@octocodeai/freellama-mcp-server 0.1.0 → 0.5.0

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;
       }
@@ -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);
           }
@@ -7189,12 +7201,12 @@ var require_dist = __commonJS({
 
 // src/index.ts
 import { execFile } from "node:child_process";
-import { existsSync as existsSync4, readdirSync } from "node:fs";
+import { existsSync as existsSync5, 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 path5 from "node:path";
+import { fileURLToPath as fileURLToPath3 } from "node:url";
 import { promisify } from "node:util";
 
 // ../../node_modules/zod/v3/external.js
@@ -7675,8 +7687,8 @@ 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
@@ -7731,18 +7743,18 @@ 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 +7764,9 @@ 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 +7775,14 @@ 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 +7804,11 @@ 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 +8242,7 @@ 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 +8256,7 @@ 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 +8269,7 @@ 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 +8295,7 @@ 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 +8305,7 @@ 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 +8318,7 @@ 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 +8328,7 @@ 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 +8338,7 @@ 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 +8348,7 @@ 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 +8358,7 @@ 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 +8368,7 @@ 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 +8380,7 @@ 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 +8392,7 @@ 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 +8404,7 @@ 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 +8418,7 @@ 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 +8428,7 @@ 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 +8439,7 @@ 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 +8450,7 @@ 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 +8461,7 @@ 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 +8471,7 @@ 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 +8481,7 @@ 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 +8491,7 @@ 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 +8501,7 @@ 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 +8511,7 @@ 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 +8521,13 @@ 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 +8803,7 @@ 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 +8814,7 @@ 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 +8828,7 @@ 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 +8842,7 @@ 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 +8852,7 @@ 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 +8861,13 @@ 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 +9032,7 @@ 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 +9045,7 @@ 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 +9058,7 @@ 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 +9068,13 @@ 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 +9232,7 @@ 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 +9246,7 @@ 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 +9259,14 @@ 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 +9452,7 @@ 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 +9475,7 @@ var ZodArray = class _ZodArray extends ZodType {
           exact: true,
           message: def.exactLength.message
         });
-        status.dirty();
+        status2.dirty();
       }
     }
     if (def.minLength !== null) {
@@ -9476,7 +9488,7 @@ var ZodArray = class _ZodArray extends ZodType {
           exact: false,
           message: def.minLength.message
         });
-        status.dirty();
+        status2.dirty();
       }
     }
     if (def.maxLength !== null) {
@@ -9489,20 +9501,20 @@ 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 +9603,7 @@ 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 +9638,7 @@ var ZodObject = class _ZodObject extends ZodType {
             code: ZodIssueCode.unrecognized_keys,
             keys: extraKeys
           });
-          status.dirty();
+          status2.dirty();
         }
       } else if (unknownKeys === "strip") {
       } else {
@@ -9660,10 +9672,10 @@ 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 +10153,7 @@ 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 +10166,9 @@ 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 +10206,7 @@ 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 +10234,7 @@ 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 +10244,10 @@ 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 +10279,7 @@ 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 +10299,9 @@ 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 +10332,7 @@ 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 +10359,11 @@ 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 +10374,11 @@ 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 +10392,7 @@ 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 +10412,7 @@ var ZodSet = class _ZodSet extends ZodType {
           exact: false,
           message: def.minSize.message
         });
-        status.dirty();
+        status2.dirty();
       }
     }
     if (def.maxSize !== null) {
@@ -10413,7 +10425,7 @@ var ZodSet = class _ZodSet extends ZodType {
           exact: false,
           message: def.maxSize.message
         });
-        status.dirty();
+        status2.dirty();
       }
     }
     const valueType = this._def.valueType;
@@ -10423,10 +10435,10 @@ 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 +10769,15 @@ 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 +10789,7 @@ 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 +10800,12 @@ 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 +10816,7 @@ 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 +10841,17 @@ 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 +10869,13 @@ 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 +11054,7 @@ 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 +11065,7 @@ 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 +11085,7 @@ 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 +11445,10 @@ 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 +11768,11 @@ 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 +15183,11 @@ 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 +17486,8 @@ 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
@@ -19694,8 +19706,8 @@ 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 +19723,15 @@ 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({
@@ -21438,9 +21450,10 @@ import path from "node:path";
 import { fileURLToPath } from "node:url";
 import { homedir } from "node:os";
 function findRepoRoot() {
+  const packageRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
   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") === packageRoot) return dir;
     const parent = path.dirname(dir);
     if (parent === dir) break;
     dir = parent;
@@ -21467,7 +21480,7 @@ 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);
@@ -21521,16 +21534,159 @@ 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",
+  "PYTHONPATH"
+]);
+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 { createScope, getScope, forkScope, deleteScope, warmModelRequest, doctor, machine, health, status, usage, createSession, deleteSession, killSession, listModels, route, runTask, runTaskRequest, runTaskBatchRequest, listTaskJobs, getTaskJob, cancelTaskJob, removeTaskJob } = native;
 var { version: SERVER_VERSION } = createRequire(import.meta.url)("../package.json");
 
-// src/helpers.ts
-import { existsSync as existsSync2, realpathSync, statSync } from "node:fs";
+// 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";
+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 {
+    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;
+  }
+}
+async function answers(endpoint, timeoutMs) {
+  try {
+    await fetch(`${endpoint.replace(/\/$/, "")}${HEALTH_PATH}`, { signal: AbortSignal.timeout(timeoutMs) });
+    return true;
+  } catch {
+    return false;
+  }
+}
+function serveBinaryCandidates() {
+  const exe = process.platform === "win32" ? "freellama.exe" : "freellama";
+  const candidates = [
+    path2.join(REPO_ROOT, "target", "release", exe),
+    path2.join(REPO_ROOT, "target", "debug", exe)
+  ];
+  const requireFromHere = createRequire2(fileURLToPath2(import.meta.url));
+  const suffixes = process.platform === "linux" ? ["-gnu", "-musl"] : process.platform === "win32" ? ["-msvc"] : [""];
+  for (const suffix of suffixes) {
+    try {
+      const manifest = requireFromHere.resolve(`@octocodeai/freellama-native-${process.platform}-${process.arch}${suffix}/package.json`);
+      candidates.push(path2.join(path2.dirname(manifest), exe));
+    } catch {
+    }
+  }
+  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 ${spawned.signalCode}).`);
+    } finally {
+      starting = null;
+    }
+  })();
+  await starting;
+}
+
+// src/helpers.ts
+import { existsSync as existsSync3, realpathSync, statSync } from "node:fs";
+import path3 from "node:path";
 function extractExistingWorkspacePath(command, workspace) {
   let root;
   try {
@@ -21545,35 +21701,36 @@ function extractExistingWorkspacePath(command, workspace) {
     if (index > 0 && patternFlags.has(tokens[index - 1] ?? "")) continue;
     const token = raw.replace(/^["']|["']$/g, "").replace(/[|;&)]+$/g, "").replace(/:\d+(?::\d+)?$/, "");
     if (!token || token.startsWith("-") || /[*?$`]/.test(token)) continue;
-    const candidate = path2.isAbsolute(token) ? token : path2.resolve(root, token);
-    if (!existsSync2(candidate)) continue;
+    const candidate = path3.isAbsolute(token) ? token : path3.resolve(root, token);
+    if (!existsSync3(candidate)) continue;
     try {
       const resolved = realpathSync(candidate);
-      const relative = path2.relative(root, resolved);
-      const withinRoot = relative && relative !== ".." && !relative.startsWith(`..${path2.sep}`);
+      const relative = path3.relative(root, resolved);
+      const withinRoot = relative && relative !== ".." && !relative.startsWith(`..${path3.sep}`);
       if (withinRoot && statSync(resolved).isFile()) return relative;
     } catch {
     }
   }
   return null;
 }
-async function ollamaFetch(endpoint, path5, init = {}) {
+async function ollamaFetch(endpoint, path6, init = {}) {
   const base = (endpoint ?? DEFAULT_OLLAMA_ENDPOINT).replace(/\/$/, "");
   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;
@@ -21582,6 +21739,26 @@ async function ollamaFetch(endpoint, path5, init = {}) {
     clearTimeout(timeout);
   }
 }
+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 {
+      }
+    }
+  }
+  return text + decoder.decode();
+}
 function summarizeOllamaPullStream(text) {
   const events = [];
   for (const line of text.split(/\r?\n/)) {
@@ -21612,10 +21789,10 @@ function summarizeOllamaPullStream(text) {
   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;
+  const status2 = typeof last.error === "string" && last.error ? "error" : last.status;
   return {
     ...last,
-    ...status === void 0 ? {} : { status },
+    ...status2 === void 0 ? {} : { status: status2 },
     progress: {
       events: events.length,
       completed,
@@ -21625,19 +21802,21 @@ function summarizeOllamaPullStream(text) {
     }
   };
 }
-async function ollamaPull(endpoint, model, timeoutSeconds) {
+async function ollamaPull(endpoint, model, timeoutSeconds, options = {}) {
   const raw = await ollamaFetch(endpoint, "/api/pull", {
     method: "POST",
-    body: { name: model, stream: true },
+    body: { model, stream: true },
     timeoutMs: (timeoutSeconds ?? DEFAULT_PULL_TIMEOUT_SECONDS) * 1e3,
-    parse: "ndjson"
+    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" };
 }
-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",
@@ -21649,17 +21828,15 @@ 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 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 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.');
+var minConfidenceParam = external_exports.enum(["low", "medium"]).optional();
 function belowConfidence(decision, minConfidence) {
   if (!minConfidence) return null;
   const actual = typeof decision.confidence === "string" ? decision.confidence : "low";
@@ -21679,22 +21856,49 @@ 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,
+  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(),
   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(),
   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(),
@@ -21716,7 +21920,9 @@ var doctorResultSchema = external_exports.object({
 }).passthrough();
 var sessionResultSchema = external_exports.object({
   session_id: external_exports.string().optional(),
-  deleted: external_exports.boolean().optional()
+  deleted: external_exports.boolean().optional(),
+  killed: external_exports.boolean().optional(),
+  runner_stop_confirmed: external_exports.boolean().optional()
 }).passthrough();
 var manageResultSchema = external_exports.object({
   status: external_exports.string().optional(),
@@ -21728,7 +21934,9 @@ var deleteResultSchema = external_exports.object({
 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()
+  page: external_exports.unknown().optional(),
+  knowledge: external_exports.unknown().optional(),
+  library: external_exports.unknown().optional()
 }).passthrough();
 var taskResultSchema = external_exports.object({
   selected_model: external_exports.string().optional(),
@@ -21763,22 +21971,22 @@ 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: {
@@ -21816,9 +22024,52 @@ function resultSummary(value) {
   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(message.content.trim());
+    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 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 ?? (options.legacyJson ? serialize(value) : resultSummary(value));
   return {
-    content: [{ type: "text", text: options.legacyJson ? serialize(value) : resultSummary(value) }],
+    content: [{ type: "text", text }],
     structuredContent: value
   };
 }
@@ -21839,13 +22090,31 @@ ${clipText(raw, 2e3)}`
   }
   return structuredResult(value);
 }
+function isStructuredSuccess(result) {
+  return !("isError" in result && result.isError) && "structuredContent" in result && result.structuredContent !== void 0;
+}
 function errorResult(error2) {
   const message = error2 instanceof Error ? error2.message : String(error2);
+  let receipt;
+  try {
+    const parsed = JSON.parse(message);
+    if (parsed !== null && typeof parsed === "object" && !Array.isArray(parsed) && typeof parsed.error === "string") {
+      receipt = parsed;
+    }
+  } catch {
+  }
   const managedServeUnavailable = /(?:connection refused|connect error|failed to connect|error sending request)/i.test(message) && /(?:_freellama|127\.0\.0\.1:11435|localhost:11435)/i.test(message);
+  const ollamaUnavailable = !managedServeUnavailable && /(?:connection refused|connect error|failed to connect|error sending request)/i.test(message) && /\/api\/(?:tags|ps|chat|generate|embed|show)/.test(message);
   const actionableMessage = managedServeUnavailable ? `${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 };
+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 } : {}
+  };
 }
 var adapterCallSchema = external_exports.object({
   raw_name: external_exports.string().optional(),
@@ -21900,11 +22169,64 @@ function summarizeEmbeddings(payload) {
 }
 
 // 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;
+    if (!/<p\b[^>]*class="max-w-lg/.test(block)) 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() ?? "";
     const capabilities = [...block.matchAll(/text-(?:indigo-600|cyan-500)[^>]*>([a-z]+)<\/span>/g)].map(
       (m) => m[1]
@@ -21918,7 +22240,8 @@ function parseModelSearch(html) {
       capabilities,
       pulls: stat("Pulls"),
       tags: stat("Tag"),
-      cloudOnly: capabilities.includes("cloud")
+      cloudAvailable: capabilities.includes("cloud"),
+      cloudOnly: null
     });
   }
   return results;
@@ -21926,38 +22249,330 @@ 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/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;
+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 {
+      return await work;
+    } finally {
+      this.pending.delete(ref.family);
+    }
+  }
+  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 ${response.status}.`);
+    }
+    if (!(response.headers.get("content-type") ?? "").includes("text/html")) {
+      await response.body?.cancel();
+      throw new LibraryFetchError("parse_error", "Ollama library did not return HTML.");
+    }
+    if (Number(response.headers.get("content-length")) > MAX_PAGE_BYTES) {
+      await response.body?.cancel();
+      throw new LibraryFetchError("unavailable", "Ollama library page exceeds the 2 MiB limit.");
+    }
+    const reader = response.body?.getReader();
+    if (!reader) throw new LibraryFetchError("parse_error", "Ollama library returned an empty body.");
+    const chunks = [];
+    let bytes = 0;
+    try {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        bytes += value.byteLength;
+        if (bytes > MAX_PAGE_BYTES) throw new LibraryFetchError("unavailable", "Ollama library page exceeds the 2 MiB limit.");
+        chunks.push(value);
+      }
+      return Buffer.concat(chunks).toString("utf8");
+    } finally {
+      await reader.cancel();
+    }
+  }
+  async load(ref) {
+    let result;
+    try {
+      const pages = await Promise.allSettled([this.page(ref.url), this.page(`${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. Local metadata remains available."
+      };
+    }
+    const now = (this.options.now ?? Date.now)();
+    return {
+      ...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 {
+    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 } } : {}
+  };
+}
+function provenance(record3) {
+  const { card: _card, tags: _tags, readmeTruncated: _truncated, ...source } = record3;
+  return source;
+}
+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 (data.remote_host || data.remote_model || /-cloud$/i.test(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 {
+          localError = "The selected Ollama backend's inventory could not be verified; public guidance was withheld.";
+        }
+      }
+      const record3 = includeLibrary && !localError && !local.remote_host && !local.remote_model && !/-cloud$/i.test(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 results;
+}
+
 // src/delegate.ts
-import { existsSync as existsSync3, readFileSync } from "node:fs";
-import path3 from "node:path";
+import { existsSync as existsSync4, readFileSync } 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,
-    path3.join(REPO_ROOT, "benchmark/evidence/model-evidence.json")
+    path4.join(REPO_ROOT, "benchmark/evidence/model-evidence.json")
   ].filter((x) => Boolean(x));
   for (const file of candidates) {
     try {
-      if (!existsSync3(file)) continue;
+      if (!existsSync4(file)) continue;
       const parsed = JSON.parse(readFileSync(file, "utf8"));
       return parsed.models ?? {};
     } catch {
@@ -22025,11 +22640,135 @@ function assessDelegatedAnswer(question, evidenceCount, model) {
   };
 }
 
+// src/context-tools.ts
+var routeDefaults = external_exports.object({
+  task: taskParam.removeDefault().optional(),
+  objective: objectiveParam,
+  model: external_exports.string().min(1).optional(),
+  contextTokens: external_exports.number().int().positive().optional(),
+  executionPreference: executionPreferenceParam,
+  minPlacementEvidence: minPlacementEvidenceParam,
+  minConfidence: minConfidenceParam,
+  requiredCapabilities: requiredCapabilitiesParam
+}).strict();
+var limits = external_exports.object({
+  maxMessages: external_exports.number().int().positive().optional(),
+  maxBytes: external_exports.number().int().positive().optional(),
+  maxEstimatedTokens: external_exports.number().int().positive().optional(),
+  ttlSeconds: external_exports.number().int().positive().optional()
+}).strict();
+function scopeBody(messages, defaults, bounds) {
+  return {
+    messages,
+    route_defaults: defaults === void 0 ? void 0 : {
+      task: defaults.task === void 0 ? void 0 : canonicalTaskKind(defaults.task),
+      objective: defaults.objective,
+      model: defaults.model,
+      context_tokens: defaults.contextTokens,
+      execution_preference: defaults.executionPreference,
+      min_placement_evidence: defaults.minPlacementEvidence,
+      min_confidence: defaults.minConfidence,
+      required_capabilities: defaults.requiredCapabilities
+    },
+    limits: bounds === void 0 ? void 0 : {
+      max_messages: bounds.maxMessages,
+      max_bytes: bounds.maxBytes,
+      max_estimated_tokens: bounds.maxEstimatedTokens,
+      ttl_seconds: bounds.ttlSeconds
+    }
+  };
+}
+function registerContextTools(server2) {
+  const createScope2 = withServe(createScope);
+  const getScope2 = withServe(getScope);
+  const forkScope2 = withServe(forkScope);
+  const deleteScope2 = withServe(deleteScope);
+  const warmModel = withServe(warmModelRequest);
+  server2.registerTool("scope", {
+    description: "Use when: saved history. Do not use when: KV control. Returns: process-local scope metadata; history is opt-in.",
+    inputSchema: external_exports.object({
+      action: external_exports.enum(["create", "get", "fork", "delete"]),
+      endpoint: endpointParam,
+      scopeId: external_exports.string().uuid().optional(),
+      revision: external_exports.number().int().nonnegative().optional(),
+      messages: messagesParam.unwrap().min(0).optional(),
+      routeDefaults: routeDefaults.optional(),
+      limits: limits.optional(),
+      includeMessages: external_exports.boolean().optional()
+    }).strict(),
+    outputSchema: objectResultSchema,
+    annotations: { destructiveHint: true }
+  }, async ({ action, endpoint, scopeId, revision, messages, routeDefaults: defaults, limits: bounds, includeMessages }) => {
+    try {
+      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 ((defaults !== void 0 || bounds !== void 0) && action !== "create" && action !== "fork") throw new Error("routeDefaults and limits are valid only for create/fork.");
+      if (action === "delete") {
+        await deleteScope2(endpoint, scopeId);
+        return structuredResult({ deleted: true, scope_id: scopeId });
+      }
+      return parsedResult(action === "create" ? await createScope2(endpoint, scopeBody(messages, defaults, bounds)) : action === "fork" ? await forkScope2(endpoint, scopeId, { ...scopeBody(void 0, defaults, bounds), revision }) : await getScope2(endpoint, scopeId, includeMessages));
+    } catch (error2) {
+      return errorResult(error2);
+    }
+  });
+  server2.registerTool("warm_model", {
+    description: "Use when: preload an installed model. Do not use when: generation or pull. Returns: managed residency receipt or job ID.",
+    inputSchema: external_exports.object({
+      endpoint: endpointParam,
+      model: external_exports.string().min(1),
+      task: taskParam.removeDefault().optional(),
+      contextTokens: external_exports.number().int().positive().optional(),
+      executionPreference: executionPreferenceParam,
+      minPlacementEvidence: minPlacementEvidenceParam,
+      priority: external_exports.enum(["interactive", "normal", "background"]).optional(),
+      keepAlive: external_exports.string().min(1).optional().describe("Omitted = adaptive finite TTL."),
+      maxWaitSeconds: external_exports.number().int().positive().optional(),
+      timeoutSeconds: external_exports.number().int().positive().optional(),
+      defer: external_exports.boolean().optional()
+    }).strict(),
+    outputSchema: objectResultSchema,
+    annotations: { destructiveHint: false }
+  }, async ({ endpoint, model, task, contextTokens, executionPreference, minPlacementEvidence, priority, keepAlive, maxWaitSeconds, timeoutSeconds, defer }) => {
+    try {
+      if (task === "embedding") throw new Error("warm_model supports chat models; embedding warming is unsupported.");
+      return parsedResult(await warmModel(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
+      }));
+    } catch (error2) {
+      return errorResult(error2);
+    }
+  });
+}
+
 // src/index.ts
 var execFileAsync = promisify(execFile);
+var createSession2 = withServe(createSession);
+var deleteSession2 = withServe(deleteSession);
+var killSession2 = withServe(killSession);
+var listModels2 = withServe(listModels);
+var ollamaLibrary = new OllamaLibraryClient();
+var route2 = withServe(route);
+var runTaskRequest2 = withServe(runTaskRequest);
+var runTaskBatchRequest2 = withServe(runTaskBatchRequest);
+var listTaskJobs2 = withServe(listTaskJobs);
+var getTaskJob2 = withServe(getTaskJob);
+var cancelTaskJob2 = withServe(cancelTaskJob);
+var removeTaskJob2 = withServe(removeTaskJob);
 var EXTERNAL_COST = configuredExternalCost();
 function withTaskTelemetry(result) {
-  if (!("structuredContent" in result)) return result;
+  if (!isStructuredSuccess(result)) return result;
   const payload = result.structuredContent;
   const metrics = payload.metrics;
   return structuredResult({
@@ -22039,10 +22778,10 @@ function withTaskTelemetry(result) {
       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 (!("structuredContent" in result)) return result;
+  if (!isStructuredSuccess(result)) return result;
   const payload = result.structuredContent;
   const rows = Array.isArray(payload.results) ? payload.results : [];
   let inputTokens = 0;
@@ -22067,7 +22806,7 @@ function withBatchTelemetry(result) {
     ...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) });
 }
 function pageLiveList(items, limit, cursor, identity) {
   const pageSize = limit ?? 20;
@@ -22095,9 +22834,9 @@ function pageLiveList(items, limit, cursor, identity) {
 }
 var liveDelegates = /* @__PURE__ */ new Set();
 function killLiveDelegates() {
-  for (const child of liveDelegates) {
+  for (const child2 of liveDelegates) {
     try {
-      child.kill("SIGKILL");
+      child2.kill("SIGKILL");
     } catch {
     }
   }
@@ -22109,63 +22848,107 @@ for (const stream of [process.stdout, process.stderr]) {
     throw error2;
   });
 }
-process.on("exit", killLiveDelegates);
+process.on("exit", () => {
+  killLiveDelegates();
+  stopAutostartedServe();
+});
+process.stdin.on("close", () => {
+  killLiveDelegates();
+  stopAutostartedServe();
+  process.exit(0);
+});
+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);
+  };
+}
 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 INSTRUCTIONS = `models{view:"installed"}; models{view:"resident"} for residency.
+run_task preview never executes; code_review aliases coding. Preview routing only; requiredCapabilities:["tools"] previews tools; omit preview and supply the payload. Caller executes tools.
+quality needs policy/model; medium needs policy+benchmark; observed needs /api/ps proof.
+Scoped run_task needs scopeId+scopeRevision. Scopes store history; sessions affinity, not messages/KV. 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 PACKAGED_DOCS_DIR = path4.resolve(path4.dirname(fileURLToPath2(import.meta.url)), "..", "docs");
-if (!existsSync4(PACKAGED_DOCS_DIR)) {
+var PACKAGED_DOCS_DIR = path5.resolve(path5.dirname(fileURLToPath3(import.meta.url)), "..", "docs");
+if (!existsSync5(PACKAGED_DOCS_DIR)) {
   throw new Error(`FreeLlama MCP documentation is missing at ${PACKAGED_DOCS_DIR}; run the package build.`);
 }
-var packagedDocs = readdirSync(PACKAGED_DOCS_DIR).filter((name) => name.endsWith(".md")).sort();
+var PACKAGED_DEV_DOCS_DIR = path5.join(PACKAGED_DOCS_DIR, "dev");
+var packagedDocs = [
+  ...readdirSync(PACKAGED_DOCS_DIR).filter((name) => name.endsWith(".md")).sort(),
+  ...existsSync5(PACKAGED_DEV_DOCS_DIR) ? readdirSync(PACKAGED_DEV_DOCS_DIR).filter((name) => name.endsWith(".md")).sort().map((name) => `dev/${name}`) : []
+];
+var packagedDocNames = packagedDocs.map((name) => path5.basename(name));
+if (new Set(packagedDocNames).size !== packagedDocNames.length) {
+  throw new Error(`FreeLlama MCP documentation has duplicate names across docs/ and docs/dev/: ${packagedDocs.join(", ")}`);
+}
 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$/, "")}`;
+  const base = path5.basename(name);
+  const uri = `freellama://docs/${base === "INDEX.md" ? "index" : base.replace(/\.md$/, "")}`;
   server.registerResource(
-    `freellama-docs-${name.toLowerCase().replace(/\.md$/, "")}`,
+    `freellama-docs-${base.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") }]
+      contents: [{ uri, mimeType: "text/markdown", text: await readFile(path5.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: {
+    description: "Use when: diagnosis/status/usage. Do not use when: choosing models. Returns: summary; opt into detail.",
+    inputSchema: external_exports.object({
       endpoint: ollamaEndpointParam,
       serveEndpoint: endpointParam,
-      view: external_exports.enum(["summary", "scheduler", "config", "full"]).optional().describe("summary default; scheduler/config/full are verbose")
-    },
+      view: external_exports.enum(["summary", "scheduler", "config", "full", "status", "usage"]).optional(),
+      days: external_exports.number().int().min(1).max(366).optional()
+    }).strict(),
     outputSchema: doctorResultSchema,
     annotations: { readOnlyHint: true }
   },
-  async ({ endpoint, serveEndpoint, view }) => {
+  async ({ endpoint, 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(endpoint));
-      if (!("structuredContent" in report)) return report;
+      if (!isStructuredSuccess(report)) return report;
       try {
         report.structuredContent.machine = JSON.parse(await machine(serveEndpoint));
         delete report.structuredContent.machine_unavailable;
@@ -22224,23 +23007,24 @@ server.registerTool(
 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"),
+    description: "Use when: affinity/session kill. Do not use when: job cancel/history/unload. 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: false }
+    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."));
-        return structuredResult(JSON.parse(await createSession(endpoint)));
+        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: delete requires sessionId."));
-      await deleteSession(endpoint, sessionId);
+      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);
@@ -22250,25 +23034,33 @@ server.registerTool(
 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"),
+    description: "Use when: inventory/library. Do not use when: execution/mutation. Returns: models; library finds families, 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 }
   },
-  async ({ view, model, includeVerbose, query, capabilities, order, limit, cursor, endpoint, ollamaEndpoint }) => {
+  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".'));
+      }
+      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 (selectedView !== "library" && [query, capabilities, order].some((value) => value !== void 0)) {
         return errorResult(new Error(`view "${selectedView}" does not accept library search fields.`));
       }
@@ -22294,11 +23086,11 @@ server.registerTool(
           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 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 size = typeof entry.size === "number" ? entry.size : null;
+            const size = typeof entry.resident_size === "number" ? entry.resident_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);
@@ -22331,6 +23123,17 @@ server.registerTool(
             method: "POST",
             body: { model }
           });
+          let local = data;
+          if (includeLibrary && !data.remote_host && !data.remote_model && !/-cloud$/i.test(model)) {
+            try {
+              const inventory = await ollamaFetch(ollamaEndpoint, "/api/tags");
+              const canonical = libraryReference(model)?.tag;
+              const entry = canonical && inventory.models?.find((entry2) => libraryReference(String(entry2.name ?? entry2.model))?.tag === canonical);
+              if (entry) local = { ...entry, ...data, digest: entry.digest };
+            } catch {
+            }
+          }
+          const library = includeLibrary && !local.remote_host && !local.remote_model && !/-cloud$/i.test(model) ? await ollamaLibrary.lookup(model) : void 0;
           const { license, modelfile, ...rest } = data;
           const modelInfo = rest.model_info ?? {};
           const contextEntry = Object.entries(modelInfo).find(
@@ -22339,13 +23142,23 @@ server.registerTool(
           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, endpoint, ollamaEndpoint });
-        default:
-          return parsedResult(await listModels(endpoint));
+          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);
@@ -22359,20 +23172,10 @@ async function libraryLookup({
   order,
   limit,
   cursor,
+  includeReadme,
   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 ${response.status}`);
-      return await response.text();
-    } finally {
-      clearTimeout(timer);
-    }
-  };
   const localState = async () => {
     let installed = /* @__PURE__ */ new Set();
     let memoryBytes = null;
@@ -22390,9 +23193,19 @@ async function libraryLookup({
   };
   try {
     if (model) {
-      const family = model.split(":")[0];
-      const familyPage = await fetchPage(`https://ollama.com/library/${encodeURIComponent(family)}`);
-      const { tags } = parseModelTags(familyPage, family);
+      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, memoryBytes } = await localState();
       const budget = memoryBytes ? memoryBytes * 0.6 : null;
       const annotated = tags.map((t) => ({
@@ -22401,15 +23214,23 @@ async function libraryLookup({
         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/${encodeURIComponent(family)}`,
+        sourcePage: ref.url,
+        library: {
+          status: library.status,
+          sourceUrl: library.sourceUrl,
+          tagsSourceUrl: library.tagsSourceUrl,
+          fetchedAt: library.fetchedAt,
+          expiresAt: library.expiresAt,
+          cached: library.cached,
+          ...publicModelGuidance(library, void 0, includeReadme)
+        },
         ...tags.length === 0 ? {
           tagsUnavailable: `No pullable tags found for "${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.`
         } : {},
@@ -22418,29 +23239,22 @@ async function libraryLookup({
         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 search = await ollamaLibrary.search(params);
+    const url = search.url;
+    const parsed = search.models.slice(0, limit ?? 10);
     const { installed } = await localState();
     const installedFamilies = new Set([...installed].map((n) => n.split(":")[0]));
     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,
       nextStep: 'Not pullable yet. Call again with view:"library", model:"<name>" to get tags, sizes and memory fit.',
@@ -22450,42 +23264,47 @@ async function libraryLookup({
     return errorResult(error2);
   }
 }
+registerContextTools(server);
 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: {
+    description: "Use when: supplied chat/tools/embeddings. Do not use when: files. Returns: response/receipts.",
+    inputSchema: external_exports.object({
       endpoint: endpointParam,
-      task: taskParam,
+      task: taskParam.removeDefault().optional().describe("caller owns prompts and output format."),
       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."),
+      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().describe("Chat input when messages is omitted."),
+      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; 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"
+      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().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."),
+      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().describe("Requires logprobs:true"),
+      topLogprobs: external_exports.number().int().nonnegative().optional(),
       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"),
+      priority: external_exports.enum(["interactive", "normal", "background"]).optional(),
+      maxWaitSeconds: external_exports.number().int().positive().optional(),
+      timeoutSeconds: external_exports.number().int().positive().optional(),
+      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"
       )
-    },
+    }).strict(),
     outputSchema: taskResultSchema,
     annotations: { destructiveHint: false }
   },
@@ -22495,11 +23314,14 @@ server.registerTool(
     objective,
     model,
     sessionId,
+    scopeId,
+    scopeRevision,
     contextTokens,
     executionPreference,
     minPlacementEvidence,
     requiredCapabilities,
     prompt,
+    systemPrompt,
     images,
     messages,
     input,
@@ -22512,14 +23334,23 @@ server.registerTool(
     topLogprobs,
     minConfidence,
     priority,
+    maxWaitSeconds,
+    timeoutSeconds,
+    defer,
     returnEmbeddings,
     preview
   }) => {
     try {
-      const canonicalTask = canonicalTaskKind(task);
+      const canonicalTask = canonicalTaskKind(task ?? "completion");
+      if (scopeId === void 0 !== (scopeRevision === void 0)) {
+        return errorResult(new Error("scopeId and scopeRevision must be supplied together."));
+      }
       if (preview) {
         const executionOnlyFields = [
+          ["scopeId", scopeId],
+          ["scopeRevision", scopeRevision],
           ["prompt", prompt],
+          ["systemPrompt", systemPrompt],
           ["images", images],
           ["messages", messages],
           ["input", input],
@@ -22530,7 +23361,11 @@ server.registerTool(
           ["options", options],
           ["logprobs", logprobs],
           ["topLogprobs", topLogprobs],
-          ["returnEmbeddings", returnEmbeddings]
+          ["returnEmbeddings", returnEmbeddings],
+          ["maxWaitSeconds", maxWaitSeconds],
+          ["priority", priority],
+          ["timeoutSeconds", timeoutSeconds],
+          ["defer", defer]
         ].filter(([, value]) => value !== void 0).map(([name]) => `\`${name}\``);
         if (executionOnlyFields.length > 0) {
           return errorResult(
@@ -22546,7 +23381,7 @@ server.registerTool(
       }
       if (!preview && task === "embedding") {
         if (input === void 0) return errorResult(new Error('task "embedding" requires `input`.'));
-        if ([prompt, images, messages, tools, format, think, logprobs, topLogprobs].some(
+        if ([prompt, systemPrompt, images, messages, tools, format, think, logprobs, topLogprobs].some(
           (value) => value !== void 0
         )) {
           return errorResult(
@@ -22558,10 +23393,10 @@ server.registerTool(
         if (input !== void 0 || returnEmbeddings !== void 0) {
           return errorResult(new Error('`input` and `returnEmbeddings` are valid only for task "embedding".'));
         }
-        if (prompt === void 0 && (messages === void 0 || messages.length === 0)) {
-          return errorResult(new Error(`task "${task}" requires \`prompt\` or \`messages\`.`));
+        if (prompt === void 0 && systemPrompt === void 0 && (messages === void 0 || messages.length === 0)) {
+          return errorResult(new Error(`task "${task}" requires \`prompt\`, \`systemPrompt\`, or \`messages\`.`));
         }
-        if (images !== void 0 && messages !== void 0) {
+        if (images !== void 0 && (messages !== void 0 || prompt === void 0)) {
           return errorResult(
             new Error("Top-level `images` is valid only with `prompt`; put images inside messages instead.")
           );
@@ -22572,44 +23407,48 @@ server.registerTool(
       }
       if (preview) {
         const result2 = parsedResult(
-          await route(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
+          await route2(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
         );
-        if ("structuredContent" in result2) {
+        if (isStructuredSuccess(result2)) {
           const refusal = belowConfidence(result2.structuredContent, minConfidence);
           if (refusal) return refusal;
         }
         return result2;
       }
-      if (minConfidence) {
+      if (minConfidence && scopeId === void 0) {
         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)
+          await route2(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 (!isStructuredSuccess(decision)) return decision;
+        const refusal = belowConfidence(decision.structuredContent, minConfidence);
+        if (refusal) return refusal;
       }
       const result = parsedResult(
-        await runTaskRequest(endpoint ?? DEFAULT_SERVE_ENDPOINT, {
-          task: canonicalTask,
-          objective: objective ?? "balanced",
+        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 ?? [],
+          required_capabilities: effectiveRequiredCapabilities,
           prompt,
           images,
-          messages: messages ?? [],
+          messages: taskMessages({ systemPrompt, messages, prompt, images }),
           input,
           tools,
           keep_alive: keepAlive,
           min_confidence: minConfidence,
           priority: priority ?? "normal",
-          execution_preference: executionPreference ?? "auto",
-          min_placement_evidence: minPlacementEvidence ?? "configured",
+          max_wait_seconds: maxWaitSeconds,
+          timeout_seconds: timeoutSeconds,
+          defer: defer ?? false,
+          execution_preference: executionPreference,
+          min_placement_evidence: minPlacementEvidence,
           request_options: {
             format,
             think,
@@ -22619,7 +23458,8 @@ server.registerTool(
           }
         })
       );
-      if (!returnEmbeddings && "structuredContent" in result) {
+      if (isStructuredSuccess(result) && result.structuredContent.deferred === true) return result;
+      if (!returnEmbeddings && isStructuredSuccess(result)) {
         const trimmed = summarizeEmbeddings(result.structuredContent);
         if (trimmed) return withTaskTelemetry(structuredResult(trimmed));
       }
@@ -22629,15 +23469,50 @@ server.registerTool(
     }
   }
 );
+server.registerTool(
+  "task_jobs",
+  {
+    description: "Use when: task jobs. Do not use when: new work/session kill. Returns: 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);
+    }
+  }
+);
 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: {
+    description: "Use when: independent:true work. Do not use when: dependencies. Returns: ordered receipts; maxParallelism caps dispatch.",
+    inputSchema: external_exports.object({
       tasks: external_exports.array(batchItemParam).min(1).max(64),
       maxParallelism: external_exports.number().int().positive().max(64).optional(),
       endpoint: endpointParam
-    },
+    }).strict(),
     outputSchema: batchResultSchema,
     annotations: { destructiveHint: false }
   },
@@ -22645,36 +23520,46 @@ server.registerTool(
     try {
       for (const item of tasks) {
         const task = item.task;
+        if (task.scopeId === void 0 !== (task.scopeRevision === void 0)) {
+          return errorResult(new Error(`batch item ${item.id}: scopeId and scopeRevision must be supplied together.`));
+        }
         if (task.task === "embedding") {
-          if (task.input === void 0 || task.prompt !== void 0 || task.messages !== void 0 || task.tools !== void 0) {
+          if (task.input === void 0 || task.prompt !== void 0 || task.systemPrompt !== 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.`));
           }
-        } 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.`));
+        } else if (task.input !== void 0 || task.prompt === void 0 && task.systemPrompt === void 0 && task.messages === void 0) {
+          return errorResult(new Error(`batch item ${item.id}: chat work requires prompt, systemPrompt, or messages and accepts no input.`));
+        }
+        if (task.images !== void 0 && (task.messages !== void 0 || task.prompt === void 0)) {
+          return errorResult(new Error(`batch item ${item.id}: top-level images requires prompt without messages; put images inside messages instead.`));
         }
         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, {
+      return withBatchTelemetry(parsedResult(await runTaskBatchRequest2(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",
+              task: task.task === "code_review" ? "coding" : task.task ?? (task.scopeId === void 0 ? "completion" : void 0),
+              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 ?? "auto",
-              min_placement_evidence: task.minPlacementEvidence ?? "configured",
-              required_capabilities: task.requiredCapabilities ?? [],
+              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,
               prompt: task.prompt,
               images: task.images,
-              messages: task.messages ?? [],
+              messages: taskMessages(task),
               input: task.input,
               tools: task.tools,
               keep_alive: task.keepAlive,
@@ -22699,24 +23584,28 @@ server.registerTool(
 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'),
+    description: "Use when: approved pull/stop. Do not use when: delete/discovery. 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().describe(`"pull" only. Defaults to ${DEFAULT_PULL_TIMEOUT_SECONDS}s.`)
-    },
+      timeoutSeconds: external_exports.number().int().positive().optional()
+    }).strict(),
     outputSchema: manageResultSchema,
     annotations: { destructiveHint: false }
   },
-  async ({ action, model, ollamaEndpoint, timeoutSeconds }) => {
+  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) : await ollamaFetch(ollamaEndpoint, "/api/generate", {
+      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 }
+        body: { model, keep_alive: 0 },
+        signal: extra.signal
       });
       const payload = data;
       if (typeof payload.error === "string" && payload.error) {
@@ -22731,11 +23620,11 @@ server.registerTool(
 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: {
+    description: "DESTRUCTIVE AND IRREVERSIBLE. Use when: human names exact tag. Do not use when: inferred cleanup. Returns: deleted tag.",
+    inputSchema: external_exports.object({
       model: external_exports.string().min(1),
       ollamaEndpoint: ollamaEndpointParam
-    },
+    }).strict(),
     outputSchema: deleteResultSchema,
     annotations: { destructiveHint: true }
   },
@@ -22743,7 +23632,7 @@ server.registerTool(
     try {
       await ollamaFetch(ollamaEndpoint, "/api/delete", {
         method: "DELETE",
-        body: { name: model }
+        body: { model }
       });
       return structuredResult({ deleted: model });
     } catch (error2) {
@@ -22751,57 +23640,54 @@ server.registerTool(
     }
   }
 );
+async function routedResearchModel(endpoint, executionPreference, minPlacementEvidence) {
+  const decision = JSON.parse(
+    await route2(endpoint, "coding", "balanced", void 0, void 0, void 0, [], void 0, executionPreference, minPlacementEvidence)
+  );
+  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;
+}
 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"),
+    description: "Use when: workspace lookup. Do not use when: mutation/external facts. 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,
       executionPreference: executionPreferenceParam,
       minPlacementEvidence: minPlacementEvidenceParam,
-      legacyText: external_exports.boolean().optional().describe("false default: compact text cue; true: legacy serialized JSON text"),
+      legacyText: external_exports.boolean().optional(),
       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."),
+        contextTokens: external_exports.number().int().positive().optional(),
+        outputTokens: external_exports.number().int().positive().optional(),
         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"'),
+        keepAlive: external_exports.string().min(1).optional(),
         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.")
-    },
+        toolTimeoutSeconds: external_exports.number().positive().optional()
+      }).strict().optional()
+    }).strict(),
     outputSchema: researchResultSchema,
     annotations: { destructiveHint: false }
   },
-  async ({ question, workspacePath, adapter, model, endpoint, executionPreference, minPlacementEvidence, legacyText, agent }) => {
+  async ({ question, workspacePath, adapter, model, endpoint, executionPreference, minPlacementEvidence, legacyText, agent }, extra) => {
     const chosenAdapter = adapter ?? DEFAULT_RESEARCH_ADAPTER;
-    const chosenModel = model ?? DEFAULT_DELEGATE_MODEL;
     let resolvedWorkspace;
+    let chosenModel;
     try {
       resolvedWorkspace = await assertAllowedWorkspace(workspacePath);
+      chosenModel = model ?? DEFAULT_DELEGATE_MODEL ?? await routedResearchModel(endpoint, executionPreference, minPlacementEvidence);
     } catch (error2) {
-      return errorResult(error2);
+      return researchErrorResult(error2);
     }
     const known = MODEL_EVIDENCE[chosenModel];
     if (known?.grade === "unusable") {
@@ -22816,23 +23702,24 @@ server.registerTool(
         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.`
       });
     }
-    const dir = await mkdtemp(path4.join(tmpdir(), "freellama-delegate-"));
-    const promptFile = path4.join(dir, "prompt.md");
-    const resultFile = path4.join(dir, "result.json");
+    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 (!existsSync4(adapter2)) {
-        return errorResult(
+      if (!existsSync5(adapter2)) {
+        return researchErrorResult(
           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], {
+        signal: extra.signal,
         env: {
-          ...process.env,
+          ...delegateEnvironment(),
           FREELLAMA_TARGET_MODEL: chosenModel,
           FREELLAMA_AGENT_MANAGED_ENDPOINT: endpoint ?? DEFAULT_SERVE_ENDPOINT,
           FREELLAMA_AGENT_EXECUTION_PREFERENCE: executionPreference ?? "auto",
@@ -22849,20 +23736,7 @@ server.registerTool(
           ...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 } : {}
+          ...agent?.toolTimeoutSeconds !== void 0 ? { FREELLAMA_AGENT_TOOL_TIMEOUT_SECONDS: String(agent.toolTimeoutSeconds) } : {}
         },
         timeout: DEFAULT_DELEGATE_TIMEOUT_SECONDS * 1e3,
         // The answer is read from `resultFile`, never from stdout — but execFile's default 1 MB
@@ -22888,7 +23762,7 @@ server.registerTool(
       try {
         result = parseAdapterResult(await readFile(resultFile, "utf8"));
       } catch {
-        return errorResult(
+        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.`
           )
@@ -22905,7 +23779,7 @@ server.registerTool(
           // 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,
+          path: target ? path5.relative(resolvedWorkspace, target) : detail ? extractExistingWorkspacePath(detail, resolvedWorkspace) : null,
           detail,
           detail_truncated: rawDetail !== detail
         };
@@ -22915,21 +23789,16 @@ server.registerTool(
       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
-      }));
+      const citations = succeeded.map((step) => ({ step: step.step, tool: step.tool, path: step.path }));
       if (adapterError) {
-        return errorResult(
-          new Error(
-            `research adapter failed: ${result.final_answer}` + (evidenceText ? `
+        const diagnostic = `research adapter failed: ${result.final_answer}` + (evidenceText ? `
 Evidence collected before the failure:
-${evidenceText}` : "")
-          )
-        );
+${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.`;
@@ -22959,9 +23828,15 @@ ${evidenceText}` : "")
         evidence,
         summary
       };
-      return structuredResult(payload, { legacyJson: legacyText === true });
+      const cited = [...new Set(citations.map((citation) => citation.path).filter(Boolean))];
+      const answerText = `${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, legacyText === true ? { legacyJson: true } : { text: answerText });
     } catch (error2) {
-      return errorResult(error2);
+      return researchErrorResult(error2);
     } finally {
       await rm(dir, { recursive: true, force: true });
     }