@elizaos/plugin-training 2.0.3-beta.6 → 2.0.3-beta.7

package/dist/core/cerebras-eval-model.d.ts

@@ -0,0 +1,54 @@
+import { type JudgeResponse } from "@elizaos/scenario-runner/cerebras-judge";
+export interface CerebrasChatRequest {
+    prompt: string;
+    systemPrompt?: string;
+    temperature?: number;
+    maxTokens?: number;
+    reasoningEffort?: "low" | "medium" | "high";
+}
+export interface CerebrasChatUsage {
+    promptTokens?: number;
+    completionTokens?: number;
+    totalTokens?: number;
+    cachedTokens?: number;
+}
+export interface CerebrasChatResponse {
+    text: string;
+    usage?: CerebrasChatUsage;
+    raw?: unknown;
+}
+export type EvalModelClient = (req: CerebrasChatRequest) => Promise<CerebrasChatResponse>;
+export declare function getEvalModelClient(): EvalModelClient;
+export declare function getTrainingModelClient(): EvalModelClient;
+/**
+ * Cerebras-only judge helper. Routes through the shared `CerebrasJudge`
+ * transport (tolerant parsing, 429/5xx retry, json_object opt-in). The
+ * Cerebras eval provider is the only one configured here; callers that
+ * need Anthropic should use `getEvalModelClient()` directly.
+ *
+ * Returns the raw model text for backward compatibility with existing
+ * callers. New callers should consume `judgeWithCerebrasShared()` (below)
+ * to get the canonical {raw, json, score?, verdict?, reason?} shape.
+ */
+export declare function judgeWithCerebras(prompt: string, options?: {
+    maxTokens?: number;
+    temperature?: number;
+    systemPrompt?: string;
+}): Promise<string>;
+/**
+ * New canonical entry: returns the full JudgeResponse for callers that
+ * want the parsed score/verdict/reason without re-parsing the raw text.
+ */
+export declare function judgeWithCerebrasShared(prompt: string, options?: {
+    maxTokens?: number;
+    temperature?: number;
+    systemPrompt?: string;
+}): Promise<JudgeResponse>;
+export declare function getTrainingUseModelAdapter(): (input: {
+    prompt: string;
+    temperature?: number;
+    maxTokens?: number;
+}) => Promise<string>;
+export declare function isCerebrasEvalEnabled(): boolean;
+export declare function isCerebrasTrainingEnabled(): boolean;
+//# sourceMappingURL=cerebras-eval-model.d.ts.map

package/dist/core/cerebras-eval-model.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"cerebras-eval-model.d.ts","sourceRoot":"","sources":["../../src/core/cerebras-eval-model.ts"],"names":[],"mappings":"AAYA,OAAO,EAEL,KAAK,aAAa,EACnB,MAAM,yCAAyC,CAAC;AAUjD,MAAM,WAAW,mBAAmB;IAClC,MAAM,EAAE,MAAM,CAAC;IACf,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,eAAe,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;CAC7C;AAED,MAAM,WAAW,iBAAiB;IAChC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,oBAAoB;IACnC,IAAI,EAAE,MAAM,CAAC;IACb,KAAK,CAAC,EAAE,iBAAiB,CAAC;IAC1B,GAAG,CAAC,EAAE,OAAO,CAAC;CACf;AAED,MAAM,MAAM,eAAe,GAAG,CAC5B,GAAG,EAAE,mBAAmB,KACrB,OAAO,CAAC,oBAAoB,CAAC,CAAC;AAgRnC,wBAAgB,kBAAkB,IAAI,eAAe,CAGpD;AAED,wBAAgB,sBAAsB,IAAI,eAAe,CAGxD;AAED;;;;;;;;;GASG;AACH,wBAAsB,iBAAiB,CACrC,MAAM,EAAE,MAAM,EACd,OAAO,CAAC,EAAE;IAAE,SAAS,CAAC,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAAC,YAAY,CAAC,EAAE,MAAM,CAAA;CAAE,GAC5E,OAAO,CAAC,MAAM,CAAC,CAGjB;AAED;;;GAGG;AACH,wBAAsB,uBAAuB,CAC3C,MAAM,EAAE,MAAM,EACd,OAAO,CAAC,EAAE;IAAE,SAAS,CAAC,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAAC,YAAY,CAAC,EAAE,MAAM,CAAA;CAAE,GAC5E,OAAO,CAAC,aAAa,CAAC,CA0BxB;AAID,wBAAgB,0BAA0B,IAAI,CAAC,KAAK,EAAE;IACpD,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB,KAAK,OAAO,CAAC,MAAM,CAAC,CAUpB;AAED,wBAAgB,qBAAqB,IAAI,OAAO,CAM/C;AAED,wBAAgB,yBAAyB,IAAI,OAAO,CAMnD"}

package/dist/core/cerebras-eval-model.js

@@ -0,0 +1,249 @@
+import {
+  CerebrasJudge
+} from "@elizaos/scenario-runner/cerebras-judge";
+function readEnv(...keys) {
+  for (const key of keys) {
+    const value = process.env[key]?.trim();
+    if (value && value.length > 0) return value;
+  }
+  return void 0;
+}
+function resolveCerebrasApiKey(role) {
+  const apiKey = readEnv(
+    role === "eval" ? "EVAL_CEREBRAS_API_KEY" : "TRAIN_CEREBRAS_API_KEY",
+    "CEREBRAS_API_KEY",
+    "ELIZA_E2E_CEREBRAS_API_KEY"
+  );
+  if (!apiKey) {
+    throw new Error(
+      `[${role}-model] CEREBRAS_API_KEY is not set. Eval/training runs require Cerebras credentials. Set CEREBRAS_API_KEY in eliza/.env.`
+    );
+  }
+  return apiKey;
+}
+function resolveBaseUrl() {
+  return readEnv("CEREBRAS_BASE_URL") ?? "https://api.cerebras.ai/v1";
+}
+function resolveEvalModel() {
+  return readEnv("EVAL_MODEL", "EVAL_MODEL_NAME") ?? readEnv("CEREBRAS_MODEL") ?? "gpt-oss-120b";
+}
+function resolveTrainingModel() {
+  return readEnv("TRAIN_MODEL", "TRAINING_MODEL", "TRAIN_MODEL_NAME") ?? readEnv("CEREBRAS_MODEL") ?? "gpt-oss-120b";
+}
+function resolveProvider(role) {
+  return readEnv(
+    role === "eval" ? "EVAL_MODEL_PROVIDER" : "TRAIN_MODEL_PROVIDER",
+    role === "eval" ? "EVAL_PROVIDER" : "TRAINING_PROVIDER"
+  ) ?? "cerebras";
+}
+function resolveAnthropicApiKey(role) {
+  const apiKey = readEnv("ANTHROPIC_API_KEY");
+  if (!apiKey) {
+    throw new Error(
+      `[${role}-model] ANTHROPIC_API_KEY is not set; required when ${role === "eval" ? "EVAL_MODEL_PROVIDER" : "TRAIN_MODEL_PROVIDER"}=anthropic.`
+    );
+  }
+  return apiKey;
+}
+function resolveAnthropicModel(role) {
+  const explicitAnthropic = readEnv("ANTHROPIC_LARGE_MODEL");
+  if (explicitAnthropic) return explicitAnthropic;
+  if (role === "eval") {
+    return readEnv("EVAL_ANTHROPIC_MODEL", "EVAL_MODEL_NAME") ?? "claude-haiku-4-5-20251001";
+  }
+  return readEnv("TRAIN_ANTHROPIC_MODEL", "TRAIN_MODEL_NAME") ?? "claude-haiku-4-5-20251001";
+}
+function resolveConfig(role) {
+  const provider = resolveProvider(role);
+  if (provider === "cerebras") {
+    return {
+      apiKey: resolveCerebrasApiKey(role),
+      baseUrl: resolveBaseUrl(),
+      model: role === "eval" ? resolveEvalModel() : resolveTrainingModel(),
+      role,
+      providerName: "cerebras"
+    };
+  }
+  if (provider === "anthropic") {
+    return {
+      apiKey: resolveAnthropicApiKey(role),
+      baseUrl: "https://api.anthropic.com/v1",
+      model: resolveAnthropicModel(role),
+      role,
+      providerName: "anthropic"
+    };
+  }
+  throw new Error(
+    `[${role}-model] unknown provider "${provider}"; supported: cerebras, anthropic. Set ${role === "eval" ? "EVAL_MODEL_PROVIDER" : "TRAIN_MODEL_PROVIDER"}=cerebras|anthropic.`
+  );
+}
+const RETRYABLE_STATUS = /* @__PURE__ */ new Set([429, 500, 502, 503, 504]);
+const MAX_CHAT_ATTEMPTS = 5;
+async function fetchChatWithRetry(config, body) {
+  let lastError;
+  for (let attempt = 1; attempt <= MAX_CHAT_ATTEMPTS; attempt += 1) {
+    try {
+      const response = await fetch(`${config.baseUrl}/chat/completions`, {
+        method: "POST",
+        headers: {
+          "Content-Type": "application/json",
+          Authorization: `Bearer ${config.apiKey}`
+        },
+        body: JSON.stringify(body)
+      });
+      if (response.ok || !RETRYABLE_STATUS.has(response.status)) {
+        return response;
+      }
+      lastError = new Error(
+        `[${config.role}-model] cerebras transient ${response.status}`
+      );
+      await response.text().catch(() => void 0);
+    } catch (err) {
+      lastError = err;
+    }
+    if (attempt < MAX_CHAT_ATTEMPTS) {
+      const backoffMs = Math.min(8e3, 400 * 2 ** (attempt - 1));
+      await new Promise((r) => setTimeout(r, backoffMs));
+    }
+  }
+  throw lastError instanceof Error ? lastError : new Error(`[${config.role}-model] chat request failed`);
+}
+async function callCerebras(config, req) {
+  const messages = [];
+  if (req.systemPrompt && req.systemPrompt.length > 0) {
+    messages.push({ role: "system", content: req.systemPrompt });
+  }
+  messages.push({ role: "user", content: req.prompt });
+  const body = {
+    model: config.model,
+    messages,
+    temperature: req.temperature ?? 0,
+    max_tokens: req.maxTokens ?? 1024
+  };
+  if (/(^|\/)gpt-oss/.test(config.model)) {
+    body.reasoning_effort = req.reasoningEffort ?? "low";
+  }
+  const response = await fetchChatWithRetry(config, body);
+  if (!response.ok) {
+    const errBody = await response.text();
+    throw new Error(
+      `[${config.role}-model] cerebras error ${response.status}: ${errBody.slice(0, 300)}`
+    );
+  }
+  const data = await response.json();
+  return {
+    text: data.choices?.[0]?.message?.content ?? "",
+    usage: data.usage ? {
+      promptTokens: data.usage.prompt_tokens,
+      completionTokens: data.usage.completion_tokens,
+      totalTokens: data.usage.total_tokens,
+      cachedTokens: data.usage.prompt_tokens_details?.cached_tokens
+    } : void 0,
+    raw: data
+  };
+}
+async function callAnthropic(config, req) {
+  const body = {
+    model: config.model,
+    max_tokens: req.maxTokens ?? 1024,
+    temperature: req.temperature ?? 0,
+    messages: [{ role: "user", content: req.prompt }]
+  };
+  if (req.systemPrompt && req.systemPrompt.length > 0) {
+    body.system = req.systemPrompt;
+  }
+  const response = await fetch(`${config.baseUrl}/messages`, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      "x-api-key": config.apiKey,
+      "anthropic-version": "2023-06-01"
+    },
+    body: JSON.stringify(body)
+  });
+  if (!response.ok) {
+    const errBody = await response.text();
+    throw new Error(
+      `[${config.role}-model] anthropic error ${response.status}: ${errBody.slice(0, 300)}`
+    );
+  }
+  const data = await response.json();
+  const text = (data.content ?? []).filter((c) => c.type === "text" || !c.type && typeof c.text === "string").map((c) => c.text ?? "").join("");
+  return {
+    text,
+    usage: data.usage ? {
+      promptTokens: data.usage.input_tokens,
+      completionTokens: data.usage.output_tokens,
+      totalTokens: (data.usage.input_tokens ?? 0) + (data.usage.output_tokens ?? 0),
+      cachedTokens: data.usage.cache_read_input_tokens
+    } : void 0,
+    raw: data
+  };
+}
+function dispatch(config, req) {
+  return config.providerName === "anthropic" ? callAnthropic(config, req) : callCerebras(config, req);
+}
+function getEvalModelClient() {
+  const config = resolveConfig("eval");
+  return (req) => dispatch(config, req);
+}
+function getTrainingModelClient() {
+  const config = resolveConfig("training");
+  return (req) => dispatch(config, req);
+}
+async function judgeWithCerebras(prompt, options) {
+  const response = await judgeWithCerebrasShared(prompt, options);
+  return response.raw;
+}
+async function judgeWithCerebrasShared(prompt, options) {
+  const provider = resolveProvider("eval");
+  if (provider !== "cerebras") {
+    const client = getEvalModelClient();
+    const result = await client({
+      prompt,
+      systemPrompt: options?.systemPrompt,
+      temperature: options?.temperature ?? 0,
+      maxTokens: options?.maxTokens ?? 700
+    });
+    return { raw: result.text, json: null };
+  }
+  const judge = new CerebrasJudge({
+    apiKey: resolveCerebrasApiKey("eval"),
+    baseUrl: resolveBaseUrl(),
+    model: resolveEvalModel()
+  });
+  return judge.judge(prompt, {
+    systemPrompt: options?.systemPrompt,
+    temperature: options?.temperature ?? 0,
+    maxTokens: options?.maxTokens ?? 700
+  });
+}
+function getTrainingUseModelAdapter() {
+  const client = getTrainingModelClient();
+  return async (input) => {
+    const result = await client({
+      prompt: input.prompt,
+      temperature: input.temperature,
+      maxTokens: input.maxTokens
+    });
+    return result.text;
+  };
+}
+function isCerebrasEvalEnabled() {
+  const provider = resolveProvider("eval");
+  return provider === "cerebras" && !!readEnv("CEREBRAS_API_KEY", "EVAL_CEREBRAS_API_KEY");
+}
+function isCerebrasTrainingEnabled() {
+  const provider = resolveProvider("training");
+  return provider === "cerebras" && !!readEnv("CEREBRAS_API_KEY", "TRAIN_CEREBRAS_API_KEY");
+}
+export {
+  getEvalModelClient,
+  getTrainingModelClient,
+  getTrainingUseModelAdapter,
+  isCerebrasEvalEnabled,
+  isCerebrasTrainingEnabled,
+  judgeWithCerebras,
+  judgeWithCerebrasShared
+};
+//# sourceMappingURL=cerebras-eval-model.js.map

package/dist/core/cerebras-eval-model.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/core/cerebras-eval-model.ts"],"sourcesContent":["// Cerebras / Anthropic eval+training LLM client for app-training.\n//\n// Routes optimizer scoring + variant generation through a real provider so\n// the agent under optimization is never used to grade itself. Mirrors the\n// app-lifeops eval-model helper but lives in app-training so production code\n// here never imports across another package's `test/` boundary.\n//\n// Judge-shaped calls (`judgeWithCerebras` / `judgeWithCerebrasShared`)\n// route their transport through the shared `CerebrasJudge` class in\n// scenario-runner so all four Cerebras judges in the repo share retry +\n// parsing logic.\n\nimport {\n  CerebrasJudge,\n  type JudgeResponse,\n} from \"@elizaos/scenario-runner/cerebras-judge\";\n\ninterface ResolvedClientConfig {\n  apiKey: string;\n  baseUrl: string;\n  model: string;\n  role: \"eval\" | \"training\";\n  providerName: \"cerebras\" | \"anthropic\";\n}\n\nexport interface CerebrasChatRequest {\n  prompt: string;\n  systemPrompt?: string;\n  temperature?: number;\n  maxTokens?: number;\n  reasoningEffort?: \"low\" | \"medium\" | \"high\";\n}\n\nexport interface CerebrasChatUsage {\n  promptTokens?: number;\n  completionTokens?: number;\n  totalTokens?: number;\n  cachedTokens?: number;\n}\n\nexport interface CerebrasChatResponse {\n  text: string;\n  usage?: CerebrasChatUsage;\n  raw?: unknown;\n}\n\nexport type EvalModelClient = (\n  req: CerebrasChatRequest,\n) => Promise<CerebrasChatResponse>;\n\nfunction readEnv(...keys: string[]): string | undefined {\n  for (const key of keys) {\n    const value = process.env[key]?.trim();\n    if (value && value.length > 0) return value;\n  }\n  return undefined;\n}\n\nfunction resolveCerebrasApiKey(role: \"eval\" | \"training\"): string {\n  const apiKey = readEnv(\n    role === \"eval\" ? \"EVAL_CEREBRAS_API_KEY\" : \"TRAIN_CEREBRAS_API_KEY\",\n    \"CEREBRAS_API_KEY\",\n    \"ELIZA_E2E_CEREBRAS_API_KEY\",\n  );\n  if (!apiKey) {\n    throw new Error(\n      `[${role}-model] CEREBRAS_API_KEY is not set. ` +\n        `Eval/training runs require Cerebras credentials. ` +\n        `Set CEREBRAS_API_KEY in eliza/.env.`,\n    );\n  }\n  return apiKey;\n}\n\nfunction resolveBaseUrl(): string {\n  return readEnv(\"CEREBRAS_BASE_URL\") ?? \"https://api.cerebras.ai/v1\";\n}\n\nfunction resolveEvalModel(): string {\n  return (\n    readEnv(\"EVAL_MODEL\", \"EVAL_MODEL_NAME\") ??\n    readEnv(\"CEREBRAS_MODEL\") ??\n    \"gpt-oss-120b\"\n  );\n}\n\nfunction resolveTrainingModel(): string {\n  return (\n    readEnv(\"TRAIN_MODEL\", \"TRAINING_MODEL\", \"TRAIN_MODEL_NAME\") ??\n    readEnv(\"CEREBRAS_MODEL\") ??\n    \"gpt-oss-120b\"\n  );\n}\n\nfunction resolveProvider(role: \"eval\" | \"training\"): string {\n  return (\n    readEnv(\n      role === \"eval\" ? \"EVAL_MODEL_PROVIDER\" : \"TRAIN_MODEL_PROVIDER\",\n      role === \"eval\" ? \"EVAL_PROVIDER\" : \"TRAINING_PROVIDER\",\n    ) ?? \"cerebras\"\n  );\n}\n\nfunction resolveAnthropicApiKey(role: \"eval\" | \"training\"): string {\n  const apiKey = readEnv(\"ANTHROPIC_API_KEY\");\n  if (!apiKey) {\n    throw new Error(\n      `[${role}-model] ANTHROPIC_API_KEY is not set; required when ${\n        role === \"eval\" ? \"EVAL_MODEL_PROVIDER\" : \"TRAIN_MODEL_PROVIDER\"\n      }=anthropic.`,\n    );\n  }\n  return apiKey;\n}\n\nfunction resolveAnthropicModel(role: \"eval\" | \"training\"): string {\n  const explicitAnthropic = readEnv(\"ANTHROPIC_LARGE_MODEL\");\n  if (explicitAnthropic) return explicitAnthropic;\n  if (role === \"eval\") {\n    return (\n      readEnv(\"EVAL_ANTHROPIC_MODEL\", \"EVAL_MODEL_NAME\") ??\n      \"claude-haiku-4-5-20251001\"\n    );\n  }\n  return (\n    readEnv(\"TRAIN_ANTHROPIC_MODEL\", \"TRAIN_MODEL_NAME\") ??\n    \"claude-haiku-4-5-20251001\"\n  );\n}\n\nfunction resolveConfig(role: \"eval\" | \"training\"): ResolvedClientConfig {\n  const provider = resolveProvider(role);\n  if (provider === \"cerebras\") {\n    return {\n      apiKey: resolveCerebrasApiKey(role),\n      baseUrl: resolveBaseUrl(),\n      model: role === \"eval\" ? resolveEvalModel() : resolveTrainingModel(),\n      role,\n      providerName: \"cerebras\",\n    };\n  }\n  if (provider === \"anthropic\") {\n    return {\n      apiKey: resolveAnthropicApiKey(role),\n      baseUrl: \"https://api.anthropic.com/v1\",\n      model: resolveAnthropicModel(role),\n      role,\n      providerName: \"anthropic\",\n    };\n  }\n  throw new Error(\n    `[${role}-model] unknown provider \"${provider}\"; supported: cerebras, anthropic. ` +\n      `Set ${role === \"eval\" ? \"EVAL_MODEL_PROVIDER\" : \"TRAIN_MODEL_PROVIDER\"}=cerebras|anthropic.`,\n  );\n}\n\nconst RETRYABLE_STATUS = new Set([429, 500, 502, 503, 504]);\nconst MAX_CHAT_ATTEMPTS = 5;\n\n/**\n * POST the chat-completions request, retrying transient failures (429 + 5xx +\n * network errors) with exponential backoff. A single optimizer run fans out\n * into dozens of provider calls, so one transient 500 (common on serverless\n * gpt-oss-120b relays) would otherwise abort the whole generation. A\n * non-retryable status is returned to the caller unchanged for its own\n * error-body handling.\n */\nasync function fetchChatWithRetry(\n  config: ResolvedClientConfig,\n  body: Record<string, unknown>,\n): Promise<Response> {\n  let lastError: unknown;\n  for (let attempt = 1; attempt <= MAX_CHAT_ATTEMPTS; attempt += 1) {\n    try {\n      const response = await fetch(`${config.baseUrl}/chat/completions`, {\n        method: \"POST\",\n        headers: {\n          \"Content-Type\": \"application/json\",\n          Authorization: `Bearer ${config.apiKey}`,\n        },\n        body: JSON.stringify(body),\n      });\n      if (response.ok || !RETRYABLE_STATUS.has(response.status)) {\n        return response;\n      }\n      lastError = new Error(\n        `[${config.role}-model] cerebras transient ${response.status}`,\n      );\n      // Drain the body so the socket can be reused before we retry.\n      await response.text().catch(() => undefined);\n    } catch (err) {\n      lastError = err;\n    }\n    if (attempt < MAX_CHAT_ATTEMPTS) {\n      const backoffMs = Math.min(8000, 400 * 2 ** (attempt - 1));\n      await new Promise((r) => setTimeout(r, backoffMs));\n    }\n  }\n  throw lastError instanceof Error\n    ? lastError\n    : new Error(`[${config.role}-model] chat request failed`);\n}\n\nasync function callCerebras(\n  config: ResolvedClientConfig,\n  req: CerebrasChatRequest,\n): Promise<CerebrasChatResponse> {\n  const messages: Array<{ role: \"system\" | \"user\"; content: string }> = [];\n  if (req.systemPrompt && req.systemPrompt.length > 0) {\n    messages.push({ role: \"system\", content: req.systemPrompt });\n  }\n  messages.push({ role: \"user\", content: req.prompt });\n\n  const body: Record<string, unknown> = {\n    model: config.model,\n    messages,\n    temperature: req.temperature ?? 0,\n    max_tokens: req.maxTokens ?? 1024,\n  };\n  // gpt-oss exposes a `reasoning_effort` knob; match it on the bare id and on\n  // the `<vendor>/gpt-oss-*` form used by OpenAI-compatible relays.\n  if (/(^|\\/)gpt-oss/.test(config.model)) {\n    body.reasoning_effort = req.reasoningEffort ?? \"low\";\n  }\n\n  const response = await fetchChatWithRetry(config, body);\n  if (!response.ok) {\n    const errBody = await response.text();\n    throw new Error(\n      `[${config.role}-model] cerebras error ${response.status}: ${errBody.slice(0, 300)}`,\n    );\n  }\n  const data = (await response.json()) as {\n    choices?: Array<{ message?: { content?: string } }>;\n    usage?: {\n      prompt_tokens?: number;\n      completion_tokens?: number;\n      total_tokens?: number;\n      prompt_tokens_details?: { cached_tokens?: number };\n    };\n  };\n  return {\n    text: data.choices?.[0]?.message?.content ?? \"\",\n    usage: data.usage\n      ? {\n          promptTokens: data.usage.prompt_tokens,\n          completionTokens: data.usage.completion_tokens,\n          totalTokens: data.usage.total_tokens,\n          cachedTokens: data.usage.prompt_tokens_details?.cached_tokens,\n        }\n      : undefined,\n    raw: data,\n  };\n}\n\nasync function callAnthropic(\n  config: ResolvedClientConfig,\n  req: CerebrasChatRequest,\n): Promise<CerebrasChatResponse> {\n  const body: Record<string, unknown> = {\n    model: config.model,\n    max_tokens: req.maxTokens ?? 1024,\n    temperature: req.temperature ?? 0,\n    messages: [{ role: \"user\", content: req.prompt }],\n  };\n  if (req.systemPrompt && req.systemPrompt.length > 0) {\n    body.system = req.systemPrompt;\n  }\n  const response = await fetch(`${config.baseUrl}/messages`, {\n    method: \"POST\",\n    headers: {\n      \"Content-Type\": \"application/json\",\n      \"x-api-key\": config.apiKey,\n      \"anthropic-version\": \"2023-06-01\",\n    },\n    body: JSON.stringify(body),\n  });\n  if (!response.ok) {\n    const errBody = await response.text();\n    throw new Error(\n      `[${config.role}-model] anthropic error ${response.status}: ${errBody.slice(0, 300)}`,\n    );\n  }\n  const data = (await response.json()) as {\n    content?: Array<{ type?: string; text?: string }>;\n    usage?: {\n      input_tokens?: number;\n      output_tokens?: number;\n      cache_read_input_tokens?: number;\n      cache_creation_input_tokens?: number;\n    };\n  };\n  const text = (data.content ?? [])\n    .filter((c) => c.type === \"text\" || (!c.type && typeof c.text === \"string\"))\n    .map((c) => c.text ?? \"\")\n    .join(\"\");\n  return {\n    text,\n    usage: data.usage\n      ? {\n          promptTokens: data.usage.input_tokens,\n          completionTokens: data.usage.output_tokens,\n          totalTokens:\n            (data.usage.input_tokens ?? 0) + (data.usage.output_tokens ?? 0),\n          cachedTokens: data.usage.cache_read_input_tokens,\n        }\n      : undefined,\n    raw: data,\n  };\n}\n\nfunction dispatch(\n  config: ResolvedClientConfig,\n  req: CerebrasChatRequest,\n): Promise<CerebrasChatResponse> {\n  return config.providerName === \"anthropic\"\n    ? callAnthropic(config, req)\n    : callCerebras(config, req);\n}\n\nexport function getEvalModelClient(): EvalModelClient {\n  const config = resolveConfig(\"eval\");\n  return (req) => dispatch(config, req);\n}\n\nexport function getTrainingModelClient(): EvalModelClient {\n  const config = resolveConfig(\"training\");\n  return (req) => dispatch(config, req);\n}\n\n/**\n * Cerebras-only judge helper. Routes through the shared `CerebrasJudge`\n * transport (tolerant parsing, 429/5xx retry, json_object opt-in). The\n * Cerebras eval provider is the only one configured here; callers that\n * need Anthropic should use `getEvalModelClient()` directly.\n *\n * Returns the raw model text for backward compatibility with existing\n * callers. New callers should consume `judgeWithCerebrasShared()` (below)\n * to get the canonical {raw, json, score?, verdict?, reason?} shape.\n */\nexport async function judgeWithCerebras(\n  prompt: string,\n  options?: { maxTokens?: number; temperature?: number; systemPrompt?: string },\n): Promise<string> {\n  const response = await judgeWithCerebrasShared(prompt, options);\n  return response.raw;\n}\n\n/**\n * New canonical entry: returns the full JudgeResponse for callers that\n * want the parsed score/verdict/reason without re-parsing the raw text.\n */\nexport async function judgeWithCerebrasShared(\n  prompt: string,\n  options?: { maxTokens?: number; temperature?: number; systemPrompt?: string },\n): Promise<JudgeResponse> {\n  const provider = resolveProvider(\"eval\");\n  if (provider !== \"cerebras\") {\n    // Caller asked for the eval-as-judge route but the eval provider is\n    // pinned to a non-Cerebras model. Fall back to the eval client so the\n    // judge still runs (cross-grader rule), but skip the shared CerebrasJudge\n    // transport — only Cerebras is supported there today.\n    const client = getEvalModelClient();\n    const result = await client({\n      prompt,\n      systemPrompt: options?.systemPrompt,\n      temperature: options?.temperature ?? 0,\n      maxTokens: options?.maxTokens ?? 700,\n    });\n    return { raw: result.text, json: null };\n  }\n  const judge = new CerebrasJudge({\n    apiKey: resolveCerebrasApiKey(\"eval\"),\n    baseUrl: resolveBaseUrl(),\n    model: resolveEvalModel(),\n  });\n  return judge.judge(prompt, {\n    systemPrompt: options?.systemPrompt,\n    temperature: options?.temperature ?? 0,\n    maxTokens: options?.maxTokens ?? 700,\n  });\n}\n\n// Adapter shaped like runtime.useModel(\"TEXT_LARGE\", { prompt, ... }) so\n// optimizer / prompt-compare consumers can drop it in unchanged.\nexport function getTrainingUseModelAdapter(): (input: {\n  prompt: string;\n  temperature?: number;\n  maxTokens?: number;\n}) => Promise<string> {\n  const client = getTrainingModelClient();\n  return async (input) => {\n    const result = await client({\n      prompt: input.prompt,\n      temperature: input.temperature,\n      maxTokens: input.maxTokens,\n    });\n    return result.text;\n  };\n}\n\nexport function isCerebrasEvalEnabled(): boolean {\n  const provider = resolveProvider(\"eval\");\n  return (\n    provider === \"cerebras\" &&\n    !!readEnv(\"CEREBRAS_API_KEY\", \"EVAL_CEREBRAS_API_KEY\")\n  );\n}\n\nexport function isCerebrasTrainingEnabled(): boolean {\n  const provider = resolveProvider(\"training\");\n  return (\n    provider === \"cerebras\" &&\n    !!readEnv(\"CEREBRAS_API_KEY\", \"TRAIN_CEREBRAS_API_KEY\")\n  );\n}\n"],"mappings":"AAYA;AAAA,EACE;AAAA,OAEK;AAmCP,SAAS,WAAW,MAAoC;AACtD,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,QAAQ,IAAI,GAAG,GAAG,KAAK;AACrC,QAAI,SAAS,MAAM,SAAS,EAAG,QAAO;AAAA,EACxC;AACA,SAAO;AACT;AAEA,SAAS,sBAAsB,MAAmC;AAChE,QAAM,SAAS;AAAA,IACb,SAAS,SAAS,0BAA0B;AAAA,IAC5C;AAAA,IACA;AAAA,EACF;AACA,MAAI,CAAC,QAAQ;AACX,UAAM,IAAI;AAAA,MACR,IAAI,IAAI;AAAA,IAGV;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,iBAAyB;AAChC,SAAO,QAAQ,mBAAmB,KAAK;AACzC;AAEA,SAAS,mBAA2B;AAClC,SACE,QAAQ,cAAc,iBAAiB,KACvC,QAAQ,gBAAgB,KACxB;AAEJ;AAEA,SAAS,uBAA+B;AACtC,SACE,QAAQ,eAAe,kBAAkB,kBAAkB,KAC3D,QAAQ,gBAAgB,KACxB;AAEJ;AAEA,SAAS,gBAAgB,MAAmC;AAC1D,SACE;AAAA,IACE,SAAS,SAAS,wBAAwB;AAAA,IAC1C,SAAS,SAAS,kBAAkB;AAAA,EACtC,KAAK;AAET;AAEA,SAAS,uBAAuB,MAAmC;AACjE,QAAM,SAAS,QAAQ,mBAAmB;AAC1C,MAAI,CAAC,QAAQ;AACX,UAAM,IAAI;AAAA,MACR,IAAI,IAAI,uDACN,SAAS,SAAS,wBAAwB,sBAC5C;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,sBAAsB,MAAmC;AAChE,QAAM,oBAAoB,QAAQ,uBAAuB;AACzD,MAAI,kBAAmB,QAAO;AAC9B,MAAI,SAAS,QAAQ;AACnB,WACE,QAAQ,wBAAwB,iBAAiB,KACjD;AAAA,EAEJ;AACA,SACE,QAAQ,yBAAyB,kBAAkB,KACnD;AAEJ;AAEA,SAAS,cAAc,MAAiD;AACtE,QAAM,WAAW,gBAAgB,IAAI;AACrC,MAAI,aAAa,YAAY;AAC3B,WAAO;AAAA,MACL,QAAQ,sBAAsB,IAAI;AAAA,MAClC,SAAS,eAAe;AAAA,MACxB,OAAO,SAAS,SAAS,iBAAiB,IAAI,qBAAqB;AAAA,MACnE;AAAA,MACA,cAAc;AAAA,IAChB;AAAA,EACF;AACA,MAAI,aAAa,aAAa;AAC5B,WAAO;AAAA,MACL,QAAQ,uBAAuB,IAAI;AAAA,MACnC,SAAS;AAAA,MACT,OAAO,sBAAsB,IAAI;AAAA,MACjC;AAAA,MACA,cAAc;AAAA,IAChB;AAAA,EACF;AACA,QAAM,IAAI;AAAA,IACR,IAAI,IAAI,6BAA6B,QAAQ,0CACpC,SAAS,SAAS,wBAAwB,sBAAsB;AAAA,EAC3E;AACF;AAEA,MAAM,mBAAmB,oBAAI,IAAI,CAAC,KAAK,KAAK,KAAK,KAAK,GAAG,CAAC;AAC1D,MAAM,oBAAoB;AAU1B,eAAe,mBACb,QACA,MACmB;AACnB,MAAI;AACJ,WAAS,UAAU,GAAG,WAAW,mBAAmB,WAAW,GAAG;AAChE,QAAI;AACF,YAAM,WAAW,MAAM,MAAM,GAAG,OAAO,OAAO,qBAAqB;AAAA,QACjE,QAAQ;AAAA,QACR,SAAS;AAAA,UACP,gBAAgB;AAAA,UAChB,eAAe,UAAU,OAAO,MAAM;AAAA,QACxC;AAAA,QACA,MAAM,KAAK,UAAU,IAAI;AAAA,MAC3B,CAAC;AACD,UAAI,SAAS,MAAM,CAAC,iBAAiB,IAAI,SAAS,MAAM,GAAG;AACzD,eAAO;AAAA,MACT;AACA,kBAAY,IAAI;AAAA,QACd,IAAI,OAAO,IAAI,8BAA8B,SAAS,MAAM;AAAA,MAC9D;AAEA,YAAM,SAAS,KAAK,EAAE,MAAM,MAAM,MAAS;AAAA,IAC7C,SAAS,KAAK;AACZ,kBAAY;AAAA,IACd;AACA,QAAI,UAAU,mBAAmB;AAC/B,YAAM,YAAY,KAAK,IAAI,KAAM,MAAM,MAAM,UAAU,EAAE;AACzD,YAAM,IAAI,QAAQ,CAAC,MAAM,WAAW,GAAG,SAAS,CAAC;AAAA,IACnD;AAAA,EACF;AACA,QAAM,qBAAqB,QACvB,YACA,IAAI,MAAM,IAAI,OAAO,IAAI,6BAA6B;AAC5D;AAEA,eAAe,aACb,QACA,KAC+B;AAC/B,QAAM,WAAgE,CAAC;AACvE,MAAI,IAAI,gBAAgB,IAAI,aAAa,SAAS,GAAG;AACnD,aAAS,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,aAAa,CAAC;AAAA,EAC7D;AACA,WAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,IAAI,OAAO,CAAC;AAEnD,QAAM,OAAgC;AAAA,IACpC,OAAO,OAAO;AAAA,IACd;AAAA,IACA,aAAa,IAAI,eAAe;AAAA,IAChC,YAAY,IAAI,aAAa;AAAA,EAC/B;AAGA,MAAI,gBAAgB,KAAK,OAAO,KAAK,GAAG;AACtC,SAAK,mBAAmB,IAAI,mBAAmB;AAAA,EACjD;AAEA,QAAM,WAAW,MAAM,mBAAmB,QAAQ,IAAI;AACtD,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,UAAU,MAAM,SAAS,KAAK;AACpC,UAAM,IAAI;AAAA,MACR,IAAI,OAAO,IAAI,0BAA0B,SAAS,MAAM,KAAK,QAAQ,MAAM,GAAG,GAAG,CAAC;AAAA,IACpF;AAAA,EACF;AACA,QAAM,OAAQ,MAAM,SAAS,KAAK;AASlC,SAAO;AAAA,IACL,MAAM,KAAK,UAAU,CAAC,GAAG,SAAS,WAAW;AAAA,IAC7C,OAAO,KAAK,QACR;AAAA,MACE,cAAc,KAAK,MAAM;AAAA,MACzB,kBAAkB,KAAK,MAAM;AAAA,MAC7B,aAAa,KAAK,MAAM;AAAA,MACxB,cAAc,KAAK,MAAM,uBAAuB;AAAA,IAClD,IACA;AAAA,IACJ,KAAK;AAAA,EACP;AACF;AAEA,eAAe,cACb,QACA,KAC+B;AAC/B,QAAM,OAAgC;AAAA,IACpC,OAAO,OAAO;AAAA,IACd,YAAY,IAAI,aAAa;AAAA,IAC7B,aAAa,IAAI,eAAe;AAAA,IAChC,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,IAAI,OAAO,CAAC;AAAA,EAClD;AACA,MAAI,IAAI,gBAAgB,IAAI,aAAa,SAAS,GAAG;AACnD,SAAK,SAAS,IAAI;AAAA,EACpB;AACA,QAAM,WAAW,MAAM,MAAM,GAAG,OAAO,OAAO,aAAa;AAAA,IACzD,QAAQ;AAAA,IACR,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,aAAa,OAAO;AAAA,MACpB,qBAAqB;AAAA,IACvB;AAAA,IACA,MAAM,KAAK,UAAU,IAAI;AAAA,EAC3B,CAAC;AACD,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,UAAU,MAAM,SAAS,KAAK;AACpC,UAAM,IAAI;AAAA,MACR,IAAI,OAAO,IAAI,2BAA2B,SAAS,MAAM,KAAK,QAAQ,MAAM,GAAG,GAAG,CAAC;AAAA,IACrF;AAAA,EACF;AACA,QAAM,OAAQ,MAAM,SAAS,KAAK;AASlC,QAAM,QAAQ,KAAK,WAAW,CAAC,GAC5B,OAAO,CAAC,MAAM,EAAE,SAAS,UAAW,CAAC,EAAE,QAAQ,OAAO,EAAE,SAAS,QAAS,EAC1E,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,EACvB,KAAK,EAAE;AACV,SAAO;AAAA,IACL;AAAA,IACA,OAAO,KAAK,QACR;AAAA,MACE,cAAc,KAAK,MAAM;AAAA,MACzB,kBAAkB,KAAK,MAAM;AAAA,MAC7B,cACG,KAAK,MAAM,gBAAgB,MAAM,KAAK,MAAM,iBAAiB;AAAA,MAChE,cAAc,KAAK,MAAM;AAAA,IAC3B,IACA;AAAA,IACJ,KAAK;AAAA,EACP;AACF;AAEA,SAAS,SACP,QACA,KAC+B;AAC/B,SAAO,OAAO,iBAAiB,cAC3B,cAAc,QAAQ,GAAG,IACzB,aAAa,QAAQ,GAAG;AAC9B;AAEO,SAAS,qBAAsC;AACpD,QAAM,SAAS,cAAc,MAAM;AACnC,SAAO,CAAC,QAAQ,SAAS,QAAQ,GAAG;AACtC;AAEO,SAAS,yBAA0C;AACxD,QAAM,SAAS,cAAc,UAAU;AACvC,SAAO,CAAC,QAAQ,SAAS,QAAQ,GAAG;AACtC;AAYA,eAAsB,kBACpB,QACA,SACiB;AACjB,QAAM,WAAW,MAAM,wBAAwB,QAAQ,OAAO;AAC9D,SAAO,SAAS;AAClB;AAMA,eAAsB,wBACpB,QACA,SACwB;AACxB,QAAM,WAAW,gBAAgB,MAAM;AACvC,MAAI,aAAa,YAAY;AAK3B,UAAM,SAAS,mBAAmB;AAClC,UAAM,SAAS,MAAM,OAAO;AAAA,MAC1B;AAAA,MACA,cAAc,SAAS;AAAA,MACvB,aAAa,SAAS,eAAe;AAAA,MACrC,WAAW,SAAS,aAAa;AAAA,IACnC,CAAC;AACD,WAAO,EAAE,KAAK,OAAO,MAAM,MAAM,KAAK;AAAA,EACxC;AACA,QAAM,QAAQ,IAAI,cAAc;AAAA,IAC9B,QAAQ,sBAAsB,MAAM;AAAA,IACpC,SAAS,eAAe;AAAA,IACxB,OAAO,iBAAiB;AAAA,EAC1B,CAAC;AACD,SAAO,MAAM,MAAM,QAAQ;AAAA,IACzB,cAAc,SAAS;AAAA,IACvB,aAAa,SAAS,eAAe;AAAA,IACrC,WAAW,SAAS,aAAa;AAAA,EACnC,CAAC;AACH;AAIO,SAAS,6BAIM;AACpB,QAAM,SAAS,uBAAuB;AACtC,SAAO,OAAO,UAAU;AACtB,UAAM,SAAS,MAAM,OAAO;AAAA,MAC1B,QAAQ,MAAM;AAAA,MACd,aAAa,MAAM;AAAA,MACnB,WAAW,MAAM;AAAA,IACnB,CAAC;AACD,WAAO,OAAO;AAAA,EAChB;AACF;AAEO,SAAS,wBAAiC;AAC/C,QAAM,WAAW,gBAAgB,MAAM;AACvC,SACE,aAAa,cACb,CAAC,CAAC,QAAQ,oBAAoB,uBAAuB;AAEzD;AAEO,SAAS,4BAAqC;AACnD,QAAM,WAAW,gBAAgB,UAAU;AAC3C,SACE,aAAa,cACb,CAAC,CAAC,QAAQ,oBAAoB,wBAAwB;AAE1D;","names":[]}

package/dist/core/cli.d.ts

@@ -0,0 +1,15 @@
+/**
+ * CLI entry point for the training data pipeline.
+ *
+ * Usage (from repo root):
+ *   bun run eliza/plugins/plugin-training/src/core/cli.ts generate --variants 5 --output ./training-data
+ *   bun run eliza/plugins/plugin-training/src/core/cli.ts validate --input ./training-data/raw_samples.json
+ *   bun run eliza/plugins/plugin-training/src/core/cli.ts export-trajectories --output ./training-data/trajectories.jsonl
+ * Or: `cd eliza/packages/agent && bun run training:cli` (delegates to this file).
+ */
+import { type ListTrainingCollectionsResult, type TrainingCollectionPreflightSummary, type TrainingCollectionRunOptions, type TrainingCollectionRunResult } from "./training-collection-runner.js";
+export declare function buildRunCollectionOptionsFromCliArgs(args: string[]): TrainingCollectionRunOptions;
+export declare function formatTrainingCollectionPreflightSummary(preflight: TrainingCollectionPreflightSummary): string[];
+export declare function formatListTrainingCollectionsSummary(result: ListTrainingCollectionsResult): string[];
+export declare function formatRunCollectionSummary(result: TrainingCollectionRunResult): string[];
+//# sourceMappingURL=cli.d.ts.map

package/dist/core/cli.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"cli.d.ts","sourceRoot":"","sources":["../../src/core/cli.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAqCH,OAAO,EAEL,KAAK,6BAA6B,EAGlC,KAAK,kCAAkC,EACvC,KAAK,4BAA4B,EACjC,KAAK,2BAA2B,EACjC,MAAM,iCAAiC,CAAC;AAufzC,wBAAgB,oCAAoC,CAClD,IAAI,EAAE,MAAM,EAAE,GACb,4BAA4B,CA4Q9B;AAED,wBAAgB,wCAAwC,CACtD,SAAS,EAAE,kCAAkC,GAC5C,MAAM,EAAE,CAeV;AAwCD,wBAAgB,oCAAoC,CAClD,MAAM,EAAE,6BAA6B,GACpC,MAAM,EAAE,CAyGV;AAuBD,wBAAgB,0BAA0B,CACxC,MAAM,EAAE,2BAA2B,GAClC,MAAM,EAAE,CAkHV"}