@acentric/llm 0.0.0-stage → 0.1.0

package/dist/usage.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"usage.d.ts","sourceRoot":"","sources":["../src/usage.ts"],"names":[],"mappings":"AAEA,4FAA4F;AAC5F,MAAM,MAAM,KAAK,GAAG;IAClB,yCAAyC;IACzC,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAC7B,8DAA8D;IAC9D,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,gFAAgF;IAChF,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;IACzB,+EAA+E;IAC/E,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAC;IAC5B,iEAAiE;IACjE,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAC7B;;;;OAIG;IACH,QAAQ,CAAC,IAAI,CAAC,EAAE,SAAS,CAAC;CAC3B,CAAC;AAEF,wDAAwD;AACxD,MAAM,MAAM,SAAS,GAAG;IACtB,0FAA0F;IAC1F,QAAQ,CAAC,MAAM,EAAE,UAAU,GAAG,WAAW,CAAC;IAC1C,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAC7B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB,CAAC"}

package/dist/usage.js

@@ -0,0 +1,3 @@
+// Tokens and cost, the same for every provider.
+export {};
+//# sourceMappingURL=usage.js.map

package/dist/usage.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"usage.js","sourceRoot":"","sources":["../src/usage.ts"],"names":[],"mappings":"AAAA,gDAAgD"}

package/dist/version.d.ts

@@ -0,0 +1,2 @@
+export declare const LLM_VERSION = "0.1.0";
+//# sourceMappingURL=version.d.ts.map

package/dist/version.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"version.d.ts","sourceRoot":"","sources":["../src/version.ts"],"names":[],"mappings":"AAKA,eAAO,MAAM,WAAW,UAAU,CAAC"}

package/dist/version.js

@@ -0,0 +1,6 @@
+// This package's version. The client sends it to the gateway, so a gateway
+// running an older copy can refuse a newer client's request plainly instead
+// of failing on a shape it does not know. A test keeps it equal to
+// package.json's version.
+export const LLM_VERSION = '0.1.0';
+//# sourceMappingURL=version.js.map

package/dist/version.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"version.js","sourceRoot":"","sources":["../src/version.ts"],"names":[],"mappings":"AAAA,2EAA2E;AAC3E,4EAA4E;AAC5E,mEAAmE;AACnE,0BAA0B;AAE1B,MAAM,CAAC,MAAM,WAAW,GAAG,OAAO,CAAC"}

package/package.json

@@ -1,6 +1,36 @@
 {
   "name": "@acentric/llm",
-  "version": "0.0.0-stage",
-  "stub": true,
-  "description": "Temporary package placeholder for staged publishing"
+  "version": "0.1.0",
+  "description": "Calls LLM providers from an agent's handlers, directly or through the acentric gateway, with readers for replies that an agent's pure half can use.",
+  "type": "module",
+  "sideEffects": false,
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "default": "./dist/index.js"
+    },
+    "./client": {
+      "types": "./dist/client.d.ts",
+      "default": "./dist/client.js"
+    }
+  },
+  "files": [
+    "dist",
+    "src"
+  ],
+  "publishConfig": {
+    "access": "public"
+  },
+  "dependencies": {
+    "@acentric/types": "0.1.0"
+  },
+  "devDependencies": {
+    "@acentric/runner-memory": "0.1.0",
+    "@types/node": "^26.6.5",
+    "esbuild": "^0.28.2"
+  },
+  "scripts": {
+    "build": "tsc --build",
+    "clean": "tsc --build --clean"
+  }
 }

package/src/catalog/anthropic.ts

@@ -0,0 +1,83 @@
+// Claude 5 models. Standard prices in US dollars per million tokens, from
+// https://platform.claude.com/docs/en/about-claude/pricing and the model pages,
+// as of October 2026. A cache write costs 1.25 times the input price for five
+// minutes and twice for an hour; reads are priced per model. Only Haiku 5.5
+// has a long-context price: for the whole request, output included, once the
+// prompt is over 100,000 tokens.
+
+import type { ModelCost, ModelInfo } from '../model.js';
+
+const LIMITS = { contextWindow: 1_000_000, maxOutputTokens: 128_000 } as const;
+const INPUT = ['text', 'image', 'file'] as const;
+
+function cost(
+  input: number,
+  output: number,
+  cacheWrite: number,
+  cacheWrite1h: number,
+  cacheRead: number,
+): ModelCost {
+  return { input, output, cacheWrite, cacheWrite1h, cacheRead };
+}
+
+export const ANTHROPIC_MODELS = [
+  {
+    provider: 'anthropic',
+    id: 'claude-fable-5-1',
+    name: 'Claude Fable 5.1',
+    ...LIMITS,
+    input: INPUT,
+    pricing: { base: cost(10, 50, 12.5, 20, 0.25) },
+  },
+  {
+    provider: 'anthropic',
+    id: 'claude-opus-5-5',
+    name: 'Claude Opus 5.5',
+    ...LIMITS,
+    input: INPUT,
+    pricing: { base: cost(4, 20, 5, 8, 0.2) },
+  },
+  {
+    provider: 'anthropic',
+    id: 'claude-sonnet-5-5',
+    name: 'Claude Sonnet 5.5',
+    ...LIMITS,
+    input: INPUT,
+    pricing: { base: cost(2, 10, 2.5, 4, 0.1) },
+  },
+  {
+    provider: 'anthropic',
+    id: 'claude-haiku-5-5',
+    name: 'Claude Haiku 5.5',
+    ...LIMITS,
+    input: INPUT,
+    pricing: {
+      base: cost(0.1, 0.5, 0.125, 0.2, 0.01),
+      tiers: [{ aboveInputTokens: 100_000, cost: cost(0.5, 2.5, 0.625, 1, 0.05) }],
+    },
+  },
+  {
+    provider: 'anthropic',
+    id: 'claude-fable-5',
+    name: 'Claude Fable 5',
+    ...LIMITS,
+    input: INPUT,
+    pricing: { base: cost(10, 50, 12.5, 20, 1) },
+  },
+  {
+    provider: 'anthropic',
+    id: 'claude-opus-5',
+    name: 'Claude Opus 5',
+    ...LIMITS,
+    input: INPUT,
+    pricing: { base: cost(5, 25, 6.25, 10, 0.5) },
+  },
+  {
+    provider: 'anthropic',
+    id: 'claude-sonnet-5',
+    name: 'Claude Sonnet 5',
+    ...LIMITS,
+    input: INPUT,
+    pricing: { base: cost(2, 10, 2.5, 4, 0.2) },
+  },
+] as const satisfies readonly ModelInfo<'anthropic'>[];

package/src/catalog/fireworks.ts

@@ -0,0 +1,121 @@
+// Fireworks serverless models. Standard prices in US dollars per million
+// tokens, from https://docs.fireworks.ai/serverless/pricing and the model
+// pages, as of October 2026. Fireworks has no long-context prices, charges
+// nothing for cache writes and documents no output limit per model. The
+// "fast" routers are model IDs of their own, with their own prices.
+
+import type { ModelPricing, ModelInfo } from '../model.js';
+
+const VISION = ['text', 'image'] as const;
+const TEXT = ['text'] as const;
+const MILLION = 1_048_576;
+
+function cost(input: number, cacheRead: number, output: number): ModelPricing {
+  return { base: { input, cacheRead, output } };
+}
+
+export const FIREWORKS_MODELS = [
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/kimi-k3',
+    name: 'Kimi K3',
+    contextWindow: MILLION,
+    input: VISION,
+    pricing: cost(3, 0.3, 15),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/routers/kimi-k3-fast',
+    name: 'Kimi K3 Fast',
+    contextWindow: MILLION,
+    input: VISION,
+    pricing: cost(4.5, 0.45, 22.5),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/ember-1',
+    name: 'Ember-1',
+    contextWindow: MILLION,
+    input: VISION,
+    pricing: cost(3, 0.3, 15),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/deepseek-v4p1-flash',
+    name: 'DeepSeek V4.1 Flash',
+    contextWindow: MILLION,
+    input: VISION,
+    pricing: cost(0.3, 0.006, 1.2),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/glm-5p3',
+    name: 'GLM 5.3',
+    contextWindow: MILLION,
+    input: TEXT,
+    pricing: cost(1.4, 0.26, 4.4),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/routers/glm-5p3-fast',
+    name: 'GLM 5.3 Fast',
+    contextWindow: MILLION,
+    input: TEXT,
+    pricing: cost(2.1, 0.39, 6.6),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/glm-5p3-flash',
+    name: 'GLM 5.3 Flash',
+    contextWindow: MILLION,
+    input: VISION,
+    pricing: cost(0.15, 0.03, 0.5),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/qwen3p8-max',
+    name: 'Qwen 3.8 Max',
+    input: VISION,
+    pricing: cost(2, 0.25, 6),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/minimax-m3',
+    name: 'MiniMax M3',
+    contextWindow: 512_000,
+    input: TEXT,
+    pricing: cost(0.3, 0.06, 1.2),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/gpt-oss-120b',
+    name: 'gpt-oss-120b',
+    contextWindow: 131_072,
+    input: TEXT,
+    pricing: cost(0.15, 0.015, 0.6),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/nemotron-lightning-3p5-30b-a3b',
+    name: 'Nemotron Lightning 3.5 30B A3B',
+    contextWindow: 262_144,
+    input: TEXT,
+    pricing: cost(0.05, 0.01, 0.2),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/nemotron-3-ultra-nvfp4',
+    name: 'Nemotron 3 Ultra (Preview)',
+    contextWindow: 262_144,
+    input: TEXT,
+    pricing: cost(0.6, 0.12, 2.4),
+  },
+  {
+    provider: 'fireworks',
+    id: 'accounts/fireworks/models/inkling',
+    name: 'Inkling',
+    contextWindow: MILLION,
+    input: VISION,
+    pricing: cost(1, 0.17, 4.05),
+  },
+] as const satisfies readonly ModelInfo<'fireworks'>[];

package/src/catalog/openai.ts

@@ -0,0 +1,133 @@
+// OpenAI models. Standard prices in US dollars per million tokens, and limits,
+// from https://developers.openai.com/api/docs/pricing and the model pages, as
+// of October 2026. Long-context prices are for the whole request once the
+// prompt is over 272,000 tokens.
+
+import type { ModelCost, ModelInfo, ModelPricing } from '../model.js';
+
+function longContext(base: ModelCost, cost: ModelCost): ModelPricing {
+  return { base, tiers: [{ aboveInputTokens: 272_000, cost }] };
+}
+
+const GPT_6_LIMITS = { contextWindow: 1_050_000, maxOutputTokens: 128_000 } as const;
+const VISION = ['text', 'image'] as const;
+
+export const OPENAI_MODELS = [
+  {
+    provider: 'openai',
+    id: 'gpt-6-astra',
+    name: 'GPT-6 Astra',
+    ...GPT_6_LIMITS,
+    input: VISION,
+    pricing: longContext(
+      { input: 10, cacheRead: 1, cacheWrite: 12.5, output: 50 },
+      { input: 20, cacheRead: 2, output: 75 },
+    ),
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-6.1-sol',
+    name: 'GPT-6.1 Sol',
+    ...GPT_6_LIMITS,
+    input: VISION,
+    pricing: longContext(
+      { input: 2, cacheRead: 0.1, cacheWrite: 2.5, output: 10 },
+      { input: 4, cacheRead: 0.2, output: 15 },
+    ),
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-6-sol',
+    name: 'GPT-6 Sol',
+    ...GPT_6_LIMITS,
+    input: VISION,
+    pricing: longContext(
+      { input: 2, cacheRead: 0.2, cacheWrite: 2.5, output: 10 },
+      { input: 4, cacheRead: 0.4, output: 15 },
+    ),
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-6-luna',
+    name: 'GPT-6 Luna',
+    ...GPT_6_LIMITS,
+    input: VISION,
+    pricing: longContext(
+      { input: 0.1, cacheRead: 0.01, cacheWrite: 0.125, output: 0.5 },
+      { input: 0.2, cacheRead: 0.02, output: 0.75 },
+    ),
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-5.6-sol',
+    name: 'GPT-5.6 Sol',
+    ...GPT_6_LIMITS,
+    input: VISION,
+    pricing: longContext(
+      { input: 4, cacheRead: 0.4, cacheWrite: 5, output: 20 },
+      { input: 8, cacheRead: 0.8, output: 30 },
+    ),
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-5.6-terra',
+    name: 'GPT-5.6 Terra',
+    ...GPT_6_LIMITS,
+    input: VISION,
+    pricing: longContext(
+      { input: 2, cacheRead: 0.2, cacheWrite: 2.5, output: 12 },
+      { input: 4, cacheRead: 0.4, output: 18 },
+    ),
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-5.6-luna',
+    name: 'GPT-5.6 Luna',
+    ...GPT_6_LIMITS,
+    input: VISION,
+    pricing: longContext(
+      { input: 0.2, cacheRead: 0.02, cacheWrite: 0.25, output: 1.2 },
+      { input: 0.4, cacheRead: 0.04, output: 1.8 },
+    ),
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-5.5',
+    name: 'GPT-5.5',
+    ...GPT_6_LIMITS,
+    input: VISION,
+    pricing: longContext(
+      { input: 5, cacheRead: 0.5, output: 30 },
+      { input: 10, cacheRead: 1, output: 45 },
+    ),
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-5.4',
+    name: 'GPT-5.4',
+    ...GPT_6_LIMITS,
+    input: VISION,
+    pricing: longContext(
+      { input: 2.5, cacheRead: 0.25, output: 15 },
+      { input: 5, cacheRead: 0.5, output: 22.5 },
+    ),
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-5.4-mini',
+    name: 'GPT-5.4 mini',
+    contextWindow: 400_000,
+    maxOutputTokens: 128_000,
+    input: VISION,
+    pricing: { base: { input: 0.75, cacheRead: 0.075, output: 4.5 } },
+  },
+  {
+    provider: 'openai',
+    id: 'gpt-4.1',
+    name: 'GPT-4.1',
+    contextWindow: 1_047_576,
+    maxOutputTokens: 32_768,
+    input: VISION,
+    pricing: { base: { input: 2, cacheRead: 0.5, output: 8 } },
+  },
+] as const satisfies readonly ModelInfo<'openai'>[];

package/src/client/anthropic/errors.ts

@@ -0,0 +1,88 @@
+// Anthropic failures read into LlmErrors. Anthropic answers
+// `{ type: "error", error: { type, message, details? }, request_id }`; a 413
+// and errors from its edge have other shapes, or a text body.
+
+import { LlmError } from '../../failure.js';
+import type { LlmFailureKind } from '../../failure.js';
+import { isObject } from '../../json.js';
+import { retryAfter } from '../http.js';
+import type { HttpAnswer } from '../http.js';
+import { cut } from '../util.js';
+
+const KIND_BY_TYPE: { readonly [type: string]: LlmFailureKind } = {
+  authentication_error: 'auth',
+  permission_error: 'auth',
+  billing_error: 'quota_exceeded',
+  not_found_error: 'not_found',
+  rate_limit_error: 'rate_limited',
+  request_too_large: 'invalid_request',
+  overloaded_error: 'overloaded',
+  timeout_error: 'timeout',
+  api_error: 'provider_error',
+};
+
+/** A validation message starts with the field at fault: `messages.1.content.0: …`. */
+const FIELD_PREFIX = /^([A-Za-z_][\w-]*(?:\.[\w-]+)*): /;
+
+export function httpFailure(answer: HttpAnswer): LlmError {
+  const { body, status, headers } = answer;
+  const error = isObject(body) && isObject(body['error']) ? body['error'] : undefined;
+  const type = typeof error?.['type'] === 'string' ? error['type'] : undefined;
+  const message = typeof error?.['message'] === 'string' ? error['message'] : undefined;
+  const details = isObject(error?.['details']) ? error['details'] : undefined;
+  const code = typeof details?.['error_code'] === 'string' ? details['error_code'] : undefined;
+  const param = message?.match(FIELD_PREFIX)?.[1];
+  const after = retryAfter(headers);
+  const shouldRetry = headers.get('x-should-retry');
+  return new LlmError(
+    message ??
+      (typeof body === 'string' && body !== '' ? body : undefined) ??
+      `the Anthropic call failed with HTTP ${status}`,
+    {
+      kind: classify(status, type, message, code),
+      provider: 'anthropic',
+      httpStatus: status,
+      providerError: cut(error ?? body),
+      ...(code === undefined ? {} : { providerCode: code }),
+      ...(type === undefined ? {} : { providerType: type }),
+      ...(param === undefined ? {} : { providerParam: param }),
+      ...(after === undefined ? {} : { retryAfterMs: after }),
+      ...(shouldRetry === 'true'
+        ? { retryable: true }
+        : shouldRetry === 'false'
+          ? { retryable: false }
+          : {}),
+    },
+  );
+}
+
+function classify(
+  status: number,
+  type: string | undefined,
+  message: string | undefined,
+  code: string | undefined,
+): LlmFailureKind {
+  // A usage tier's monthly spend cap: a rate limit that lasts until next month.
+  if (code === 'enforced_spend_limit_reached') return 'quota_exceeded';
+  if (message !== undefined && type === 'invalid_request_error') {
+    // "prompt is too long: 205024 tokens > 200000 maximum", with no code.
+    if (/^prompt is too long/i.test(message)) return 'context_overflow';
+    if (
+      /reached your specified (workspace )?API usage limits|credit balance is too low/i.test(
+        message,
+      )
+    ) {
+      return 'quota_exceeded';
+    }
+  }
+  const byType = type === undefined ? undefined : KIND_BY_TYPE[type];
+  if (byType !== undefined) return byType;
+  if (status === 401 || status === 403) return 'auth';
+  if (status === 402) return 'quota_exceeded';
+  if (status === 404) return 'not_found';
+  if (status === 429) return 'rate_limited';
+  if (status === 503 || status === 529) return 'overloaded';
+  if (status === 504) return 'timeout';
+  if (status >= 500) return 'provider_error';
+  return 'invalid_request';
+}