-        },
-        evidence,
-        summary
-      };
-      return structuredResult(payload, { legacyJson: legacyText === true });
+    }
+  );
+}
+
+// src/tools/task-jobs.ts
+function registerTaskJobsTool(server2) {
+  server2.registerTool(
+    "task_jobs",
+    {
+      description: "Use when: following work submitted with run_task {defer:true}: list, get (result), cancel, or remove a job. Do not use when: submitting new work or killing a whole session. Returns: job status/result; remove cancels first.",
+      inputSchema: external_exports.object({
+        endpoint: endpointParam,
+        action: external_exports.enum(["list", "get", "cancel", "remove"]),
+        jobId: external_exports.string().uuid().optional(),
+        returnEmbeddings: external_exports.boolean().optional()
+      }).strict(),
+      outputSchema: objectResultSchema,
+      annotations: { destructiveHint: true }
+    },
+    async ({ endpoint, action, jobId, returnEmbeddings }) => {
+      try {
+        if (action === "list" ? jobId !== void 0 : jobId === void 0) {
+          return errorResult(new Error("Use jobId only with get/cancel/remove; all require a jobId."));
+        }
+        if (returnEmbeddings !== void 0 && action !== "get") {
+          return errorResult(new Error("returnEmbeddings is valid only with get."));
+        }
+        const result = parsedResult(action === "list" ? await listTaskJobs2(endpoint) : action === "get" ? await getTaskJob2(endpoint, jobId) : action === "cancel" ? await cancelTaskJob2(endpoint, jobId) : await removeTaskJob2(endpoint, jobId));
+        if (!returnEmbeddings && isStructuredSuccess(result)) {
+          const job = result.structuredContent.job;
+          if (job?.result && typeof job.result === "object") {
+            const trimmed = summarizeEmbeddings(job.result);
+            if (trimmed) return structuredResult({ ...result.structuredContent, job: { ...job, result: trimmed } });
+          }
+        }
+        return result;
+      } catch (error2) {
+        return errorResult(error2);
+      }
+    }
+  );
+}
+
+// src/tools/ollama-lifecycle.ts
+var manageResultSchema = external_exports.object({
+  status: external_exports.string().optional(),
+  progress: external_exports.unknown().optional()
+}).passthrough();
+var deleteResultSchema = external_exports.object({
+  deleted: external_exports.string()
+}).passthrough();
+function summarizeOllamaPullStream(text) {
+  const events = [];
+  for (const line of text.split(/\r?\n/)) {
+    const trimmed = line.trim();
+    if (!trimmed) continue;
+    try {
+      const parsed = JSON.parse(trimmed);
+      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
+        events.push(parsed);
+      }
+    } catch {
+    }
+  }
+  if (events.length === 0) {
+    if (!text.trim()) return { status: "empty" };
+    try {
+      const parsed = JSON.parse(text);
+      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
+        return parsed;
+      }
+    } catch {
+      return { status: "unparsed", raw: clipText(text, 500) };
+    }
+    return { status: "unparsed", raw: clipText(text, 500) };
+  }
+  const last = events[events.length - 1] ?? {};
+  const withBytes = [...events].reverse().find((event) => typeof event.total === "number" && event.total > 0);
+  const completed = typeof withBytes?.completed === "number" ? withBytes.completed : null;
+  const total = typeof withBytes?.total === "number" ? withBytes.total : null;
+  const percent = completed != null && total != null && total > 0 ? Math.round(completed / total * 1e3) / 10 : null;
+  const status2 = typeof last.error === "string" && last.error ? "error" : last.status;
+  return {
+    ...last,
+    ...status2 === void 0 ? {} : { status: status2 },
+    progress: {
+      events: events.length,
+      completed,
+      total,
+      percent,
+      lastStatus: last.status ?? null
+    }
+  };
+}
+async function ollamaPull(endpoint, model, timeoutSeconds, options = {}) {
+  const raw = await ollamaFetch(endpoint, "/api/pull", {
+    method: "POST",
+    body: { model, stream: true },
+    timeoutMs: (timeoutSeconds ?? DEFAULT_PULL_TIMEOUT_SECONDS) * 1e3,
+    parse: "ndjson",
+    signal: options.signal,
+    onLine: options.onProgress
+  });
+  if (typeof raw === "string") return summarizeOllamaPullStream(raw);
+  if (raw && typeof raw === "object" && !Array.isArray(raw)) return raw;
+  return { status: "empty" };
+}
+function pullProgressReporter(extra) {
+  const token = extra._meta?.progressToken;
+  if (token === void 0) return void 0;
+  let lastSent = 0;
+  let step = 0;
+  return (event) => {
+    const now = Date.now();
+    if (now - lastSent < 500 && event.status !== "success") return;
+    lastSent = now;
+    step += 1;
+    const completed = typeof event.completed === "number" ? event.completed : void 0;
+    const total = typeof event.total === "number" ? event.total : void 0;
+    void extra.sendNotification({
+      method: "notifications/progress",
+      params: {
+        progressToken: token,
+        progress: completed ?? step,
+        ...completed !== void 0 && total !== void 0 ? { total } : {},
+        message: String(event.status ?? "pulling")
+      }
+    }).catch(() => void 0);
+  };
+}
+function registerOllamaLifecycleTools(server2) {
+  server2.registerTool(
+    "ollama_manage",
+    {
+      description: "Use when: a human approved pulling one exact tag (action:pull) or unloading a loaded model (action:stop). Do not use when: deleting (ollama_delete), searching (models), or preloading (run_task {warm:true}). Returns: receipt; timeoutSeconds is pull-only.",
+      inputSchema: external_exports.object({
+        action: external_exports.enum(["pull", "stop"]),
+        model: external_exports.string().min(1),
+        ollamaEndpoint: ollamaEndpointParam,
+        timeoutSeconds: external_exports.number().int().positive().optional()
+      }).strict(),
+      outputSchema: manageResultSchema,
+      annotations: { destructiveHint: false }
+    },
+    async ({ action, model, ollamaEndpoint, timeoutSeconds }, extra) => {
+      try {
+        if (action === "stop" && timeoutSeconds !== void 0) {
+          return errorResult(new Error('`timeoutSeconds` is valid only for action "pull".'));
+        }
+        const data = action === "pull" ? await ollamaPull(ollamaEndpoint, model, timeoutSeconds, {
+          signal: extra.signal,
+          onProgress: pullProgressReporter(extra)
+        }) : await ollamaFetch(ollamaEndpoint, "/api/generate", {
+          method: "POST",
+          body: { model, keep_alive: 0 },
+          signal: extra.signal
+        });
+        const payload = data;
+        if (typeof payload.error === "string" && payload.error) {
+          return errorResult(new Error(`${action} ${model} failed: ${payload.error}`));
+        }
+        return structuredResult(payload);
+      } catch (error2) {
+        return errorResult(error2);
+      }
+    }
+  );
+  server2.registerTool(
+    "ollama_delete",
+    {
+      description: "DESTRUCTIVE AND IRREVERSIBLE. Use when: a human explicitly names the exact installed tag to delete. Do not use when: you inferred the cleanup yourself, or only want to unload (ollama_manage stop). Returns: the deleted tag.",
+      inputSchema: external_exports.object({
+        model: external_exports.string().min(1),
+        ollamaEndpoint: ollamaEndpointParam
+      }).strict(),
+      outputSchema: deleteResultSchema,
+      annotations: { destructiveHint: true }
+    },
+    async ({ model, ollamaEndpoint }) => {
+      try {
+        await ollamaFetch(ollamaEndpoint, "/api/delete", {
+          method: "DELETE",
+          body: { model }
+        });
+        return structuredResult({ deleted: model });
+      } catch (error2) {
+        return errorResult(error2);
+      }
+    }
+  );
+}
+
+// src/tools/delegate-research.ts
+import { existsSync as existsSync5 } from "node:fs";
+import { mkdtemp, readFile as readFile2, rm, writeFile } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import path5 from "node:path";
+
+// src/delegate.ts
+import { existsSync as existsSync4, readFileSync, realpathSync, statSync } from "node:fs";
+import path4 from "node:path";
+function researchErrorResult(error2, diagnostic) {
+  const result = errorResult(error2);
+  const summary = diagnostic ?? result.content[0].text;
+  return {
+    ...result,
+    content: [{ type: "text", text: summary }],
+    structuredContent: {
+      ...result.structuredContent ?? { error: summary },
+      answer: "",
+      summary,
+      verification: {
+        recommendation: "escalate",
+        grounded: false,
+        why: "The research request failed; no usable answer was produced."
+      }
+    }
+  };
+}
+var MODEL_EVIDENCE = (() => {
+  const configured = process.env.FREELLAMA_MCP_MODEL_EVIDENCE;
+  const candidates = [
+    configured,
+    path4.join(REPO_ROOT, "packages/mcp/evidence/model-evidence.json")
+  ].filter((x) => Boolean(x));
+  for (const file of candidates) {
+    try {
+      if (!existsSync4(file)) continue;
+      const parsed = JSON.parse(readFileSync(file, "utf8"));
+      return parsed.models ?? {};
     } catch (error2) {
-      return errorResult(error2);
-    } finally {
-      await rm(dir, { recursive: true, force: true });
+      process.stderr.write(
+        `freellama-mcp: ignoring unreadable model evidence ${file}: ${error2 instanceof Error ? error2.message : String(error2)}
+`
+      );
     }
   }
+  return {};
+})();
+var STRONG_MODEL_HINT = "Re-run with a model graded strong in packages/mcp/evidence/model-evidence.json";
+function assessDelegatedAnswer(question, evidenceCount, model) {
+  const grounded = evidenceCount > 0;
+  const evidence = MODEL_EVIDENCE[model];
+  if (evidence?.grade === "unusable") {
+    return {
+      recommendation: "escalate",
+      grounded,
+      why: `${model} is not viable for research on this machine (${evidence.note}). It answers fast and confidently while being wrong \u2014 a fast wrong answer is not a speed win. ${STRONG_MODEL_HINT}, or answer it yourself.`,
+      measuredBaseRate: `${model}: ${evidence.note}`
+    };
+  }
+  if (!evidence) {
+    return {
+      recommendation: "verify",
+      grounded,
+      why: `${model} has no measured accuracy in this repo's benchmarks, so no base rate applies to this answer. Treat it as unverified until it has been evaluated \u2014 accuracy fell off a cliff below ~12B in the models that were measured.`,
+      measuredBaseRate: "no measured base rate for this model"
+    };
+  }
+  if (evidence.grade === "weak" && grounded) {
+    return {
+      recommendation: "verify",
+      grounded,
+      why: `${model} holds up on simple single-file lookups but not beyond (${evidence.note}). Check the evidence trail, or ${STRONG_MODEL_HINT.toLowerCase()} if the answer matters.`,
+      measuredBaseRate: `${model}: ${evidence.note}`
+    };
+  }
+  if (!grounded) {
+    return {
+      recommendation: "escalate",
+      grounded: false,
+      why: "The model answered without reading a single file, so this is parametric recall, not research \u2014 the failure mode where `run_task` was verified inventing wrong facts about this project's own architecture. Re-ask with a narrower question, or answer it yourself.",
+      measuredBaseRate: "ungrounded answers have no measured accuracy \u2014 they were never the tested path"
+    };
+  }
+  const judgmentSignals = /\b(should (we|i|you)|better|best|worth|review|assess|evaluate|improve|opinion|recommend|why is|is it (good|safe|correct)|design|refactor)\b/i;
+  if (judgmentSignals.test(question)) {
+    return {
+      recommendation: "verify",
+      grounded: true,
+      why: "The question reads as a judgment call (keyword heuristic). Measured accuracy on judgment calls was well below grounded lookups, and the tone is identical either way. Check the evidence trail against the claim before acting on it.",
+      measuredBaseRate: `${model}: ${evidence.note}; measured on lookups, not judgment calls`
+    };
+  }
+  if (evidenceCount > 5) {
+    return {
+      recommendation: "verify",
+      grounded: true,
+      why: `${evidenceCount} tool calls is outside the 1-5 file envelope this tool was measured on. Wide searches are where it drifts; spot-check the evidence trail.`,
+      measuredBaseRate: `${model}: ${evidence.note}; measured within a 1-5 file scope`
+    };
+  }
+  return {
+    recommendation: "accept",
+    grounded: true,
+    why: `Grounded in ${evidenceCount} tool call(s) within the measured 1-5 file envelope, the question is lookup-shaped, and ${model} is measured strong for this (${evidence.note}).`,
+    measuredBaseRate: `${model}: ${evidence.note}`
+  };
+}
+function extractExistingWorkspacePath(command, workspace) {
+  let root;
+  try {
+    root = realpathSync(workspace);
+  } catch {
+    return null;
+  }
+  const tokens = command.match(/(?:[^\s"'\\]+|"(?:\\.|[^"])*"|'[^']*')+/g) ?? [];
+  const patternFlags = /* @__PURE__ */ new Set(["-name", "-iname", "-path", "-ipath", "-regex", "-iregex", "-wholename"]);
+  for (let index = 0; index < tokens.length; index += 1) {
+    const raw = tokens[index] ?? "";
+    if (index > 0 && patternFlags.has(tokens[index - 1] ?? "")) continue;
+    const token = raw.replace(/^["']|["']$/g, "").replace(/[|;&)]+$/g, "").replace(/:\d+(?::\d+)?$/, "");
+    if (!token || token.startsWith("-") || /[*?$`]/.test(token)) continue;
+    const candidate = path4.isAbsolute(token) ? token : path4.resolve(root, token);
+    if (!existsSync4(candidate)) continue;
+    try {
+      const resolved = realpathSync(candidate);
+      const relative = path4.relative(root, resolved);
+      const withinRoot = relative && relative !== ".." && !relative.startsWith(`..${path4.sep}`);
+      if (withinRoot && statSync(resolved).isFile()) return relative;
+    } catch {
+    }
+  }
+  return null;
+}
+var adapterCallSchema = external_exports.object({
+  raw_name: external_exports.string().optional(),
+  status: external_exports.string().optional(),
+  arguments: external_exports.object({
+    tool: external_exports.string().optional(),
+    command: external_exports.string().optional(),
+    queries: external_exports.object({ path: external_exports.string().optional() }).passthrough().optional()
+  }).passthrough().optional()
+}).passthrough();
+var adapterResultSchema = external_exports.object({
+  final_answer: external_exports.string(),
+  tool_calls: external_exports.array(adapterCallSchema).default([]),
+  usage: external_exports.object({
+    input_tokens: external_exports.number().nullable().optional(),
+    output_tokens: external_exports.number().nullable().optional()
+  }).passthrough().default({}),
+  model_metadata: external_exports.object({
+    context_compactions: external_exports.number().int().nonnegative().optional(),
+    runtime_config: external_exports.record(external_exports.string(), external_exports.unknown()).optional(),
+    context_management: external_exports.object({
+      token_counting: external_exports.enum(["configured_estimate", "model_calibrated_estimate"]),
+      estimate_scale: external_exports.number().positive(),
+      calibration_samples: external_exports.number().int().nonnegative(),
+      pinned_overflow: external_exports.enum(["error", "clip"]),
+      compactions: external_exports.number().int().nonnegative()
+    }).passthrough().optional()
+  }).passthrough().optional()
+});
+function parseAdapterResult(raw) {
+  return adapterResultSchema.parse(JSON.parse(raw));
+}
+
+// src/tools/delegate-research.ts
+var researchResultSchema = external_exports.object({
+  answer: external_exports.string(),
+  summary: external_exports.string(),
+  citations: external_exports.unknown().optional(),
+  verification: external_exports.unknown(),
+  usage: external_exports.unknown().optional(),
+  telemetry: external_exports.unknown().optional()
+}).passthrough();
+async function routedResearchModel(endpoint) {
+  const decision = JSON.parse(
+    await route2(endpoint, "coding", "balanced", void 0, void 0, void 0, [], void 0, void 0, void 0)
+  );
+  if (typeof decision.selected_model !== "string" || !decision.selected_model) {
+    throw new Error(
+      "No research model: pass `model`, set FREELLAMA_MCP_DEFAULT_MODEL, or install a coding model (models{view:'installed'} shows what routing can use)."
+    );
+  }
+  return decision.selected_model;
+}
+function registerDelegateResearchTool(server2) {
+  server2.registerTool(
+    "delegate_research",
+    {
+      description: "Use when: a bounded read-only question about files under workspacePath (where is X, what calls Y). Do not use when: editing, external facts, or final judgment. Returns: citations/answer; discard verification.recommendation=escalate, otherwise verify citations.",
+      inputSchema: external_exports.object({
+        question: external_exports.string().min(1),
+        workspacePath: external_exports.string().min(1),
+        adapter: external_exports.enum(["bash", "octocode"]).optional(),
+        model: external_exports.string().min(1).optional(),
+        endpoint: endpointParam,
+        agent: external_exports.object({
+          maxTurns: external_exports.number().int().positive().optional(),
+          contextTokens: external_exports.number().int().positive().optional(),
+          outputTokens: external_exports.number().int().positive().optional()
+        }).strict().optional()
+      }).strict(),
+      outputSchema: researchResultSchema,
+      annotations: { destructiveHint: false }
+    },
+    async ({ question, workspacePath, adapter, model, endpoint, agent }, extra) => {
+      const chosenAdapter = adapter ?? DEFAULT_RESEARCH_ADAPTER;
+      let resolvedWorkspace;
+      let chosenModel;
+      try {
+        resolvedWorkspace = await assertAllowedWorkspace(workspacePath);
+        chosenModel = model ?? DEFAULT_DELEGATE_MODEL ?? await routedResearchModel(endpoint);
+      } catch (error2) {
+        return researchErrorResult(error2);
+      }
+      const known = MODEL_EVIDENCE[chosenModel];
+      if (known?.grade === "unusable") {
+        return structuredResult({
+          adapter: chosenAdapter,
+          verification: assessDelegatedAnswer(question, 0, chosenModel),
+          answer: "",
+          toolCallCount: 0,
+          usage: { inputTokens: null, outputTokens: null },
+          telemetry: costTelemetry({ inputTokens: null, outputTokens: null }, EXTERNAL_COST),
+          evidence: [],
+          summary: `Refused before running: ${chosenModel} is measured unusable for research here (${known.note}). ${STRONG_MODEL_HINT}, or answer it yourself.`
+        });
+      }
+      const dir = await mkdtemp(path5.join(tmpdir(), "freellama-delegate-"));
+      const promptFile = path5.join(dir, "prompt.md");
+      const resultFile = path5.join(dir, "result.json");
+      try {
+        await writeFile(promptFile, `${question}
+`, "utf8");
+        const adapter2 = RESEARCH_ADAPTERS[chosenAdapter];
+        if (!existsSync5(adapter2)) {
+          return researchErrorResult(
+            new Error(
+              `research adapter not found at ${adapter2}. It is compiled to <package>/dist/research by the package build. Reinstall, or run \`yarn build\` from packages/mcp/ to rebuild it.`
+            )
+          );
+        }
+        const running = runDelegate(process.execPath, [adapter2], {
+          signal: extra.signal,
+          env: {
+            ...delegateEnvironment(),
+            FREELLAMA_TARGET_MODEL: chosenModel,
+            FREELLAMA_AGENT_MANAGED_ENDPOINT: endpoint ?? DEFAULT_SERVE_ENDPOINT,
+            FREELLAMA_AGENT_TOKEN_CALIBRATION_DIR: DEFAULT_TOKEN_CALIBRATION_DIR,
+            FREELLAMA_BENCH_WORKSPACE: resolvedWorkspace,
+            FREELLAMA_BENCH_PROMPT: promptFile,
+            FREELLAMA_AGENT_RESULT: resultFile,
+            FREELLAMA_AGENT_MAX_TURNS: String(agent?.maxTurns ?? DEFAULT_DELEGATE_MAX_TURNS),
+            ...agent?.contextTokens !== void 0 ? { FREELLAMA_AGENT_NUM_CTX: String(agent.contextTokens) } : {},
+            ...agent?.outputTokens !== void 0 ? { FREELLAMA_AGENT_NUM_PREDICT: String(agent.outputTokens) } : {}
+          },
+          timeoutMs: DEFAULT_DELEGATE_TIMEOUT_SECONDS * 1e3,
+          // The answer is read from `resultFile`, never from stdout — but execFile's usual 1 MB
+          // maxBuffer still applies to the agent's own progress logging, and overflowing it kills
+          // the subprocess and surfaces as a research failure with no explanation. Give the logs
+          // room; nothing here is proportional to the size of the answer.
+          maxBuffer: 32 * 1024 * 1024
+          // On timeout, abort, or overflow runDelegate SIGKILLs the whole process group: SIGTERM can
+          // be swallowed by an adapter blocked in a long model call, and killing the adapter alone
+          // would orphan `npx octocode` and sandbox shells.
+        });
+        liveDelegates.add(running.child);
+        let adapterError = null;
+        try {
+          await running.done;
+        } catch (error2) {
+          adapterError = error2;
+        } finally {
+          liveDelegates.delete(running.child);
+          killDelegateTree(running.child);
+        }
+        let result;
+        try {
+          result = parseAdapterResult(await readFile2(resultFile, "utf8"));
+        } catch {
+          return researchErrorResult(
+            adapterError ?? new Error(
+              `research adapter exited without writing a readable result file \u2014 it was killed before it could report. Check that the model is loadable and that ${DEFAULT_DELEGATE_TIMEOUT_SECONDS}s is enough for this question.`
+            )
+          );
+        }
+        const evidence = result.tool_calls.map((call, index) => {
+          const target = call.arguments?.queries?.path;
+          const rawDetail = call.arguments?.command ?? null;
+          const detail = rawDetail ? clipText(rawDetail, 400) : null;
+          return {
+            step: index + 1,
+            tool: call.arguments?.tool ?? call.raw_name ?? "?",
+            // The adapters record "ok" | "error" | "repeat" per call. Carrying it through is what
+            // lets a reader tell a run that read three files from one that failed three commands —
+            // indistinguishable in the trail before, and graded identically.
+            status: call.status ?? "ok",
+            path: target ? path5.relative(resolvedWorkspace, target) : detail ? extractExistingWorkspacePath(detail, resolvedWorkspace) : null,
+            detail,
+            detail_truncated: rawDetail !== detail
+          };
+        });
+        const succeeded = evidence.filter((step) => step.status === "ok");
+        const failed = evidence.length - succeeded.length;
+        const evidenceText = evidence.map(
+          (step) => `  ${step.step}. ${step.tool}${step.status === "ok" ? "" : ` [${step.status}]`}${step.path ? ` -> ${step.path}` : ""}${step.detail ? `: ${step.detail}` : ""}`
+        ).join("\n");
+        const citations = succeeded.map((step) => ({ step: step.step, tool: step.tool, path: step.path }));
+        if (adapterError) {
+          const diagnostic = `research adapter failed: ${result.final_answer}` + (evidenceText ? `
+Evidence collected before the failure:
+${evidenceText}` : "");
+          const refusal = external_exports.object({ receipt: external_exports.object({ error: external_exports.string() }).passthrough() }).safeParse(result.model_metadata?.transport_error);
+          if (refusal.success) {
+            return researchErrorResult(new Error(JSON.stringify(refusal.data.receipt)), diagnostic);
+          }
+          return researchErrorResult(new Error(diagnostic));
+        }
+        const verification = assessDelegatedAnswer(question, succeeded.length, chosenModel);
+        const summary = `Delegated answer ready: ${result.tool_calls.length} tool call(s)` + (failed > 0 ? `, ${failed} of which did not succeed` : "") + `; ${result.usage.input_tokens ?? "?"} input / ${result.usage.output_tokens ?? "?"} output local tokens; verification=${verification.recommendation}. Read structuredContent.answer and citations.`;
+        const payload = {
+          adapter: chosenAdapter,
+          verification,
+          answer: result.final_answer,
+          // recommendation + why live under `verification`; `citations` completes the triple so a
+          // caller never has to read `summary` to act on the result.
+          citations,
+          toolCallCount: result.tool_calls.length,
+          successfulToolCallCount: succeeded.length,
+          usage: {
+            inputTokens: result.usage.input_tokens,
+            outputTokens: result.usage.output_tokens
+          },
+          telemetry: costTelemetry({
+            inputTokens: result.usage.input_tokens ?? null,
+            outputTokens: result.usage.output_tokens ?? null
+          }, EXTERNAL_COST),
+          contextManagement: result.model_metadata?.context_management ?? null,
+          execution: { receipts: result.model_metadata?.execution_receipts ?? [] },
+          evidence,
+          summary
+        };
+        const cited = [...new Set(citations.map((citation) => citation.path).filter(Boolean))];
+        const answerText = `${clipText(result.final_answer, ANSWER_TEXT_MAX_CHARS)}
+
+verification: ${verification.recommendation}` + (cited.length ? `
+cited: ${cited.slice(0, 20).join(", ")}` : "") + `
+(${chosenModel}, ${result.tool_calls.length} tool call(s))`;
+        return structuredResult(payload, { text: answerText });
+      } catch (error2) {
+        return researchErrorResult(error2);
+      } finally {
+        await rm(dir, { recursive: true, force: true });
+      }
+    }
+  );
+}
+
+// src/index.ts
+installProcessLifecycle();
+var INSTRUCTIONS = `models{view:"installed"}; models{view:"resident"}.
+run_task preview never executes; code_review aliases coding. requiredCapabilities:["tools"] previews a tool route; omit preview and supply the payload. Caller runs tool calls.
+quality objective needs a task policy or explicit model; medium confidence needs policy+benchmark; observed placement needs /api/ps proof.
+Scoped run_task needs scopeId+scopeRevision. defer:true: job.id\u2192task_jobs.jobId unchanged. Restart loses scopes/sessions/jobs.
+Check isError first; prefer structuredContent, else content[].text. page.next_cursor\u2192cursor unchanged.
+caller owns task decomposition/prompts/format/verification; findings are candidates, not accepted defects. operator owns endpoints, exact --cpu-model assignments/lifecycle. Ollama plus the OS/driver run physical CPU/GPU.
+ask approval for one exact tag and reported size before ollama_manage pull; search or recommendation is never download permission. Stop/delete need exact-tag approval.
+serve :11435; default-endpoint autostart optional. Docs: freellama://docs/index.`;
+var server = new McpServer(
+  { name: "freellama", version: SERVER_VERSION },
+  { instructions: INSTRUCTIONS }
 );
-var transport = new StdioServerTransport();
+registerDocsResources(server);
+registerDoctorTool(server);
+registerSessionTool(server);
+registerModelsTool(server);
+registerScopeTool(server);
+registerRunTaskTool(server);
+registerTaskJobsTool(server);
+registerOllamaLifecycleTools(server);
+registerDelegateResearchTool(server);
+var transport = withoutSchemaDialect(new StdioServerTransport());
 await server.connect(transport);