@tuned-tensor/cli 0.26.2 → 0.27.0

package/dist/index.js

@@ -1050,9 +1050,9 @@ async function launchBackgroundSupervisor(args2) {
       stdio: "ignore"
     });
     if (child.pid === void 0) {
-      await new Promise((resolve13, reject) => {
+      await new Promise((resolve14, reject) => {
         child.once("error", reject);
-        child.once("spawn", () => resolve13());
+        child.once("spawn", () => resolve14());
       });
     }
   } catch (error) {
@@ -1117,15 +1117,15 @@ async function superviseBackgroundRun(options) {
       detached: process.platform !== "win32",
       stdio: ["ignore", resultFd, logFd]
     });
-    const closed = new Promise((resolve13) => {
+    const closed = new Promise((resolve14) => {
       child.once("error", (error) => {
         state.spawnError = error;
-        resolve13();
+        resolve14();
       });
       child.once("close", (code2, signal) => {
         state.exitCode = code2;
         state.exitSignal = signal;
-        resolve13();
+        resolve14();
       });
     });
     if (child.pid !== void 0) {
@@ -1705,7 +1705,7 @@ function formatCents(cents) {
   return `$${(cents / 100).toFixed(2)}`;
 }
 function sleep(ms) {
-  return new Promise((resolve13) => setTimeout(resolve13, ms));
+  return new Promise((resolve14) => setTimeout(resolve14, ms));
 }
 function isTransientError(error) {
   if (error instanceof ApiError) return error.status === 429 || error.status >= 500;
@@ -2215,10 +2215,10 @@ function registerLabelCommands(parent) {
       console.log(`Follow progress with \`tt cloud label watch ${shortId(job.id)}\`.`);
       return;
     }
-    const finished = await watchJob(job.id, opts);
-    if (isJsonMode()) return printJson(finished);
+    const finished2 = await watchJob(job.id, opts);
+    if (isJsonMode()) return printJson(finished2);
     printSuccess(
-      `Labeling ${finished.status === "awaiting_review" ? "complete" : finished.status}: ${finished.labeled_count}/${finished.row_count} rows labeled.`
+      `Labeling ${finished2.status === "awaiting_review" ? "complete" : finished2.status}: ${finished2.labeled_count}/${finished2.row_count} rows labeled.`
     );
     console.log(
       `Review with \`tt cloud label rows ${shortId(job.id)} --status labeled\`, then \`tt cloud label download ${shortId(job.id)}\` to write local training files.`
@@ -3577,7 +3577,7 @@ async function fetchHuggingFaceSearch(kind, query, limit, filters, options) {
   let lastError = new Error("Hugging Face search failed.");
   for (let attempt = 0; attempt < 2; attempt += 1) {
     if (attempt > 0) {
-      await new Promise((resolve13) => setTimeout(resolve13, options.huggingFaceRetryDelayMs ?? 750));
+      await new Promise((resolve14) => setTimeout(resolve14, options.huggingFaceRetryDelayMs ?? 750));
     }
     try {
       const payload = await fetchHuggingFaceJson(url, options);
@@ -3958,7 +3958,7 @@ function createTunedTensorTools(api, options = {}) {
             dry_run: "tt pipeline run --dry-run --spec tunedtensor.json",
             run: "tt pipeline run --spec tunedtensor.json"
           },
-          note: "Runs locally on CPU, Apple MPS or CUDA; runtime.gpu (AWS) is not used. The tuned model directory loads in Python with laya.load(<model_dir>).",
+          note: "Runs locally on CPU, Apple MPS or CUDA; runtime.gpu (AWS) is not used. The tuned model directory loads in Python with laya.load(<model_dir>). For paid serverless CPU hosting, the user can run tt cloud endpoints create --model <model_dir> --name <name>; see docs/local-runtime/hosting.md. This uploads weights to the TT account, returns an authenticated inference URL and scoped key, and charges prepaid credits per successful request.",
           ...host ? { host } : {}
         };
       }
@@ -4602,11 +4602,11 @@ function createLocalAgentClient(options) {
     const runs = options.backgroundRuns?.store;
     if (!runs) return void 0;
     try {
-      const finished = (await runs.list()).filter((run) => run.requested_by.thread_id === threadId && !isActiveBackgroundRun(run) && !run.reported_at);
-      if (finished.length === 0) return void 0;
-      await runs.markReported(finished.map((run) => run.id));
+      const finished2 = (await runs.list()).filter((run) => run.requested_by.thread_id === threadId && !isActiveBackgroundRun(run) && !run.reported_at);
+      if (finished2.length === 0) return void 0;
+      await runs.markReported(finished2.map((run) => run.id));
       return `[Background run outcomes, recorded by tt]
-${finished.map((run) => JSON.stringify(describeBackgroundRun(run))).join("\n")}
+${finished2.map((run) => JSON.stringify(describeBackgroundRun(run))).join("\n")}
 These runs finished since your last turn. Mention them briefly when relevant; use get_background_run for logs.`;
     } catch {
       return void 0;
@@ -6787,12 +6787,12 @@ async function promptKeyChoice(message, input = stdin, output = stdout) {
       terminal.setRawMode(true);
       input.resume();
       output.write(`${message} ${hint()}`);
-      const choice = await new Promise((resolve13) => {
+      const choice = await new Promise((resolve14) => {
         const finish = (decision) => {
           input.removeListener("keypress", onKeypress);
           input.removeListener("end", onEnd);
           input.removeListener("close", onEnd);
-          resolve13(decision);
+          resolve14(decision);
         };
         const onKeypress = (text3, key) => {
           const decided = keyChoiceFor(key, text3);
@@ -12272,31 +12272,166 @@ Eval Results (${data3._evals.length}):`);
   });
 }
 
+// src/commands/endpoints.ts
+import { createHash as createHash8, randomUUID as randomUUID10 } from "crypto";
+import { createReadStream as createReadStream2, readFileSync as readFileSync11, statSync as statSync4, openSync as openSync3, readSync, closeSync as closeSync3 } from "fs";
+import { open as open4 } from "fs/promises";
+import { finished } from "stream/promises";
+import { resolve as resolve10, join as join15 } from "path";
+import { z } from "zod";
+async function inspectCheckpoint(modelPath) {
+  const directory = resolve10(modelPath);
+  const config = JSON.parse(readFileSync11(join15(directory, "rl_agent_config.json"), "utf8"));
+  if (config.encoder !== "answerdotai/ModernBERT-large" || config.head_layers !== 2 || config.max_len !== 512 || config.head_max_len !== 192) {
+    throw new Error("Hosting currently supports only the pinned English Laya checkpoint and TT decision fine-tunes.");
+  }
+  const kind = config.tuned_tensor ? "tuned" : "base";
+  if (kind === "tuned" && (JSON.stringify(config.temperature) !== "[1,1,1]" || config.temperature_by_options)) {
+    throw new Error("Custom calibration is not supported by hosted decision inference.");
+  }
+  const file = join15(directory, "model.safetensors");
+  const size = statSync4(file).size;
+  if (size <= 8 || size > 2e9) throw new Error("Weights must be a safetensors file smaller than 2 GB. GPU and adapter models are not supported.");
+  const fd = openSync3(file, "r");
+  try {
+    const length = Buffer.alloc(8);
+    if (readSync(fd, length, 0, 8, 0) !== 8) throw new Error("Invalid safetensors header.");
+    const headerSize = Number(length.readBigUInt64LE());
+    if (!Number.isSafeInteger(headerSize) || headerSize < 2 || headerSize > 1e6 || headerSize + 8 >= size) throw new Error("Invalid safetensors header.");
+    const header = Buffer.alloc(headerSize);
+    if (readSync(fd, header, 0, headerSize, 8) !== headerSize) throw new Error("Truncated safetensors header.");
+    const tensors = JSON.parse(header.toString("utf8"));
+    for (const prefix of ["encoder.", "type_emb.", "scorer.", "act_head."]) {
+      if (!Object.keys(tensors).some((key) => key.startsWith(prefix))) throw new Error("Not a complete Laya checkpoint.");
+    }
+  } finally {
+    closeSync3(fd);
+  }
+  const hash = createHash8("sha256");
+  for await (const chunk of createReadStream2(file)) hash.update(chunk);
+  return { file, weights_bytes: size, weights_sha256: hash.digest("hex"), checkpoint_kind: kind };
+}
+function readInferencePayload(file) {
+  const path = resolve10(file);
+  if (statSync4(path).size > 131072) throw new Error("Inference JSON must be at most 128 KiB.");
+  const data3 = JSON.parse(readFileSync11(path, "utf8"));
+  return z.record(z.string(), z.json()).refine((value) => Object.keys(value).length > 0).parse(data3);
+}
+async function uploadWeights(endpoint, file) {
+  const url = new URL(endpoint.upload.url);
+  if (url.protocol !== "https:" || url.username || url.password || !/\.s3\.[a-z0-9-]+\.amazonaws\.com$/.test(url.hostname)) {
+    throw new Error("The server returned an invalid model upload URL.");
+  }
+  const handle = await open4(file, "r");
+  const stream = handle.createReadStream();
+  const completion = finished(stream);
+  void completion.catch(() => {
+  });
+  try {
+    const response = await fetch(url, {
+      method: "PUT",
+      headers: endpoint.upload.headers,
+      body: stream,
+      duplex: "half",
+      redirect: "error",
+      signal: AbortSignal.timeout(60 * 6e4)
+    });
+    if (!response.ok) throw new Error(`Model upload failed (${response.status}). Delete endpoint ${endpoint.id} and retry.`);
+  } finally {
+    stream.destroy();
+    await completion.catch(() => {
+    });
+  }
+}
+function show(endpoint, opts) {
+  const result2 = { ...endpoint, invocation_url: new URL(endpoint.invocation_path, getBaseUrl(opts)).toString() };
+  if (isJsonMode()) return printJson(result2);
+  printDetail([
+    ["Endpoint", result2.id],
+    ["Name", result2.name],
+    ["Status", result2.status],
+    ["Inference URL", result2.invocation_url],
+    ["Price per successful request", `${result2.price_cents} cents`],
+    ...result2.inference_key ? [["Inference key (save securely)", result2.inference_key]] : []
+  ]);
+}
+function registerEndpointCommands(parent) {
+  const endpoints = parent.command("endpoints").description("Host small CPU decision models using prepaid TT credits");
+  endpoints.command("create").description("Upload Laya weights and deploy a serverless endpoint").requiredOption("--model <directory>", "Local Laya checkpoint (e.g. a decision run's train/model)").requiredOption("--name <name>", "Endpoint display name").action(async (options) => {
+    const opts = parent.optsWithGlobals();
+    const { file, ...artifact } = await inspectCheckpoint(options.model);
+    const { data: data3 } = await post("/endpoints", { name: options.name, format: "laya-safetensors-v1", ...artifact }, opts);
+    if (!isJsonMode()) console.error(`Endpoint ${data3.id} created. Uploading ${(artifact.weights_bytes / 1e9).toFixed(2)} GB of weights...`);
+    try {
+      await uploadWeights(data3, file);
+      const { data: deployed } = await post(`/endpoints/${data3.id}/deploy`, {}, opts);
+      show({ ...deployed, inference_key: data3.inference_key }, opts);
+    } catch (error) {
+      throw new Error(`Endpoint ${data3.id}: ${error instanceof Error ? error.message : "upload/deployment failed"}. Use tt cloud endpoints get ${data3.id} or delete ${data3.id}. Rotate its key if the create response was lost.`, { cause: error });
+    }
+  });
+  endpoints.command("list").action(async () => {
+    const opts = parent.optsWithGlobals();
+    const { data: data3 } = await get("/endpoints", void 0, opts);
+    if (isJsonMode()) return printJson(data3);
+    for (const row of data3) show(row, opts);
+  });
+  endpoints.command("get <id>").description("Check deployment or deletion status").action(async (id) => {
+    z.uuid().parse(id);
+    const opts = parent.optsWithGlobals();
+    show((await get(`/endpoints/${id}`, void 0, opts)).data, opts);
+  });
+  endpoints.command("delete <id>").description("Stop serving and delete cloud weights and resources").action(async (id) => {
+    z.uuid().parse(id);
+    const opts = parent.optsWithGlobals();
+    show((await del(`/endpoints/${id}`, opts)).data, opts);
+  });
+  endpoints.command("rotate-key <id>").description("Revoke the old inference key and return a new key").action(async (id) => {
+    z.uuid().parse(id);
+    const { data: data3 } = await post(`/endpoints/${id}/key`, {}, parent.optsWithGlobals());
+    if (isJsonMode()) return printJson(data3);
+    printDetail([["Inference key (save securely)", data3.inference_key]]);
+  });
+  endpoints.command("invoke <id>").requiredOption("--payload <file>", "Complete model request as a JSON object (up to 128 KiB)").option("--request-id <uuid>", "Reuse this ID to retry without a second charge").action(async (id, options) => {
+    z.uuid().parse(id);
+    const requestId = options.requestId ?? randomUUID10();
+    z.uuid().parse(requestId);
+    const payload = readInferencePayload(options.payload);
+    if (!isJsonMode()) console.error(`Request ID: ${requestId} (reuse with --request-id when retrying)`);
+    try {
+      const { data: data3 } = await post(`/endpoints/${id}/invoke`, payload, parent.optsWithGlobals(), { "Idempotency-Key": requestId });
+      printJson({ request_id: requestId, data: data3 });
+    } catch (error) {
+      throw new Error(`Request ${requestId}: ${error instanceof Error ? error.message : "inference failed"}. Retry with --request-id ${requestId}.`, { cause: error });
+    }
+  });
+}
+
 // src/commands/models.ts
 import { spawn as spawn3, execFileSync } from "child_process";
-import { createHash as createHash8 } from "crypto";
+import { createHash as createHash9 } from "crypto";
 import {
   existsSync as existsSync10,
   mkdirSync as mkdirSync5,
-  statSync as statSync4,
+  statSync as statSync5,
   createWriteStream,
   unlinkSync,
-  readFileSync as readFileSync12,
+  readFileSync as readFileSync13,
   writeFileSync as writeFileSync6,
   readdirSync,
   rmSync
 } from "fs";
 import { homedir as homedir3 } from "os";
-import { dirname as dirname13, join as join15, resolve as resolve11, basename as basename9, normalize, isAbsolute as isAbsolute6 } from "path";
+import { dirname as dirname13, join as join16, resolve as resolve12, basename as basename9, normalize, isAbsolute as isAbsolute6 } from "path";
 import { Readable, Transform } from "stream";
 import { pipeline } from "stream/promises";
 
 // src/commands/init.ts
-import { existsSync as existsSync9, readFileSync as readFileSync11, writeFileSync as writeFileSync5 } from "fs";
-import { resolve as resolve10 } from "path";
+import { existsSync as existsSync9, readFileSync as readFileSync12, writeFileSync as writeFileSync5 } from "fs";
+import { resolve as resolve11 } from "path";
 var DEFAULT_SPEC_FILE2 = "tunedtensor.json";
 function loadSpec(filePath) {
-  const resolved = resolve10(filePath || DEFAULT_SPEC_FILE2);
+  const resolved = resolve11(filePath || DEFAULT_SPEC_FILE2);
   if (!existsSync9(resolved)) {
     printError(
       `Spec file not found: ${filePath || DEFAULT_SPEC_FILE2}
@@ -12304,13 +12439,13 @@ Run \`tt init\` to create one.`
     );
     process.exit(1);
   }
-  const raw = JSON.parse(readFileSync11(resolved, "utf-8"));
+  const raw = JSON.parse(readFileSync12(resolved, "utf-8"));
   return raw;
 }
 
 // src/serve-manager.ts
 import { spawn as spawn2 } from "child_process";
-import { randomUUID as randomUUID10 } from "crypto";
+import { randomUUID as randomUUID11 } from "crypto";
 import { appendFileSync, mkdirSync as mkdirSync4 } from "fs";
 import { createServer } from "http";
 import { createServer as createNetServer } from "net";
@@ -12327,10 +12462,10 @@ function parsePositiveInteger(value, fallback) {
   return Number.isInteger(value) && value >= 0 ? value : fallback;
 }
 function sleep2(ms) {
-  return new Promise((resolve13) => setTimeout(resolve13, ms));
+  return new Promise((resolve14) => setTimeout(resolve14, ms));
 }
 function readRequestBody(req) {
-  return new Promise((resolve13, reject) => {
+  return new Promise((resolve14, reject) => {
     if (Number(req.headers["content-length"]) > MAX_SERVE_REQUEST_BYTES) {
       reject(new RequestTooLargeError());
       return;
@@ -12360,7 +12495,7 @@ function readRequestBody(req) {
     };
     const onEnd = () => {
       cleanup();
-      resolve13(Buffer.concat(chunks, bytes));
+      resolve14(Buffer.concat(chunks, bytes));
     };
     req.on("data", onData);
     req.on("end", onEnd);
@@ -12437,7 +12572,7 @@ function errorSummary(response, fallback) {
   return typeof message === "string" ? message : fallback;
 }
 async function allocatePort(host) {
-  return new Promise((resolve13, reject) => {
+  return new Promise((resolve14, reject) => {
     const server = createNetServer();
     server.on("error", reject);
     server.listen(0, host, () => {
@@ -12447,7 +12582,7 @@ async function allocatePort(host) {
         return;
       }
       const port = address.port;
-      server.close(() => resolve13(port));
+      server.close(() => resolve14(port));
     });
   });
 }
@@ -12488,11 +12623,11 @@ var ManagedModelServer = class {
         sendJson(res, 500, { error: { message: err instanceof Error ? err.message : String(err) } });
       });
     });
-    await new Promise((resolve13, reject) => {
+    await new Promise((resolve14, reject) => {
       this.server?.once("error", reject);
       this.server?.listen(this.options.publicPort, this.options.publicHost, () => {
         this.server?.off("error", reject);
-        resolve13();
+        resolve14();
       });
     });
   }
@@ -12502,15 +12637,15 @@ var ManagedModelServer = class {
     if (!this.server) return;
     const server = this.server;
     this.server = null;
-    await new Promise((resolve13, reject) => {
-      server.close((err) => err ? reject(err) : resolve13());
+    await new Promise((resolve14, reject) => {
+      server.close((err) => err ? reject(err) : resolve14());
     });
   }
   async run() {
     await this.start();
-    await new Promise((resolve13, reject) => {
+    await new Promise((resolve14, reject) => {
       this.server?.once("error", reject);
-      this.server?.once("close", resolve13);
+      this.server?.once("close", resolve14);
     });
   }
   async handle(req, res) {
@@ -12591,7 +12726,7 @@ var ManagedModelServer = class {
     });
   }
   async processGeneration(route, body, res, receivedAt) {
-    const requestId = `ttreq-${randomUUID10()}`;
+    const requestId = `ttreq-${randomUUID11()}`;
     const startedAt = Date.now();
     const queuedMs = startedAt - receivedAt;
     this.activeRequests += 1;
@@ -12697,7 +12832,7 @@ var ManagedModelServer = class {
     try {
       child.kill("SIGTERM");
       await Promise.race([
-        new Promise((resolve13) => child.once("exit", () => resolve13())),
+        new Promise((resolve14) => child.once("exit", () => resolve14())),
         sleep2(5e3).then(() => {
           if (child.exitCode == null && child.signalCode == null) child.kill("SIGKILL");
         })
@@ -13328,8 +13463,8 @@ function validateDownloadFilename(filename) {
 function resolveOutputPath(output, filename) {
   validateDownloadFilename(filename);
   if (!output) return filename;
-  if (existsSync10(output) && statSync4(output).isDirectory()) {
-    return join15(output, filename);
+  if (existsSync10(output) && statSync5(output).isDirectory()) {
+    return join16(output, filename);
   }
   return output;
 }
@@ -13436,15 +13571,15 @@ async function downloadUrlToFile(url, outputPath) {
   return contentLength;
 }
 function defaultCacheDir() {
-  const root = process.env.XDG_CACHE_HOME || join15(homedir3(), ".cache");
-  return join15(root, "tuned-tensor", "models");
+  const root = process.env.XDG_CACHE_HOME || join16(homedir3(), ".cache");
+  return join16(root, "tuned-tensor", "models");
 }
 function runtimeDir(cacheDir) {
-  return join15(cacheDir, "_runtime");
+  return join16(cacheDir, "_runtime");
 }
 function runtimePythonPath(cacheDir) {
   const dir = runtimeDir(cacheDir);
-  return process.platform === "win32" ? join15(dir, "Scripts", "python.exe") : join15(dir, "bin", "python");
+  return process.platform === "win32" ? join16(dir, "Scripts", "python.exe") : join16(dir, "bin", "python");
 }
 function resolveServePython(cacheDir, explicitPython) {
   if (explicitPython) return explicitPython;
@@ -13517,7 +13652,7 @@ print(json.dumps({"missing": missing}))
   }
 }
 function setupRuntimeCommands(python, venvDir, deps) {
-  const venvPython = process.platform === "win32" ? join15(venvDir, "Scripts", "python.exe") : join15(venvDir, "bin", "python");
+  const venvPython = process.platform === "win32" ? join16(venvDir, "Scripts", "python.exe") : join16(venvDir, "bin", "python");
   return [
     [python, "-m", "venv", venvDir],
     [venvPython, "-m", "pip", "install", "--upgrade", "pip"],
@@ -13528,7 +13663,7 @@ function safeSegment(value) {
   return value.replace(/[^a-zA-Z0-9._-]+/g, "-").replace(/^-+|-+$/g, "") || "model";
 }
 function directoryHasFiles(path) {
-  return existsSync10(path) && statSync4(path).isDirectory() && readdirSync(path).length > 0;
+  return existsSync10(path) && statSync5(path).isDirectory() && readdirSync(path).length > 0;
 }
 function validateArchiveEntryPath(entry) {
   const normalized = normalize(entry).replace(/^[\\/]+/, "");
@@ -13555,9 +13690,9 @@ function extractArchive(archivePath, targetDir, force) {
   return targetDir;
 }
 function localArchiveCacheDir(archivePath, cacheDir) {
-  const resolved = resolve11(archivePath);
-  const digest = createHash8("sha256").update(resolved).digest("hex").slice(0, 12);
-  return join15(cacheDir, `local-${safeSegment(basename9(archivePath))}-${digest}`);
+  const resolved = resolve12(archivePath);
+  const digest = createHash9("sha256").update(resolved).digest("hex").slice(0, 12);
+  return join16(cacheDir, `local-${safeSegment(basename9(archivePath))}-${digest}`);
 }
 function buildSystemPrompt(spec) {
   const parts = [];
@@ -13573,40 +13708,40 @@ ${spec.constraints.map((c) => `- ${c}`).join("\n")}`);
   return parts.join("\n\n");
 }
 function findSpecPath(explicitPath, modelPath) {
-  const candidates = explicitPath ? [explicitPath] : [join15(process.cwd(), DEFAULT_SPEC_FILE2), join15(modelPath, DEFAULT_SPEC_FILE2)];
+  const candidates = explicitPath ? [explicitPath] : [join16(process.cwd(), DEFAULT_SPEC_FILE2), join16(modelPath, DEFAULT_SPEC_FILE2)];
   for (const candidate of candidates) {
-    const resolved = resolve11(candidate);
-    if (existsSync10(resolved) && statSync4(resolved).isFile()) return resolved;
+    const resolved = resolve12(candidate);
+    if (existsSync10(resolved) && statSync5(resolved).isFile()) return resolved;
   }
   return null;
 }
 function loadSystemPromptFromSpec(specPath) {
-  const spec = JSON.parse(readFileSync12(specPath, "utf8"));
+  const spec = JSON.parse(readFileSync13(specPath, "utf8"));
   return buildSystemPrompt(spec);
 }
 function loadJsonSchemaForServe(schemaPath) {
-  const resolved = resolve11(schemaPath);
-  if (!existsSync10(resolved) || !statSync4(resolved).isFile()) {
+  const resolved = resolve12(schemaPath);
+  if (!existsSync10(resolved) || !statSync5(resolved).isFile()) {
     throw new Error(`JSON schema file not found: ${schemaPath}`);
   }
   try {
-    const schema = JSON.parse(readFileSync12(resolved, "utf8"));
+    const schema = JSON.parse(readFileSync13(resolved, "utf8"));
     return JSON.stringify(schema);
   } catch (err) {
     throw new Error(`JSON schema file is not valid JSON: ${schemaPath}`);
   }
 }
 function writeReferenceServerScript(cacheDir) {
-  const scriptDir = join15(cacheDir, "_server");
+  const scriptDir = join16(cacheDir, "_server");
   mkdirSync5(scriptDir, { recursive: true });
-  const scriptPath = join15(scriptDir, "openai_reference_server.py");
+  const scriptPath = join16(scriptDir, "openai_reference_server.py");
   writeFileSync6(scriptPath, REFERENCE_SERVER_SCRIPT, "utf8");
   return scriptPath;
 }
 async function resolveServeTarget(target, opts, cacheDir, forceDownload) {
   if (existsSync10(target)) {
-    const resolved = resolve11(target);
-    const stats = statSync4(resolved);
+    const resolved = resolve12(target);
+    const stats = statSync5(resolved);
     if (stats.isDirectory()) {
       return {
         modelPath: resolved,
@@ -13615,7 +13750,7 @@ async function resolveServeTarget(target, opts, cacheDir, forceDownload) {
       };
     }
     if (stats.isFile() && (resolved.endsWith(".tar.gz") || resolved.endsWith(".tgz"))) {
-      const extractDir2 = join15(localArchiveCacheDir(resolved, cacheDir), "model");
+      const extractDir2 = join16(localArchiveCacheDir(resolved, cacheDir), "model");
       return {
         modelPath: extractArchive(resolved, extractDir2, forceDownload),
         modelName: basename9(resolved).replace(/\.t(ar\.)?gz$/, ""),
@@ -13629,9 +13764,9 @@ async function resolveServeTarget(target, opts, cacheDir, forceDownload) {
   const { data: data3 } = await get(`/models/${fullId}/download`, void 0, opts);
   validateDownloadFilename(fullId);
   validateDownloadFilename(data3.filename);
-  const modelCacheDir = join15(cacheDir, fullId);
-  const archivePath = join15(modelCacheDir, data3.filename);
-  const extractDir = join15(modelCacheDir, "model");
+  const modelCacheDir = join16(cacheDir, fullId);
+  const archivePath = join16(modelCacheDir, data3.filename);
+  const extractDir = join16(modelCacheDir, "model");
   if (!existsSync10(archivePath) || forceDownload) {
     await downloadUrlToFile(data3.url, archivePath);
   }
@@ -13736,8 +13871,8 @@ function ollamaSlug(name) {
 }
 function firstExisting(candidates) {
   for (const candidate of candidates) {
-    if (candidate && existsSync10(candidate) && statSync4(candidate).isFile()) {
-      return resolve11(candidate);
+    if (candidate && existsSync10(candidate) && statSync5(candidate).isFile()) {
+      return resolve12(candidate);
     }
   }
   return null;
@@ -13747,14 +13882,14 @@ function llamaCppDir(explicit) {
 }
 function resolveConvertScript(opts, required) {
   if (opts.convertScript) {
-    const resolved = resolve11(opts.convertScript);
+    const resolved = resolve12(opts.convertScript);
     if (required && !existsSync10(resolved)) {
       throw new Error(`Conversion script not found: ${opts.convertScript}`);
     }
     return resolved;
   }
   const dir = llamaCppDir(opts.llamaCpp);
-  const candidates = dir ? [join15(dir, "convert_hf_to_gguf.py"), join15(dir, "convert-hf-to-gguf.py")] : [];
+  const candidates = dir ? [join16(dir, "convert_hf_to_gguf.py"), join16(dir, "convert-hf-to-gguf.py")] : [];
   const found = firstExisting(candidates);
   if (found) return found;
   if (required) {
@@ -13762,11 +13897,11 @@ function resolveConvertScript(opts, required) {
       "Could not find convert_hf_to_gguf.py. Pass --llama-cpp <dir> (a llama.cpp checkout) or --convert-script <path>. Get it from https://github.com/ggml-org/llama.cpp."
     );
   }
-  return dir ? join15(dir, "convert_hf_to_gguf.py") : "convert_hf_to_gguf.py";
+  return dir ? join16(dir, "convert_hf_to_gguf.py") : "convert_hf_to_gguf.py";
 }
 function resolveQuantizeBin(opts, required) {
   if (opts.quantizeBin) {
-    const resolved = resolve11(opts.quantizeBin);
+    const resolved = resolve12(opts.quantizeBin);
     if (required && !existsSync10(resolved)) {
       throw new Error(`llama-quantize binary not found: ${opts.quantizeBin}`);
     }
@@ -13775,10 +13910,10 @@ function resolveQuantizeBin(opts, required) {
   const dir = llamaCppDir(opts.llamaCpp);
   if (dir) {
     const candidates = [
-      join15(dir, "build", "bin", "llama-quantize"),
-      join15(dir, "build", "bin", "quantize"),
-      join15(dir, "llama-quantize"),
-      join15(dir, "quantize")
+      join16(dir, "build", "bin", "llama-quantize"),
+      join16(dir, "build", "bin", "quantize"),
+      join16(dir, "llama-quantize"),
+      join16(dir, "quantize")
     ];
     const found = firstExisting(candidates);
     if (found) return found;
@@ -13926,7 +14061,7 @@ function registerModelsCommands(parent) {
     }
     const quantPlan = planQuant(String(cmdOpts.quant));
     const printOnly = Boolean(cmdOpts.printCommand);
-    const cacheDir = resolve11(cmdOpts.cacheDir || defaultCacheDir());
+    const cacheDir = resolve12(cmdOpts.cacheDir || defaultCacheDir());
     mkdirSync5(cacheDir, { recursive: true });
     const serveTarget = await resolveServeTarget(
       target,
@@ -13935,9 +14070,9 @@ function registerModelsCommands(parent) {
       Boolean(cmdOpts.forceDownload)
     );
     const slug = ollamaSlug(serveTarget.modelName);
-    const outputDir = resolve11(cmdOpts.output || join15(process.cwd(), `${slug}-gguf`));
-    const finalPath = join15(outputDir, `${slug}.${quantPlan.quant}.gguf`);
-    const intermediatePath = quantPlan.requiresQuantize ? join15(outputDir, `${slug}.f16.gguf`) : void 0;
+    const outputDir = resolve12(cmdOpts.output || join16(process.cwd(), `${slug}-gguf`));
+    const finalPath = join16(outputDir, `${slug}.${quantPlan.quant}.gguf`);
+    const intermediatePath = quantPlan.requiresQuantize ? join16(outputDir, `${slug}.f16.gguf`) : void 0;
     const convertScript = resolveConvertScript(cmdOpts, !printOnly);
     const convertOutfile = intermediatePath ?? finalPath;
     const steps = [
@@ -13975,7 +14110,7 @@ function registerModelsCommands(parent) {
         }
         if (specPath) systemPrompt2 = loadSystemPromptFromSpec(specPath);
       }
-      modelfilePath = join15(outputDir, "Modelfile");
+      modelfilePath = join16(outputDir, "Modelfile");
       modelfileContent = buildModelfile(basename9(finalPath), systemPrompt2);
       const createModel = cmdOpts.ollamaCreate !== false;
       if (createModel) {
@@ -14062,7 +14197,7 @@ function registerModelsCommands(parent) {
     }
   });
   models.command("setup-runtime").description("Install an isolated Python runtime for local model serving").option("--python <path>", "Python 3.10-3.12 executable to create the runtime").option("--cache-dir <path>", "Cache directory for the managed runtime").option("-f, --force", "Recreate the runtime if it already exists").option("--print-command", "Print the setup commands without running them").action(async (cmdOpts) => {
-    const cacheDir = resolve11(cmdOpts.cacheDir || defaultCacheDir());
+    const cacheDir = resolve12(cmdOpts.cacheDir || defaultCacheDir());
     const venvDir = runtimeDir(cacheDir);
     const venvPython = runtimePythonPath(cacheDir);
     const python = pickRuntimePython(cmdOpts.python);
@@ -14108,7 +14243,7 @@ function registerModelsCommands(parent) {
   });
   models.command("serve").description("Serve a downloaded model with an OpenAI-compatible local API").argument("<target>", "Model ID/prefix, model directory, or .tar.gz artifact").option("--spec <path>", "Behavior spec JSON to apply as the default system prompt").option("--no-spec-prompt", "Do not auto-apply a behavior spec system prompt").option("--host <host>", "Host to bind", "127.0.0.1").option("--port <port>", "Port to bind", "8000").option("--python <path>", "Python executable").option("--cache-dir <path>", "Cache directory for downloaded/extracted models").option("--device <device>", "Inference device: auto, cpu, cuda, or mps", "auto").option("--force-download", "Re-download and re-extract model artifacts").option("--max-tokens <n>", "Default max completion tokens", "512").option("--temperature <n>", "Default sampling temperature", "0.7").option("--json-schema <path>", "JSON Schema file to enforce by default for chat completions").option("--json-repair-attempts <n>", "Number of JSON/schema repair retries", "1").option("--trust-remote-code", "Pass trust_remote_code=True to transformers").option("--managed", "Run a local lifecycle manager in front of the model server").option("--idle-timeout <seconds>", "Managed mode idle timeout before stopping the model", "300").option("--restart-after-requests <n>", "Managed mode restart threshold; 0 disables request-count restarts", "100").option("--gate-field <field>", "Response JSON field to log as the managed serving gate result", "should_process").option("--log-file <path>", "Write managed serving JSONL request logs to a file").option("--print-command", "Print the underlying Python command without starting it").action(async (target, cmdOpts) => {
     const opts = parent.optsWithGlobals();
-    const cacheDir = resolve11(cmdOpts.cacheDir || defaultCacheDir());
+    const cacheDir = resolve12(cmdOpts.cacheDir || defaultCacheDir());
     mkdirSync5(cacheDir, { recursive: true });
     const python = resolveServePython(cacheDir, cmdOpts.python);
     if (!cmdOpts.printCommand) ensureServingRuntime(python, cacheDir);
@@ -14157,7 +14292,7 @@ function registerModelsCommands(parent) {
       idle_timeout_seconds: idleTimeoutSeconds,
       restart_after_requests: restartAfterRequests,
       gate_field: String(cmdOpts.gateField),
-      log_file: cmdOpts.logFile ? resolve11(cmdOpts.logFile) : void 0
+      log_file: cmdOpts.logFile ? resolve12(cmdOpts.logFile) : void 0
     } : void 0;
     if (cmdOpts.printCommand) return printServeCommand(python, args2, env, managedConfig);
     if (!isJsonMode()) {
@@ -14194,7 +14329,7 @@ function registerModelsCommands(parent) {
         idleTimeoutSeconds,
         restartAfterRequests,
         gateField: String(cmdOpts.gateField),
-        logFile: cmdOpts.logFile ? resolve11(cmdOpts.logFile) : void 0
+        logFile: cmdOpts.logFile ? resolve12(cmdOpts.logFile) : void 0
       });
       return;
     }
@@ -14220,15 +14355,15 @@ function registerModelsCommands(parent) {
 }
 
 // src/commands/push.ts
-import { readFileSync as readFileSync13, writeFileSync as writeFileSync7 } from "fs";
-import { resolve as resolve12 } from "path";
+import { readFileSync as readFileSync14, writeFileSync as writeFileSync7 } from "fs";
+import { resolve as resolve13 } from "path";
 function persistRemoteId(filePath, id) {
-  const raw = JSON.parse(readFileSync13(filePath, "utf-8"));
+  const raw = JSON.parse(readFileSync14(filePath, "utf-8"));
   raw.id = id;
   writeFileSync7(filePath, JSON.stringify(raw, null, 2) + "\n");
 }
 async function pushLocalSpec(file, opts) {
-  const filePath = resolve12(file);
+  const filePath = resolve13(file);
   const spec = loadSpec(file);
   const evalCaseErrors = validateEvalCases(spec);
   if (evalCaseErrors.length > 0) {
@@ -14287,6 +14422,7 @@ function registerCloudCommands(parent) {
   registerLabelCommands(cloud);
   registerRunsCommands(cloud);
   registerModelsCommands(cloud);
+  registerEndpointCommands(cloud);
 }
 function weeklyRows(data3) {
   if (typeof data3.weekly_remaining_percent !== "number" || !Number.isFinite(data3.weekly_remaining_percent) || typeof data3.weekly_cost_limit_usd !== "number" || !Number.isFinite(data3.weekly_cost_limit_usd) || data3.weekly_cost_limit_usd <= 0) return [];
@@ -14326,14 +14462,14 @@ Use \`tt balance\` for cloud training credits.`);
 
 // src/commands/upgrade.ts
 import { spawn as spawn5 } from "child_process";
-import { accessSync, constants as constants2, existsSync as existsSync11, readFileSync as readFileSync15, realpathSync } from "fs";
-import { dirname as dirname15, join as join17 } from "path";
+import { accessSync, constants as constants2, existsSync as existsSync11, readFileSync as readFileSync16, realpathSync } from "fs";
+import { dirname as dirname15, join as join18 } from "path";
 import chalk14 from "chalk";
 
 // src/update-check.ts
 import { spawn as spawn4 } from "child_process";
-import { mkdirSync as mkdirSync6, readFileSync as readFileSync14, writeFileSync as writeFileSync8 } from "fs";
-import { dirname as dirname14, join as join16 } from "path";
+import { mkdirSync as mkdirSync6, readFileSync as readFileSync15, writeFileSync as writeFileSync8 } from "fs";
+import { dirname as dirname14, join as join17 } from "path";
 import chalk13 from "chalk";
 var CLI_PACKAGE_NAME = "@tuned-tensor/cli";
 var PACKAGE_LATEST_URL = "https://registry.npmjs.org/@tuned-tensor%2fcli/latest";
@@ -14382,14 +14518,14 @@ function compareVersions(left, right) {
   return comparePrerelease(left.prerelease, right.prerelease);
 }
 function getCliUpdateCacheFile(env = process.env) {
-  return join16(getTunedTensorHome(env), "update-check.json");
+  return join17(getTunedTensorHome(env), "update-check.json");
 }
 function cliUpdateChecksDisabled(env = process.env) {
   return [env.TT_NO_UPDATE_CHECK, env.NO_UPDATE_NOTIFIER, env.CI].some((value) => value !== void 0 && value !== "" && value !== "0" && value !== "false");
 }
 function readUpdateCache(file) {
   try {
-    const payload = JSON.parse(readFileSync14(file, "utf8"));
+    const payload = JSON.parse(readFileSync15(file, "utf8"));
     if (typeof payload.checkedAt !== "number" || typeof payload.latestVersion !== "string" || !parseVersion(payload.latestVersion)) return null;
     return { checkedAt: payload.checkedAt, latestVersion: payload.latestVersion };
   } catch {
@@ -14441,10 +14577,10 @@ async function fetchLatestCliVersion(options = {}) {
       return null;
     }
   })();
-  const deadline = new Promise((resolve13) => {
+  const deadline = new Promise((resolve14) => {
     timeout = setTimeout(() => {
       controller.abort();
-      resolve13(null);
+      resolve14(null);
     }, timeoutMs);
   });
   try {
@@ -14529,7 +14665,7 @@ function resolveGlobalInstall(entrypoint, platform = process.platform) {
   return { prefix: match[1], packageDir: dirname15(dirname15(script)) };
 }
 function prefixIsWritable(prefix, platform) {
-  const modules = platform === "win32" ? join17(prefix, "node_modules") : join17(prefix, "lib", "node_modules");
+  const modules = platform === "win32" ? join18(prefix, "node_modules") : join18(prefix, "lib", "node_modules");
   try {
     accessSync(existsSync11(modules) ? modules : prefix, constants2.W_OK);
     return true;
@@ -14539,7 +14675,7 @@ function prefixIsWritable(prefix, platform) {
 }
 function installedVersion(packageDir) {
   try {
-    const manifest = JSON.parse(readFileSync15(join17(packageDir, "package.json"), "utf8"));
+    const manifest = JSON.parse(readFileSync16(join18(packageDir, "package.json"), "utf8"));
     return typeof manifest.version === "string" ? manifest.version : void 0;
   } catch {
     return void 0;
@@ -14549,7 +14685,7 @@ function quoteForCmd(arg) {
   return /[\s"&|<>^]/.test(arg) ? `"${arg.replace(/"/g, '""')}"` : arg;
 }
 function defaultNpmRunner(platform) {
-  return async (args2) => await new Promise((resolve13, reject) => {
+  return async (args2) => await new Promise((resolve14, reject) => {
     const windows = platform === "win32";
     const child = spawn5(
       windows ? "npm.cmd" : "npm",
@@ -14561,7 +14697,7 @@ function defaultNpmRunner(platform) {
       }
     );
     child.once("error", reject);
-    child.once("close", (code2) => resolve13(code2 ?? 1));
+    child.once("close", (code2) => resolve14(code2 ?? 1));
   });
 }
 function formatCommand(args2) {
@@ -14802,7 +14938,7 @@ async function runSelfCommand(args2, options = {}) {
   if (!entrypoint) {
     throw new Error("Cannot locate the tt CLI entrypoint.");
   }
-  return await new Promise((resolve13, reject) => {
+  return await new Promise((resolve14, reject) => {
     const child = spawn6(process.execPath, [entrypoint, ...options.captureJson ? ["--json"] : [], ...args2], {
       cwd: options.cwd,
       env: options.env ?? process.env,
@@ -14849,7 +14985,7 @@ async function runSelfCommand(args2, options = {}) {
     child.once("close", (exitCode, signal) => {
       cleanup();
       const signalNumber = signal ? osConstants2.signals[signal] : void 0;
-      resolve13({
+      resolve14({
         exitCode: exitCode ?? (typeof signalNumber === "number" ? 128 + signalNumber : 1),
         signal,
         ...options.captureJson ? { stdout: stdout3 } : {}
@@ -15084,7 +15220,7 @@ ${formatCliUpdateNotice(update)}
 }
 
 // src/index.ts
-runCli("0.26.2").catch((err) => {
+runCli("0.27.0").catch((err) => {
   reportError(err);
   process.exit(1);
 });