@octocodeai/freellama-mcp-server 0.5.0 → 0.6.1

package/dist/index.js

@@ -3837,7 +3837,7 @@ var require_fast_uri = __commonJS({
         normalizeString(uri, options);
       } else if (typeof uri === "object") {
         uri = /** @type {T} */
-        parse3(serialize2(uri, options), options);
+        parse3(serialize(uri, options), options);
       }
       return uri;
     }
@@ -3872,13 +3872,13 @@ var require_fast_uri = __commonJS({
         throw new Error(resolved.error);
       }
       schemelessOptions.skipEscape = true;
-      return serialize2(resolved, schemelessOptions);
+      return serialize(resolved, schemelessOptions);
     }
     function resolveComponent(base, relative, options, skipNormalization) {
       const target = {};
       if (!skipNormalization) {
-        base = parse3(serialize2(base, options), options);
-        relative = parse3(serialize2(relative, options), options);
+        base = parse3(serialize(base, options), options);
+        relative = parse3(serialize(relative, options), options);
       }
       options = options || {};
       if (!options.tolerant && relative.scheme) {
@@ -3932,7 +3932,7 @@ var require_fast_uri = __commonJS({
       const normalizedB = normalizeComparableURI(uriB, options);
       return normalizedA !== void 0 && normalizedB !== void 0 && normalizedA === normalizedB;
     }
-    function serialize2(cmpts, opts) {
+    function serialize(cmpts, opts) {
       const component = {
         host: cmpts.host,
         scheme: cmpts.scheme,
@@ -4186,7 +4186,7 @@ var require_fast_uri = __commonJS({
     function normalizeStringWithStatus(uri, opts) {
       const { parsed, malformedAuthorityOrPort, malformedPercentEncoding, malformedSchemeSpecific, malformedHost, malformedScheme } = parseWithStatus(uri, opts);
       return {
-        normalized: malformedAuthorityOrPort || malformedPercentEncoding || malformedSchemeSpecific || malformedHost || malformedScheme ? uri : serialize2(parsed, opts),
+        normalized: malformedAuthorityOrPort || malformedPercentEncoding || malformedSchemeSpecific || malformedHost || malformedScheme ? uri : serialize(parsed, opts),
         malformedAuthorityOrPort,
         malformedPercentEncoding,
         malformedSchemeSpecific,
@@ -4200,7 +4200,7 @@ var require_fast_uri = __commonJS({
       }
       let value;
       try {
-        value = typeof uri === "string" ? uri : serialize2(uri, opts);
+        value = typeof uri === "string" ? uri : serialize(uri, opts);
       } catch {
         return void 0;
       }
@@ -4213,7 +4213,7 @@ var require_fast_uri = __commonJS({
       resolve,
       resolveComponent,
       equal,
-      serialize: serialize2,
+      serialize,
       parse: parse3
     };
     module.exports = fastUri;
@@ -7199,16 +7199,6 @@ var require_dist = __commonJS({
   }
 });
 
-// src/index.ts
-import { execFile } from "node:child_process";
-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 path5 from "node:path";
-import { fileURLToPath as fileURLToPath3 } from "node:url";
-import { promisify } from "node:util";
-
 // ../../node_modules/zod/v3/external.js
 var external_exports = {};
 __export(external_exports, {
@@ -7700,15 +7690,15 @@ var makeIssue = (params) => {
       message: issueData.message
     };
   }
-  let errorMessage = "";
+  let errorMessage2 = "";
   const maps = errorMaps.filter((m) => !!m).slice().reverse();
   for (const map of maps) {
-    errorMessage = map(fullIssue, { data, defaultError: errorMessage }).message;
+    errorMessage2 = map(fullIssue, { data, defaultError: errorMessage2 }).message;
   }
   return {
     ...issueData,
     path: fullPath,
-    message: errorMessage
+    message: errorMessage2
   };
 };
 var EMPTY_PATH = [];
@@ -17546,19 +17536,19 @@ var getRefs = (options) => {
 };
 
 // ../../node_modules/zod-to-json-schema/dist/esm/errorMessages.js
-function addErrorMessage(res, key, errorMessage, refs) {
+function addErrorMessage(res, key, errorMessage2, refs) {
   if (!refs?.errorMessages)
     return;
-  if (errorMessage) {
+  if (errorMessage2) {
     res.errorMessage = {
       ...res.errorMessage,
-      [key]: errorMessage
+      [key]: errorMessage2
     };
   }
 }
-function setResponseValueAndErrors(res, key, value, errorMessage, refs) {
+function setResponseValueAndErrors(res, key, value, errorMessage2, refs) {
   res[key] = value;
-  addErrorMessage(res, key, errorMessage, refs);
+  addErrorMessage(res, key, errorMessage2, refs);
 }
 
 // ../../node_modules/zod-to-json-schema/dist/esm/getRelativePath.js
@@ -18869,8 +18859,8 @@ var Protocol = class {
                   if (queuedMessage.type === "response") {
                     resolver(message);
                   } else {
-                    const errorMessage = message;
-                    const error2 = new McpError(errorMessage.error.code, errorMessage.error.message, errorMessage.error.data);
+                    const errorMessage2 = message;
+                    const error2 = new McpError(errorMessage2.error.code, errorMessage2.error.message, errorMessage2.error.data);
                     resolver(error2);
                   }
                 } else {
@@ -20161,23 +20151,23 @@ var Server = class extends Protocol {
       const wrappedHandler = async (request, extra) => {
         const validatedRequest = safeParse2(CallToolRequestSchema, request);
         if (!validatedRequest.success) {
-          const errorMessage = validatedRequest.error instanceof Error ? validatedRequest.error.message : String(validatedRequest.error);
-          throw new McpError(ErrorCode.InvalidParams, `Invalid tools/call request: ${errorMessage}`);
+          const errorMessage2 = validatedRequest.error instanceof Error ? validatedRequest.error.message : String(validatedRequest.error);
+          throw new McpError(ErrorCode.InvalidParams, `Invalid tools/call request: ${errorMessage2}`);
         }
         const { params } = validatedRequest.data;
         const result = await Promise.resolve(handler(request, extra));
         if (params.task) {
           const taskValidationResult = safeParse2(CreateTaskResultSchema, result);
           if (!taskValidationResult.success) {
-            const errorMessage = taskValidationResult.error instanceof Error ? taskValidationResult.error.message : String(taskValidationResult.error);
-            throw new McpError(ErrorCode.InvalidParams, `Invalid task creation result: ${errorMessage}`);
+            const errorMessage2 = taskValidationResult.error instanceof Error ? taskValidationResult.error.message : String(taskValidationResult.error);
+            throw new McpError(ErrorCode.InvalidParams, `Invalid task creation result: ${errorMessage2}`);
           }
           return taskValidationResult.data;
         }
         const validationResult = safeParse2(CallToolResultSchema, result);
         if (!validationResult.success) {
-          const errorMessage = validationResult.error instanceof Error ? validationResult.error.message : String(validationResult.error);
-          throw new McpError(ErrorCode.InvalidParams, `Invalid tools/call result: ${errorMessage}`);
+          const errorMessage2 = validationResult.error instanceof Error ? validationResult.error.message : String(validationResult.error);
+          throw new McpError(ErrorCode.InvalidParams, `Invalid tools/call result: ${errorMessage2}`);
         }
         return validationResult.data;
       };
@@ -20671,12 +20661,12 @@ var McpServer = class {
    * @param errorMessage - The error message.
    * @returns The tool error result.
    */
-  createToolError(errorMessage) {
+  createToolError(errorMessage2) {
     return {
       content: [
         {
           type: "text",
-          text: errorMessage
+          text: errorMessage2
         }
       ],
       isError: true
@@ -20694,8 +20684,8 @@ var McpServer = class {
     const parseResult = await safeParseAsync2(schemaToParse, args);
     if (!parseResult.success) {
       const error2 = "error" in parseResult ? parseResult.error : "Unknown error";
-      const errorMessage = getParseErrorMessage(error2);
-      throw new McpError(ErrorCode.InvalidParams, `Input validation error: Invalid arguments for tool ${toolName}: ${errorMessage}`);
+      const errorMessage2 = getParseErrorMessage(error2);
+      throw new McpError(ErrorCode.InvalidParams, `Input validation error: Invalid arguments for tool ${toolName}: ${errorMessage2}`);
     }
     return parseResult.data;
   }
@@ -20719,8 +20709,8 @@ var McpServer = class {
     const parseResult = await safeParseAsync2(outputObj, result.structuredContent);
     if (!parseResult.success) {
       const error2 = "error" in parseResult ? parseResult.error : "Unknown error";
-      const errorMessage = getParseErrorMessage(error2);
-      throw new McpError(ErrorCode.InvalidParams, `Output validation error: Invalid structured content for tool ${toolName}: ${errorMessage}`);
+      const errorMessage2 = getParseErrorMessage(error2);
+      throw new McpError(ErrorCode.InvalidParams, `Output validation error: Invalid structured content for tool ${toolName}: ${errorMessage2}`);
     }
   }
   /**
@@ -20932,8 +20922,8 @@ var McpServer = class {
         const parseResult = await safeParseAsync2(argsObj, request.params.arguments);
         if (!parseResult.success) {
           const error2 = "error" in parseResult ? parseResult.error : "Unknown error";
-          const errorMessage = getParseErrorMessage(error2);
-          throw new McpError(ErrorCode.InvalidParams, `Invalid arguments for prompt ${request.params.name}: ${errorMessage}`);
+          const errorMessage2 = getParseErrorMessage(error2);
+          throw new McpError(ErrorCode.InvalidParams, `Invalid arguments for prompt ${request.params.name}: ${errorMessage2}`);
         }
         const args = parseResult.data;
         const cb = prompt.callback;
@@ -21443,33 +21433,47 @@ var StdioServerTransport = class {
   }
 };
 
+// src/native.ts
+import { createRequire } from "node:module";
+import native from "../native/index.js";
+var { createScope, getScope, forkScope, deleteScope, warmModelRequest, doctor, machine, health, status, usage, createSession, deleteSession, killSession, listModels, route, runTaskRequest, runTaskBatchRequest, listTaskJobs, getTaskJob, cancelTaskJob, removeTaskJob } = native;
+var { version: SERVER_VERSION } = createRequire(import.meta.url)("../package.json");
+
+// src/lifecycle.ts
+import { spawn as spawn2 } from "node:child_process";
+
+// src/serve.ts
+import { spawn } from "node:child_process";
+import { existsSync as existsSync2 } from "node:fs";
+import { createRequire as createRequire2 } from "node:module";
+import path2 from "node:path";
+import { fileURLToPath as fileURLToPath2 } from "node:url";
+
 // src/config.ts
 import { existsSync } from "node:fs";
 import { realpath } from "node:fs/promises";
 import path from "node:path";
 import { fileURLToPath } from "node:url";
 import { homedir } from "node:os";
+var PACKAGE_ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
 function findRepoRoot() {
-  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")) && path.join(dir, "packages", "mcp") === packageRoot) return dir;
+    if (existsSync(path.join(dir, "Cargo.toml")) && path.join(dir, "packages", "mcp") === PACKAGE_ROOT) return dir;
     const parent = path.dirname(dir);
     if (parent === dir) break;
     dir = parent;
   }
   return void 0;
 }
-var REPO_ROOT = findRepoRoot() ?? path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
-var PACKAGE_ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
+var CHECKOUT_ROOT = findRepoRoot();
+var REPO_ROOT = CHECKOUT_ROOT ?? PACKAGE_ROOT;
 function adapterPath(file) {
-  const bundled = path.join(PACKAGE_ROOT, "adapters", file);
-  if (existsSync(bundled)) return bundled;
-  return path.join(REPO_ROOT, "benchmark/local/scripts", file);
+  return path.join(PACKAGE_ROOT, "dist", "research", file);
 }
 var RESEARCH_ADAPTERS = {
-  bash: adapterPath("bash_agent.py"),
-  octocode: adapterPath("octocode_agent.py")
+  bash: adapterPath("bash-agent.js"),
+  octocode: adapterPath("octocode-agent.js")
 };
 var DEFAULT_RESEARCH_ADAPTER = process.env.FREELLAMA_MCP_DEFAULT_ADAPTER === "octocode" ? "octocode" : "bash";
 function envInt(name, fallback) {
@@ -21494,7 +21498,7 @@ function parseAllowedResearchRoots(raw, delimiter = path.delimiter) {
   return raw.split(delimiter).filter(Boolean).map((root) => path.resolve(root));
 }
 var ALLOWED_RESEARCH_ROOTS = parseAllowedResearchRoots(
-  process.env.FREELLAMA_MCP_ALLOWED_ROOTS ?? findRepoRoot() ?? ""
+  process.env.FREELLAMA_MCP_ALLOWED_ROOTS ?? CHECKOUT_ROOT ?? ""
 );
 var resolvedRootsPromise = null;
 function allowedResearchRoots() {
@@ -21566,8 +21570,7 @@ var DELEGATE_ENV_ALLOWLIST = /* @__PURE__ */ new Set([
   "REQUESTS_CA_BUNDLE",
   "XDG_CACHE_HOME",
   "XDG_DATA_HOME",
-  "npm_config_cache",
-  "PYTHONPATH"
+  "npm_config_cache"
 ]);
 function delegateEnvironment(source = process.env) {
   const env = {};
@@ -21577,18 +21580,7 @@ function delegateEnvironment(source = process.env) {
   return env;
 }
 
-// src/native.ts
-import { createRequire } from "node:module";
-import native from "../native/index.js";
-var { createScope, getScope, forkScope, deleteScope, warmModelRequest, doctor, machine, health, status, usage, createSession, deleteSession, killSession, listModels, route, runTask, runTaskRequest, runTaskBatchRequest, listTaskJobs, getTaskJob, cancelTaskJob, removeTaskJob } = native;
-var { version: SERVER_VERSION } = createRequire(import.meta.url)("../package.json");
-
 // 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;
@@ -21637,6 +21629,12 @@ function serveBinaryCandidates() {
   if (process.env.FREELLAMA_SERVE_BINARY) candidates.unshift(process.env.FREELLAMA_SERVE_BINARY);
   return candidates;
 }
+function startServeInBackground() {
+  ensureServe(void 0).catch((error2) => {
+    process.stderr.write(`[freellama serve] background start failed: ${error2 instanceof Error ? error2.message : String(error2)}
+`);
+  });
+}
 function stopAutostartedServe() {
   if (child && child.exitCode === null) child.kill("SIGTERM");
   child = null;
@@ -21684,35 +21682,159 @@ async function ensureServe(endpoint) {
   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 {
-    root = realpathSync(workspace);
-  } catch {
-    return null;
-  }
-  const tokens = command.match(/(?:[^\s"'\\]+|"(?:\\.|[^"])*"|'[^']*')+/g) ?? [];
-  const patternFlags = /* @__PURE__ */ new Set(["-name", "-iname", "-path", "-ipath", "-regex", "-iregex", "-wholename"]);
-  for (let index = 0; index < tokens.length; index += 1) {
-    const raw = tokens[index] ?? "";
-    if (index > 0 && patternFlags.has(tokens[index - 1] ?? "")) continue;
-    const token = raw.replace(/^["']|["']$/g, "").replace(/[|;&)]+$/g, "").replace(/:\d+(?::\d+)?$/, "");
-    if (!token || token.startsWith("-") || /[*?$`]/.test(token)) continue;
-    const candidate = path3.isAbsolute(token) ? token : path3.resolve(root, token);
-    if (!existsSync3(candidate)) continue;
+// src/lifecycle.ts
+var liveDelegates = /* @__PURE__ */ new Set();
+var DETACH_DELEGATES = process.platform !== "win32";
+function killDelegateTree(child2) {
+  if (DETACH_DELEGATES && child2.pid !== void 0) {
     try {
-      const resolved = realpathSync(candidate);
-      const relative = path3.relative(root, resolved);
-      const withinRoot = relative && relative !== ".." && !relative.startsWith(`..${path3.sep}`);
-      if (withinRoot && statSync(resolved).isFile()) return relative;
+      process.kill(-child2.pid, "SIGKILL");
+      return;
     } catch {
     }
   }
-  return null;
+  try {
+    child2.kill("SIGKILL");
+  } catch {
+  }
+}
+function runDelegate(file, args, options) {
+  const child2 = spawn2(file, args, { env: options.env, detached: DETACH_DELEGATES, stdio: ["ignore", "pipe", "pipe"] });
+  const done = new Promise((resolve, reject) => {
+    const out = [];
+    const err = [];
+    let outLen = 0;
+    let errLen = 0;
+    let failure;
+    let timedOut = false;
+    let settled = false;
+    const fail = (error2) => {
+      if (failure === void 0) failure = error2;
+      killDelegateTree(child2);
+    };
+    const timer = setTimeout(() => {
+      timedOut = true;
+      killDelegateTree(child2);
+    }, options.timeoutMs);
+    const onAbort = () => {
+      const abort = new Error("The operation was aborted", { cause: options.signal?.reason });
+      abort.name = "AbortError";
+      abort.code = "ABORT_ERR";
+      fail(abort);
+    };
+    if (options.signal?.aborted) onAbort();
+    else options.signal?.addEventListener("abort", onAbort, { once: true });
+    const collect = (chunks, name) => (chunk) => {
+      const length = name === "stdout" ? outLen += chunk.length : errLen += chunk.length;
+      if (length > options.maxBuffer) {
+        const overflow = new RangeError(`${name} maxBuffer length exceeded`);
+        overflow.code = "ERR_CHILD_PROCESS_STDIO_MAXBUFFER";
+        fail(overflow);
+        return;
+      }
+      chunks.push(chunk);
+    };
+    child2.stdout.on("data", collect(out, "stdout"));
+    child2.stderr.on("data", collect(err, "stderr"));
+    const finish = (code, signal, spawnError) => {
+      if (settled) return;
+      settled = true;
+      clearTimeout(timer);
+      options.signal?.removeEventListener("abort", onAbort);
+      const stdout = Buffer.concat(out).toString("utf8");
+      const stderr = Buffer.concat(err).toString("utf8");
+      if (spawnError ?? failure) {
+        reject(spawnError ?? failure);
+        return;
+      }
+      if (code === 0 && signal === null) {
+        resolve({ stdout, stderr });
+        return;
+      }
+      const error2 = new Error(`Command failed: ${[file, ...args].join(" ")}
+${stderr}`);
+      Object.assign(error2, { code, killed: timedOut, signal, stdout, stderr });
+      reject(error2);
+    };
+    child2.once("error", (error2) => finish(null, null, error2));
+    child2.once("close", (code, signal) => finish(code, signal));
+  });
+  return { child: child2, done };
+}
+function killLiveDelegates() {
+  for (const child2 of liveDelegates) killDelegateTree(child2);
+  liveDelegates.clear();
+}
+function installProcessLifecycle() {
+  for (const stream of [process.stdout, process.stderr]) {
+    stream.on("error", (error2) => {
+      if (error2.code === "EPIPE") process.exit(0);
+      throw error2;
+    });
+  }
+  process.on("exit", () => {
+    killLiveDelegates();
+    stopAutostartedServe();
+  });
+  process.stdin.on("close", () => {
+    killLiveDelegates();
+    stopAutostartedServe();
+    process.exit(0);
+  });
+  for (const signal of ["SIGINT", "SIGTERM", "SIGHUP"]) {
+    process.on(signal, () => {
+      killLiveDelegates();
+      process.exit(0);
+    });
+  }
+}
+
+// src/schema-dialect.ts
+function stripSchemaDialect(message) {
+  const tools = message.result?.tools;
+  if (!Array.isArray(tools)) return message;
+  for (const tool of tools) {
+    delete tool.inputSchema?.$schema;
+    delete tool.outputSchema?.$schema;
+  }
+  return message;
+}
+function withoutSchemaDialect(transport2) {
+  const send = transport2.send.bind(transport2);
+  transport2.send = (message, options) => send(stripSchemaDialect(message), options);
+  return transport2;
+}
+
+// src/docs-resources.ts
+import { existsSync as existsSync3, readdirSync } from "node:fs";
+import { readFile } from "node:fs/promises";
+import path3 from "node:path";
+import { fileURLToPath as fileURLToPath3 } from "node:url";
+function registerDocsResources(server2) {
+  const PACKAGED_DOCS_DIR = path3.resolve(path3.dirname(fileURLToPath3(import.meta.url)), "..", "docs");
+  if (!existsSync3(path3.join(PACKAGED_DOCS_DIR, "README.md"))) {
+    throw new Error(`FreeLlama MCP documentation index is missing at ${PACKAGED_DOCS_DIR}/README.md; run the package build.`);
+  }
+  const packagedDocs = readdirSync(PACKAGED_DOCS_DIR).filter((name) => name.endsWith(".md")).sort();
+  for (const name of packagedDocs) {
+    const isIndex = name === "README.md";
+    const slug = isIndex ? "index" : name.replace(/\.md$/, "");
+    const uri = `freellama://docs/${slug}`;
+    server2.registerResource(
+      `freellama-docs-${slug.toLowerCase()}`,
+      uri,
+      {
+        mimeType: "text/markdown",
+        description: isIndex ? "Index of packaged FreeLlama operator and agent documentation. Read this first, then fetch one relevant document." : `Packaged FreeLlama documentation: ${name}.`
+      },
+      async () => ({
+        contents: [{ uri, mimeType: "text/markdown", text: await readFile(path3.join(PACKAGED_DOCS_DIR, name), "utf8") }]
+      })
+    );
+  }
 }
+
+// src/helpers.ts
 async function ollamaFetch(endpoint, path6, init = {}) {
   const base = (endpoint ?? DEFAULT_OLLAMA_ENDPOINT).replace(/\/$/, "");
   const controller = new AbortController();
@@ -21759,62 +21881,6 @@ async function readNdjson(body, onLine) {
   }
   return text + decoder.decode();
 }
-function summarizeOllamaPullStream(text) {
-  const events = [];
-  for (const line of text.split(/\r?\n/)) {
-    const trimmed = line.trim();
-    if (!trimmed) continue;
-    try {
-      const parsed = JSON.parse(trimmed);
-      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-        events.push(parsed);
-      }
-    } catch {
-    }
-  }
-  if (events.length === 0) {
-    if (!text.trim()) return { status: "empty" };
-    try {
-      const parsed = JSON.parse(text);
-      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-        return parsed;
-      }
-    } catch {
-      return { status: "unparsed", raw: clipText(text, 500) };
-    }
-    return { status: "unparsed", raw: clipText(text, 500) };
-  }
-  const last = events[events.length - 1] ?? {};
-  const withBytes = [...events].reverse().find((event) => typeof event.total === "number" && event.total > 0);
-  const completed = typeof withBytes?.completed === "number" ? withBytes.completed : null;
-  const total = typeof withBytes?.total === "number" ? withBytes.total : null;
-  const percent = completed != null && total != null && total > 0 ? Math.round(completed / total * 1e3) / 10 : null;
-  const status2 = typeof last.error === "string" && last.error ? "error" : last.status;
-  return {
-    ...last,
-    ...status2 === void 0 ? {} : { status: status2 },
-    progress: {
-      events: events.length,
-      completed,
-      total,
-      percent,
-      lastStatus: last.status ?? null
-    }
-  };
-}
-async function ollamaPull(endpoint, model, timeoutSeconds, options = {}) {
-  const raw = await ollamaFetch(endpoint, "/api/pull", {
-    method: "POST",
-    body: { model, stream: true },
-    timeoutMs: (timeoutSeconds ?? DEFAULT_PULL_TIMEOUT_SECONDS) * 1e3,
-    parse: "ndjson",
-    signal: options.signal,
-    onLine: options.onProgress
-  });
-  if (typeof raw === "string") return summarizeOllamaPullStream(raw);
-  if (raw && typeof raw === "object" && !Array.isArray(raw)) return raw;
-  return { status: "empty" };
-}
 var endpointParam = external_exports.string().min(1).optional();
 var ollamaEndpointParam = external_exports.string().min(1).optional();
 var TASK_KINDS = [
@@ -21835,19 +21901,7 @@ function canonicalTaskKind(task) {
 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();
-function belowConfidence(decision, minConfidence) {
-  if (!minConfidence) return null;
-  const actual = typeof decision.confidence === "string" ? decision.confidence : "low";
-  if ((CONFIDENCE_RANK[actual] ?? 0) >= (CONFIDENCE_RANK[minConfidence] ?? 0)) return null;
-  return errorResult(
-    new Error(
-      `Route refused: confidence is "${actual}" (evidence: ${decision.evidence}), below the requested minimum "${minConfidence}". Selected model would have been "${decision.selected_model}" for reasons [${decision.reasons?.join(", ")}].
-This is a fail-closed refusal, not a failure: the local router cannot justify this pick. Escalate to your own model, pass an explicit \`model\`, or run \`bench-all\` and configure a task policy to raise the evidence level.`
-    )
-  );
-}
 var REQUIRED_CAPABILITIES = [
   "completion",
   "tools",
@@ -21880,7 +21934,8 @@ function taskMessages({ systemPrompt, messages, prompt, images }) {
   return [{ role: "system", content: systemPrompt }, ...conversation];
 }
 var batchTaskParam = external_exports.object({
-  task: taskParam.removeDefault().optional().describe("caller owns prompts and output format."),
+  // Nested `task` shares the top-level description; repeating it re-sends the text every request.
+  task: taskParam.removeDefault().optional(),
   objective: objectiveParam,
   model: external_exports.string().min(1).optional(),
   sessionId: external_exports.string().uuid().optional(),
@@ -21893,6 +21948,7 @@ var batchTaskParam = external_exports.object({
   priority: external_exports.enum(["interactive", "normal", "background"]).optional(),
   maxWaitSeconds: external_exports.number().int().positive().optional(),
   timeoutSeconds: external_exports.number().int().positive().optional(),
+  stopOnRepetition: external_exports.boolean().optional().describe("Stream and stop once output repeats itself; keeps all generated text"),
   prompt: external_exports.string().min(1).optional(),
   systemPrompt: systemPromptParam,
   messages: messagesParam,
@@ -21907,56 +21963,27 @@ var batchTaskParam = external_exports.object({
   topLogprobs: external_exports.number().int().nonnegative().optional(),
   minConfidence: minConfidenceParam
 }).strict();
+function fresh(schema) {
+  const def = { ...schema._def };
+  for (const [key, value] of Object.entries(def)) {
+    if (value instanceof external_exports.ZodType) def[key] = fresh(value);
+    else if (Array.isArray(value) && value.length > 0 && value.every((v) => v instanceof external_exports.ZodType)) {
+      def[key] = value.map((v) => fresh(v));
+    }
+  }
+  if (schema instanceof external_exports.ZodObject) {
+    const shape = Object.fromEntries(Object.entries(schema.shape).map(([k, v]) => [k, fresh(v)]));
+    def.shape = () => shape;
+  }
+  return new schema.constructor(def);
+}
+var BATCH_TASK_FIELDS = new Set(Object.keys(batchTaskParam.shape));
 var batchItemParam = external_exports.object({
   id: external_exports.string().min(1),
   independent: external_exports.literal(true),
-  task: batchTaskParam
+  task: fresh(batchTaskParam)
 }).strict();
 var objectResultSchema = external_exports.object({}).passthrough();
-var doctorResultSchema = external_exports.object({
-  summary: external_exports.string().optional(),
-  endpoint: external_exports.string().optional(),
-  scheduler: external_exports.unknown().optional()
-}).passthrough();
-var sessionResultSchema = external_exports.object({
-  session_id: external_exports.string().optional(),
-  deleted: external_exports.boolean().optional(),
-  killed: external_exports.boolean().optional(),
-  runner_stop_confirmed: external_exports.boolean().optional()
-}).passthrough();
-var manageResultSchema = external_exports.object({
-  status: external_exports.string().optional(),
-  progress: external_exports.unknown().optional()
-}).passthrough();
-var deleteResultSchema = external_exports.object({
-  deleted: external_exports.string()
-}).passthrough();
-var modelsResultSchema = external_exports.object({
-  models: external_exports.array(external_exports.record(external_exports.unknown())).optional(),
-  tags: external_exports.array(external_exports.record(external_exports.unknown())).optional(),
-  page: external_exports.unknown().optional(),
-  knowledge: external_exports.unknown().optional(),
-  library: external_exports.unknown().optional()
-}).passthrough();
-var taskResultSchema = external_exports.object({
-  selected_model: external_exports.string().optional(),
-  context_window_fit: external_exports.string().optional(),
-  execution: external_exports.unknown().optional(),
-  response: external_exports.unknown().optional(),
-  telemetry: external_exports.unknown().optional()
-}).passthrough();
-var batchResultSchema = external_exports.object({
-  results: external_exports.unknown().optional(),
-  telemetry: external_exports.unknown().optional()
-}).passthrough();
-var researchResultSchema = external_exports.object({
-  answer: external_exports.string(),
-  summary: external_exports.string(),
-  citations: external_exports.unknown().optional(),
-  verification: external_exports.unknown(),
-  usage: external_exports.unknown().optional(),
-  telemetry: external_exports.unknown().optional()
-}).passthrough();
 function configuredExternalCost(env = process.env) {
   const model = env.FREELLAMA_EXTERNAL_COST_MODEL;
   const inputRaw = env.FREELLAMA_EXTERNAL_COST_INPUT_USD_PER_M;
@@ -22013,11 +22040,6 @@ function clipText(text, limit) {
   const head = Math.floor(usable * 2 / 3);
   return text.slice(0, head) + marker + text.slice(-(usable - head));
 }
-var PRETTY_PRINT_MAX_BYTES = 8 * 1024;
-function serialize(value) {
-  const compact = JSON.stringify(value);
-  return compact.length > PRETTY_PRINT_MAX_BYTES ? compact : JSON.stringify(value, null, 2);
-}
 function resultSummary(value) {
   const explicit = value.summary;
   if (typeof explicit === "string" && explicit.trim()) return clipText(explicit.trim(), 500);
@@ -22036,7 +22058,7 @@ function taskAnswerText(payload) {
   const message = record2(response.message);
   if (message) {
     const parts = [];
-    if (typeof message.content === "string" && message.content.trim()) parts.push(message.content.trim());
+    if (typeof message.content === "string" && message.content.trim()) parts.push(answerContent(message.content, payload));
     if (Array.isArray(message.tool_calls) && message.tool_calls.length > 0) {
       parts.push(`tool_calls: ${JSON.stringify(message.tool_calls)}`);
     }
@@ -22052,6 +22074,15 @@ function taskAnswerText(payload) {
   }
   return void 0;
 }
+function answerContent(content, payload) {
+  const check2 = record2(payload.output_check);
+  const keep = check2?.unique_prefix_chars;
+  if (check2?.repetition_loop !== true || typeof keep !== "number") return content.trim();
+  const characters = Array.from(content);
+  const trimmed = characters.slice(0, keep).join("").trim();
+  return `${trimmed}
+[freellama: trimmed ${characters.length - keep} repeated characters; the model looped until num_predict \u2014 treat as low confidence]`;
+}
 function batchAnswerText(payload) {
   if (!Array.isArray(payload.results)) return void 0;
   const perItem = Math.max(512, Math.floor(ANSWER_TEXT_MAX_CHARS / Math.max(payload.results.length, 1)));
@@ -22067,7 +22098,7 @@ ${clipText(answer, perItem)}`;
   return lines.join("\n\n");
 }
 function structuredResult(value, options = {}) {
-  const text = options.text ?? (options.legacyJson ? serialize(value) : resultSummary(value));
+  const text = options.text ?? resultSummary(value);
   return {
     content: [{ type: "text", text }],
     structuredContent: value
@@ -22093,6 +22124,19 @@ ${clipText(raw, 2e3)}`
 function isStructuredSuccess(result) {
   return !("isError" in result && result.isError) && "structuredContent" in result && result.structuredContent !== void 0;
 }
+var escapeRegExp = (text) => text.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+var SERVE_ADDRESS_PATTERN = (() => {
+  let url;
+  try {
+    url = new URL(DEFAULT_SERVE_ENDPOINT);
+  } catch {
+    return /_freellama/i;
+  }
+  const { hostname: hostname2, port } = url;
+  const hosts = ["127.0.0.1", "localhost"].includes(hostname2) ? ["127.0.0.1", "localhost"] : [hostname2];
+  const addresses = hosts.map((host) => escapeRegExp(port ? `${host}:${port}` : host));
+  return new RegExp(`(?:_freellama|${addresses.join("|")})`, "i");
+})();
 function errorResult(error2) {
   const message = error2 instanceof Error ? error2.message : String(error2);
   let receipt;
@@ -22103,7 +22147,7 @@ function errorResult(error2) {
     }
   } catch {
   }
-  const managedServeUnavailable = /(?:connection refused|connect error|failed to connect|error sending request)/i.test(message) && /(?:_freellama|127\.0\.0\.1:11435|localhost:11435)/i.test(message);
+  const managedServeUnavailable = /(?:connection refused|connect error|failed to connect|error sending request)/i.test(message) && SERVE_ADDRESS_PATTERN.test(message);
   const ollamaUnavailable = !managedServeUnavailable && /(?:connection refused|connect error|failed to connect|error sending request)/i.test(message) && /\/api\/(?:tags|ps|chat|generate|embed|show)/.test(message);
   const actionableMessage = managedServeUnavailable ? `${message}
 
@@ -22116,37 +22160,6 @@ Ollama is not running at that address. Ask the user to start Ollama (the app, or
     ...receipt ? { structuredContent: receipt } : {}
   };
 }
-var adapterCallSchema = external_exports.object({
-  raw_name: external_exports.string().optional(),
-  status: external_exports.string().optional(),
-  arguments: external_exports.object({
-    tool: external_exports.string().optional(),
-    command: external_exports.string().optional(),
-    queries: external_exports.object({ path: external_exports.string().optional() }).passthrough().optional()
-  }).passthrough().optional()
-}).passthrough();
-var adapterResultSchema = external_exports.object({
-  final_answer: external_exports.string(),
-  tool_calls: external_exports.array(adapterCallSchema).default([]),
-  usage: external_exports.object({
-    input_tokens: external_exports.number().nullable().optional(),
-    output_tokens: external_exports.number().nullable().optional()
-  }).passthrough().default({}),
-  model_metadata: external_exports.object({
-    context_compactions: external_exports.number().int().nonnegative().optional(),
-    runtime_config: external_exports.record(external_exports.string(), external_exports.unknown()).optional(),
-    context_management: external_exports.object({
-      token_counting: external_exports.enum(["configured_estimate", "model_calibrated_estimate"]),
-      estimate_scale: external_exports.number().positive(),
-      calibration_samples: external_exports.number().int().nonnegative(),
-      pinned_overflow: external_exports.enum(["error", "clip"]),
-      compactions: external_exports.number().int().nonnegative()
-    }).passthrough().optional()
-  }).passthrough().optional()
-});
-function parseAdapterResult(raw) {
-  return adapterResultSchema.parse(JSON.parse(raw));
-}
 function summarizeEmbeddings(payload) {
   const response = payload.response;
   if (response === null || typeof response !== "object") return null;
@@ -22168,22 +22181,166 @@ 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) {
+// src/tools/doctor.ts
+var doctorResultSchema = external_exports.object({
+  summary: external_exports.string().optional(),
+  endpoint: external_exports.string().optional()
+}).passthrough();
+function registerDoctorTool(server2) {
+  server2.registerTool(
+    "doctor",
+    {
+      description: "Use when: checking whether Ollama/FreeLlama are healthy, reading live queues/limits (view:status) or token usage (view:usage). Do not use when: choosing a model (models). Returns: a short summary; pick a view for detail.",
+      inputSchema: external_exports.object({
+        ollamaEndpoint: ollamaEndpointParam,
+        serveEndpoint: endpointParam,
+        view: external_exports.enum(["summary", "full", "status", "usage"]).optional(),
+        days: external_exports.number().int().min(1).max(366).optional()
+      }).strict(),
+      outputSchema: doctorResultSchema,
+      annotations: { readOnlyHint: true }
+    },
+    async ({ ollamaEndpoint, serveEndpoint, view, days }) => {
+      try {
+        if (days !== void 0 && view !== "usage") return errorResult(new Error("days is only valid with view: usage."));
+        if (view === "status") return parsedResult(await status(serveEndpoint));
+        if (view === "usage") return parsedResult(await usage(serveEndpoint, days));
+        const report = parsedResult(await doctor(ollamaEndpoint));
+        if (!isStructuredSuccess(report)) return report;
+        try {
+          report.structuredContent.machine = JSON.parse(await machine(serveEndpoint));
+          delete report.structuredContent.machine_unavailable;
+        } catch (error2) {
+          if (report.structuredContent.machine == null) {
+            report.structuredContent.machine_unavailable = `freellama serve unreachable, so no machine profile: ${error2 instanceof Error ? error2.message : String(error2)}`;
+          }
+        }
+        try {
+          report.structuredContent.platform_health = JSON.parse(await health(serveEndpoint));
+        } catch (error2) {
+          report.structuredContent.platform_health_unavailable = `freellama serve unreachable: ${error2 instanceof Error ? error2.message : String(error2)}`;
+        }
+        const full = { ...report.structuredContent };
+        delete full.ollama_env_config;
+        if ((view ?? "summary") === "full") return structuredResult(full);
+        const running = full.running?.models ?? [];
+        const compact = {
+          status: "ok",
+          summary: `Ollama ${String(full.version?.version ?? "unknown")}; ${running.length} resident model(s).`,
+          endpoint: full.endpoint,
+          ollama: { endpoint: full.endpoint, version: full.version, resident_model_count: running.length },
+          machine: full.machine,
+          local_conservative_config_posture: full.local_conservative_config_posture,
+          host_runtime_signals: full.host_runtime_signals,
+          evidence: { runtime: "ollama_api_version_and_ps", configuration: full.ollama_env_config_source },
+          next: 'Use models{view:"installed"} to select a local model; view:"full" has Ollama configuration and admission.'
+        };
+        return structuredResult(compact);
+      } catch (error2) {
+        return errorResult(error2);
+      }
+    }
+  );
+}
+
+// src/serve-client.ts
+var createSession2 = withServe(createSession);
+var deleteSession2 = withServe(deleteSession);
+var killSession2 = withServe(killSession);
+var listModels2 = withServe(listModels);
+var route2 = withServe(route);
+var runTaskRequest2 = withServe(runTaskRequest);
+var runTaskBatchRequest2 = withServe(runTaskBatchRequest);
+var warmModelRequest2 = withServe(warmModelRequest);
+var listTaskJobs2 = withServe(listTaskJobs);
+var getTaskJob2 = withServe(getTaskJob);
+var cancelTaskJob2 = withServe(cancelTaskJob);
+var removeTaskJob2 = withServe(removeTaskJob);
+var createScope2 = withServe(createScope);
+var getScope2 = withServe(getScope);
+var forkScope2 = withServe(forkScope);
+var deleteScope2 = withServe(deleteScope);
+
+// src/tools/session.ts
+var sessionResultSchema = external_exports.object({
+  session_id: external_exports.string().optional(),
+  deleted: external_exports.boolean().optional(),
+  killed: external_exports.boolean().optional(),
+  runner_stop_confirmed: external_exports.boolean().optional()
+}).passthrough();
+function registerSessionTool(server2) {
+  server2.registerTool(
+    "session",
+    {
+      description: "Use when: several related run_task calls should stay on one model (create, pass sessionId), or to kill a session's in-flight work. Do not use when: cancelling one deferred job (task_jobs), keeping history (scope), or unloading (ollama_manage). Returns: receipt.",
+      inputSchema: external_exports.object({
+        action: external_exports.enum(["create", "delete", "kill"]),
+        sessionId: external_exports.string().uuid().optional(),
+        endpoint: endpointParam
+      }).strict(),
+      outputSchema: sessionResultSchema,
+      annotations: { destructiveHint: true }
+    },
+    async ({ action, sessionId, endpoint }) => {
+      try {
+        if (action === "create") {
+          if (sessionId !== void 0) return errorResult(new Error("sessionId is only valid for action: delete or kill."));
+          return structuredResult(JSON.parse(await createSession2(endpoint)));
+        }
+        if (sessionId === void 0) return errorResult(new Error(`action: ${action} requires sessionId.`));
+        if (action === "kill") return parsedResult(await killSession2(endpoint, sessionId));
+        await deleteSession2(endpoint, sessionId);
+        return structuredResult({ session_id: sessionId, deleted: true });
+      } catch (error2) {
+        return errorResult(error2);
+      }
+    }
+  );
+}
+
+// src/paging.ts
+import { createHash } from "node:crypto";
+function pageLiveList(items, limit, cursor, identity) {
+  const pageSize = limit ?? 20;
+  const fingerprint = createHash("sha256").update(items.map(identity).join("\n")).digest("base64url").slice(0, 16);
+  let offset = 0;
+  if (cursor) {
+    try {
+      const parsed = JSON.parse(Buffer.from(cursor, "base64url").toString("utf8"));
+      if (!Number.isInteger(parsed.offset) || parsed.offset < 0 || parsed.fingerprint !== fingerprint) {
+        throw new Error("invalid or stale cursor");
+      }
+      offset = parsed.offset;
+    } catch {
+      throw new Error("`cursor` is invalid or the model list changed; restart without cursor.");
+    }
+  }
+  const page = items.slice(offset, offset + pageSize);
+  const nextOffset = offset + page.length;
+  return {
+    items: page,
+    returned: page.length,
+    total: items.length,
+    next_cursor: nextOffset < items.length ? Buffer.from(JSON.stringify({ offset: nextOffset, fingerprint }), "utf8").toString("base64url") : null
+  };
+}
+
+// src/model-search.ts
+function decodeHtml(text) {
+  const entities = { amp: "&", lt: "<", gt: ">", quot: '"', apos: "'", nbsp: " ", middot: "\xB7", bull: "\u2022" };
+  return text.replace(/&(#x[\da-f]+|#\d+|[a-z]+);/gi, (raw, entity) => {
+    if (!entity.startsWith("#")) return entities[entity.toLowerCase()] ?? raw;
+    const code = entity[1].toLowerCase() === "x" ? Number.parseInt(entity.slice(2), 16) : Number(entity.slice(1));
+    return code > 0 && code <= 1114111 && !(code >= 55296 && code <= 57343) ? String.fromCodePoint(code) : raw;
+  });
+}
+function attribute(tag, name) {
+  return tag.match(new RegExp(`(?:^|\\s)${name}\\s*=\\s*(["'])([\\s\\S]*?)\\1`, "i"))?.[2];
+}
+function htmlText(html) {
+  return decodeHtml(html.replace(/<(script|style|svg)\b[^>]*>[\s\S]*?<\/\1>/gi, "").replace(/<\/(?:p|li|div|h[1-6]|pre|tr)>|<br\s*\/?\s*>/gi, "\n").replace(/<[^>]*>/g, " ")).split("\n").map((line) => line.replace(/\s+/g, " ").trim()).filter(Boolean).join("\n");
+}
+function divContent(html, id) {
   const tags = /<\/?div\b[^>]*>/gi;
   let start;
   let depth = 0;
@@ -22227,7 +22384,8 @@ function parseModelSearch(html) {
     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 rawDescription = block.match(/<p class="max-w-lg[^"]*">([\s\S]*?)<\/p>/)?.[1];
+    const description = rawDescription === void 0 ? "" : decodeHtml(rawDescription.replace(/<[^>]+>/g, "")).replace(/\s+/g, " ").trim();
     const capabilities = [...block.matchAll(/text-(?:indigo-600|cyan-500)[^>]*>([a-z]+)<\/span>/g)].map(
       (m) => m[1]
     );
@@ -22240,8 +22398,7 @@ function parseModelSearch(html) {
       capabilities,
       pulls: stat("Pulls"),
       tags: stat("Tag"),
-      cloudAvailable: capabilities.includes("cloud"),
-      cloudOnly: null
+      cloudAvailable: capabilities.includes("cloud")
     });
   }
   return results;
@@ -22284,6 +22441,10 @@ var CACHE_TTL_MS = 60 * 60 * 1e3;
 var FAILURE_TTL_MS = 60 * 1e3;
 var MAX_PAGE_BYTES = 2 * 1024 * 1024;
 var MAX_README_CHARS = 64 * 1024;
+var errorMessage = (error2) => error2 instanceof Error ? error2.message : String(error2);
+function isRemoteModel(entry, name) {
+  return Boolean(entry.remote_host || entry.remote_model) || /-cloud$/i.test(name);
+}
 function libraryReference(model) {
   const match = model.match(/^(?:library\/)?([a-zA-Z0-9][\w.-]*(?:\/[a-zA-Z0-9][\w.-]*)?)(?::([\w.-]+))?$/);
   if (!match || match[1].split("/").some((part) => part === "." || part === "..")) return null;
@@ -22402,7 +22563,7 @@ var OllamaLibraryClient = class {
     } 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."
+        message: error2 instanceof LibraryFetchError ? error2.message : `Ollama library lookup failed or timed out (${errorMessage(error2)}). Local metadata remains available.`
       };
     }
     const now = (this.options.now ?? Date.now)();
@@ -22481,7 +22642,7 @@ function modelKnowledge(model, data, record3, includeReadme = false, supportedFe
     quality: "metadata_only",
     note: "Supported features and advertised uses do not establish task quality. Keep exact-model benchmark and policy evidence separate."
   };
-  if (data.remote_host || data.remote_model || /-cloud$/i.test(model)) {
+  if (isRemoteModel(data, model)) {
     return { ...local, localModel: false, library: { status: "remote_model", message: "This Ollama entry points to a remote model." } };
   }
   if (!record3) return { ...local, localModel: true, library: { status: "not_requested" } };
@@ -22531,11 +22692,11 @@ async function enrichInstalledModels(models, client, includeLibrary, loadTags) {
             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.";
+        } catch (error2) {
+          localError = `The selected Ollama backend's inventory could not be verified (${errorMessage(error2)}); public guidance was withheld.`;
         }
       }
-      const record3 = includeLibrary && !localError && !local.remote_host && !local.remote_model && !/-cloud$/i.test(name) ? await client.lookup(name) : void 0;
+      const record3 = includeLibrary && !localError && !isRemoteModel(local, name) ? await client.lookup(name) : void 0;
       const knowledge = modelKnowledge(name, local, record3, false, featureSource);
       results[index] = { ...model, knowledge: localError ? { ...knowledge, library: { status: "unverified", message: localError } } : knowledge };
     }
@@ -22543,229 +22704,277 @@ async function enrichInstalledModels(models, client, includeLibrary, loadTags) {
   return results;
 }
 
-// src/delegate.ts
-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."
-      }
+// src/tools/models.ts
+var modelsResultSchema = external_exports.object({
+  models: external_exports.array(external_exports.record(external_exports.unknown())).optional(),
+  tags: external_exports.array(external_exports.record(external_exports.unknown())).optional(),
+  page: external_exports.unknown().optional(),
+  knowledge: external_exports.unknown().optional(),
+  library: external_exports.unknown().optional()
+}).passthrough();
+var ollamaLibrary = new OllamaLibraryClient();
+var HOST_MEMORY_FIT_FRACTION = 0.6;
+async function libraryLookup({
+  model,
+  query,
+  capabilities,
+  order,
+  limit,
+  cursor,
+  includeReadme,
+  endpoint,
+  ollamaEndpoint
+}) {
+  const localState = async () => {
+    let installed = null;
+    let installedUnavailable;
+    let memoryBytes = null;
+    try {
+      const tags = await ollamaFetch(ollamaEndpoint, "/api/tags");
+      installed = new Set((tags.models ?? []).map((m) => m.name ?? ""));
+    } catch (error2) {
+      installedUnavailable = `Local inventory could not be read (${error2 instanceof Error ? error2.message : String(error2)}); installed is null.`;
     }
-  };
-}
-var MODEL_EVIDENCE = (() => {
-  const configured = process.env.FREELLAMA_MCP_MODEL_EVIDENCE;
-  const candidates = [
-    configured,
-    path4.join(REPO_ROOT, "benchmark/evidence/model-evidence.json")
-  ].filter((x) => Boolean(x));
-  for (const file of candidates) {
     try {
-      if (!existsSync4(file)) continue;
-      const parsed = JSON.parse(readFileSync(file, "utf8"));
-      return parsed.models ?? {};
+      const profile = JSON.parse(await machine(endpoint));
+      memoryBytes = profile.memory_bytes ?? profile.unified_memory_bytes ?? null;
     } catch {
     }
+    return { installed, installedUnavailable, memoryBytes };
+  };
+  if (!model && cursor !== void 0) {
+    return errorResult(new Error("Library step 1 has no `cursor`; it pages step-2 tags. Use `limit` for search, or pass `model`."));
   }
-  return {};
-})();
-function assessDelegatedAnswer(question, evidenceCount, model) {
-  const grounded = evidenceCount > 0;
-  const evidence = MODEL_EVIDENCE[model];
-  if (evidence?.grade === "unusable") {
-    return {
-      recommendation: "escalate",
-      grounded,
-      why: `${model} is not viable for research on this machine (${evidence.note}). It answers fast and confidently while being wrong \u2014 a fast wrong answer is not a speed win. Re-run with a ~27B model (see README), or answer it yourself.`,
-      measuredBaseRate: `${model}: ${evidence.note}`
-    };
-  }
-  if (!evidence) {
-    return {
-      recommendation: "verify",
-      grounded,
-      why: `${model} has no measured accuracy in this repo's benchmarks, so no base rate applies to this answer. Treat it as unverified until it has been evaluated \u2014 accuracy fell off a cliff below ~12B in the models that were measured.`,
-      measuredBaseRate: "no measured base rate for this model"
-    };
-  }
-  if (evidence.grade === "weak" && grounded) {
-    return {
-      recommendation: "verify",
-      grounded,
-      why: `${model} holds up on simple single-file lookups but not beyond (${evidence.note}). Check the evidence trail, or re-run on a ~27B model (see README) if the answer matters.`,
-      measuredBaseRate: `${model}: ${evidence.note}`
-    };
-  }
-  if (!grounded) {
-    return {
-      recommendation: "escalate",
-      grounded: false,
-      why: "The model answered without reading a single file, so this is parametric recall, not research \u2014 the failure mode where `run_task` was verified inventing wrong facts about this project's own architecture. Re-ask with a narrower question, or answer it yourself.",
-      measuredBaseRate: "ungrounded answers have no measured accuracy \u2014 they were never the tested path"
-    };
-  }
-  const judgmentSignals = /\b(should (we|i|you)|better|best|worth|review|assess|evaluate|improve|opinion|recommend|why is|is it (good|safe|correct)|design|refactor)\b/i;
-  if (judgmentSignals.test(question)) {
-    return {
-      recommendation: "verify",
-      grounded: true,
-      why: "The question reads as a judgment call (keyword heuristic), which is the ~67%-accurate bucket rather than the 98.9% one \u2014 and the tone is identical either way. Check the evidence trail against the claim before acting on it.",
-      measuredBaseRate: `${model}: ~67% on judgment calls vs 98.9% on grounded lookups`
-    };
+  try {
+    if (model) {
+      const ref = libraryReference(model);
+      if (!ref) throw new Error("Use an exact Ollama family or namespace/model name, not a URL or path.");
+      const { family } = ref;
+      const library = await ollamaLibrary.lookup(model);
+      if (library.status !== "available") return structuredResult({
+        family,
+        sourcePage: ref.url,
+        tags: [],
+        tagsUnavailable: library.message,
+        library,
+        recommendation: null
+      });
+      const tags = library.tags;
+      const { installed: installed2, installedUnavailable: installedUnavailable2, memoryBytes } = await localState();
+      const budget = memoryBytes ? memoryBytes * HOST_MEMORY_FIT_FRACTION : null;
+      const annotated = tags.map((t) => ({
+        ...t,
+        installed: installed2 ? installed2.has(t.tag) : null,
+        fitsInMemory: budget && t.sizeBytes ? t.sizeBytes <= budget : null,
+        fitScope: "host_memory_budget_only"
+      }));
+      const runnable = annotated.filter((t) => t.fitsInMemory === true && !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: ref.url,
+        library: {
+          status: library.status,
+          sourceUrl: library.sourceUrl,
+          tagsSourceUrl: library.tagsSourceUrl,
+          fetchedAt: library.fetchedAt,
+          expiresAt: library.expiresAt,
+          cached: library.cached,
+          ...publicModelGuidance(library, void 0, includeReadme)
+        },
+        ...installedUnavailable2 ? { installedUnavailable: installedUnavailable2 } : {},
+        machineMemoryBytes: memoryBytes,
+        fitBudgetBytes: budget,
+        tags: tagPage.items,
+        page: { returned: tagPage.returned, total: tagPage.total, next_cursor: tagPage.next_cursor },
+        recommendationUnavailable: budget ? void 0 : "No machine profile (freellama serve unreachable), so memory fit could not be checked and no tag is recommended. Start serve, or read the sizes yourself.",
+        recommendation: best ? 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 search = await ollamaLibrary.search(params);
+    const url = search.url;
+    const parsed = search.models.slice(0, limit ?? 10);
+    const { installed, installedUnavailable } = await localState();
+    const installedFamilies = installed ? new Set([...installed].map((n) => n.split(":")[0])) : null;
+    return structuredResult({
+      query: url,
+      fetchedAt: search.fetchedAt,
+      ...parsed.length === 0 ? { searchUnavailable: "No result cards were recognized. The query may have no matches or Ollama's markup may have changed; inspect the source URL." } : {},
+      order: order ?? "popular",
+      count: parsed.length,
+      ...installedUnavailable ? { installedUnavailable } : {},
+      nextStep: 'Not pullable yet. Call again with view:"library", model:"<name>" to get tags, sizes and memory fit.',
+      models: parsed.map((m) => ({ ...m, installed: installedFamilies ? installedFamilies.has(m.name) : null }))
+    });
+  } catch (error2) {
+    return errorResult(error2);
   }
-  if (evidenceCount > 5) {
-    return {
-      recommendation: "verify",
-      grounded: true,
-      why: `${evidenceCount} tool calls is outside the 1-5 file envelope this tool was measured on. Wide searches are where it drifts; spot-check the evidence trail.`,
-      measuredBaseRate: `${model}: 98.9% on grounded lookups, measured within a 1-5 file scope`
-    };
-  }
-  return {
-    recommendation: "accept",
-    grounded: true,
-    why: `Grounded in ${evidenceCount} tool call(s) within the measured 1-5 file envelope, the question is lookup-shaped, and ${model} is measured strong for this (${evidence.note}).`,
-    measuredBaseRate: `${model}: 98.9% on grounded lookups (100+ questions)`
-  };
 }
-
-// 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
+function registerModelsTool(server2) {
+  server2.registerTool(
+    "models",
+    {
+      description: "Use when: listing installed or loaded models with capabilities, policy rank and memory fit, inspecting one model, or searching ollama.com (view:library). Do not use when: running work (run_task) or pulling (ollama_manage). Returns: models; library lists families first, then one family's tags.",
+      inputSchema: external_exports.object({
+        view: external_exports.enum(["installed", "resident", "detail", "raw", "library"]).optional(),
+        model: external_exports.string().min(1).optional(),
+        includeVerbose: external_exports.boolean().optional(),
+        includeLibrary: external_exports.boolean().optional(),
+        includeReadme: external_exports.boolean().optional(),
+        query: external_exports.string().min(1).optional(),
+        capabilities: external_exports.array(external_exports.enum(["vision", "tools", "thinking", "embedding", "cloud"])).min(1).max(5).optional(),
+        order: external_exports.enum(["popular", "newest"]).optional(),
+        limit: external_exports.number().int().positive().max(50).optional(),
+        cursor: external_exports.string().min(1).optional(),
+        endpoint: endpointParam,
+        ollamaEndpoint: ollamaEndpointParam
+      }).strict(),
+      outputSchema: modelsResultSchema,
+      annotations: { readOnlyHint: true }
     },
-    limits: bounds === void 0 ? void 0 : {
-      max_messages: bounds.maxMessages,
-      max_bytes: bounds.maxBytes,
-      max_estimated_tokens: bounds.maxEstimatedTokens,
-      ttl_seconds: bounds.ttlSeconds
+    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.`));
+        }
+        if (selectedView !== "library" && selectedView !== "raw" && [limit, cursor].some((value) => value !== void 0)) {
+          return errorResult(new Error(`view "${selectedView}" accepts neither pagination nor library search fields.`));
+        }
+        if (selectedView !== "detail" && includeVerbose !== void 0) {
+          return errorResult(new Error('`includeVerbose` is valid only for view "detail".'));
+        }
+        if (selectedView !== "detail" && selectedView !== "library" && model !== void 0) {
+          return errorResult(new Error('`model` is valid only for views "detail" and "library".'));
+        }
+        if (selectedView === "library" && model && [query, capabilities, order].some((value) => value !== void 0)) {
+          return errorResult(
+            new Error("Library step 2 accepts only `model` plus endpoint overrides; omit step-1 search fields.")
+          );
+        }
+        switch (selectedView) {
+          case "raw": {
+            const raw = await ollamaFetch(ollamaEndpoint, "/api/tags");
+            const models = Array.isArray(raw.models) ? raw.models : [];
+            const page = pageLiveList(models, limit, cursor, (model2) => String(model2.name ?? JSON.stringify(model2)));
+            return structuredResult({ ...raw, models: page.items, page: { returned: page.returned, total: page.total, next_cursor: page.next_cursor } });
+          }
+          case "resident": {
+            const managed = parsedResult(await listModels2(endpoint));
+            if (!isStructuredSuccess(managed)) return managed;
+            const data = managed.structuredContent;
+            const models = (data.models ?? []).filter((entry) => entry.resident === true).map((entry) => {
+              const { resident_view: view2, ...rest } = entry;
+              if (!view2) return rest;
+              return {
+                ...rest,
+                placement: {
+                  gpu_percent: view2.gpu_percent,
+                  assigned: view2.assigned,
+                  processor: view2.processor_label,
+                  ...view2.warning !== void 0 ? { warning: view2.warning } : {}
+                }
+              };
+            });
+            return structuredResult({ ...data, models });
+          }
+          case "detail": {
+            if (!model) {
+              return errorResult(new Error('view "detail" needs `model` set to an installed tag.'));
+            }
+            const data = await ollamaFetch(ollamaEndpoint, "/api/show", {
+              method: "POST",
+              body: { model }
+            });
+            let local = data;
+            if (includeLibrary && !isRemoteModel(data, model)) {
+              try {
+                const inventory = await ollamaFetch(ollamaEndpoint, "/api/tags");
+                const canonical = libraryReference(model)?.tag;
+                const entry = canonical && inventory.models?.find((entry2) => libraryReference(String(entry2.name ?? entry2.model))?.tag === canonical);
+                if (entry) local = { ...entry, ...data, digest: entry.digest };
+              } catch {
+              }
+            }
+            const library = includeLibrary && !isRemoteModel(local, model) ? await ollamaLibrary.lookup(model) : void 0;
+            const { license, modelfile, ...rest } = data;
+            const modelInfo = rest.model_info ?? {};
+            const contextEntry = Object.entries(modelInfo).find(
+              ([key, value]) => key.endsWith(".context_length") && typeof value === "number"
+            );
+            return structuredResult({
+              ...rest,
+              max_context_length: contextEntry?.[1] ?? null,
+              knowledge: modelKnowledge(model, local, library, includeReadme),
+              ...includeVerbose ? { license, modelfile } : {}
+            });
+          }
+          case "library":
+            return await libraryLookup({ model, query, capabilities, order, limit, cursor, includeReadme, endpoint, ollamaEndpoint });
+          default: {
+            const result = parsedResult(await listModels2(endpoint));
+            if (!isStructuredSuccess(result)) return result;
+            const data = result.structuredContent;
+            return structuredResult({ ...data, models: await enrichInstalledModels(
+              data.models ?? [],
+              ollamaLibrary,
+              includeLibrary ?? false,
+              async (upstream) => (await ollamaFetch(upstream ?? ollamaEndpoint, "/api/tags")).models ?? []
+            ) });
+          }
+        }
+      } catch (error2) {
+        return errorResult(error2);
+      }
     }
-  };
+  );
 }
-function registerContextTools(server2) {
-  const createScope2 = withServe(createScope);
-  const getScope2 = withServe(getScope);
-  const forkScope2 = withServe(forkScope);
-  const deleteScope2 = withServe(deleteScope);
-  const warmModel = withServe(warmModelRequest);
+
+// src/tools/scope.ts
+function registerScopeTool(server2) {
   server2.registerTool("scope", {
-    description: "Use when: saved history. Do not use when: KV control. Returns: process-local scope metadata; history is opt-in.",
+    description: "Use when: a multi-turn local conversation should keep its message history server-side (create/get/fork/delete), then continue with run_task {scopeId, scopeRevision}. Do not use when: you want model affinity (session) or runner KV control. Returns: process-local scope metadata; message history only with includeMessages:true.",
     inputSchema: external_exports.object({
       action: external_exports.enum(["create", "get", "fork", "delete"]),
       endpoint: endpointParam,
       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 }) => {
+  }, async ({ action, endpoint, scopeId, revision, messages, 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
-      }));
+      return parsedResult(action === "create" ? await createScope2(endpoint, { messages }) : action === "fork" ? await forkScope2(endpoint, scopeId, { revision }) : await getScope2(endpoint, scopeId, includeMessages));
     } catch (error2) {
       return errorResult(error2);
     }
   });
 }
 
-// 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);
+// src/telemetry.ts
 var EXTERNAL_COST = configuredExternalCost();
 function withTaskTelemetry(result) {
   if (!isStructuredSuccess(result)) return result;
@@ -22808,55 +23017,469 @@ function withBatchTelemetry(result) {
     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;
-  const fingerprint = createHash("sha256").update(items.map(identity).join("\n")).digest("base64url").slice(0, 16);
-  let offset = 0;
-  if (cursor) {
+
+// src/tools/run-task.ts
+var taskResultSchema = external_exports.object({
+  selected_model: external_exports.string().optional(),
+  context_window_fit: external_exports.string().optional(),
+  execution: external_exports.unknown().optional(),
+  response: external_exports.unknown().optional(),
+  telemetry: external_exports.unknown().optional()
+}).passthrough();
+var batchResultSchema = external_exports.object({
+  results: external_exports.unknown().optional(),
+  telemetry: external_exports.unknown().optional()
+}).passthrough();
+function validateTaskPayload(task, label) {
+  const fail = (message) => label ? `${label}: ${message}` : message;
+  if (task.scopeId === void 0 !== (task.scopeRevision === void 0)) {
+    return fail("scopeId and scopeRevision must be supplied together.");
+  }
+  if (task.topLogprobs !== void 0 && task.logprobs !== true) return fail("`topLogprobs` requires `logprobs:true`.");
+  if (task.task === "embedding") {
+    if (task.input === void 0) return fail('task "embedding" requires `input`.');
+    const chatControls = [
+      task.prompt,
+      task.systemPrompt,
+      task.images,
+      task.messages,
+      task.tools,
+      task.format,
+      task.think,
+      task.logprobs,
+      task.topLogprobs
+    ];
+    if (chatControls.some((value) => value !== void 0)) {
+      return fail('task "embedding" accepts `input`, `options`, and routing fields; chat controls are not valid.');
+    }
+    return null;
+  }
+  if (task.input !== void 0 || task.returnEmbeddings !== void 0) {
+    return fail('`input` and `returnEmbeddings` are valid only for task "embedding".');
+  }
+  if (task.prompt === void 0 && task.systemPrompt === void 0 && (task.messages === void 0 || task.messages.length === 0)) {
+    return fail(`task "${task.task ?? "completion"}" requires \`prompt\`, \`systemPrompt\`, or \`messages\`.`);
+  }
+  if (task.images !== void 0 && (task.messages !== void 0 || task.prompt === void 0)) {
+    return fail("Top-level `images` is valid only with `prompt`; put images inside messages instead.");
+  }
+  return null;
+}
+async function runBatch(tasks, maxParallelism, endpoint) {
+  for (const item of tasks) {
+    const problem = validateTaskPayload(item.task, `batch item ${item.id}`);
+    if (problem) return errorResult(new Error(problem));
+  }
+  return withBatchTelemetry(parsedResult(await runTaskBatchRequest2(endpoint ?? DEFAULT_SERVE_ENDPOINT, {
+    tasks: tasks.map(({ id, independent, task }) => ({
+      id,
+      independent,
+      task: {
+        task: task.task === void 0 ? task.scopeId === void 0 ? "completion" : void 0 : canonicalTaskKind(task.task),
+        objective: task.objective ?? (task.scopeId === void 0 ? "balanced" : void 0),
+        model: task.model,
+        session_id: task.sessionId,
+        scope_id: task.scopeId,
+        scope_revision: task.scopeRevision,
+        context_tokens: task.contextTokens,
+        execution_preference: task.executionPreference,
+        min_placement_evidence: task.minPlacementEvidence,
+        required_capabilities: task.requiredCapabilities,
+        priority: task.priority ?? "normal",
+        max_wait_seconds: task.maxWaitSeconds,
+        timeout_seconds: task.timeoutSeconds,
+        stop_on_repetition: task.stopOnRepetition,
+        prompt: task.prompt,
+        images: task.images,
+        messages: taskMessages(task),
+        input: task.input,
+        tools: task.tools,
+        keep_alive: task.keepAlive,
+        min_confidence: task.minConfidence,
+        request_options: {
+          format: task.format,
+          think: task.think,
+          options: task.options,
+          logprobs: task.logprobs,
+          top_logprobs: task.topLogprobs
+        }
+      }
+    })),
+    max_parallelism: maxParallelism
+  })));
+}
+function registerRunTaskTool(server2) {
+  server2.registerTool(
+    "run_task",
+    {
+      description: "Use when: offloading caller-supplied work (chat, small tasks, code, tool calls, OCR/vision, embeddings) to a local model. Modes: single (`prompt`/`messages`, or `input` for embedding); `tasks` = up to 64 independent items, ordered results, `maxParallelism` caps dispatch; `preview:true` decides only; `warm:true` preloads `model`. Images: base64 `images` + the model's own prompt; `stopOnRepetition` ends a looping answer early, keeping its text. Do not use when: it needs file access (pass content, or delegate_research) or high-stakes judgment. Returns: answer text; receipts in structuredContent.",
+      inputSchema: external_exports.object({
+        endpoint: endpointParam,
+        task: taskParam.removeDefault().optional().describe("caller owns prompts and output format."),
+        objective: objectiveParam,
+        model: external_exports.string().min(1).optional(),
+        sessionId: external_exports.string().uuid().optional(),
+        scopeId: external_exports.string().uuid().optional(),
+        scopeRevision: external_exports.number().int().nonnegative().optional(),
+        contextTokens: external_exports.number().int().positive().optional().describe("Total input + output window (num_ctx)."),
+        executionPreference: executionPreferenceParam,
+        minPlacementEvidence: minPlacementEvidenceParam,
+        requiredCapabilities: requiredCapabilitiesParam,
+        prompt: external_exports.string().min(1).optional(),
+        systemPrompt: systemPromptParam,
+        images: external_exports.array(external_exports.string().min(1)).min(1).optional().describe("base64, no data-URI prefix; prompt mode only; needs `model` or [policies.vision]"),
+        messages: messagesParam.describe(
+          "Caller system prompts; no injected task instructions. Overrides prompt; extra fields preserved."
+        ),
+        input: external_exports.union([external_exports.string().min(1), external_exports.array(external_exports.string().min(1)).min(1)]).optional(),
+        tools: localToolsParam,
+        keepAlive: external_exports.string().min(1).optional().describe('"0" unloads now, "-1" pins; omitted = adaptive finite TTL'),
+        format: external_exports.union([external_exports.literal("json"), external_exports.record(external_exports.unknown())]).optional(),
+        think: external_exports.union([external_exports.boolean(), external_exports.enum(["low", "medium", "high"])]).optional(),
+        options: external_exports.record(external_exports.unknown()).optional().describe("num_ctx/contextTokens and num_gpu/placement are routing-owned; num_predict caps output."),
+        logprobs: external_exports.boolean().optional(),
+        topLogprobs: external_exports.number().int().nonnegative().optional(),
+        minConfidence: minConfidenceParam,
+        priority: external_exports.enum(["interactive", "normal", "background"]).optional(),
+        maxWaitSeconds: external_exports.number().int().positive().optional(),
+        timeoutSeconds: external_exports.number().int().positive().optional(),
+        stopOnRepetition: external_exports.boolean().optional().describe("Stream and stop once output repeats itself (a loop to num_predict); keeps all generated text"),
+        defer: external_exports.boolean().optional(),
+        returnEmbeddings: external_exports.boolean().optional(),
+        preview: external_exports.boolean().optional().describe(
+          "true = routing fields only; rejects payloads and runtime controls; never executes"
+        ),
+        warm: external_exports.boolean().optional().describe("Load exact `model`; no prompt"),
+        tasks: external_exports.array(batchItemParam).min(1).max(64).optional().describe("Batch mode: independent items; per-item fields go inside each item's `task`"),
+        maxParallelism: external_exports.number().int().positive().max(64).optional().describe("Batch mode only")
+      }).strict(),
+      outputSchema: taskResultSchema.merge(batchResultSchema),
+      annotations: { destructiveHint: false }
+    },
+    async ({
+      endpoint,
+      task,
+      objective,
+      model,
+      sessionId,
+      scopeId,
+      scopeRevision,
+      contextTokens,
+      executionPreference,
+      minPlacementEvidence,
+      requiredCapabilities,
+      prompt,
+      systemPrompt,
+      images,
+      messages,
+      input,
+      tools,
+      keepAlive,
+      format,
+      think,
+      options,
+      logprobs,
+      topLogprobs,
+      minConfidence,
+      priority,
+      maxWaitSeconds,
+      timeoutSeconds,
+      stopOnRepetition,
+      defer,
+      returnEmbeddings,
+      preview,
+      warm,
+      tasks,
+      maxParallelism
+    }) => {
+      try {
+        const singleFields = {
+          task,
+          objective,
+          model,
+          sessionId,
+          scopeId,
+          scopeRevision,
+          contextTokens,
+          executionPreference,
+          minPlacementEvidence,
+          requiredCapabilities,
+          prompt,
+          systemPrompt,
+          images,
+          messages,
+          input,
+          tools,
+          keepAlive,
+          format,
+          think,
+          options,
+          logprobs,
+          topLogprobs,
+          minConfidence,
+          priority,
+          maxWaitSeconds,
+          timeoutSeconds,
+          stopOnRepetition,
+          defer,
+          returnEmbeddings,
+          preview,
+          warm
+        };
+        if (tasks !== void 0) {
+          const stray = Object.entries(singleFields).filter(([, value]) => value !== void 0).map(([name]) => name);
+          const unsupported = stray.filter((name) => !BATCH_TASK_FIELDS.has(name)).map((name) => `\`${name}\``);
+          if (unsupported.length > 0) {
+            return errorResult(new Error(`${unsupported.join(", ")} ${unsupported.length === 1 ? "is" : "are"} not supported with \`tasks\`.`));
+          }
+          const movable = stray.map((name) => `\`${name}\``);
+          if (movable.length > 0) {
+            return errorResult(new Error(`With \`tasks\`, put ${movable.join(", ")} inside each item's \`task\`.`));
+          }
+          return await runBatch(tasks, maxParallelism, endpoint);
+        }
+        if (maxParallelism !== void 0) return errorResult(new Error("`maxParallelism` is valid only with `tasks`."));
+        if (warm) {
+          const allowed = /* @__PURE__ */ new Set([
+            "warm",
+            "model",
+            "task",
+            "contextTokens",
+            "executionPreference",
+            "minPlacementEvidence",
+            "priority",
+            "keepAlive",
+            "maxWaitSeconds",
+            "timeoutSeconds",
+            "defer"
+          ]);
+          const stray = Object.entries(singleFields).filter(([name, value]) => value !== void 0 && !allowed.has(name)).map(([name]) => `\`${name}\``);
+          if (stray.length > 0) return errorResult(new Error(`\`warm:true\` loads a model without generation; remove ${stray.join(", ")}.`));
+          if (model === void 0) return errorResult(new Error("`warm:true` requires an exact installed `model`."));
+          if (task === "embedding") return errorResult(new Error("`warm:true` supports chat models; embedding warming is unsupported."));
+          return parsedResult(await warmModelRequest2(endpoint, {
+            model,
+            task: task === void 0 ? void 0 : canonicalTaskKind(task),
+            context_tokens: contextTokens,
+            execution_preference: executionPreference,
+            min_placement_evidence: minPlacementEvidence,
+            priority,
+            keep_alive: keepAlive,
+            max_wait_seconds: maxWaitSeconds,
+            timeout_seconds: timeoutSeconds,
+            defer
+          }));
+        }
+        const canonicalTask = canonicalTaskKind(task ?? "completion");
+        if (preview) {
+          const executionOnlyFields = [
+            ["scopeId", scopeId],
+            ["scopeRevision", scopeRevision],
+            ["prompt", prompt],
+            ["systemPrompt", systemPrompt],
+            ["images", images],
+            ["messages", messages],
+            ["input", input],
+            ["tools", tools],
+            ["keepAlive", keepAlive],
+            ["format", format],
+            ["think", think],
+            ["options", options],
+            ["logprobs", logprobs],
+            ["topLogprobs", topLogprobs],
+            ["returnEmbeddings", returnEmbeddings],
+            ["maxWaitSeconds", maxWaitSeconds],
+            ["priority", priority],
+            ["timeoutSeconds", timeoutSeconds],
+            ["stopOnRepetition", stopOnRepetition],
+            ["defer", defer]
+          ].filter(([, value]) => value !== void 0).map(([name]) => `\`${name}\``);
+          if (executionOnlyFields.length > 0) {
+            return errorResult(
+              new Error(
+                `\`preview:true\` accepts routing fields only; remove execution-only fields ${executionOnlyFields.join(", ")}. Use \`requiredCapabilities:["tools"]\` to preview tool capability, then make a separate execution call with the payload.`
+              )
+            );
+          }
+        }
+        const effectiveRequiredCapabilities = tools === void 0 ? requiredCapabilities : withRequiredCapability(requiredCapabilities, "tools");
+        if (preview) {
+          return parsedResult(
+            await route2(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
+          );
+        }
+        const problem = validateTaskPayload({
+          task,
+          scopeId,
+          scopeRevision,
+          prompt,
+          systemPrompt,
+          messages,
+          images,
+          input,
+          tools,
+          format,
+          think,
+          logprobs,
+          topLogprobs,
+          returnEmbeddings
+        });
+        if (problem) return errorResult(new Error(problem));
+        if (minConfidence && defer && scopeId === void 0) {
+          const decision = parsedResult(
+            await route2(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
+          );
+          if (!isStructuredSuccess(decision)) return decision;
+        }
+        const result = parsedResult(
+          await runTaskRequest2(endpoint ?? DEFAULT_SERVE_ENDPOINT, {
+            task: scopeId !== void 0 && task === void 0 ? void 0 : canonicalTask,
+            objective: objective ?? (scopeId === void 0 ? "balanced" : void 0),
+            model,
+            session_id: sessionId,
+            scope_id: scopeId,
+            scope_revision: scopeRevision,
+            context_tokens: contextTokens,
+            required_capabilities: effectiveRequiredCapabilities,
+            prompt,
+            images,
+            messages: taskMessages({ systemPrompt, messages, prompt, images }),
+            input,
+            tools,
+            keep_alive: keepAlive,
+            min_confidence: minConfidence,
+            priority: priority ?? "normal",
+            max_wait_seconds: maxWaitSeconds,
+            timeout_seconds: timeoutSeconds,
+            stop_on_repetition: stopOnRepetition,
+            defer: defer ?? false,
+            execution_preference: executionPreference,
+            min_placement_evidence: minPlacementEvidence,
+            request_options: {
+              format,
+              think,
+              options,
+              logprobs,
+              top_logprobs: topLogprobs
+            }
+          })
+        );
+        if (isStructuredSuccess(result) && result.structuredContent.deferred === true) return result;
+        if (!returnEmbeddings && isStructuredSuccess(result)) {
+          const trimmed = summarizeEmbeddings(result.structuredContent);
+          if (trimmed) return withTaskTelemetry(structuredResult(trimmed));
+        }
+        return withTaskTelemetry(result);
+      } catch (error2) {
+        return errorResult(error2);
+      }
+    }
+  );
+}
+
+// src/tools/task-jobs.ts
+function registerTaskJobsTool(server2) {
+  server2.registerTool(
+    "task_jobs",
+    {
+      description: "Use when: following work submitted with run_task {defer:true}: list, get (result), cancel, or remove a job. Do not use when: submitting new work or killing a whole session. Returns: job status/result; remove cancels first.",
+      inputSchema: external_exports.object({
+        endpoint: endpointParam,
+        action: external_exports.enum(["list", "get", "cancel", "remove"]),
+        jobId: external_exports.string().uuid().optional(),
+        returnEmbeddings: external_exports.boolean().optional()
+      }).strict(),
+      outputSchema: objectResultSchema,
+      annotations: { destructiveHint: true }
+    },
+    async ({ endpoint, action, jobId, returnEmbeddings }) => {
+      try {
+        if (action === "list" ? jobId !== void 0 : jobId === void 0) {
+          return errorResult(new Error("Use jobId only with get/cancel/remove; all require a jobId."));
+        }
+        if (returnEmbeddings !== void 0 && action !== "get") {
+          return errorResult(new Error("returnEmbeddings is valid only with get."));
+        }
+        const result = parsedResult(action === "list" ? await listTaskJobs2(endpoint) : action === "get" ? await getTaskJob2(endpoint, jobId) : action === "cancel" ? await cancelTaskJob2(endpoint, jobId) : await removeTaskJob2(endpoint, jobId));
+        if (!returnEmbeddings && isStructuredSuccess(result)) {
+          const job = result.structuredContent.job;
+          if (job?.result && typeof job.result === "object") {
+            const trimmed = summarizeEmbeddings(job.result);
+            if (trimmed) return structuredResult({ ...result.structuredContent, job: { ...job, result: trimmed } });
+          }
+        }
+        return result;
+      } catch (error2) {
+        return errorResult(error2);
+      }
+    }
+  );
+}
+
+// src/tools/ollama-lifecycle.ts
+var manageResultSchema = external_exports.object({
+  status: external_exports.string().optional(),
+  progress: external_exports.unknown().optional()
+}).passthrough();
+var deleteResultSchema = external_exports.object({
+  deleted: external_exports.string()
+}).passthrough();
+function summarizeOllamaPullStream(text) {
+  const events = [];
+  for (const line of text.split(/\r?\n/)) {
+    const trimmed = line.trim();
+    if (!trimmed) continue;
     try {
-      const parsed = JSON.parse(Buffer.from(cursor, "base64url").toString("utf8"));
-      if (!Number.isInteger(parsed.offset) || parsed.offset < 0 || parsed.fingerprint !== fingerprint) {
-        throw new Error("invalid or stale cursor");
+      const parsed = JSON.parse(trimmed);
+      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
+        events.push(parsed);
       }
-      offset = parsed.offset;
     } catch {
-      throw new Error("`cursor` is invalid or the model list changed; restart without cursor.");
     }
   }
-  const page = items.slice(offset, offset + pageSize);
-  const nextOffset = offset + page.length;
-  return {
-    items: page,
-    returned: page.length,
-    total: items.length,
-    next_cursor: nextOffset < items.length ? Buffer.from(JSON.stringify({ offset: nextOffset, fingerprint }), "utf8").toString("base64url") : null
-  };
-}
-var liveDelegates = /* @__PURE__ */ new Set();
-function killLiveDelegates() {
-  for (const child2 of liveDelegates) {
+  if (events.length === 0) {
+    if (!text.trim()) return { status: "empty" };
     try {
-      child2.kill("SIGKILL");
+      const parsed = JSON.parse(text);
+      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
+        return parsed;
+      }
     } catch {
+      return { status: "unparsed", raw: clipText(text, 500) };
     }
+    return { status: "unparsed", raw: clipText(text, 500) };
   }
-  liveDelegates.clear();
+  const last = events[events.length - 1] ?? {};
+  const withBytes = [...events].reverse().find((event) => typeof event.total === "number" && event.total > 0);
+  const completed = typeof withBytes?.completed === "number" ? withBytes.completed : null;
+  const total = typeof withBytes?.total === "number" ? withBytes.total : null;
+  const percent = completed != null && total != null && total > 0 ? Math.round(completed / total * 1e3) / 10 : null;
+  const status2 = typeof last.error === "string" && last.error ? "error" : last.status;
+  return {
+    ...last,
+    ...status2 === void 0 ? {} : { status: status2 },
+    progress: {
+      events: events.length,
+      completed,
+      total,
+      percent,
+      lastStatus: last.status ?? null
+    }
+  };
 }
-for (const stream of [process.stdout, process.stderr]) {
-  stream.on("error", (error2) => {
-    if (error2.code === "EPIPE") process.exit(0);
-    throw error2;
+async function ollamaPull(endpoint, model, timeoutSeconds, options = {}) {
+  const raw = await ollamaFetch(endpoint, "/api/pull", {
+    method: "POST",
+    body: { model, stream: true },
+    timeoutMs: (timeoutSeconds ?? DEFAULT_PULL_TIMEOUT_SECONDS) * 1e3,
+    parse: "ndjson",
+    signal: options.signal,
+    onLine: options.onProgress
   });
+  if (typeof raw === "string") return summarizeOllamaPullStream(raw);
+  if (raw && typeof raw === "object" && !Array.isArray(raw)) return raw;
+  return { status: "empty" };
 }
-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;
@@ -22880,769 +23503,245 @@ function pullProgressReporter(extra) {
     }).catch(() => void 0);
   };
 }
-for (const signal of ["SIGINT", "SIGTERM", "SIGHUP"]) {
-  process.on(signal, () => {
-    killLiveDelegates();
-    process.exit(0);
-  });
-}
-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 = path5.resolve(path5.dirname(fileURLToPath3(import.meta.url)), "..", "docs");
-if (!existsSync5(PACKAGED_DOCS_DIR)) {
-  throw new Error(`FreeLlama MCP documentation is missing at ${PACKAGED_DOCS_DIR}; run the package build.`);
-}
-var PACKAGED_DEV_DOCS_DIR = path5.join(PACKAGED_DOCS_DIR, "dev");
-var packagedDocs = [
-  ...readdirSync(PACKAGED_DOCS_DIR).filter((name) => name.endsWith(".md")).sort(),
-  ...existsSync5(PACKAGED_DEV_DOCS_DIR) ? readdirSync(PACKAGED_DEV_DOCS_DIR).filter((name) => name.endsWith(".md")).sort().map((name) => `dev/${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 ${PACKAGED_DOCS_DIR}/INDEX.md; run the package build.`);
-}
-for (const name of packagedDocs) {
-  const base = path5.basename(name);
-  const uri = `freellama://docs/${base === "INDEX.md" ? "index" : base.replace(/\.md$/, "")}`;
-  server.registerResource(
-    `freellama-docs-${base.toLowerCase().replace(/\.md$/, "")}`,
-    uri,
+function registerOllamaLifecycleTools(server2) {
+  server2.registerTool(
+    "ollama_manage",
     {
-      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}.`
+      description: "Use when: a human approved pulling one exact tag (action:pull) or unloading a loaded model (action:stop). Do not use when: deleting (ollama_delete), searching (models), or preloading (run_task {warm:true}). Returns: receipt; timeoutSeconds is pull-only.",
+      inputSchema: external_exports.object({
+        action: external_exports.enum(["pull", "stop"]),
+        model: external_exports.string().min(1),
+        ollamaEndpoint: ollamaEndpointParam,
+        timeoutSeconds: external_exports.number().int().positive().optional()
+      }).strict(),
+      outputSchema: manageResultSchema,
+      annotations: { destructiveHint: false }
     },
-    async () => ({
-      contents: [{ uri, mimeType: "text/markdown", text: await readFile(path5.join(PACKAGED_DOCS_DIR, name), "utf8") }]
-    })
-  );
-}
-server.registerTool(
-  "doctor",
-  {
-    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", "status", "usage"]).optional(),
-      days: external_exports.number().int().min(1).max(366).optional()
-    }).strict(),
-    outputSchema: doctorResultSchema,
-    annotations: { readOnlyHint: true }
-  },
-  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 (!isStructuredSuccess(report)) return report;
+    async ({ action, model, ollamaEndpoint, timeoutSeconds }, extra) => {
       try {
-        report.structuredContent.machine = JSON.parse(await machine(serveEndpoint));
-        delete report.structuredContent.machine_unavailable;
-      } catch (error2) {
-        if (report.structuredContent.machine == null) {
-          report.structuredContent.machine_unavailable = `freellama serve unreachable, so no machine profile: ${error2 instanceof Error ? error2.message : String(error2)}`;
+        if (action === "stop" && timeoutSeconds !== void 0) {
+          return errorResult(new Error('`timeoutSeconds` is valid only for action "pull".'));
+        }
+        const data = action === "pull" ? await ollamaPull(ollamaEndpoint, model, timeoutSeconds, {
+          signal: extra.signal,
+          onProgress: pullProgressReporter(extra)
+        }) : await ollamaFetch(ollamaEndpoint, "/api/generate", {
+          method: "POST",
+          body: { model, keep_alive: 0 },
+          signal: extra.signal
+        });
+        const payload = data;
+        if (typeof payload.error === "string" && payload.error) {
+          return errorResult(new Error(`${action} ${model} failed: ${payload.error}`));
         }
-      }
-      try {
-        report.structuredContent.platform_health = JSON.parse(await health(serveEndpoint));
+        return structuredResult(payload);
       } catch (error2) {
-        report.structuredContent.platform_health_unavailable = `freellama serve unreachable: ${error2 instanceof Error ? error2.message : String(error2)}`;
+        return errorResult(error2);
       }
-      const full = { ...report.structuredContent };
-      const flatConfig = full.ollama_env_config;
-      delete full.ollama_env_config;
-      if (view === "config") {
-        return structuredResult({
-          status: "ok",
-          summary: "Categorized Ollama configuration; values are source-qualified, not endpoint/PID proof.",
-          endpoint: full.endpoint,
-          ollama_config: full.ollama_config,
-          ollama_env_config_source: full.ollama_env_config_source,
-          ollama_env_config_warning: full.ollama_env_config_warning,
-          host_runtime_signals: full.host_runtime_signals
+    }
+  );
+  server2.registerTool(
+    "ollama_delete",
+    {
+      description: "DESTRUCTIVE AND IRREVERSIBLE. Use when: a human explicitly names the exact installed tag to delete. Do not use when: you inferred the cleanup yourself, or only want to unload (ollama_manage stop). Returns: the deleted tag.",
+      inputSchema: external_exports.object({
+        model: external_exports.string().min(1),
+        ollamaEndpoint: ollamaEndpointParam
+      }).strict(),
+      outputSchema: deleteResultSchema,
+      annotations: { destructiveHint: true }
+    },
+    async ({ model, ollamaEndpoint }) => {
+      try {
+        await ollamaFetch(ollamaEndpoint, "/api/delete", {
+          method: "DELETE",
+          body: { model }
         });
+        return structuredResult({ deleted: model });
+      } catch (error2) {
+        return errorResult(error2);
       }
-      if ((view ?? "summary") === "full") return structuredResult(full);
-      const running = full.running?.models ?? [];
-      const platformHealth = full.platform_health;
-      const compact = {
-        status: "ok",
-        summary: `Ollama ${String(full.version?.version ?? "unknown")}; ${running.length} resident model(s).`,
-        endpoint: full.endpoint,
-        ollama: { endpoint: full.endpoint, version: full.version, resident_model_count: running.length },
-        machine: full.machine,
-        local_conservative_config_posture: full.local_conservative_config_posture,
-        host_runtime_signals: full.host_runtime_signals,
-        ...view === "scheduler" ? {
-          scheduler: {
-            admission: platformHealth?.admission ?? null,
-            ollama_num_parallel: flatConfig?.OLLAMA_NUM_PARALLEL,
-            ollama_max_queue: flatConfig?.OLLAMA_MAX_QUEUE,
-            proof_level: "configured_and_snapshot_only; measure concurrent execution before claiming throughput"
-          }
-        } : {},
-        evidence: { runtime: "ollama_api_version_and_ps", configuration: full.ollama_env_config_source },
-        next: view === "scheduler" ? "Use run_task preview for a per-task advisory receipt." : 'Use models{view:"installed"} to select a local model.'
-      };
-      return structuredResult(compact);
-    } catch (error2) {
-      return errorResult(error2);
     }
-  }
-);
-server.registerTool(
-  "session",
-  {
-    description: "Use when: 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: true }
-  },
-  async ({ action, sessionId, endpoint }) => {
-    try {
-      if (action === "create") {
-        if (sessionId !== void 0) return errorResult(new Error("sessionId is only valid for action: delete or kill."));
-        return structuredResult(JSON.parse(await createSession2(endpoint)));
-      }
-      if (sessionId === void 0) return errorResult(new Error(`action: ${action} requires sessionId.`));
-      if (action === "kill") return parsedResult(await killSession2(endpoint, sessionId));
-      await deleteSession2(endpoint, sessionId);
-      return structuredResult({ session_id: sessionId, deleted: true });
-    } catch (error2) {
-      return errorResult(error2);
-    }
-  }
-);
-server.registerTool(
-  "models",
-  {
-    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, 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.`));
-      }
-      if (selectedView !== "library" && selectedView !== "raw" && [limit, cursor].some((value) => value !== void 0)) {
-        return errorResult(new Error(`view "${selectedView}" accepts neither pagination nor library search fields.`));
-      }
-      if (selectedView !== "detail" && includeVerbose !== void 0) {
-        return errorResult(new Error('`includeVerbose` is valid only for view "detail".'));
-      }
-      if (selectedView !== "detail" && selectedView !== "library" && model !== void 0) {
-        return errorResult(new Error('`model` is valid only for views "detail" and "library".'));
-      }
-      if (selectedView === "library" && model && [query, capabilities, order].some((value) => value !== void 0)) {
-        return errorResult(
-          new Error("Library step 2 accepts only `model` plus endpoint overrides; omit step-1 search fields.")
-        );
-      }
-      switch (selectedView) {
-        case "raw": {
-          const raw = await ollamaFetch(ollamaEndpoint, "/api/tags");
-          const models = Array.isArray(raw.models) ? raw.models : [];
-          const page = pageLiveList(models, limit, cursor, (model2) => String(model2.name ?? JSON.stringify(model2)));
-          return structuredResult({ ...raw, models: page.items, page: { returned: page.returned, total: page.total, next_cursor: page.next_cursor } });
-        }
-        case "resident": {
-          const managed = parsedResult(await 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.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);
-            const execution = entry.execution;
-            const assignedCpu = execution?.placement === "cpu";
-            const observedProcessor = execution?.observation?.processor;
-            const processor = observedProcessor === "cpu" || observedProcessor === "gpu" || observedProcessor === "mixed" ? observedProcessor : gpuPercent >= 100 ? "gpu" : gpuPercent <= 0 ? "cpu" : "mixed";
-            const mismatch = execution?.observation?.status === "mismatch";
-            return {
-              ...entry,
-              placement: {
-                gpu_percent: gpuPercent,
-                assigned: assignedCpu,
-                processor: processor === "gpu" ? "100% GPU" : processor === "cpu" ? "100% CPU" : `${gpuPercent}% GPU / ${100 - gpuPercent}% CPU`,
-                ...mismatch ? {
-                  warning: `Configured ${execution?.placement ?? "unknown"} backend disagrees with Ollama /api/ps: ${processor} was physically observed. This sample is excluded from adaptive routing feedback.`
-                } : processor === "mixed" ? {
-                  warning: 'Partially offloaded to CPU \u2014 expect a large slowdown. Free VRAM (`ollama_manage` action "stop") or lower the context length.'
-                } : {}
-              }
-            };
-          });
-          return structuredResult({ ...data, models });
-        }
-        case "detail": {
-          if (!model) {
-            return errorResult(new Error('view "detail" needs `model` set to an installed tag.'));
-          }
-          const data = await ollamaFetch(ollamaEndpoint, "/api/show", {
-            method: "POST",
-            body: { model }
-          });
-          let local = data;
-          if (includeLibrary && !data.remote_host && !data.remote_model && !/-cloud$/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(
-            ([key, value]) => key.endsWith(".context_length") && typeof value === "number"
-          );
-          return structuredResult({
-            ...rest,
-            max_context_length: contextEntry?.[1] ?? null,
-            knowledge: modelKnowledge(model, local, library, includeReadme),
-            ...includeVerbose ? { license, modelfile } : {}
-          });
-        }
-        case "library":
-          return await libraryLookup({ model, query, capabilities, order, limit, cursor, includeReadme, endpoint, ollamaEndpoint });
-        default: {
-          const result = parsedResult(await listModels2(endpoint));
-          if (!isStructuredSuccess(result)) return result;
-          const data = result.structuredContent;
-          return structuredResult({ ...data, models: await enrichInstalledModels(
-            data.models ?? [],
-            ollamaLibrary,
-            includeLibrary ?? false,
-            async (upstream) => (await ollamaFetch(upstream ?? ollamaEndpoint, "/api/tags")).models ?? []
-          ) });
-        }
+  );
+}
+
+// src/tools/delegate-research.ts
+import { existsSync as existsSync5 } from "node:fs";
+import { mkdtemp, readFile as readFile2, rm, writeFile } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import path5 from "node:path";
+
+// src/delegate.ts
+import { existsSync as existsSync4, readFileSync, realpathSync, statSync } from "node:fs";
+import path4 from "node:path";
+function researchErrorResult(error2, diagnostic) {
+  const result = errorResult(error2);
+  const summary = diagnostic ?? result.content[0].text;
+  return {
+    ...result,
+    content: [{ type: "text", text: summary }],
+    structuredContent: {
+      ...result.structuredContent ?? { error: summary },
+      answer: "",
+      summary,
+      verification: {
+        recommendation: "escalate",
+        grounded: false,
+        why: "The research request failed; no usable answer was produced."
       }
-    } catch (error2) {
-      return errorResult(error2);
-    }
-  }
-);
-async function libraryLookup({
-  model,
-  query,
-  capabilities,
-  order,
-  limit,
-  cursor,
-  includeReadme,
-  endpoint,
-  ollamaEndpoint
-}) {
-  const localState = async () => {
-    let installed = /* @__PURE__ */ new Set();
-    let memoryBytes = null;
-    try {
-      const tags = await ollamaFetch(ollamaEndpoint, "/api/tags");
-      installed = new Set((tags.models ?? []).map((m) => m.name ?? ""));
-    } catch {
-    }
-    try {
-      const profile = JSON.parse(await machine(endpoint));
-      memoryBytes = profile.memory_bytes ?? profile.unified_memory_bytes ?? null;
-    } catch {
     }
-    return { installed, memoryBytes };
   };
-  try {
-    if (model) {
-      const 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) => ({
-        ...t,
-        installed: installed2.has(t.tag),
-        fitsInMemory: budget && t.sizeBytes ? t.sizeBytes <= budget : null,
-        fitScope: "host_memory_budget_only"
-      }));
-      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: 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.`
-        } : {},
-        machineMemoryBytes: memoryBytes,
-        fitBudgetBytes: budget,
-        tags: tagPage.items,
-        page: { returned: tagPage.returned, total: tagPage.total, next_cursor: tagPage.next_cursor },
-        recommendationUnavailable: budget ? void 0 : "No machine profile (freellama serve unreachable), so memory fit could not be checked and no tag is recommended. Start serve, or read the sizes yourself.",
-        recommendation: best ? 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 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.',
-      models: parsed.map((m) => ({ ...m, installed: installedFamilies.has(m.name) }))
-    });
-  } catch (error2) {
-    return errorResult(error2);
-  }
 }
-registerContextTools(server);
-server.registerTool(
-  "run_task",
-  {
-    description: "Use when: supplied chat/tools/embeddings. Do not use when: files. Returns: response/receipts.",
-    inputSchema: external_exports.object({
-      endpoint: endpointParam,
-      task: taskParam.removeDefault().optional().describe("caller owns prompts and output format."),
-      objective: objectiveParam,
-      model: external_exports.string().min(1).optional(),
-      sessionId: external_exports.string().uuid().optional(),
-      scopeId: external_exports.string().uuid().optional(),
-      scopeRevision: external_exports.number().int().nonnegative().optional(),
-      contextTokens: external_exports.number().int().positive().optional().describe("Total input + output window (num_ctx)."),
-      executionPreference: executionPreferenceParam,
-      minPlacementEvidence: minPlacementEvidenceParam,
-      requiredCapabilities: requiredCapabilitiesParam,
-      prompt: external_exports.string().min(1).optional(),
-      systemPrompt: systemPromptParam,
-      images: external_exports.array(external_exports.string().min(1)).min(1).optional().describe("base64, no data-URI prefix; prompt mode only; requires an explicit tested vision model"),
-      messages: messagesParam.describe(
-        "Caller system prompts; no injected task instructions. Overrides prompt; extra fields preserved."
-      ),
-      input: external_exports.union([external_exports.string().min(1), external_exports.array(external_exports.string().min(1)).min(1)]).optional(),
-      tools: localToolsParam,
-      keepAlive: external_exports.string().min(1).optional().describe('"0" unloads now, "-1" pins; omitted = adaptive finite TTL'),
-      format: external_exports.union([external_exports.literal("json"), external_exports.record(external_exports.unknown())]).optional(),
-      think: external_exports.union([external_exports.boolean(), external_exports.enum(["low", "medium", "high"])]).optional(),
-      options: external_exports.record(external_exports.unknown()).optional().describe("num_ctx/contextTokens and num_gpu/placement are routing-owned; num_predict caps output."),
-      logprobs: external_exports.boolean().optional(),
-      topLogprobs: external_exports.number().int().nonnegative().optional(),
-      minConfidence: minConfidenceParam,
-      priority: external_exports.enum(["interactive", "normal", "background"]).optional(),
-      maxWaitSeconds: external_exports.number().int().positive().optional(),
-      timeoutSeconds: external_exports.number().int().positive().optional(),
-      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 }
-  },
-  async ({
-    endpoint,
-    task,
-    objective,
-    model,
-    sessionId,
-    scopeId,
-    scopeRevision,
-    contextTokens,
-    executionPreference,
-    minPlacementEvidence,
-    requiredCapabilities,
-    prompt,
-    systemPrompt,
-    images,
-    messages,
-    input,
-    tools,
-    keepAlive,
-    format,
-    think,
-    options,
-    logprobs,
-    topLogprobs,
-    minConfidence,
-    priority,
-    maxWaitSeconds,
-    timeoutSeconds,
-    defer,
-    returnEmbeddings,
-    preview
-  }) => {
+var MODEL_EVIDENCE = (() => {
+  const configured = process.env.FREELLAMA_MCP_MODEL_EVIDENCE;
+  const candidates = [
+    configured,
+    path4.join(REPO_ROOT, "packages/mcp/evidence/model-evidence.json")
+  ].filter((x) => Boolean(x));
+  for (const file of candidates) {
     try {
-      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],
-          ["tools", tools],
-          ["keepAlive", keepAlive],
-          ["format", format],
-          ["think", think],
-          ["options", options],
-          ["logprobs", logprobs],
-          ["topLogprobs", topLogprobs],
-          ["returnEmbeddings", returnEmbeddings],
-          ["maxWaitSeconds", maxWaitSeconds],
-          ["priority", priority],
-          ["timeoutSeconds", timeoutSeconds],
-          ["defer", defer]
-        ].filter(([, value]) => value !== void 0).map(([name]) => `\`${name}\``);
-        if (executionOnlyFields.length > 0) {
-          return errorResult(
-            new Error(
-              `\`preview:true\` accepts routing fields only; remove execution-only fields ${executionOnlyFields.join(", ")}. Use \`requiredCapabilities:["tools"]\` to preview tool capability, then make a separate execution call with the payload.`
-            )
-          );
-        }
-      }
-      const effectiveRequiredCapabilities = tools === void 0 ? requiredCapabilities : withRequiredCapability(requiredCapabilities, "tools");
-      if (topLogprobs !== void 0 && logprobs !== true) {
-        return errorResult(new Error("`topLogprobs` requires `logprobs:true`."));
-      }
-      if (!preview && task === "embedding") {
-        if (input === void 0) return errorResult(new Error('task "embedding" requires `input`.'));
-        if ([prompt, systemPrompt, images, messages, tools, format, think, logprobs, topLogprobs].some(
-          (value) => value !== void 0
-        )) {
-          return errorResult(
-            new Error('task "embedding" accepts `input`, `options`, and routing fields; chat controls are not valid.')
-          );
-        }
-      }
-      if (!preview && task !== "embedding") {
-        if (input !== void 0 || returnEmbeddings !== void 0) {
-          return errorResult(new Error('`input` and `returnEmbeddings` are valid only for task "embedding".'));
-        }
-        if (prompt === void 0 && systemPrompt === void 0 && (messages === void 0 || messages.length === 0)) {
-          return errorResult(new Error(`task "${task}" requires \`prompt\`, \`systemPrompt\`, or \`messages\`.`));
-        }
-        if (images !== void 0 && (messages !== void 0 || prompt === void 0)) {
-          return errorResult(
-            new Error("Top-level `images` is valid only with `prompt`; put images inside messages instead.")
-          );
-        }
-        if (images !== void 0 && model === void 0) {
-          return errorResult(new Error("Images require an explicit tested vision-capable `model`."));
-        }
-      }
-      if (preview) {
-        const result2 = parsedResult(
-          await route2(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
-        );
-        if (isStructuredSuccess(result2)) {
-          const refusal = belowConfidence(result2.structuredContent, minConfidence);
-          if (refusal) return refusal;
-        }
-        return result2;
-      }
-      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 route2(endpoint, canonicalTask, objective, model, sessionId, contextTokens, effectiveRequiredCapabilities, minConfidence, executionPreference, minPlacementEvidence)
-        );
-        if (!isStructuredSuccess(decision)) return decision;
-        const refusal = belowConfidence(decision.structuredContent, minConfidence);
-        if (refusal) return refusal;
-      }
-      const result = parsedResult(
-        await runTaskRequest2(endpoint ?? DEFAULT_SERVE_ENDPOINT, {
-          task: scopeId !== void 0 && task === void 0 ? void 0 : canonicalTask,
-          objective: objective ?? (scopeId === void 0 ? "balanced" : void 0),
-          model,
-          session_id: sessionId,
-          scope_id: scopeId,
-          scope_revision: scopeRevision,
-          context_tokens: contextTokens,
-          required_capabilities: effectiveRequiredCapabilities,
-          prompt,
-          images,
-          messages: taskMessages({ systemPrompt, messages, prompt, images }),
-          input,
-          tools,
-          keep_alive: keepAlive,
-          min_confidence: minConfidence,
-          priority: priority ?? "normal",
-          max_wait_seconds: maxWaitSeconds,
-          timeout_seconds: timeoutSeconds,
-          defer: defer ?? false,
-          execution_preference: executionPreference,
-          min_placement_evidence: minPlacementEvidence,
-          request_options: {
-            format,
-            think,
-            options,
-            logprobs,
-            top_logprobs: topLogprobs
-          }
-        })
-      );
-      if (isStructuredSuccess(result) && result.structuredContent.deferred === true) return result;
-      if (!returnEmbeddings && isStructuredSuccess(result)) {
-        const trimmed = summarizeEmbeddings(result.structuredContent);
-        if (trimmed) return withTaskTelemetry(structuredResult(trimmed));
-      }
-      return withTaskTelemetry(result);
+      if (!existsSync4(file)) continue;
+      const parsed = JSON.parse(readFileSync(file, "utf8"));
+      return parsed.models ?? {};
     } catch (error2) {
-      return errorResult(error2);
+      process.stderr.write(
+        `freellama-mcp: ignoring unreadable model evidence ${file}: ${error2 instanceof Error ? error2.message : String(error2)}
+`
+      );
     }
   }
-);
-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);
-    }
+  return {};
+})();
+var STRONG_MODEL_HINT = "Re-run with a model graded strong in packages/mcp/evidence/model-evidence.json";
+function assessDelegatedAnswer(question, evidenceCount, model) {
+  const grounded = evidenceCount > 0;
+  const evidence = MODEL_EVIDENCE[model];
+  if (evidence?.grade === "unusable") {
+    return {
+      recommendation: "escalate",
+      grounded,
+      why: `${model} is not viable for research on this machine (${evidence.note}). It answers fast and confidently while being wrong \u2014 a fast wrong answer is not a speed win. ${STRONG_MODEL_HINT}, or answer it yourself.`,
+      measuredBaseRate: `${model}: ${evidence.note}`
+    };
   }
-);
-server.registerTool(
-  "run_task_batch",
-  {
-    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 }
-  },
-  async ({ tasks, maxParallelism, endpoint }) => {
-    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.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.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 ${item.id}: images require an explicit tested vision model.`));
-        }
-      }
-      return withBatchTelemetry(parsedResult(await runTaskBatchRequest2(endpoint ?? DEFAULT_SERVE_ENDPOINT, {
-        tasks: tasks.map((item) => {
-          const task = item.task;
-          return {
-            id: item.id,
-            independent: item.independent,
-            task: {
-              task: task.task === "code_review" ? "coding" : task.task ?? (task.scopeId === void 0 ? "completion" : void 0),
-              objective: task.objective ?? (task.scopeId === void 0 ? "balanced" : void 0),
-              model: task.model,
-              session_id: task.sessionId,
-              scope_id: task.scopeId,
-              scope_revision: task.scopeRevision,
-              context_tokens: task.contextTokens,
-              execution_preference: task.executionPreference,
-              min_placement_evidence: task.minPlacementEvidence,
-              required_capabilities: task.requiredCapabilities,
-              priority: task.priority ?? "normal",
-              max_wait_seconds: task.maxWaitSeconds,
-              timeout_seconds: task.timeoutSeconds,
-              prompt: task.prompt,
-              images: task.images,
-              messages: taskMessages(task),
-              input: task.input,
-              tools: task.tools,
-              keep_alive: task.keepAlive,
-              min_confidence: task.minConfidence,
-              request_options: {
-                format: task.format,
-                think: task.think,
-                options: task.options,
-                logprobs: task.logprobs,
-                top_logprobs: task.topLogprobs
-              }
-            }
-          };
-        }),
-        max_parallelism: maxParallelism
-      })));
-    } catch (error2) {
-      return errorResult(error2);
-    }
+  if (!evidence) {
+    return {
+      recommendation: "verify",
+      grounded,
+      why: `${model} has no measured accuracy in this repo's benchmarks, so no base rate applies to this answer. Treat it as unverified until it has been evaluated \u2014 accuracy fell off a cliff below ~12B in the models that were measured.`,
+      measuredBaseRate: "no measured base rate for this model"
+    };
   }
-);
-server.registerTool(
-  "ollama_manage",
-  {
-    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()
-    }).strict(),
-    outputSchema: manageResultSchema,
-    annotations: { destructiveHint: false }
-  },
-  async ({ action, model, ollamaEndpoint, timeoutSeconds }, extra) => {
-    try {
-      if (action === "stop" && timeoutSeconds !== void 0) {
-        return errorResult(new Error('`timeoutSeconds` is valid only for action "pull".'));
-      }
-      const data = action === "pull" ? await ollamaPull(ollamaEndpoint, model, timeoutSeconds, {
-        signal: extra.signal,
-        onProgress: pullProgressReporter(extra)
-      }) : await ollamaFetch(ollamaEndpoint, "/api/generate", {
-        method: "POST",
-        body: { model, keep_alive: 0 },
-        signal: extra.signal
-      });
-      const payload = data;
-      if (typeof payload.error === "string" && payload.error) {
-        return errorResult(new Error(`${action} ${model} failed: ${payload.error}`));
-      }
-      return structuredResult(payload);
-    } catch (error2) {
-      return errorResult(error2);
-    }
+  if (evidence.grade === "weak" && grounded) {
+    return {
+      recommendation: "verify",
+      grounded,
+      why: `${model} holds up on simple single-file lookups but not beyond (${evidence.note}). Check the evidence trail, or ${STRONG_MODEL_HINT.toLowerCase()} if the answer matters.`,
+      measuredBaseRate: `${model}: ${evidence.note}`
+    };
   }
-);
-server.registerTool(
-  "ollama_delete",
-  {
-    description: "DESTRUCTIVE AND IRREVERSIBLE. Use when: human names exact tag. Do not use when: inferred cleanup. Returns: deleted tag.",
-    inputSchema: external_exports.object({
-      model: external_exports.string().min(1),
-      ollamaEndpoint: ollamaEndpointParam
-    }).strict(),
-    outputSchema: deleteResultSchema,
-    annotations: { destructiveHint: true }
-  },
-  async ({ model, ollamaEndpoint }) => {
+  if (!grounded) {
+    return {
+      recommendation: "escalate",
+      grounded: false,
+      why: "The model answered without reading a single file, so this is parametric recall, not research \u2014 the failure mode where `run_task` was verified inventing wrong facts about this project's own architecture. Re-ask with a narrower question, or answer it yourself.",
+      measuredBaseRate: "ungrounded answers have no measured accuracy \u2014 they were never the tested path"
+    };
+  }
+  const judgmentSignals = /\b(should (we|i|you)|better|best|worth|review|assess|evaluate|improve|opinion|recommend|why is|is it (good|safe|correct)|design|refactor)\b/i;
+  if (judgmentSignals.test(question)) {
+    return {
+      recommendation: "verify",
+      grounded: true,
+      why: "The question reads as a judgment call (keyword heuristic). Measured accuracy on judgment calls was well below grounded lookups, and the tone is identical either way. Check the evidence trail against the claim before acting on it.",
+      measuredBaseRate: `${model}: ${evidence.note}; measured on lookups, not judgment calls`
+    };
+  }
+  if (evidenceCount > 5) {
+    return {
+      recommendation: "verify",
+      grounded: true,
+      why: `${evidenceCount} tool calls is outside the 1-5 file envelope this tool was measured on. Wide searches are where it drifts; spot-check the evidence trail.`,
+      measuredBaseRate: `${model}: ${evidence.note}; measured within a 1-5 file scope`
+    };
+  }
+  return {
+    recommendation: "accept",
+    grounded: true,
+    why: `Grounded in ${evidenceCount} tool call(s) within the measured 1-5 file envelope, the question is lookup-shaped, and ${model} is measured strong for this (${evidence.note}).`,
+    measuredBaseRate: `${model}: ${evidence.note}`
+  };
+}
+function extractExistingWorkspacePath(command, workspace) {
+  let root;
+  try {
+    root = realpathSync(workspace);
+  } catch {
+    return null;
+  }
+  const tokens = command.match(/(?:[^\s"'\\]+|"(?:\\.|[^"])*"|'[^']*')+/g) ?? [];
+  const patternFlags = /* @__PURE__ */ new Set(["-name", "-iname", "-path", "-ipath", "-regex", "-iregex", "-wholename"]);
+  for (let index = 0; index < tokens.length; index += 1) {
+    const raw = tokens[index] ?? "";
+    if (index > 0 && patternFlags.has(tokens[index - 1] ?? "")) continue;
+    const token = raw.replace(/^["']|["']$/g, "").replace(/[|;&)]+$/g, "").replace(/:\d+(?::\d+)?$/, "");
+    if (!token || token.startsWith("-") || /[*?$`]/.test(token)) continue;
+    const candidate = path4.isAbsolute(token) ? token : path4.resolve(root, token);
+    if (!existsSync4(candidate)) continue;
     try {
-      await ollamaFetch(ollamaEndpoint, "/api/delete", {
-        method: "DELETE",
-        body: { model }
-      });
-      return structuredResult({ deleted: model });
-    } catch (error2) {
-      return errorResult(error2);
+      const resolved = realpathSync(candidate);
+      const relative = path4.relative(root, resolved);
+      const withinRoot = relative && relative !== ".." && !relative.startsWith(`..${path4.sep}`);
+      if (withinRoot && statSync(resolved).isFile()) return relative;
+    } catch {
     }
   }
-);
-async function routedResearchModel(endpoint, executionPreference, minPlacementEvidence) {
+  return null;
+}
+var adapterCallSchema = external_exports.object({
+  raw_name: external_exports.string().optional(),
+  status: external_exports.string().optional(),
+  arguments: external_exports.object({
+    tool: external_exports.string().optional(),
+    command: external_exports.string().optional(),
+    queries: external_exports.object({ path: external_exports.string().optional() }).passthrough().optional()
+  }).passthrough().optional()
+}).passthrough();
+var adapterResultSchema = external_exports.object({
+  final_answer: external_exports.string(),
+  tool_calls: external_exports.array(adapterCallSchema).default([]),
+  usage: external_exports.object({
+    input_tokens: external_exports.number().nullable().optional(),
+    output_tokens: external_exports.number().nullable().optional()
+  }).passthrough().default({}),
+  model_metadata: external_exports.object({
+    context_compactions: external_exports.number().int().nonnegative().optional(),
+    runtime_config: external_exports.record(external_exports.string(), external_exports.unknown()).optional(),
+    context_management: external_exports.object({
+      token_counting: external_exports.enum(["configured_estimate", "model_calibrated_estimate"]),
+      estimate_scale: external_exports.number().positive(),
+      calibration_samples: external_exports.number().int().nonnegative(),
+      pinned_overflow: external_exports.enum(["error", "clip"]),
+      compactions: external_exports.number().int().nonnegative()
+    }).passthrough().optional()
+  }).passthrough().optional()
+});
+function parseAdapterResult(raw) {
+  return adapterResultSchema.parse(JSON.parse(raw));
+}
+
+// src/tools/delegate-research.ts
+var researchResultSchema = external_exports.object({
+  answer: external_exports.string(),
+  summary: external_exports.string(),
+  citations: external_exports.unknown().optional(),
+  verification: external_exports.unknown(),
+  usage: external_exports.unknown().optional(),
+  telemetry: external_exports.unknown().optional()
+}).passthrough();
+async function routedResearchModel(endpoint) {
   const decision = JSON.parse(
-    await route2(endpoint, "coding", "balanced", void 0, void 0, void 0, [], void 0, executionPreference, minPlacementEvidence)
+    await route2(endpoint, "coding", "balanced", void 0, void 0, void 0, [], void 0, void 0, void 0)
   );
   if (typeof decision.selected_model !== "string" || !decision.selected_model) {
     throw new Error(
@@ -23651,196 +23750,202 @@ async function routedResearchModel(endpoint, executionPreference, minPlacementEv
   }
   return decision.selected_model;
 }
-server.registerTool(
-  "delegate_research",
-  {
-    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(),
-      agent: external_exports.object({
-        maxTurns: external_exports.number().int().positive().optional(),
-        contextTokens: external_exports.number().int().positive().optional(),
-        outputTokens: external_exports.number().int().positive().optional(),
-        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(),
-        requestTimeoutSeconds: external_exports.number().positive().optional(),
-        toolTimeoutSeconds: external_exports.number().positive().optional()
-      }).strict().optional()
-    }).strict(),
-    outputSchema: researchResultSchema,
-    annotations: { destructiveHint: false }
-  },
-  async ({ question, workspacePath, adapter, model, endpoint, executionPreference, minPlacementEvidence, legacyText, agent }, extra) => {
-    const chosenAdapter = adapter ?? DEFAULT_RESEARCH_ADAPTER;
-    let resolvedWorkspace;
-    let chosenModel;
-    try {
-      resolvedWorkspace = await assertAllowedWorkspace(workspacePath);
-      chosenModel = model ?? DEFAULT_DELEGATE_MODEL ?? await routedResearchModel(endpoint, executionPreference, minPlacementEvidence);
-    } catch (error2) {
-      return researchErrorResult(error2);
-    }
-    const known = MODEL_EVIDENCE[chosenModel];
-    if (known?.grade === "unusable") {
-      return structuredResult({
-        adapter: chosenAdapter,
-        verification: assessDelegatedAnswer(question, 0, chosenModel),
-        answer: "",
-        toolCallCount: 0,
-        usage: { inputTokens: null, outputTokens: null },
-        telemetry: costTelemetry({ inputTokens: null, outputTokens: null }, EXTERNAL_COST),
-        evidence: [],
-        summary: `Refused before running: ${chosenModel} is measured unusable for research here (${known.note}). Re-run with a ~27B model (see README), or answer it yourself.`
-      });
-    }
-    const dir = await mkdtemp(path5.join(tmpdir(), "freellama-delegate-"));
-    const promptFile = path5.join(dir, "prompt.md");
-    const resultFile = path5.join(dir, "result.json");
-    try {
-      await writeFile(promptFile, `${question}
-`, "utf8");
-      const adapter2 = RESEARCH_ADAPTERS[chosenAdapter];
-      if (!existsSync5(adapter2)) {
-        return researchErrorResult(
-          new Error(
-            `research adapter not found at ${adapter2}. 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: {
-          ...delegateEnvironment(),
-          FREELLAMA_TARGET_MODEL: chosenModel,
-          FREELLAMA_AGENT_MANAGED_ENDPOINT: endpoint ?? DEFAULT_SERVE_ENDPOINT,
-          FREELLAMA_AGENT_EXECUTION_PREFERENCE: executionPreference ?? "auto",
-          FREELLAMA_AGENT_MIN_PLACEMENT_EVIDENCE: minPlacementEvidence ?? "configured",
-          FREELLAMA_AGENT_TOKEN_CALIBRATION_DIR: DEFAULT_TOKEN_CALIBRATION_DIR,
-          FREELLAMA_BENCH_WORKSPACE: resolvedWorkspace,
-          FREELLAMA_BENCH_PROMPT: promptFile,
-          FREELLAMA_AGENT_RESULT: resultFile,
-          FREELLAMA_AGENT_MAX_TURNS: String(agent?.maxTurns ?? DEFAULT_DELEGATE_MAX_TURNS),
-          ...agent?.contextTokens !== void 0 ? { FREELLAMA_AGENT_NUM_CTX: String(agent.contextTokens) } : {},
-          ...agent?.outputTokens !== void 0 ? { FREELLAMA_AGENT_NUM_PREDICT: String(agent.outputTokens) } : {},
-          ...agent?.temperature !== void 0 ? { FREELLAMA_AGENT_TEMPERATURE: String(agent.temperature) } : {},
-          ...agent?.seed !== void 0 ? { FREELLAMA_AGENT_SEED: String(agent.seed) } : {},
-          ...agent?.think !== void 0 ? { FREELLAMA_AGENT_THINK: String(agent.think) } : {},
-          ...agent?.keepAlive !== void 0 ? { FREELLAMA_AGENT_KEEP_ALIVE: agent.keepAlive } : {},
-          ...agent?.requestTimeoutSeconds !== void 0 ? { FREELLAMA_AGENT_REQUEST_TIMEOUT_SECONDS: String(agent.requestTimeoutSeconds) } : {},
-          ...agent?.toolTimeoutSeconds !== void 0 ? { FREELLAMA_AGENT_TOOL_TIMEOUT_SECONDS: String(agent.toolTimeoutSeconds) } : {}
-        },
-        timeout: DEFAULT_DELEGATE_TIMEOUT_SECONDS * 1e3,
-        // The answer is read from `resultFile`, never from stdout — but execFile's default 1 MB
-        // maxBuffer still applies to the agent's own progress logging, and overflowing it kills
-        // the subprocess and surfaces as a research failure with no explanation. Give the logs
-        // room; nothing here is proportional to the size of the answer.
-        maxBuffer: 32 * 1024 * 1024,
-        // On timeout, don't negotiate: SIGTERM can be swallowed by a python process blocked in a
-        // long model call, which would leave the child (and its VRAM) alive past the deadline the
-        // caller was promised.
-        killSignal: "SIGKILL"
-      });
-      liveDelegates.add(running.child);
-      let adapterError = null;
+function registerDelegateResearchTool(server2) {
+  server2.registerTool(
+    "delegate_research",
+    {
+      description: "Use when: a bounded read-only question about files under workspacePath (where is X, what calls Y). Do not use when: editing, external facts, or final judgment. Returns: citations/answer; discard verification.recommendation=escalate, otherwise verify citations.",
+      inputSchema: external_exports.object({
+        question: external_exports.string().min(1),
+        workspacePath: external_exports.string().min(1),
+        adapter: external_exports.enum(["bash", "octocode"]).optional(),
+        model: external_exports.string().min(1).optional(),
+        endpoint: endpointParam,
+        agent: external_exports.object({
+          maxTurns: external_exports.number().int().positive().optional(),
+          contextTokens: external_exports.number().int().positive().optional(),
+          outputTokens: external_exports.number().int().positive().optional()
+        }).strict().optional()
+      }).strict(),
+      outputSchema: researchResultSchema,
+      annotations: { destructiveHint: false }
+    },
+    async ({ question, workspacePath, adapter, model, endpoint, agent }, extra) => {
+      const chosenAdapter = adapter ?? DEFAULT_RESEARCH_ADAPTER;
+      let resolvedWorkspace;
+      let chosenModel;
       try {
-        await running;
+        resolvedWorkspace = await assertAllowedWorkspace(workspacePath);
+        chosenModel = model ?? DEFAULT_DELEGATE_MODEL ?? await routedResearchModel(endpoint);
       } catch (error2) {
-        adapterError = error2;
-      } finally {
-        liveDelegates.delete(running.child);
+        return researchErrorResult(error2);
       }
-      let result;
-      try {
-        result = parseAdapterResult(await readFile(resultFile, "utf8"));
-      } catch {
-        return researchErrorResult(
-          adapterError ?? new Error(
-            `research adapter exited without writing a readable result file \u2014 it was killed before it could report. Check that the model is loadable and that ${DEFAULT_DELEGATE_TIMEOUT_SECONDS}s is enough for this question.`
-          )
-        );
+      const known = MODEL_EVIDENCE[chosenModel];
+      if (known?.grade === "unusable") {
+        return structuredResult({
+          adapter: chosenAdapter,
+          verification: assessDelegatedAnswer(question, 0, chosenModel),
+          answer: "",
+          toolCallCount: 0,
+          usage: { inputTokens: null, outputTokens: null },
+          telemetry: costTelemetry({ inputTokens: null, outputTokens: null }, EXTERNAL_COST),
+          evidence: [],
+          summary: `Refused before running: ${chosenModel} is measured unusable for research here (${known.note}). ${STRONG_MODEL_HINT}, or answer it yourself.`
+        });
       }
-      const evidence = result.tool_calls.map((call, index) => {
-        const target = call.arguments?.queries?.path;
-        const rawDetail = call.arguments?.command ?? null;
-        const detail = rawDetail ? clipText(rawDetail, 400) : null;
-        return {
-          step: index + 1,
-          tool: call.arguments?.tool ?? call.raw_name ?? "?",
-          // The adapters record "ok" | "error" | "repeat" per call. Carrying it through is what
-          // lets a reader tell a run that read three files from one that failed three commands —
-          // indistinguishable in the trail before, and graded identically.
-          status: call.status ?? "ok",
-          path: target ? path5.relative(resolvedWorkspace, target) : detail ? extractExistingWorkspacePath(detail, resolvedWorkspace) : null,
-          detail,
-          detail_truncated: rawDetail !== detail
-        };
-      });
-      const succeeded = evidence.filter((step) => step.status === "ok");
-      const failed = evidence.length - succeeded.length;
-      const evidenceText = evidence.map(
-        (step) => `  ${step.step}. ${step.tool}${step.status === "ok" ? "" : ` [${step.status}]`}${step.path ? ` -> ${step.path}` : ""}${step.detail ? `: ${clipText(step.detail, 400)}` : ""}`
-      ).join("\n");
-      const citations = succeeded.map((step) => ({ step: step.step, tool: step.tool, path: step.path }));
-      if (adapterError) {
-        const diagnostic = `research adapter failed: ${result.final_answer}` + (evidenceText ? `
+      const dir = await mkdtemp(path5.join(tmpdir(), "freellama-delegate-"));
+      const promptFile = path5.join(dir, "prompt.md");
+      const resultFile = path5.join(dir, "result.json");
+      try {
+        await writeFile(promptFile, `${question}
+`, "utf8");
+        const adapter2 = RESEARCH_ADAPTERS[chosenAdapter];
+        if (!existsSync5(adapter2)) {
+          return researchErrorResult(
+            new Error(
+              `research adapter not found at ${adapter2}. It is compiled to <package>/dist/research by the package build. Reinstall, or run \`yarn build\` from packages/mcp/ to rebuild it.`
+            )
+          );
+        }
+        const running = runDelegate(process.execPath, [adapter2], {
+          signal: extra.signal,
+          env: {
+            ...delegateEnvironment(),
+            FREELLAMA_TARGET_MODEL: chosenModel,
+            FREELLAMA_AGENT_MANAGED_ENDPOINT: endpoint ?? DEFAULT_SERVE_ENDPOINT,
+            FREELLAMA_AGENT_TOKEN_CALIBRATION_DIR: DEFAULT_TOKEN_CALIBRATION_DIR,
+            FREELLAMA_BENCH_WORKSPACE: resolvedWorkspace,
+            FREELLAMA_BENCH_PROMPT: promptFile,
+            FREELLAMA_AGENT_RESULT: resultFile,
+            FREELLAMA_AGENT_MAX_TURNS: String(agent?.maxTurns ?? DEFAULT_DELEGATE_MAX_TURNS),
+            ...agent?.contextTokens !== void 0 ? { FREELLAMA_AGENT_NUM_CTX: String(agent.contextTokens) } : {},
+            ...agent?.outputTokens !== void 0 ? { FREELLAMA_AGENT_NUM_PREDICT: String(agent.outputTokens) } : {}
+          },
+          timeoutMs: DEFAULT_DELEGATE_TIMEOUT_SECONDS * 1e3,
+          // The answer is read from `resultFile`, never from stdout — but execFile's usual 1 MB
+          // maxBuffer still applies to the agent's own progress logging, and overflowing it kills
+          // the subprocess and surfaces as a research failure with no explanation. Give the logs
+          // room; nothing here is proportional to the size of the answer.
+          maxBuffer: 32 * 1024 * 1024
+          // On timeout, abort, or overflow runDelegate SIGKILLs the whole process group: SIGTERM can
+          // be swallowed by an adapter blocked in a long model call, and killing the adapter alone
+          // would orphan `npx octocode` and sandbox shells.
+        });
+        liveDelegates.add(running.child);
+        let adapterError = null;
+        try {
+          await running.done;
+        } catch (error2) {
+          adapterError = error2;
+        } finally {
+          liveDelegates.delete(running.child);
+          killDelegateTree(running.child);
+        }
+        let result;
+        try {
+          result = parseAdapterResult(await readFile2(resultFile, "utf8"));
+        } catch {
+          return researchErrorResult(
+            adapterError ?? new Error(
+              `research adapter exited without writing a readable result file \u2014 it was killed before it could report. Check that the model is loadable and that ${DEFAULT_DELEGATE_TIMEOUT_SECONDS}s is enough for this question.`
+            )
+          );
+        }
+        const evidence = result.tool_calls.map((call, index) => {
+          const target = call.arguments?.queries?.path;
+          const rawDetail = call.arguments?.command ?? null;
+          const detail = rawDetail ? clipText(rawDetail, 400) : null;
+          return {
+            step: index + 1,
+            tool: call.arguments?.tool ?? call.raw_name ?? "?",
+            // The adapters record "ok" | "error" | "repeat" per call. Carrying it through is what
+            // lets a reader tell a run that read three files from one that failed three commands —
+            // indistinguishable in the trail before, and graded identically.
+            status: call.status ?? "ok",
+            path: target ? path5.relative(resolvedWorkspace, target) : detail ? extractExistingWorkspacePath(detail, resolvedWorkspace) : null,
+            detail,
+            detail_truncated: rawDetail !== detail
+          };
+        });
+        const succeeded = evidence.filter((step) => step.status === "ok");
+        const failed = evidence.length - succeeded.length;
+        const evidenceText = evidence.map(
+          (step) => `  ${step.step}. ${step.tool}${step.status === "ok" ? "" : ` [${step.status}]`}${step.path ? ` -> ${step.path}` : ""}${step.detail ? `: ${step.detail}` : ""}`
+        ).join("\n");
+        const citations = succeeded.map((step) => ({ step: step.step, tool: step.tool, path: step.path }));
+        if (adapterError) {
+          const diagnostic = `research adapter failed: ${result.final_answer}` + (evidenceText ? `
 Evidence collected before the failure:
 ${evidenceText}` : "");
-        const refusal = external_exports.object({ receipt: external_exports.object({ error: external_exports.string() }).passthrough() }).safeParse(result.model_metadata?.transport_error);
-        if (refusal.success) {
-          return researchErrorResult(new Error(JSON.stringify(refusal.data.receipt)), diagnostic);
-        }
-        return researchErrorResult(new Error(diagnostic));
-      }
-      const verification = assessDelegatedAnswer(question, succeeded.length, chosenModel);
-      const summary = `Delegated answer ready: ${result.tool_calls.length} tool call(s)` + (failed > 0 ? `, ${failed} of which did not succeed` : "") + `; ${result.usage.input_tokens ?? "?"} input / ${result.usage.output_tokens ?? "?"} output local tokens; verification=${verification.recommendation}. Read structuredContent.answer and citations.`;
-      const payload = {
-        adapter: chosenAdapter,
-        verification,
-        answer: result.final_answer,
-        // recommendation + why live under `verification`; `citations` completes the triple so a
-        // caller never has to read `summary` to act on the result.
-        citations,
-        toolCallCount: result.tool_calls.length,
-        successfulToolCallCount: succeeded.length,
-        usage: {
-          inputTokens: result.usage.input_tokens,
-          outputTokens: result.usage.output_tokens
-        },
-        telemetry: costTelemetry({
-          inputTokens: result.usage.input_tokens ?? null,
-          outputTokens: result.usage.output_tokens ?? null
-        }, EXTERNAL_COST),
-        contextManagement: result.model_metadata?.context_management ?? null,
-        execution: {
-          preference: executionPreference ?? "auto",
-          minPlacementEvidence: minPlacementEvidence ?? "configured",
-          receipts: result.model_metadata?.execution_receipts ?? []
-        },
-        evidence,
-        summary
-      };
-      const cited = [...new Set(citations.map((citation) => citation.path).filter(Boolean))];
-      const answerText = `${clipText(result.final_answer, ANSWER_TEXT_MAX_CHARS)}
+          const refusal = external_exports.object({ receipt: external_exports.object({ error: external_exports.string() }).passthrough() }).safeParse(result.model_metadata?.transport_error);
+          if (refusal.success) {
+            return researchErrorResult(new Error(JSON.stringify(refusal.data.receipt)), diagnostic);
+          }
+          return researchErrorResult(new Error(diagnostic));
+        }
+        const verification = assessDelegatedAnswer(question, succeeded.length, chosenModel);
+        const summary = `Delegated answer ready: ${result.tool_calls.length} tool call(s)` + (failed > 0 ? `, ${failed} of which did not succeed` : "") + `; ${result.usage.input_tokens ?? "?"} input / ${result.usage.output_tokens ?? "?"} output local tokens; verification=${verification.recommendation}. Read structuredContent.answer and citations.`;
+        const payload = {
+          adapter: chosenAdapter,
+          verification,
+          answer: result.final_answer,
+          // recommendation + why live under `verification`; `citations` completes the triple so a
+          // caller never has to read `summary` to act on the result.
+          citations,
+          toolCallCount: result.tool_calls.length,
+          successfulToolCallCount: succeeded.length,
+          usage: {
+            inputTokens: result.usage.input_tokens,
+            outputTokens: result.usage.output_tokens
+          },
+          telemetry: costTelemetry({
+            inputTokens: result.usage.input_tokens ?? null,
+            outputTokens: result.usage.output_tokens ?? null
+          }, EXTERNAL_COST),
+          contextManagement: result.model_metadata?.context_management ?? null,
+          execution: { receipts: result.model_metadata?.execution_receipts ?? [] },
+          evidence,
+          summary
+        };
+        const cited = [...new Set(citations.map((citation) => citation.path).filter(Boolean))];
+        const answerText = `${clipText(result.final_answer, ANSWER_TEXT_MAX_CHARS)}
 
 verification: ${verification.recommendation}` + (cited.length ? `
 cited: ${cited.slice(0, 20).join(", ")}` : "") + `
 (${chosenModel}, ${result.tool_calls.length} tool call(s))`;
-      return structuredResult(payload, legacyText === true ? { legacyJson: true } : { text: answerText });
-    } catch (error2) {
-      return researchErrorResult(error2);
-    } finally {
-      await rm(dir, { recursive: true, force: true });
+        return structuredResult(payload, { text: answerText });
+      } catch (error2) {
+        return researchErrorResult(error2);
+      } finally {
+        await rm(dir, { recursive: true, force: true });
+      }
     }
-  }
+  );
+}
+
+// src/index.ts
+installProcessLifecycle();
+var INSTRUCTIONS = `models{view:"installed"}; models{view:"resident"}.
+run_task preview never executes; code_review aliases coding. requiredCapabilities:["tools"] previews a tool route; omit preview and supply the payload. Caller runs tool calls.
+quality objective needs a task policy or explicit model; medium confidence needs policy+benchmark; observed placement needs /api/ps proof.
+Scoped run_task needs scopeId+scopeRevision. defer:true: job.id\u2192task_jobs.jobId unchanged. Restart loses scopes/sessions/jobs.
+Check isError first; prefer structuredContent, else content[].text. page.next_cursor\u2192cursor unchanged.
+caller owns task decomposition/prompts/format/verification; findings are candidates, not accepted defects. operator owns endpoints, exact --cpu-model assignments/lifecycle. Ollama plus the OS/driver run physical CPU/GPU.
+ask approval for one exact tag and reported size before ollama_manage pull; search or recommendation is never download permission. Stop/delete need exact-tag approval.
+serve :11435; default-endpoint autostart optional. Docs: freellama://docs/index.`;
+var server = new McpServer(
+  { name: "freellama", version: SERVER_VERSION },
+  { instructions: INSTRUCTIONS }
 );
-var transport = new StdioServerTransport();
+registerDocsResources(server);
+registerDoctorTool(server);
+registerSessionTool(server);
+registerModelsTool(server);
+registerScopeTool(server);
+registerRunTaskTool(server);
+registerTaskJobsTool(server);
+registerOllamaLifecycleTools(server);
+registerDelegateResearchTool(server);
+var transport = withoutSchemaDialect(new StdioServerTransport());
 await server.connect(transport);
+startServeInBackground();