vern-llm 0.2.1 → 0.3.0

package/README.md

@@ -1,6 +1,28 @@
-# vern-llm
+<p align="center">
+  <img src="https://raw.githubusercontent.com/LakBud/vernLLM/main/apps/docs/public/logo.png" alt="vern-llm logo" width="96" />
+</p>
 
-Retry + timeout + cache wrapper for OpenAI-compatible chat completion calls (OpenAI, Groq, Anthropic via adapter).
+<h1 align="center">vern-llm</h1>
+
+<p align="center">
+  <a href="https://github.com/LakBud/vernLLM">GitHub</a> ·
+  <a href="https://vern-llm.vercel.app">Documentation</a> ·
+  <a href="https://www.npmjs.com/package/vern-llm">npm</a>
+</p>
+
+<p align="center">
+  <a href="https://www.npmjs.com/package/vern-llm"><img src="https://img.shields.io/npm/v/vern-llm.svg" alt="npm version" /></a>
+  <a href="https://www.npmjs.com/package/vern-llm"><img src="https://img.shields.io/npm/dm/vern-llm.svg" alt="npm downloads" /></a>
+  <a href="https://bundlephobia.com/package/vern-llm"><img src="https://img.shields.io/bundlephobia/minzip/vern-llm.svg" alt="bundle size" /></a>
+  <a href="https://github.com/LakBud/vernLLM/actions/workflows/test.yml"><img src="https://github.com/LakBud/vernLLM/actions/workflows/test.yml/badge.svg" alt="test status" /></a>
+  <a href="https://github.com/LakBud/vernLLM/blob/main/LICENSE.md"><img src="https://img.shields.io/npm/l/vern-llm.svg" alt="license" /></a>
+  <img src="https://img.shields.io/node/v/vern-llm.svg" alt="node version" />
+  <img src="https://img.shields.io/badge/TypeScript-strict-3178C6?logo=typescript&logoColor=white" alt="TypeScript" />
+</p>
+
+Production-ready resilience for LLM calls; Retries, timeouts, caching, and circuit breaking for any OpenAI-compatible chat completion API (OpenAI, Groq, Anthropic, Gemini, Bedrock, and more via adapters).
+
+**Full documentation: [vern-llm.vercel.app](https://vern-llm.vercel.app)** — installation, structured output, caching, circuit breaker, every adapter, and the complete API reference all live there and are kept up to date. This README is a quick pitch, not the manual.
 
 ## Install
 
@@ -8,300 +30,37 @@ Retry + timeout + cache wrapper for OpenAI-compatible chat completion calls (Ope
 pnpm add vern-llm openai
 ```
 
-## Basic usage
+## Quick start
 
 ```ts
 import OpenAI from 'openai';
 import { VernLLM } from 'vern-llm';
 
-const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
-
 const llm = new VernLLM({
-  client: openai,
+  client: new OpenAI({ apiKey: process.env.OPENAI_API_KEY }),
   model: 'gpt-4o',
+  maxRetries: 3,
+  timeoutMs: 10_000,
+  circuitBreaker: true,
 });
 
-const parsed = await llm.call({
+const result = await llm.call({
   systemPrompt: 'Return JSON: { "skills": string[] }',
   userContent: 'Extract skills from: ...',
 });
-
-// jsonMode: false returns the raw string, no parsing
-const text = await llm.call({
-  systemPrompt: 'Summarize this.',
-  userContent: '...',
-  jsonMode: false,
-});
-```
-
-`VernLLM` is also exported as `RetryLLM` — same class, same behavior, pick whichever name reads better in your codebase:
-
-```ts
-import { RetryLLM } from 'vern-llm';
-
-const llm = new RetryLLM({ client: openai, model: 'gpt-4o' });
-```
-
-## Structured output with Zod
-
-Pass a `schema` and get a typed, validated result. Works with any validator exposing `safeParse` (Zod v3/v4).
-
-```ts
-import { z } from 'zod';
-
-const CandidateSchema = z.object({
-  name: z.string(),
-  skills: z.array(z.string()),
-});
-
-// result is typed as z.infer<typeof CandidateSchema>
-const result = await llm.call({
-  systemPrompt: 'Extract the candidate name and skills as JSON.',
-  userContent: resumeText,
-  schema: CandidateSchema,
-});
-```
-
-On a schema mismatch, `call` throws `LLMError('validation')` with `.issues` set to the validator's error object, without burning a retry (validation failures are deterministic).
-
-## Provider-native JSON Schema mode
-
-`schema` above only validates client-side, after the model has already responded. `jsonSchema` sends the schema to the provider (OpenAI/Groq `response_format: { type: 'json_schema' }`), constraining generation itself — the model can't produce a shape that violates it. Combine both for provider-level constraint plus typed client-side inference:
-
-```ts
-const result = await llm.call({
-  systemPrompt: 'Extract the candidate name and skills.',
-  userContent: resumeText,
-  jsonSchema: {
-    name: 'Candidate',
-    schema: {
-      type: 'object',
-      properties: {
-        name: { type: 'string' },
-        skills: { type: 'array', items: { type: 'string' } },
-      },
-      required: ['name', 'skills'],
-    },
-  },
-  schema: CandidateSchema, // optional client-side type/validation on top
-});
-```
-
-`jsonSchema` implies JSON mode. The Anthropic adapter has no native structured-output equivalent, so it embeds the schema in the system prompt as an instruction instead of provider-enforcing it.
-
-## Per-call model override and reasoning effort
-
-```ts
-const llm = new VernLLM({ client: openai, model: 'gpt-4o-mini' }); // default model
-
-await llm.call({
-  systemPrompt: '...',
-  userContent: '...',
-  model: 'o3', // overrides the instance default for this call only
-  reasoningEffort: 'high', // passed through as `reasoning_effort`; ignored by models that don't support it
-});
 ```
 
-`reasoningEffort` is dropped by the Anthropic adapter — Claude's extended thinking uses a token budget, not an effort tier, so there's no faithful 1:1 mapping.
-
-## Caching + usage metering
-
-```ts
-import { InMemoryCacheAdapter } from 'vern-llm';
-
-const llm = new VernLLM({
-  client: openai,
-  model: 'gpt-4o',
-  cache: new InMemoryCacheAdapter(), // swap for your own Redis/Upstash adapter
-});
-
-const result = await llm.cachedCall({
-  cacheKey: `cv:${cvId}`,
-  ttl: 3600,
-  fn: () => llm.call({ systemPrompt, userContent }),
-  reserveUsage: () => quota.reserve(userId),
-  refundUsage: () => quota.refund(userId),
-});
-```
-
-If `refundUsage` itself throws, that failure is logged and swallowed — the original error from `fn` (or the cache write) is what propagates, so a broken refund path never masks the real failure.
-
-### `cachedLLMCall` — cached + retried in one call
+## Why vern-llm?
 
-`cachedCall` is a thin cache wrapper; it doesn't apply retry/timeout/circuit-breaker behavior on its own — that's up to whatever `fn` does. If `fn` is always going to be `() => llm.call(...)`, `cachedLLMCall` wires that up for you:
-
-```ts
-const result = await llm.cachedLLMCall({
-  cacheKey: `cv:${cvId}`,
-  ttl: 3600,
-  call: { systemPrompt, userContent }, // same shape as call()'s params
-  reserveUsage: () => quota.reserve(userId),
-  refundUsage: () => quota.refund(userId),
-});
-```
-
-### Custom cache adapter
-
-```ts
-import type { CacheAdapter } from 'vern-llm';
-
-class UpstashCacheAdapter implements CacheAdapter {
-  async get(key: string) {
-    /* ... */
-  }
-  async set(key: string, value: unknown, ttl: number) {
-    /* ... */
-  }
-}
-```
-
-## Token usage tracking
-
-```ts
-const llm = new VernLLM({
-  client: openai,
-  model: 'gpt-4o',
-  onUsage: (usage) => {
-    // { promptTokens, completionTokens, totalTokens, requestId, model }
-    billing.record(usage);
-  },
-});
-```
+- **Retries with backoff**: transient failures retry automatically; validation errors and non-retryable status codes fail fast instead
+- **Structured output**: pass a Zod schema, get a typed, validated result back
+- **Provider-native JSON Schema mode**: constrain generation itself, not just validate after the fact
+- **Caching**: wrap any call with `cachedCall`/`cachedLLMCall`, bring your own cache adapter
+- **Circuit breaker**: trips after repeated failures, recovers automatically once the provider's back
+- **One interface, every provider**: OpenAI, Groq, Mistral, DeepSeek, Cerebras, Together, Fireworks, Ollama, Anthropic, Gemini, Bedrock, or raw HTTP via `fromFetch`
+- **Zero bundled deps**: `zod` and provider SDKs are peer dependencies; this package only relies on their shapes structurally
 
-## Circuit breaker
-
-Stops hammering a provider that's down instead of retrying every call.
-
-```ts
-const llm = new VernLLM({
-  client: openai,
-  model: 'gpt-4o',
-  circuitBreaker: { threshold: 5, cooldownMs: 30_000 }, // or `true` for defaults
-});
-
-llm.getCircuitState(); // 'closed' | 'open' | 'half-open' | undefined
-```
-
-Once `threshold` consecutive failures occur, further calls fail immediately with `LLMError('circuit_open')` until `cooldownMs` elapses, at which point one trial call is allowed through (half-open). A successful trial closes the circuit; a failed one reopens it.
-
-## Pluggable logger
-
-```ts
-import type { Logger } from 'vern-llm';
-
-const pinoLogger: Logger = {
-  debug: (msg) => logger.debug(msg),
-  warn: (msg) => logger.warn(msg),
-  error: (msg, meta) => logger.error(meta, msg),
-};
-
-const llm = new VernLLM({ client: openai, model: 'gpt-4o', logger: pinoLogger });
-```
-
-Defaults to a console logger; `debug` logging is gated by the `debug` option (`NODE_ENV !== 'production'` by default), `warn`/`error` always fire.
-
-## Multi-provider adapters
-
-`VernLLM` expects a client shaped like OpenAI's `chat.completions.create`. Some providers already match that shape and need no adapter; others need one that translates their native API into it.
-
-### Already OpenAI-wire-compatible — zero-transform passthrough
-
-Groq, Mistral, DeepSeek, Cerebras, Together AI, Fireworks AI, and Ollama (via its `/v1/chat/completions` endpoint) all speak the same wire format as OpenAI. These are thin named wrappers around the same passthrough — pick whichever name matches your provider for readability:
-
-```ts
-import { VernLLM, fromGroq, fromMistral, fromTogether } from 'vern-llm';
-
-const llm = new VernLLM({
-  client: fromGroq(new Groq({ apiKey: process.env.GROQ_API_KEY })),
-  model: 'llama-3.3-70b-versatile',
-});
-```
-
-`fromOpenAICompatible()` is the underlying function if your provider isn't in the named list — same thing, generic name.
-
-### Anthropic
-
-```ts
-import Anthropic from '@anthropic-ai/sdk';
-import { VernLLM, fromAnthropic } from 'vern-llm';
-
-const llm = new VernLLM({
-  client: fromAnthropic(new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY })),
-  model: 'claude-sonnet-4-6',
-});
-```
-
-No native `response_format` or `reasoning_effort` equivalent — JSON mode/schema are emulated via a system-prompt instruction (schema text embedded, not provider-enforced); `reasoning_effort` is dropped.
-
-### Gemini
-
-```ts
-import { GoogleGenerativeAI } from '@google/generative-ai';
-import { VernLLM, fromGemini } from 'vern-llm';
-
-const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY);
-const model = genAI.getGenerativeModel({ model: 'gemini-2.5-flash' });
-
-const llm = new VernLLM({ client: fromGemini(model), model: 'gemini-2.5-flash' });
-```
-
-Unlike Anthropic, `jsonSchema` maps to Gemini's native `responseSchema` + `responseMimeType: 'application/json'`, so it's actually provider-enforced here. `reasoning_effort` is dropped (Gemini's thinking models use a token budget instead).
-
-### AWS Bedrock
-
-Bedrock's SDK v3 exposes `.send(command)`, not a direct method, so wrap it:
-
-```ts
-import { BedrockRuntimeClient, ConverseCommand } from '@aws-sdk/client-bedrock-runtime';
-import { VernLLM, fromBedrock } from 'vern-llm';
-
-const client = new BedrockRuntimeClient({ region: 'us-east-1' });
-
-const llm = new VernLLM({
-  client: fromBedrock({
-    converse: (params, options) =>
-      client.send(new ConverseCommand(params), { abortSignal: options.signal }),
-  }),
-  model: 'anthropic.claude-3-5-sonnet-20241022-v2:0',
-});
-```
-
-Uses Bedrock's Converse API, which is unified across model families (Anthropic, Titan, Llama, Mistral, etc. all speak the same request/response shape), so this one adapter works regardless of which `model` you point it at. JSON mode is emulated via system prompt, same as Anthropic; `reasoning_effort` is dropped.
-
-### Anything else — `fromFetch`
-
-For a provider with no SDK, or where pulling one in isn't worth it, `fromFetch` is a raw HTTP escape hatch — supply the URL, headers, and two small mapping functions, and retries/timeouts/circuit-breaker/JSON handling all still apply:
-
-```ts
-import { VernLLM, fromFetch } from 'vern-llm';
-
-const llm = new VernLLM({
-  client: fromFetch({
-    url: 'https://api.example.com/v1/generate',
-    headers: () => ({ Authorization: `Bearer ${process.env.EXAMPLE_API_KEY}` }),
-    mapRequest: (params) => ({
-      model: params.model,
-      prompt: params.messages.map((m) => m.content).join('\n\n'),
-      max_tokens: params.max_tokens,
-    }),
-    mapResponse: (json) => ({
-      content: json.output,
-      usage: { promptTokens: json.usage?.input, completionTokens: json.usage?.output },
-    }),
-  }),
-  model: 'example-model-v1',
-});
-```
-
-Non-2xx responses throw with `.status` set, so `nonRetryableStatus` still fails fast on 401/403 as usual.
-
-## Notes
-
-- Requires Node 20+ (`AbortSignal.any`).
-- Non-retryable status codes (default `400, 401, 403`) fail fast instead of burning a retry, and are reported as `LLMError('api', status)` rather than a generic unknown error.
-- An already-aborted `signal` rejects immediately with `LLMError('aborted')` before any request is dispatched.
-- `jsonMode: false` skips JSON parsing entirely and returns the raw string.
-- `zod` and provider SDKs (`openai`, `groq-sdk`, `@anthropic-ai/sdk`) are not bundled — bring your own; this package only relies on their shapes structurally.
+See the [docs](https://vern-llm.vercel.app) for adapter setup, caching, the circuit breaker, and structured output in depth.
 
 ## Development
 
@@ -309,11 +68,13 @@ Non-2xx responses throw with `.status` set, so `nonRetryableStatus` still fails
 pnpm install
 pnpm run build            # tsdown → dist (ESM + CJS + types)
 pnpm run typecheck        # tsc --noEmit on src, since tsdown doesn't fully type-check
-pnpm run typecheck:test   # tsc --noEmit on src + test (separate tsconfig, no rootDir conflict)
 pnpm run test             # vitest run
-pnpm run test:watch       # vitest, watch mode
 pnpm run test:coverage    # vitest run --coverage (v8 provider)
 pnpm run changeset        # record a change for the next release
 ```
 
-Tests live in `tests/`, mirroring `src/`: `VernLLM.call.test.ts` and `VernLLM.schema.test.ts` cover retry/backoff/timeout/abort/schema/model-override/usage behavior, `circuitBreaker.test.ts` covers the breaker as a unit and its integration with `call()`, `cachedCall.test.ts` covers caching and usage reservation/refund, `logger.test.ts` covers the injectable logger, and `test/adapters/*.test.ts` cover each provider adapter's request/response translation against a fake client, no real API calls are made anywhere in the suite.
+Tests live in `tests/`, mirroring `src/`, and cover retry/backoff/timeout/abort/schema/model-override/usage behavior, the circuit breaker (unit + integration), caching, the injectable logger, and every provider adapter's request/response translation against a fake client — no real API calls anywhere in the suite.
+
+## License
+
+[MIT](https://github.com/LakBud/vernLLM/blob/main/LICENSE.md) © LakBud

package/dist/index.cjs

@@ -135,9 +135,12 @@ function extractStatus(err) {
 * Runs an async function and cancels it if it takes longer than the given
 * timeout. Creates an internal abort controller that fires after the
 * timeout elapses, and combines it with any external signal the caller
-* passed in so either one can cancel the underlying call. The internal
-* timer is always cleared afterward, whether the function succeeds,
-* fails, or is aborted, so nothing is left running in the background
+* passed in so either one can cancel the underlying call. If the internal
+* timeout triggers and the underlying operation aborts, the error is
+* converted into an LLMError with type "timeout". External cancellations
+* continue to propagate as aborted errors. The internal timer is always
+* cleared afterward, whether the function succeeds, fails, or is aborted,
+* so nothing is left running in the background.
 */
 async function withTimeout(fn, timeoutMs, externalSignal) {
 	const controller = new AbortController();
@@ -147,6 +150,9 @@ async function withTimeout(fn, timeoutMs, externalSignal) {
 	const signal = externalSignal ? AbortSignal.any([externalSignal, controller.signal]) : controller.signal;
 	try {
 		return await fn(signal);
+	} catch (err) {
+		if (controller.signal.aborted && !externalSignal?.aborted && err instanceof DOMException && err.name === "AbortError") throw new LLMError("Request timed out", "timeout");
+		throw err;
 	} finally {
 		clearTimeout(timer);
 	}

package/dist/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","names":["message: string","type: LLMErrorType","status?: number","issues?: unknown","err: unknown","key: string","value: T","ttl: number","options: CircuitBreakerOptions","content: string","err: unknown","fn: (signal: AbortSignal) => Promise<T>","timeoutMs: number","externalSignal?: AbortSignal","baseDelayMs: number","attempt: number","delay: number","signal?: AbortSignal","debugEnabled: boolean","message: string","meta?: Record<string, unknown>","options: VernLLMOptions","params: CallParams<T>","fn: () => Promise<T>","requestId: string","signal?: AbortSignal","lastError: unknown","error: unknown","jsonSchema: CallParams<unknown>['jsonSchema']","useJson: boolean","response: Awaited<ReturnType<LLMClient['chat']['completions']['create']>>","model: string","content: string","schema?: CallParams<T>['schema']","parsed: unknown","attempt: number","params: CachedCallParams<T>","params: Omit<CachedCallParams<T>, 'fn'> & { call: CallParams<T> }","anthropicClient: AnthropicClient","jsonInstruction: string | undefined","geminiClient: GeminiClient","generationConfig: NonNullable<\n            Parameters<GeminiClient['generateContent']>[0]['generationConfig']\n          >","bedrockClient: BedrockConverseClient","jsonInstruction: string | undefined","text","config: FetchAdapterConfig","client: unknown"],"sources":["../src/types.ts","../src/circuitBreaker.ts","../src/internal/vernLLM.utilts.ts","../src/logger.ts","../src/vernLLM.ts","../src/adapters/anthropic.ts","../src/adapters/gemini.ts","../src/adapters/bedrock.ts","../src/adapters/fetch.ts","../src/adapters/openaiCompatible.ts"],"sourcesContent":["export type LLMErrorType =\n  | 'timeout'\n  | 'api'\n  | 'parse'\n  | 'validation'\n  | 'circuit_open'\n  | 'unknown'\n  | 'aborted';\n\nexport class LLMError extends Error {\n  constructor(\n    message: string,\n    public type: LLMErrorType,\n    public status?: number,\n    public issues?: unknown,\n  ) {\n    super(message);\n    this.name = 'LLMError';\n  }\n}\n\nexport function isLLMError(err: unknown): err is LLMError {\n  return err instanceof LLMError;\n}\n\nexport interface CacheAdapter<T = unknown> {\n  get(key: string): Promise<T | null>;\n  set(key: string, value: T, ttl: number): Promise<void>;\n}\n\n/**\n * Trivial default so the package works out of the box with no external deps\n * Not shared across processes, swap in Redis/Upstash/etc for production\n */\nexport class InMemoryCacheAdapter<T = unknown> implements CacheAdapter<T> {\n  private store = new Map<string, { value: T; expiresAt: number }>();\n\n  async get(key: string): Promise<T | null> {\n    const entry = this.store.get(key);\n    if (!entry) return null;\n    if (Date.now() > entry.expiresAt) {\n      this.store.delete(key);\n      return null;\n    }\n    return entry.value;\n  }\n\n  async set(key: string, value: T, ttl: number): Promise<void> {\n    this.store.set(key, { value, expiresAt: Date.now() + ttl * 1000 });\n  }\n}\n\n/**\n * Minimal shape compatible with the OpenAI SDKs chat.completions.create,\n * so consumers can pass an OpenAI client directly\n * `response_format.json_schema` and `reasoning_effort` are optional on the wire\n * providers that don't support them will just ignore fields they don't recognize,\n * but not every SDKs TS types accept them, hence this being a structural type\n * rather than importing the SDKs own params type\n */\nexport interface LLMClient {\n  chat: {\n    completions: {\n      create(\n        params: {\n          model: string;\n          temperature: number;\n          max_tokens: number;\n          response_format?:\n            | { type: 'json_object' }\n            | {\n                type: 'json_schema';\n                json_schema: {\n                  name: string;\n                  schema: Record<string, unknown>;\n                  strict?: boolean;\n                  description?: string;\n                };\n              };\n          /** OpenAI reasoning-model param (o-series, gpt-5), ignored by providers that don't support it */\n          reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n          messages: Array<{ role: 'system' | 'user'; content: string }>;\n        },\n        options: { signal: AbortSignal },\n      ): Promise<{\n        choices?: Array<{ message?: { content?: string | null } }>;\n        usage?: {\n          prompt_tokens?: number;\n          completion_tokens?: number;\n          total_tokens?: number;\n        };\n      }>;\n    };\n  };\n}\n\nexport type ReserveUsage = () => Promise<void>;\nexport type RefundUsage = () => Promise<void>;\n\nexport interface TokenUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  requestId: string;\n  model: string;\n}\n\nexport type OnUsage = (usage: TokenUsage) => void;\n\n/**\n * Minimal structural type for a Zod-like schema, so this package doesnt need\n * a hard dependency on a specific Zod major version. Any object exposing\n * `safeParse` (Zod v3/v4, and most Zod-compatible validators) should satisfy this\n */\nexport interface SchemaLike<T> {\n  safeParse(data: unknown): { success: true; data: T } | { success: false; error: unknown };\n}\n\nexport interface VernLLMOptions {\n  client: LLMClient;\n  model: string;\n  /** Max retries after the first attempt. Default 1 (2 attempts total) */\n  maxRetries?: number;\n  /** Per-attempt timeout in ms. Default 25000 */\n  timeoutMs?: number;\n  /** Base delay for exponential backoff in ms. Default 500 */\n  baseDelayMs?: number;\n  /** Default max_tokens for calls that don't override it. Default 1000 */\n  defaultMaxTokens?: number;\n  /** Enables debug logging of raw model output. Default: NODE_ENV !== 'production' */\n  debug?: boolean;\n  /** Cache adapter for cachedCall. Defaults to an in-memory adapter */\n  cache?: CacheAdapter;\n  /** HTTP status codes that should fail fast without retrying. Default [400, 401, 403] */\n  nonRetryableStatus?: number[];\n  /** Custom JSON parser. Must return undefined/null on failure. Default: JSON.parse wrapped in try/catch */\n  parseJson?: (content: string) => unknown;\n  /** Called after every successful call with token usage, if the provider reports it */\n  onUsage?: OnUsage;\n  /** Injectable logger. Defaults to a console-based logger gated by `debug` */\n  logger?: import('./logger.js').Logger;\n  /**\n   * Enables a circuit breaker that short-circuits calls after repeated\n   * consecutive failures, instead of continuing to hammer a down provider\n   * Pass `true` for defaults, or an options object to tune threshold/cooldown\n   */\n  circuitBreaker?: boolean | import('./circuitBreaker.js').CircuitBreakerOptions;\n}\n\n/**\n * A provider-native JSON Schema for structured outputs (OpenAI/Groq\n * `response_format: { type: 'json_schema' }`) This is the wire-format\n * schema the model is constrained to generate against, distinct from\n * `schema`, which is a client-side Zod validator run on the parsed result\n * You can use one, both, or neither; using both gets you provider-level\n * constraint plus client-side type inference/validation as a safety net\n */\nexport interface JsonSchemaSpec {\n  name: string;\n  schema: Record<string, unknown>;\n  /** Enforces the schema strictly (OpenAI-specific), default true when supported */\n  strict?: boolean;\n  description?: string;\n}\n\nexport interface CallParams<T = unknown> {\n  systemPrompt: string;\n  userContent: string;\n  temperature?: number;\n  jsonMode?: boolean;\n  maxTokens?: number;\n  requestId?: string;\n  signal?: AbortSignal;\n  /** Overrides the model set on the VernLLM instance for this call only */\n  model?: string;\n  /**\n   * OpenAI-style reasoning effort for reasoning models (o-series, gpt-5, etc).\n   * Passed through as-is, providers/models that don't support it ignore it\n   */\n  reasoningEffort?: 'minimal' | 'low' | 'medium' | 'high';\n  /**\n   * Provider-native JSON Schema structured-output mode. When set, this is sent\n   * as `response_format: { type: 'json_schema', json_schema: ... }` instead of\n   * the looser `json_object` mode, constraining the models output shape at\n   * generation time (not just validating it after the fact). Implies jsonMode: true.\n   */\n  jsonSchema?: JsonSchemaSpec;\n  /**\n   * Optional Zod (or Zod-compatible) schema. When provided, the parsed JSON\n   * is validated against it; on failure an LLMError('validation') is thrown\n   * with `.issues` set to the schema's error object. Implies jsonMode: true.\n   * Can be combined with `jsonSchema` for provider-level constraint + client-side typing.\n   */\n  schema?: SchemaLike<T>;\n}\n\nexport interface CachedCallParams<T> {\n  cacheKey: string;\n  ttl: number;\n  fn: () => Promise<T>;\n  reserveUsage?: ReserveUsage;\n  refundUsage?: RefundUsage;\n}\n","import { LLMError } from './types.js';\n\nexport interface CircuitBreakerOptions {\n  /** Consecutive failures before the circuit opens, default 5 */\n  threshold?: number;\n  /** How long the circuit stays open before allowing a trial request, in ms. Default 30000 */\n  cooldownMs?: number;\n}\n\ntype CircuitState = 'closed' | 'open' | 'half-open';\n\n/**\n * Per retry VernLLM-instance circuit breaker. Tracks consecutive failures across\n * calls. Once the threshold is hit, short-circuits new calls with an\n * LLMError('circuit_open') instead of hitting the provider, until the\n * cooldown elapses and a single trial call is allowed through\n */\nexport class CircuitBreaker {\n  private state: CircuitState = 'closed';\n  private consecutiveFailures = 0;\n  private openedAt = 0;\n  private threshold: number;\n  private cooldownMs: number;\n\n  constructor(options: CircuitBreakerOptions = {}) {\n    this.threshold = options.threshold ?? 5;\n    this.cooldownMs = options.cooldownMs ?? 30_000;\n  }\n\n  /** Throws if the circuit is open and the cooldown hasnt elapsed */\n  assertClosed(): void {\n    if (this.state !== 'open') return;\n\n    const elapsed = Date.now() - this.openedAt;\n    if (elapsed >= this.cooldownMs) {\n      this.state = 'half-open';\n      return;\n    }\n\n    throw new LLMError(\n      `Circuit open — provider has failed ${this.consecutiveFailures} times in a row. Retry in ${Math.ceil((this.cooldownMs - elapsed) / 1000)}s.`,\n      'circuit_open',\n    );\n  }\n\n  recordSuccess(): void {\n    this.consecutiveFailures = 0;\n    this.state = 'closed';\n  }\n\n  recordFailure(): void {\n    this.consecutiveFailures += 1;\n\n    if (this.state === 'half-open') {\n      // Trial call failed: reopen and reset the cooldown window.\n      this.state = 'open';\n      this.openedAt = Date.now();\n      return;\n    }\n\n    if (this.consecutiveFailures >= this.threshold) {\n      this.state = 'open';\n      this.openedAt = Date.now();\n    }\n  }\n\n  getState(): CircuitState {\n    return this.state;\n  }\n}\n","import { LLMError } from '../types.js';\n\nexport function defaultParseJson(content: string): unknown {\n  try {\n    return JSON.parse(content);\n  } catch {\n    return undefined;\n  }\n}\n\n/**\n * Looks inside an unknown error value and pulls out an http status code\n * if one is present. Checks the status field first then the status code\n * field since different client libraries use different names for this.\n * Returns undefined when the error is not an object or carries no status\n */\nexport function extractStatus(err: unknown): number | undefined {\n  if (!err || typeof err !== 'object') return undefined;\n\n  const error = err as {\n    status?: unknown;\n    statusCode?: unknown;\n  };\n\n  if (typeof error.status === 'number') return error.status;\n  if (typeof error.statusCode === 'number') return error.statusCode;\n\n  return undefined;\n}\n\n/**\n * Runs an async function and cancels it if it takes longer than the given\n * timeout. Creates an internal abort controller that fires after the\n * timeout elapses, and combines it with any external signal the caller\n * passed in so either one can cancel the underlying call. The internal\n * timer is always cleared afterward, whether the function succeeds,\n * fails, or is aborted, so nothing is left running in the background\n */\nexport async function withTimeout<T>(\n  fn: (signal: AbortSignal) => Promise<T>,\n  timeoutMs: number,\n  externalSignal?: AbortSignal,\n): Promise<T> {\n  const controller = new AbortController();\n\n  const timer = setTimeout(() => {\n    controller.abort();\n  }, timeoutMs);\n\n  const signal = externalSignal\n    ? AbortSignal.any([externalSignal, controller.signal])\n    : controller.signal;\n\n  try {\n    return await fn(signal);\n  } finally {\n    clearTimeout(timer);\n  }\n}\n\n/**\n * Exponential backoff with jitter, capped at maxDelayMs.\n * Jitter avoids thundering-herd retries when many callers back off in lockstep,\n * the cap prevents unbounded delays when maxRetries is high\n */\nexport function getBackoffDelay(baseDelayMs: number, attempt: number, maxDelayMs = 10_000): number {\n  const exp = Math.min(baseDelayMs * 2 ** attempt, maxDelayMs);\n  return exp / 2 + Math.random() * (exp / 2);\n}\n\n/**\n * Pauses execution for the given delay before a retry attempt. If an\n * abort signal is provided and it fires while waiting, the pending\n * timer is cancelled immediately and the wait rejects right away with\n * an aborted error instead of continuing to sit idle until the delay\n * would have finished on its own\n */\nexport async function waitForRetry(delay: number, signal?: AbortSignal): Promise<void> {\n  await new Promise<void>((resolve, reject) => {\n    const onAbort = () => {\n      clearTimeout(timer);\n      reject(new LLMError('Operation aborted', 'aborted'));\n    };\n\n    const timer = setTimeout(() => {\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    }, delay);\n\n    signal?.addEventListener('abort', onAbort, { once: true });\n  });\n}\n","export interface Logger {\n  debug(message: string): void;\n  warn(message: string): void;\n  error(message: string, meta?: Record<string, unknown>): void;\n}\n\n/**\n * Default logger. `debug` is gated by the `debug` option on VernLLM\n * warn/error always fire since they indicate real problems (retries, cache failures)\n */\nexport class ConsoleLogger implements Logger {\n  constructor(private debugEnabled: boolean) {}\n\n  debug(message: string): void {\n    if (this.debugEnabled) console.debug(message);\n  }\n\n  warn(message: string): void {\n    console.warn(message);\n  }\n\n  error(message: string, meta?: Record<string, unknown>): void {\n    console.error(message, meta ?? '');\n  }\n}\n","import { randomUUID } from 'crypto';\n\nimport { CircuitBreaker } from './circuitBreaker.js';\nimport {\n  defaultParseJson,\n  extractStatus,\n  withTimeout,\n  getBackoffDelay,\n  waitForRetry,\n} from './internal/vernLLM.utilts.js';\nimport { ConsoleLogger, type Logger } from './logger.js';\nimport {\n  LLMError,\n  InMemoryCacheAdapter,\n  type VernLLMOptions,\n  type CallParams,\n  type CachedCallParams,\n  type CacheAdapter,\n  type LLMClient,\n} from './types.js';\n\n/**\n * A resilient wrapper around an LLM chat completions client, this is VernLLM!\n *\n * Adds retry with exponential backoff and jitter, per-attempt timeouts,\n * an optional circuit breaker, JSON parsing with optional schema\n * validation, usage tracking, and an optional response cache, all\n * configurable, all opt-in beyond sensible defaults\n */\nexport class VernLLM {\n  private readonly client: LLMClient;\n  private readonly model: string;\n\n  private readonly maxRetries: number;\n  private readonly timeoutMs: number;\n  private readonly baseDelayMs: number;\n  private readonly defaultMaxTokens: number;\n\n  private readonly cache: CacheAdapter<unknown>;\n  private readonly nonRetryableStatus: number[];\n\n  private readonly parseJson: (content: string) => unknown;\n  private readonly onUsage?: VernLLMOptions['onUsage'];\n\n  private readonly logger: Logger;\n  private readonly breaker?: CircuitBreaker;\n\n  /**\n   * @param options: Client, model, and all tunables (retries, timeout,\n   * backoff, cache, circuit breaker, logger, etc). See VernLLMOptions in `types.ts`\n   * for individual defaults\n   */\n  constructor(options: VernLLMOptions) {\n    this.client = options.client;\n    this.model = options.model;\n\n    const retryConfig = this.resolveRetryConfig(options);\n    this.maxRetries = retryConfig.maxRetries;\n    this.timeoutMs = retryConfig.timeoutMs;\n    this.baseDelayMs = retryConfig.baseDelayMs;\n    this.defaultMaxTokens = retryConfig.defaultMaxTokens;\n\n    this.cache = options.cache ?? new InMemoryCacheAdapter();\n    this.nonRetryableStatus = options.nonRetryableStatus ?? [400, 401, 403];\n\n    this.parseJson = options.parseJson ?? defaultParseJson;\n    this.onUsage = options.onUsage;\n\n    this.logger = this.resolveLogger(options);\n    this.breaker = this.resolveCircuitBreaker(options);\n  }\n\n  /**\n   * Resolves retry/timeout/token defaults from the given options,\n   * falling back to the librarys built-in defaults for anything unset\n   */\n  private resolveRetryConfig(options: VernLLMOptions) {\n    return {\n      maxRetries: options.maxRetries ?? 1,\n      timeoutMs: options.timeoutMs ?? 25_000,\n      baseDelayMs: options.baseDelayMs ?? 500,\n      defaultMaxTokens: options.defaultMaxTokens ?? 1000,\n    };\n  }\n\n  /**\n   * Returns the caller supplied logger, or a console-based logger whose\n   * debug output is gated by the `debug` option (defaulting to on\n   * outside production)\n   */\n  private resolveLogger(options: VernLLMOptions): Logger {\n    return (\n      options.logger ?? new ConsoleLogger(options.debug ?? process.env.NODE_ENV !== 'production')\n    );\n  }\n\n  /**\n   * Builds a circuit breaker if `circuitBreaker` is truthy on the\n   * options. Passing `true` uses default thresholds, passing an options\n   * object tunes them. Returns undefined when the breaker is disabled\n   */\n  private resolveCircuitBreaker(options: VernLLMOptions): CircuitBreaker | undefined {\n    if (!options.circuitBreaker) return undefined;\n\n    return new CircuitBreaker(options.circuitBreaker === true ? undefined : options.circuitBreaker);\n  }\n\n  /**\n   * Makes a single logical LLM call, transparently retrying on failure\n   * according to the configured retry policy\n   *\n   * Fails fast if the circuit breaker is open or the signal is already\n   * aborted, before any request is dispatched. On exhausting all\n   * retries, records a circuit breaker failure and rejects with a\n   * normalized LLMError\n   *\n   * @param params : System/user content plus per call overrides\n   * (model, temperature, jsonMode, schema, signal, etc)\n   * @returns The parsed and optionally schema-validated response, or\n   * the raw string content when jsonMode is disabled\n   */\n  async call<T = unknown>(params: CallParams<T>): Promise<T> {\n    this.breaker?.assertClosed();\n\n    if (params.signal?.aborted) {\n      throw new LLMError('LLM request aborted', 'aborted');\n    }\n\n    const requestId = params.requestId ?? randomUUID();\n\n    try {\n      const result = await this.retryWithBackoff(\n        () => this.executeCall(params, requestId),\n        requestId,\n        params.signal,\n      );\n\n      return result;\n    } catch (error) {\n      this.breaker?.recordFailure();\n      throw this.normalizeError(error, params.signal);\n    }\n  }\n\n  /**\n   * Runs `fn`, retrying with backoff according to `shouldRetry` policy\n   * Purely mechanical: knows nothing about LLM specifics beyond the retry\n   * predicate, so its testable independent of request/response shaping\n   */\n  private async retryWithBackoff<T>(\n    fn: () => Promise<T>,\n    requestId: string,\n    signal?: AbortSignal,\n  ): Promise<T> {\n    let lastError: unknown;\n\n    for (let attempt = 0; attempt <= this.maxRetries; attempt++) {\n      try {\n        if (attempt > 0) {\n          await this.recoverDelay(requestId, attempt, signal);\n        }\n\n        return await fn();\n      } catch (error) {\n        lastError = error;\n\n        if (!this.shouldRetry(error, signal)) {\n          break;\n        }\n      }\n    }\n\n    throw lastError;\n  }\n\n  /**\n   * Converts any thrown value into a well-typed LLMError for the public\n   * API surface. Preserves an existing LLMError as is, reports aborted\n   * signals as such, classifies errors carrying an http status as\n   * type api, and otherwise falls back to a generic unknown error.\n   */\n  private normalizeError(error: unknown, signal?: AbortSignal): LLMError {\n    if (signal?.aborted) {\n      return new LLMError('LLM request aborted', 'aborted');\n    }\n\n    if (error instanceof LLMError) {\n      return error;\n    }\n\n    const status = extractStatus(error);\n\n    if (status !== undefined) {\n      return new LLMError('LLM request failed', 'api', status);\n    }\n\n    return new LLMError('LLM request failed', 'unknown');\n  }\n\n  /**\n   * Performs a single attempt: builds the request, dispatches it with a\n   * timeout, and shapes the response. Throws on an empty response so\n   * the retry loop treats it like any other transient failure. Records\n   * usage and a circuit breaker success before returning\n   */\n  private async executeCall<T>(params: CallParams<T>, requestId: string): Promise<T> {\n    const { useJson, model, request } = this.buildRequestPayload(params);\n\n    const response = await withTimeout(\n      (attemptSignal) => this.client.chat.completions.create(request, { signal: attemptSignal }),\n      this.timeoutMs,\n      params.signal,\n    );\n\n    const content = response.choices?.[0]?.message?.content?.trim();\n\n    if (!content) {\n      throw new LLMError('Empty LLM response', 'api');\n    }\n\n    this.logger.debug(`[vern:${requestId}] output:\\n${content.slice(0, 800)}`);\n\n    this.recordUsage(response, requestId, model);\n\n    this.breaker?.recordSuccess();\n\n    if (!useJson) {\n      return content as T;\n    }\n\n    return this.parseAndValidate(content, params.schema);\n  }\n\n  /**\n   * Applies per call defaults and shapes the params into the request\n   * object expected by the underlying client, including the resolved\n   * response format. Also returns whether JSON parsing should be\n   * applied to the response and which model was ultimately used\n   */\n  private buildRequestPayload<T>(params: CallParams<T>) {\n    const {\n      systemPrompt,\n      userContent,\n      temperature = 0.2,\n      jsonMode = true,\n      maxTokens = this.defaultMaxTokens,\n      model = this.model,\n      reasoningEffort,\n      jsonSchema,\n    } = params;\n\n    const useJson = jsonMode || Boolean(jsonSchema);\n    const responseFormat = this.buildResponseFormat(jsonSchema, useJson);\n\n    const request = {\n      model,\n      temperature,\n      max_tokens: maxTokens,\n      ...(responseFormat ? { response_format: responseFormat } : {}),\n      ...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}),\n      messages: [\n        { role: 'system' as const, content: systemPrompt },\n        { role: 'user' as const, content: userContent },\n      ],\n    };\n\n    return { useJson, model, request };\n  }\n\n  /**\n   * Chooses the response format to send to the provider. A provider\n   * native json schema takes priority when supplied, constraining\n   * generation directly, otherwise falls back to the looser json\n   * object mode when JSON output is requested, or no format at all\n   * for plain text responses\n   */\n  private buildResponseFormat(jsonSchema: CallParams<unknown>['jsonSchema'], useJson: boolean) {\n    if (jsonSchema) {\n      return {\n        type: 'json_schema' as const,\n        json_schema: {\n          name: jsonSchema.name,\n          schema: jsonSchema.schema,\n          strict: jsonSchema.strict ?? true,\n          description: jsonSchema.description,\n        },\n      };\n    }\n\n    return useJson ? { type: 'json_object' as const } : undefined;\n  }\n\n  /**\n   * Reports token usage to the caller supplied onUsage callback, when\n   * both a callback was configured and the provider actually returned\n   * usage data on this response. A no op otherwise.\n   */\n  private recordUsage(\n    response: Awaited<ReturnType<LLMClient['chat']['completions']['create']>>,\n    requestId: string,\n    model: string,\n  ): void {\n    if (!response.usage || !this.onUsage) return;\n\n    this.onUsage({\n      promptTokens: response.usage.prompt_tokens ?? 0,\n      completionTokens: response.usage.completion_tokens ?? 0,\n      totalTokens: response.usage.total_tokens ?? 0,\n      requestId,\n      model,\n    });\n  }\n\n  /**\n   * Parses the raw response content as JSON and, when a schema is\n   * supplied, validates the parsed value against it. Throws a parse\n   * type LLMError on malformed JSON and a validation type LLMError,\n   * carrying the schemas issues, on a failed validation\n   */\n  private parseAndValidate<T>(content: string, schema?: CallParams<T>['schema']): T {\n    let parsed: unknown;\n\n    try {\n      parsed = this.parseJson(content);\n    } catch {\n      throw new LLMError('Invalid JSON response', 'parse');\n    }\n\n    if (parsed === null || parsed === undefined) {\n      throw new LLMError('Invalid JSON response', 'parse');\n    }\n\n    if (!schema) {\n      return parsed as T;\n    }\n\n    const result = schema.safeParse(parsed);\n\n    if (!result.success) {\n      throw new LLMError('Schema validation failed', 'validation', undefined, result.error);\n    }\n\n    return result.data;\n  }\n\n  /**\n   * Waits out the backoff delay for a given retry attempt, logging the\n   * attempt for observability before the wait begins. Rejects early if\n   * the signal aborts during the wait\n   */\n  private async recoverDelay(requestId: string, attempt: number, signal?: AbortSignal) {\n    const delay = getBackoffDelay(this.baseDelayMs, attempt);\n\n    this.logger.warn(\n      `[vern:${requestId}] recovery attempt ${attempt}/${this.maxRetries}, waiting ${delay}ms`,\n    );\n\n    await waitForRetry(delay, signal);\n  }\n\n  /**\n   * Decides whether a failed attempt is worth retrying. Never retries\n   * once the signal has aborted, never retries a parse or validation\n   * failure since those stem from the response content rather than a\n   * transient fault, and never retries a status code the caller has\n   * marked as non retryable. Retries everything else\n   */\n  private shouldRetry(error: unknown, signal?: AbortSignal): boolean {\n    if (signal?.aborted) {\n      return false;\n    }\n\n    if (error instanceof LLMError && (error.type === 'parse' || error.type === 'validation')) {\n      return false;\n    }\n\n    const status = extractStatus(error);\n\n    if (status !== undefined && this.nonRetryableStatus.includes(status)) {\n      return false;\n    }\n\n    return true;\n  }\n\n  /**\n   * Thin cache wrapper around caller supplied logic. `params.fn` is expected\n   * to be a call that itself invokes `this.call(...)` (see `cachedLLMCall`\n   * below for a convenience wrapper that wires this up automatically),\n   * `cachedCall` does not itself apply retry/timeout policy.\n   */\n  async cachedCall<T>(params: CachedCallParams<T>): Promise<T> {\n    const cached = await this.cache.get(params.cacheKey);\n\n    if (cached !== null) {\n      return cached as T;\n    }\n\n    try {\n      await params.reserveUsage?.();\n\n      const result = await params.fn();\n\n      try {\n        await this.cache.set(params.cacheKey, result, params.ttl);\n      } catch (error) {\n        this.logger.error('[VernLLM] cache write failed', {\n          message: error instanceof Error ? error.message : 'unknown',\n        });\n      }\n\n      return result;\n    } catch (error) {\n      try {\n        await params.refundUsage?.();\n      } catch (refundError) {\n        this.logger.error('[VernLLM] refundUsage failed', {\n          message: refundError instanceof Error ? refundError.message : 'unknown',\n        });\n      }\n\n      throw error;\n    }\n  }\n\n  /**\n   * Convenience wrapper composing `call` + `cachedCall`, so cached LLM calls\n   * automatically get retry/timeout/circuit-breaker behavior without callers\n   * having to remember to wire `fn: () => this.call(...)` themselves\n   */\n  async cachedLLMCall<T>(\n    params: Omit<CachedCallParams<T>, 'fn'> & { call: CallParams<T> },\n  ): Promise<T> {\n    const { call: callParams, ...cacheParams } = params;\n\n    return this.cachedCall({\n      ...cacheParams,\n      fn: () => this.call(callParams),\n    });\n  }\n\n  /**\n   * Returns the current circuit breaker state, or undefined when no\n   * circuit breaker was configured on this instance\n   */\n  getCircuitState() {\n    return this.breaker?.getState();\n  }\n}\n","import type { LLMClient } from '../types.js';\n\n/** Minimal structural type for the Anthropic SDKs `messages.create` */\nexport interface AnthropicClient {\n  messages: {\n    create(\n      params: {\n        model: string;\n        max_tokens: number;\n        temperature?: number;\n        system?: string;\n        messages: Array<{ role: 'user' | 'assistant'; content: string }>;\n      },\n      options: { signal: AbortSignal },\n    ): Promise<{\n      content: Array<{ type: string; text?: string }>;\n      usage?: { input_tokens?: number; output_tokens?: number };\n    }>;\n  };\n}\n\n/**\n * Wraps an Anthropic SDK client so it satisfies the same `LLMClient`\n * interface VernLLM uses for OpenAI/Groq. Anthropics Messages API has no\n * `response_format: json_object` equivalent, so when the caller requests\n * JSON mode, this adapter appends an instruction to the system prompt\n * asking the model to respond with JSON only\n */\nexport function fromAnthropic(anthropicClient: AnthropicClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          // Anthropics Messages API has no `response_format: json_object` or\n          // `json_schema` equivalent, so both are emulated via system prompt\n          // instructions. For json_schema, the schema itself is embedded so the\n          // model has something concrete to conform to (not provider-enforced,\n          // unlike OpenAIs native structured outputs)\n          let jsonInstruction: string | undefined;\n          if (params.response_format?.type === 'json_schema') {\n            const { name, schema } = params.response_format.json_schema;\n            jsonInstruction = `Respond with valid JSON only, no prose or markdown fences. The JSON must conform to this schema (name: \"${name}\"):\\n${JSON.stringify(schema)}`;\n          } else if (params.response_format?.type === 'json_object') {\n            jsonInstruction = 'Respond with valid JSON only, no prose or markdown fences.';\n          }\n\n          // `reasoning_effort` (OpenAI o-series/gpt-5 style) has no direct Anthropic\n          // equivalent, Claudes extended thinking uses a token budget, not a tier\n          // string, so its intentionally dropped here rather than guessed at.\n\n          const system = [systemMessage?.content, jsonInstruction].filter(Boolean).join('\\n\\n');\n\n          const response = await anthropicClient.messages.create(\n            {\n              model: params.model,\n              max_tokens: params.max_tokens,\n              temperature: params.temperature,\n              system: system || undefined,\n              messages: userMessages.map((m) => ({ role: 'user' as const, content: m.content })),\n            },\n            options,\n          );\n\n          const text = response.content.find((block) => block.type === 'text')?.text ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usage?.input_tokens,\n              completion_tokens: response.usage?.output_tokens,\n              total_tokens:\n                (response.usage?.input_tokens ?? 0) + (response.usage?.output_tokens ?? 0),\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Minimal structural type for Geminis `generateContent`, matching both the\n * legacy `@google/generative-ai` SDKs `model.generateContent(...)` and the\n * newer `@google/genai` SDKs `ai.models.generateContent({ model, ... })`\n * closely enough to adapt either — pass whichever `.generateContent` you have.\n */\nexport interface GeminiClient {\n  generateContent(\n    params: {\n      model?: string;\n      contents: Array<{ role: 'user'; parts: Array<{ text: string }> }>;\n      systemInstruction?: { parts: Array<{ text: string }> };\n      generationConfig?: {\n        temperature?: number;\n        maxOutputTokens?: number;\n        responseMimeType?: string;\n        responseSchema?: Record<string, unknown>;\n      };\n    },\n    options: { signal: AbortSignal },\n  ): Promise<{\n    candidates?: Array<{ content?: { parts?: Array<{ text?: string }> } }>;\n    usageMetadata?: {\n      promptTokenCount?: number;\n      candidatesTokenCount?: number;\n      totalTokenCount?: number;\n    };\n  }>;\n}\n\n/**\n * Wraps a Gemini client so it satisfies the `LLMClient` interface VernLLM\n * uses for OpenAI/Groq. Geminis shape differs on nearly every axis: a\n * `contents` array instead of `messages`, a separate `systemInstruction`\n * field instead of a `system` role message, `generationConfig` instead of\n * top-level `temperature`/`max_tokens`, and native JSON Schema support via\n * `responseMimeType: 'application/json'` + `responseSchema` (so `jsonSchema`\n * is provider-enforced here, unlike the Anthropic adapters prompt-embedding\n * fallback). `reasoning_effort` has no equivalent. Geminis thinking models\n * use a token budget, not an effort tier, so its dropped, same as Anthropic.\n */\nexport function fromGemini(geminiClient: GeminiClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          const wantsJson = Boolean(params.response_format);\n          const generationConfig: NonNullable<\n            Parameters<GeminiClient['generateContent']>[0]['generationConfig']\n          > = {\n            temperature: params.temperature,\n            maxOutputTokens: params.max_tokens,\n          };\n\n          if (wantsJson) {\n            generationConfig.responseMimeType = 'application/json';\n          }\n          if (params.response_format?.type === 'json_schema') {\n            generationConfig.responseSchema = params.response_format.json_schema.schema;\n          }\n\n          const response = await geminiClient.generateContent(\n            {\n              model: params.model,\n              contents: userMessages.map((m) => ({\n                role: 'user' as const,\n                parts: [{ text: m.content }],\n              })),\n              systemInstruction: systemMessage\n                ? { parts: [{ text: systemMessage.content }] }\n                : undefined,\n              generationConfig,\n            },\n            options,\n          );\n\n          const text =\n            response.candidates?.[0]?.content?.parts?.map((p) => p.text ?? '').join('') ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usageMetadata?.promptTokenCount,\n              completion_tokens: response.usageMetadata?.candidatesTokenCount,\n              total_tokens: response.usageMetadata?.totalTokenCount,\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Minimal structural type matching AWS Bedrocks Converse API. This is\n * intentionally NOT `BedrockRuntimeClient` itself, the AWS SDK v3 client\n * exposes `.send(command)`, not a direct `.converse()` method, and pulling\n * in `@aws-sdk/client-bedrock-runtime` as a dependency just for its types\n * isn't worth it for a structural adapter. Wrap your client, e.g:\n *\n * ```ts\n * import { BedrockRuntimeClient, ConverseCommand } from '@aws-sdk/client-bedrock-runtime';\n * const client = new BedrockRuntimeClient({ region: 'us-east-1' });\n * const converseClient = {\n *   converse: (params, options) =>\n *     client.send(new ConverseCommand(params), { abortSignal: options.signal }),\n * };\n * ```\n */\nexport interface BedrockConverseClient {\n  converse(\n    params: {\n      modelId: string;\n      messages: Array<{ role: 'user'; content: Array<{ text: string }> }>;\n      system?: Array<{ text: string }>;\n      inferenceConfig?: { temperature?: number; maxTokens?: number };\n    },\n    options: { signal: AbortSignal },\n  ): Promise<{\n    output?: { message?: { content?: Array<{ text?: string }> } };\n    usage?: { inputTokens?: number; outputTokens?: number; totalTokens?: number };\n  }>;\n}\n\n/**\n * Wraps a Bedrock Converse-API client so it satisfies the `LLMClient`\n * interface VernLLM uses for OpenAI/Groq. The Converse API is unified\n * across Bedrocks model families (Anthropic, Titan, Llama, Mistral, etc.),\n * so unlike raw per-model Bedrock invocation, this one adapter works\n * regardless of which underlying model `modelId` points at, as long as\n * that model supports Converse (most current-generation ones do)\n *\n * Theres no uniform native JSON Schema enforcement across families here\n * (some support it via forced tool-use, which varies per model), so\n * `jsonSchema`/`jsonMode` are emulated via a system-prompt instruction, same\n * approach as the Anthropic adapter. `reasoning_effort` has no Converse\n * equivalent and is dropped\n */\nexport function fromBedrock(bedrockClient: BedrockConverseClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          let jsonInstruction: string | undefined;\n          if (params.response_format?.type === 'json_schema') {\n            const { name, schema } = params.response_format.json_schema;\n            jsonInstruction = `Respond with valid JSON only, no prose or markdown fences. The JSON must conform to this schema (name: \"${name}\"):\\n${JSON.stringify(schema)}`;\n          } else if (params.response_format?.type === 'json_object') {\n            jsonInstruction = 'Respond with valid JSON only, no prose or markdown fences.';\n          }\n\n          const systemParts = [systemMessage?.content, jsonInstruction].filter((s): s is string =>\n            Boolean(s),\n          );\n\n          const response = await bedrockClient.converse(\n            {\n              modelId: params.model,\n              messages: userMessages.map((m) => ({\n                role: 'user' as const,\n                content: [{ text: m.content }],\n              })),\n              system: systemParts.length ? systemParts.map((text) => ({ text })) : undefined,\n              inferenceConfig: {\n                temperature: params.temperature,\n                maxTokens: params.max_tokens,\n              },\n            },\n            options,\n          );\n\n          const text = response.output?.message?.content?.map((c) => c.text ?? '').join('') ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usage?.inputTokens,\n              completion_tokens: response.usage?.outputTokens,\n              total_tokens: response.usage?.totalTokens,\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/** The chat-completion-shaped request VernLLM builds internally */\ntype ChatRequest = Parameters<LLMClient['chat']['completions']['create']>[0];\n\nexport interface FetchAdapterConfig {\n  /** Endpoint URL, or a function of the request in case it depends on model/params */\n  url: string | ((params: ChatRequest) => string);\n  /** Static headers, or a function (sync or async) for things like refreshed auth tokens */\n  headers?:\n    | Record<string, string>\n    | (() => Record<string, string> | Promise<Record<string, string>>);\n  /** HTTP method. Default 'POST' */\n  method?: string;\n  /** Maps VernLLMs internal chat-completion request into the providers raw request body */\n  mapRequest: (params: ChatRequest) => unknown;\n  /**\n   * Maps the providers raw JSON response into `{ content, usage? }`\n   * `content` is the assistants text (JSON string when JSON mode was requested)\n   */\n  mapResponse: (json: unknown) => {\n    content: string;\n    usage?: { promptTokens?: number; completionTokens?: number; totalTokens?: number };\n  };\n}\n\n/**\n * A fetch-based escape hatch for providers with no SDK, or where pulling one\n * in isnt worth it. You supply the URL, headers, and two small mapping\n * functions; this handles the HTTP call and slots the result into the same\n * `LLMClient` shape every other adapter produces, so retries, timeouts,\n * the circuit breaker, and JSON/schema handling all still work unmodified\n *\n * Non-2xx responses throw an error with `.status` set to the HTTP status\n * code, so VernLLMs `nonRetryableStatus` handling (e.g. failing fast on\n * 401/403) applies here too\n */\nexport function fromFetch(config: FetchAdapterConfig): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const url = typeof config.url === 'function' ? config.url(params) : config.url;\n          const headers =\n            typeof config.headers === 'function' ? await config.headers() : config.headers;\n\n          const res = await fetch(url, {\n            method: config.method ?? 'POST',\n            headers: { 'Content-Type': 'application/json', ...headers },\n            body: JSON.stringify(config.mapRequest(params)),\n            signal: options.signal,\n          });\n\n          if (!res.ok) {\n            const body = await res.text().catch(() => '');\n            const err = new Error(\n              `Fetch adapter request failed (${res.status}): ${body.slice(0, 500)}`,\n            ) as Error & { status?: number };\n            err.status = res.status;\n            throw err;\n          }\n\n          const json = await res.json();\n          const { content, usage } = config.mapResponse(json);\n\n          return {\n            choices: [{ message: { content } }],\n            usage: usage\n              ? {\n                  prompt_tokens: usage.promptTokens,\n                  completion_tokens: usage.completionTokens,\n                  total_tokens: usage.totalTokens,\n                }\n              : undefined,\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Passthrough adapter for any SDK/client whose `chat.completions.create`\n * already matches the OpenAI wire format 1:1 : this covers most hosted\n * inference providers, since \"OpenAI-compatible\" is a de facto standard for\n * chat completion APIs. No transformation happens here, this exists purely\n * so call sites read clearly (`fromMistral(client)` vs handing a Mistral\n * client to something typed for OpenAI) and so a real transformation could\n * be added later, per-provider, without a breaking change.\n *\n * Not every SDKs own TypeScript types line up exactly with `LLMClient`\n * (extra fields, stricter unions, etc.), so this takes `unknown` and casts:\n * the actual compatibility contract is the JSON each provider sends and\n * receives over the wire, not the SDKs TS types.\n */\nexport function fromOpenAICompatible(client: unknown): LLMClient {\n  return client as LLMClient;\n}\n\n/** Groqs SDK matches the OpenAI wire format */\nexport const fromGroq = fromOpenAICompatible;\n\n/** Mistrals `chat.completions`-shaped client (or their OpenAI-compat endpoint) */\nexport const fromMistral = fromOpenAICompatible;\n\n/** DeepSeeks API is OpenAI-compatible */\nexport const fromDeepSeek = fromOpenAICompatible;\n\n/** Cerebras inference API is OpenAI-compatible */\nexport const fromCerebras = fromOpenAICompatible;\n\n/** Together AIs API is OpenAI-compatible */\nexport const fromTogether = fromOpenAICompatible;\n\n/** Fireworks AIs API is OpenAI-compatible */\nexport const fromFireworks = fromOpenAICompatible;\n\n/**\n * Ollama exposes an OpenAI-compatible endpoint at `/v1/chat/completions`\n * (as opposed to its native `/api/chat` format, which differs). Point an\n * OpenAI SDK instances `baseURL` at your Ollama server and pass it here:\n * this does not talk to Ollamas native API directly.\n */\nexport const fromOllama = fromOpenAICompatible;\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AASA,IAAa,WAAb,cAA8B,MAAM;CAClC,YACEA,SACOC,MACAC,QACAC,QACP;AACA,QAAM,QAAQ;EA2LjB,KA/LU;EA+LT,KA9LS;EA8LR,KA7LQ;AAGP,OAAK,OAAO;CACb;AACF;AAED,SAAgB,WAAWC,KAA+B;AACxD,QAAO,eAAe;AACvB;;;;;AAWD,IAAa,uBAAb,MAA0E;CACxE,AAAQ,QAAQ,IAAI;CAEpB,MAAM,IAAIC,KAAgC;EACxC,MAAM,QAAQ,KAAK,MAAM,IAAI,IAAI;AACjC,OAAK,MAAO,QAAO;AACnB,MAAI,KAAK,KAAK,GAAG,MAAM,WAAW;AAChC,QAAK,MAAM,OAAO,IAAI;AACtB,UAAO;EACR;AACD,SAAO,MAAM;CACd;CAED,MAAM,IAAIA,KAAaC,OAAUC,KAA4B;AAC3D,OAAK,MAAM,IAAI,KAAK;GAAE;GAAO,WAAW,KAAK,KAAK,GAAG,MAAM;EAAM,EAAC;CACnE;AACF;;;;;;;;;;ACjCD,IAAa,iBAAb,MAA4B;CAC1B,AAAQ,QAAsB;CAC9B,AAAQ,sBAAsB;CAC9B,AAAQ,WAAW;CACnB,AAAQ;CACR,AAAQ;CAER,YAAYC,UAAiC,CAAE,GAAE;AAC/C,OAAK,YAAY,QAAQ,aAAa;AACtC,OAAK,aAAa,QAAQ,cAAc;CACzC;;CAGD,eAAqB;AACnB,MAAI,KAAK,UAAU,OAAQ;EAE3B,MAAM,UAAU,KAAK,KAAK,GAAG,KAAK;AAClC,MAAI,WAAW,KAAK,YAAY;AAC9B,QAAK,QAAQ;AACb;EACD;AAED,QAAM,IAAI,UACP,qCAAqC,KAAK,oBAAoB,4BAA4B,KAAK,MAAM,KAAK,aAAa,WAAW,IAAK,CAAC,KACzI;CAEH;CAED,gBAAsB;AACpB,OAAK,sBAAsB;AAC3B,OAAK,QAAQ;CACd;CAED,gBAAsB;AACpB,OAAK,uBAAuB;AAE5B,MAAI,KAAK,UAAU,aAAa;AAE9B,QAAK,QAAQ;AACb,QAAK,WAAW,KAAK,KAAK;AAC1B;EACD;AAED,MAAI,KAAK,uBAAuB,KAAK,WAAW;AAC9C,QAAK,QAAQ;AACb,QAAK,WAAW,KAAK,KAAK;EAC3B;CACF;CAED,WAAyB;AACvB,SAAO,KAAK;CACb;AACF;;;;ACnED,SAAgB,iBAAiBC,SAA0B;AACzD,KAAI;AACF,SAAO,KAAK,MAAM,QAAQ;CAC3B,QAAO;AACN;CACD;AACF;;;;;;;AAQD,SAAgB,cAAcC,KAAkC;AAC9D,MAAK,cAAc,QAAQ,SAAU;CAErC,MAAM,QAAQ;AAKd,YAAW,MAAM,WAAW,SAAU,QAAO,MAAM;AACnD,YAAW,MAAM,eAAe,SAAU,QAAO,MAAM;AAEvD;AACD;;;;;;;;;AAUD,eAAsB,YACpBC,IACAC,WACAC,gBACY;CACZ,MAAM,aAAa,IAAI;CAEvB,MAAM,QAAQ,WAAW,MAAM;AAC7B,aAAW,OAAO;CACnB,GAAE,UAAU;CAEb,MAAM,SAAS,iBACX,YAAY,IAAI,CAAC,gBAAgB,WAAW,MAAO,EAAC,GACpD,WAAW;AAEf,KAAI;AACF,SAAO,MAAM,GAAG,OAAO;CACxB,UAAS;AACR,eAAa,MAAM;CACpB;AACF;;;;;;AAOD,SAAgB,gBAAgBC,aAAqBC,SAAiB,aAAa,KAAgB;CACjG,MAAM,MAAM,KAAK,IAAI,cAAc,KAAK,SAAS,WAAW;AAC5D,QAAO,MAAM,IAAI,KAAK,QAAQ,IAAI,MAAM;AACzC;;;;;;;;AASD,eAAsB,aAAaC,OAAeC,QAAqC;AACrF,OAAM,IAAI,QAAc,CAAC,SAAS,WAAW;EAC3C,MAAM,UAAU,MAAM;AACpB,gBAAa,MAAM;AACnB,UAAO,IAAI,SAAS,qBAAqB,WAAW;EACrD;EAED,MAAM,QAAQ,WAAW,MAAM;AAC7B,WAAQ,oBAAoB,SAAS,QAAQ;AAC7C,YAAS;EACV,GAAE,MAAM;AAET,UAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAM,EAAC;CAC3D;AACF;;;;;;;;ACjFD,IAAa,gBAAb,MAA6C;CAC3C,YAAoBC,cAAuB;EAc5C,KAdqB;CAAyB;CAE7C,MAAMC,SAAuB;AAC3B,MAAI,KAAK,aAAc,SAAQ,MAAM,QAAQ;CAC9C;CAED,KAAKA,SAAuB;AAC1B,UAAQ,KAAK,QAAQ;CACtB;CAED,MAAMA,SAAiBC,MAAsC;AAC3D,UAAQ,MAAM,SAAS,QAAQ,GAAG;CACnC;AACF;;;;;;;;;;;;ACKD,IAAa,UAAb,MAAqB;CACnB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;;;;;;CAOjB,YAAYC,SAAyB;AACnC,OAAK,SAAS,QAAQ;AACtB,OAAK,QAAQ,QAAQ;EAErB,MAAM,cAAc,KAAK,mBAAmB,QAAQ;AACpD,OAAK,aAAa,YAAY;AAC9B,OAAK,YAAY,YAAY;AAC7B,OAAK,cAAc,YAAY;AAC/B,OAAK,mBAAmB,YAAY;AAEpC,OAAK,QAAQ,QAAQ,SAAS,IAAI;AAClC,OAAK,qBAAqB,QAAQ,sBAAsB;GAAC;GAAK;GAAK;EAAI;AAEvE,OAAK,YAAY,QAAQ,aAAa;AACtC,OAAK,UAAU,QAAQ;AAEvB,OAAK,SAAS,KAAK,cAAc,QAAQ;AACzC,OAAK,UAAU,KAAK,sBAAsB,QAAQ;CACnD;;;;;CAMD,AAAQ,mBAAmBA,SAAyB;AAClD,SAAO;GACL,YAAY,QAAQ,cAAc;GAClC,WAAW,QAAQ,aAAa;GAChC,aAAa,QAAQ,eAAe;GACpC,kBAAkB,QAAQ,oBAAoB;EAC/C;CACF;;;;;;CAOD,AAAQ,cAAcA,SAAiC;AACrD,SACE,QAAQ,UAAU,IAAI,cAAc,QAAQ,SAAS,QAAQ,IAAI,aAAa;CAEjF;;;;;;CAOD,AAAQ,sBAAsBA,SAAqD;AACjF,OAAK,QAAQ,eAAgB;AAE7B,SAAO,IAAI,eAAe,QAAQ,mBAAmB,gBAAmB,QAAQ;CACjF;;;;;;;;;;;;;;;CAgBD,MAAM,KAAkBC,QAAmC;AACzD,OAAK,SAAS,cAAc;AAE5B,MAAI,OAAO,QAAQ,QACjB,OAAM,IAAI,SAAS,uBAAuB;EAG5C,MAAM,YAAY,OAAO,aAAa,wBAAY;AAElD,MAAI;GACF,MAAM,SAAS,MAAM,KAAK,iBACxB,MAAM,KAAK,YAAY,QAAQ,UAAU,EACzC,WACA,OAAO,OACR;AAED,UAAO;EACR,SAAQ,OAAO;AACd,QAAK,SAAS,eAAe;AAC7B,SAAM,KAAK,eAAe,OAAO,OAAO,OAAO;EAChD;CACF;;;;;;CAOD,MAAc,iBACZC,IACAC,WACAC,QACY;EACZ,IAAIC;AAEJ,OAAK,IAAI,UAAU,GAAG,WAAW,KAAK,YAAY,UAChD,KAAI;AACF,OAAI,UAAU,EACZ,OAAM,KAAK,aAAa,WAAW,SAAS,OAAO;AAGrD,UAAO,MAAM,IAAI;EAClB,SAAQ,OAAO;AACd,eAAY;AAEZ,QAAK,KAAK,YAAY,OAAO,OAAO,CAClC;EAEH;AAGH,QAAM;CACP;;;;;;;CAQD,AAAQ,eAAeC,OAAgBF,QAAgC;AACrE,MAAI,QAAQ,QACV,QAAO,IAAI,SAAS,uBAAuB;AAG7C,MAAI,iBAAiB,SACnB,QAAO;EAGT,MAAM,SAAS,cAAc,MAAM;AAEnC,MAAI,kBACF,QAAO,IAAI,SAAS,sBAAsB,OAAO;AAGnD,SAAO,IAAI,SAAS,sBAAsB;CAC3C;;;;;;;CAQD,MAAc,YAAeH,QAAuBE,WAA+B;EACjF,MAAM,EAAE,SAAS,OAAO,SAAS,GAAG,KAAK,oBAAoB,OAAO;EAEpE,MAAM,WAAW,MAAM,YACrB,CAAC,kBAAkB,KAAK,OAAO,KAAK,YAAY,OAAO,SAAS,EAAE,QAAQ,cAAe,EAAC,EAC1F,KAAK,WACL,OAAO,OACR;EAED,MAAM,UAAU,SAAS,UAAU,IAAI,SAAS,SAAS,MAAM;AAE/D,OAAK,QACH,OAAM,IAAI,SAAS,sBAAsB;AAG3C,OAAK,OAAO,OAAO,QAAQ,UAAU,aAAa,QAAQ,MAAM,GAAG,IAAI,CAAC,EAAE;AAE1E,OAAK,YAAY,UAAU,WAAW,MAAM;AAE5C,OAAK,SAAS,eAAe;AAE7B,OAAK,QACH,QAAO;AAGT,SAAO,KAAK,iBAAiB,SAAS,OAAO,OAAO;CACrD;;;;;;;CAQD,AAAQ,oBAAuBF,QAAuB;EACpD,MAAM,EACJ,cACA,aACA,cAAc,IACd,WAAW,MACX,YAAY,KAAK,kBACjB,QAAQ,KAAK,OACb,iBACA,YACD,GAAG;EAEJ,MAAM,UAAU,YAAY,QAAQ,WAAW;EAC/C,MAAM,iBAAiB,KAAK,oBAAoB,YAAY,QAAQ;EAEpE,MAAM,UAAU;GACd;GACA;GACA,YAAY;GACZ,GAAI,iBAAiB,EAAE,iBAAiB,eAAgB,IAAG,CAAE;GAC7D,GAAI,kBAAkB,EAAE,kBAAkB,gBAAiB,IAAG,CAAE;GAChE,UAAU,CACR;IAAE,MAAM;IAAmB,SAAS;GAAc,GAClD;IAAE,MAAM;IAAiB,SAAS;GAAa,CAChD;EACF;AAED,SAAO;GAAE;GAAS;GAAO;EAAS;CACnC;;;;;;;;CASD,AAAQ,oBAAoBM,YAA+CC,SAAkB;AAC3F,MAAI,WACF,QAAO;GACL,MAAM;GACN,aAAa;IACX,MAAM,WAAW;IACjB,QAAQ,WAAW;IACnB,QAAQ,WAAW,UAAU;IAC7B,aAAa,WAAW;GACzB;EACF;AAGH,SAAO,UAAU,EAAE,MAAM,cAAwB;CAClD;;;;;;CAOD,AAAQ,YACNC,UACAN,WACAO,OACM;AACN,OAAK,SAAS,UAAU,KAAK,QAAS;AAEtC,OAAK,QAAQ;GACX,cAAc,SAAS,MAAM,iBAAiB;GAC9C,kBAAkB,SAAS,MAAM,qBAAqB;GACtD,aAAa,SAAS,MAAM,gBAAgB;GAC5C;GACA;EACD,EAAC;CACH;;;;;;;CAQD,AAAQ,iBAAoBC,SAAiBC,QAAqC;EAChF,IAAIC;AAEJ,MAAI;AACF,YAAS,KAAK,UAAU,QAAQ;EACjC,QAAO;AACN,SAAM,IAAI,SAAS,yBAAyB;EAC7C;AAED,MAAI,WAAW,QAAQ,kBACrB,OAAM,IAAI,SAAS,yBAAyB;AAG9C,OAAK,OACH,QAAO;EAGT,MAAM,SAAS,OAAO,UAAU,OAAO;AAEvC,OAAK,OAAO,QACV,OAAM,IAAI,SAAS,4BAA4B,sBAAyB,OAAO;AAGjF,SAAO,OAAO;CACf;;;;;;CAOD,MAAc,aAAaV,WAAmBW,SAAiBV,QAAsB;EACnF,MAAM,QAAQ,gBAAgB,KAAK,aAAa,QAAQ;AAExD,OAAK,OAAO,MACT,QAAQ,UAAU,qBAAqB,QAAQ,GAAG,KAAK,WAAW,YAAY,MAAM,IACtF;AAED,QAAM,aAAa,OAAO,OAAO;CAClC;;;;;;;;CASD,AAAQ,YAAYE,OAAgBF,QAA+B;AACjE,MAAI,QAAQ,QACV,QAAO;AAGT,MAAI,iBAAiB,aAAa,MAAM,SAAS,WAAW,MAAM,SAAS,cACzE,QAAO;EAGT,MAAM,SAAS,cAAc,MAAM;AAEnC,MAAI,qBAAwB,KAAK,mBAAmB,SAAS,OAAO,CAClE,QAAO;AAGT,SAAO;CACR;;;;;;;CAQD,MAAM,WAAcW,QAAyC;EAC3D,MAAM,SAAS,MAAM,KAAK,MAAM,IAAI,OAAO,SAAS;AAEpD,MAAI,WAAW,KACb,QAAO;AAGT,MAAI;AACF,SAAM,OAAO,gBAAgB;GAE7B,MAAM,SAAS,MAAM,OAAO,IAAI;AAEhC,OAAI;AACF,UAAM,KAAK,MAAM,IAAI,OAAO,UAAU,QAAQ,OAAO,IAAI;GAC1D,SAAQ,OAAO;AACd,SAAK,OAAO,MAAM,gCAAgC,EAChD,SAAS,iBAAiB,QAAQ,MAAM,UAAU,UACnD,EAAC;GACH;AAED,UAAO;EACR,SAAQ,OAAO;AACd,OAAI;AACF,UAAM,OAAO,eAAe;GAC7B,SAAQ,aAAa;AACpB,SAAK,OAAO,MAAM,gCAAgC,EAChD,SAAS,uBAAuB,QAAQ,YAAY,UAAU,UAC/D,EAAC;GACH;AAED,SAAM;EACP;CACF;;;;;;CAOD,MAAM,cACJC,QACY;EACZ,MAAM,EAAE,MAAM,WAAY,GAAG,aAAa,GAAG;AAE7C,SAAO,KAAK,WAAW;GACrB,GAAG;GACH,IAAI,MAAM,KAAK,KAAK,WAAW;EAChC,EAAC;CACH;;;;;CAMD,kBAAkB;AAChB,SAAO,KAAK,SAAS,UAAU;CAChC;AACF;;;;;;;;;;;ACpaD,SAAgB,cAAcC,iBAA6C;AACzE,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAOrE,IAAIC;AACJ,MAAI,OAAO,iBAAiB,SAAS,eAAe;GAClD,MAAM,EAAE,MAAM,QAAQ,GAAG,OAAO,gBAAgB;AAChD,sBAAmB,0GAA0G,KAAK,OAAO,KAAK,UAAU,OAAO,CAAC;EACjK,WAAU,OAAO,iBAAiB,SAAS,cAC1C,mBAAkB;EAOpB,MAAM,SAAS,CAAC,eAAe,SAAS,eAAgB,EAAC,OAAO,QAAQ,CAAC,KAAK,OAAO;EAErF,MAAM,WAAW,MAAM,gBAAgB,SAAS,OAC9C;GACE,OAAO,OAAO;GACd,YAAY,OAAO;GACnB,aAAa,OAAO;GACpB,QAAQ;GACR,UAAU,aAAa,IAAI,CAAC,OAAO;IAAE,MAAM;IAAiB,SAAS,EAAE;GAAS,GAAE;EACnF,GACD,QACD;EAED,MAAM,OAAO,SAAS,QAAQ,KAAK,CAAC,UAAU,MAAM,SAAS,OAAO,EAAE,QAAQ;AAE9E,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,OAAO;IAC/B,mBAAmB,SAAS,OAAO;IACnC,eACG,SAAS,OAAO,gBAAgB,MAAM,SAAS,OAAO,iBAAiB;GAC3E;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;ACtCD,SAAgB,WAAWC,cAAuC;AAChE,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAErE,MAAM,YAAY,QAAQ,OAAO,gBAAgB;EACjD,MAAMC,mBAEF;GACF,aAAa,OAAO;GACpB,iBAAiB,OAAO;EACzB;AAED,MAAI,UACF,kBAAiB,mBAAmB;AAEtC,MAAI,OAAO,iBAAiB,SAAS,cACnC,kBAAiB,iBAAiB,OAAO,gBAAgB,YAAY;EAGvE,MAAM,WAAW,MAAM,aAAa,gBAClC;GACE,OAAO,OAAO;GACd,UAAU,aAAa,IAAI,CAAC,OAAO;IACjC,MAAM;IACN,OAAO,CAAC,EAAE,MAAM,EAAE,QAAS,CAAC;GAC7B,GAAE;GACH,mBAAmB,gBACf,EAAE,OAAO,CAAC,EAAE,MAAM,cAAc,QAAS,CAAC,EAAE;GAEhD;EACD,GACD,QACD;EAED,MAAM,OACJ,SAAS,aAAa,IAAI,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,QAAQ,GAAG,CAAC,KAAK,GAAG,IAAI;AAEjF,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,eAAe;IACvC,mBAAmB,SAAS,eAAe;IAC3C,cAAc,SAAS,eAAe;GACvC;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;;;;ACjDD,SAAgB,YAAYC,eAAiD;AAC3E,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAErE,IAAIC;AACJ,MAAI,OAAO,iBAAiB,SAAS,eAAe;GAClD,MAAM,EAAE,MAAM,QAAQ,GAAG,OAAO,gBAAgB;AAChD,sBAAmB,0GAA0G,KAAK,OAAO,KAAK,UAAU,OAAO,CAAC;EACjK,WAAU,OAAO,iBAAiB,SAAS,cAC1C,mBAAkB;EAGpB,MAAM,cAAc,CAAC,eAAe,SAAS,eAAgB,EAAC,OAAO,CAAC,MACpE,QAAQ,EAAE,CACX;EAED,MAAM,WAAW,MAAM,cAAc,SACnC;GACE,SAAS,OAAO;GAChB,UAAU,aAAa,IAAI,CAAC,OAAO;IACjC,MAAM;IACN,SAAS,CAAC,EAAE,MAAM,EAAE,QAAS,CAAC;GAC/B,GAAE;GACH,QAAQ,YAAY,SAAS,YAAY,IAAI,CAACC,YAAU,EAAE,aAAM,GAAE;GAClE,iBAAiB;IACf,aAAa,OAAO;IACpB,WAAW,OAAO;GACnB;EACF,GACD,QACD;EAED,MAAM,OAAO,SAAS,QAAQ,SAAS,SAAS,IAAI,CAAC,MAAM,EAAE,QAAQ,GAAG,CAAC,KAAK,GAAG,IAAI;AAErF,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,OAAO;IAC/B,mBAAmB,SAAS,OAAO;IACnC,cAAc,SAAS,OAAO;GAC/B;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;AC5DD,SAAgB,UAAUC,QAAuC;AAC/D,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,aAAa,OAAO,QAAQ,aAAa,OAAO,IAAI,OAAO,GAAG,OAAO;EAC3E,MAAM,iBACG,OAAO,YAAY,aAAa,MAAM,OAAO,SAAS,GAAG,OAAO;EAEzE,MAAM,MAAM,MAAM,MAAM,KAAK;GAC3B,QAAQ,OAAO,UAAU;GACzB,SAAS;IAAE,gBAAgB;IAAoB,GAAG;GAAS;GAC3D,MAAM,KAAK,UAAU,OAAO,WAAW,OAAO,CAAC;GAC/C,QAAQ,QAAQ;EACjB,EAAC;AAEF,OAAK,IAAI,IAAI;GACX,MAAM,OAAO,MAAM,IAAI,MAAM,CAAC,MAAM,MAAM,GAAG;GAC7C,MAAM,MAAM,IAAI,OACb,gCAAgC,IAAI,OAAO,KAAK,KAAK,MAAM,GAAG,IAAI,CAAC;AAEtE,OAAI,SAAS,IAAI;AACjB,SAAM;EACP;EAED,MAAM,OAAO,MAAM,IAAI,MAAM;EAC7B,MAAM,EAAE,SAAS,OAAO,GAAG,OAAO,YAAY,KAAK;AAEnD,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,QAAS,EAAE,CAAC;GACnC,OAAO,QACH;IACE,eAAe,MAAM;IACrB,mBAAmB,MAAM;IACzB,cAAc,MAAM;GACrB;EAEN;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;;;;AC/DD,SAAgB,qBAAqBC,QAA4B;AAC/D,QAAO;AACR;;AAGD,MAAa,WAAW;;AAGxB,MAAa,cAAc;;AAG3B,MAAa,eAAe;;AAG5B,MAAa,eAAe;;AAG5B,MAAa,eAAe;;AAG5B,MAAa,gBAAgB;;;;;;;AAQ7B,MAAa,aAAa"}
+{"version":3,"file":"index.cjs","names":["message: string","type: LLMErrorType","status?: number","issues?: unknown","err: unknown","key: string","value: T","ttl: number","options: CircuitBreakerOptions","content: string","err: unknown","fn: (signal: AbortSignal) => Promise<T>","timeoutMs: number","externalSignal?: AbortSignal","baseDelayMs: number","attempt: number","delay: number","signal?: AbortSignal","debugEnabled: boolean","message: string","meta?: Record<string, unknown>","options: VernLLMOptions","params: CallParams<T>","fn: () => Promise<T>","requestId: string","signal?: AbortSignal","lastError: unknown","error: unknown","jsonSchema: CallParams<unknown>['jsonSchema']","useJson: boolean","response: Awaited<ReturnType<LLMClient['chat']['completions']['create']>>","model: string","content: string","schema?: CallParams<T>['schema']","parsed: unknown","attempt: number","params: CachedCallParams<T>","params: Omit<CachedCallParams<T>, 'fn'> & { call: CallParams<T> }","anthropicClient: AnthropicClient","jsonInstruction: string | undefined","geminiClient: GeminiClient","generationConfig: NonNullable<\n            Parameters<GeminiClient['generateContent']>[0]['generationConfig']\n          >","bedrockClient: BedrockConverseClient","jsonInstruction: string | undefined","text","config: FetchAdapterConfig","client: unknown"],"sources":["../src/types.ts","../src/circuitBreaker.ts","../src/internal/vernLLM.utilts.ts","../src/logger.ts","../src/vernLLM.ts","../src/adapters/anthropic.ts","../src/adapters/gemini.ts","../src/adapters/bedrock.ts","../src/adapters/fetch.ts","../src/adapters/openaiCompatible.ts"],"sourcesContent":["export type LLMErrorType =\n  | 'timeout'\n  | 'api'\n  | 'parse'\n  | 'validation'\n  | 'circuit_open'\n  | 'unknown'\n  | 'aborted';\n\nexport class LLMError extends Error {\n  constructor(\n    message: string,\n    public type: LLMErrorType,\n    public status?: number,\n    public issues?: unknown,\n  ) {\n    super(message);\n    this.name = 'LLMError';\n  }\n}\n\nexport function isLLMError(err: unknown): err is LLMError {\n  return err instanceof LLMError;\n}\n\nexport interface CacheAdapter<T = unknown> {\n  get(key: string): Promise<T | null>;\n  set(key: string, value: T, ttl: number): Promise<void>;\n}\n\n/**\n * Trivial default so the package works out of the box with no external deps\n * Not shared across processes, swap in Redis/Upstash/etc for production\n */\nexport class InMemoryCacheAdapter<T = unknown> implements CacheAdapter<T> {\n  private store = new Map<string, { value: T; expiresAt: number }>();\n\n  async get(key: string): Promise<T | null> {\n    const entry = this.store.get(key);\n    if (!entry) return null;\n    if (Date.now() > entry.expiresAt) {\n      this.store.delete(key);\n      return null;\n    }\n    return entry.value;\n  }\n\n  async set(key: string, value: T, ttl: number): Promise<void> {\n    this.store.set(key, { value, expiresAt: Date.now() + ttl * 1000 });\n  }\n}\n\n/**\n * Minimal shape compatible with the OpenAI SDKs chat.completions.create,\n * so consumers can pass an OpenAI client directly\n * `response_format.json_schema` and `reasoning_effort` are optional on the wire\n * providers that don't support them will just ignore fields they don't recognize,\n * but not every SDKs TS types accept them, hence this being a structural type\n * rather than importing the SDKs own params type\n */\nexport interface LLMClient {\n  chat: {\n    completions: {\n      create(\n        params: {\n          model: string;\n          temperature: number;\n          max_tokens: number;\n          response_format?:\n            | { type: 'json_object' }\n            | {\n                type: 'json_schema';\n                json_schema: {\n                  name: string;\n                  schema: Record<string, unknown>;\n                  strict?: boolean;\n                  description?: string;\n                };\n              };\n          /** OpenAI reasoning-model param (o-series, gpt-5), ignored by providers that don't support it */\n          reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n          messages: Array<{ role: 'system' | 'user'; content: string }>;\n        },\n        options: { signal: AbortSignal },\n      ): Promise<{\n        choices?: Array<{ message?: { content?: string | null } }>;\n        usage?: {\n          prompt_tokens?: number;\n          completion_tokens?: number;\n          total_tokens?: number;\n        };\n      }>;\n    };\n  };\n}\n\nexport type ReserveUsage = () => Promise<void>;\nexport type RefundUsage = () => Promise<void>;\n\nexport interface TokenUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  requestId: string;\n  model: string;\n}\n\nexport type OnUsage = (usage: TokenUsage) => void;\n\n/**\n * Minimal structural type for a Zod-like schema, so this package doesnt need\n * a hard dependency on a specific Zod major version. Any object exposing\n * `safeParse` (Zod v3/v4, and most Zod-compatible validators) should satisfy this\n */\nexport interface SchemaLike<T> {\n  safeParse(data: unknown): { success: true; data: T } | { success: false; error: unknown };\n}\n\nexport interface VernLLMOptions {\n  client: LLMClient;\n  model: string;\n  /** Max retries after the first attempt. Default 1 (2 attempts total) */\n  maxRetries?: number;\n  /** Per-attempt timeout in ms. Default 25000 */\n  timeoutMs?: number;\n  /** Base delay for exponential backoff in ms. Default 500 */\n  baseDelayMs?: number;\n  /** Default max_tokens for calls that don't override it. Default 1000 */\n  defaultMaxTokens?: number;\n  /** Enables debug logging of raw model output. Default: NODE_ENV !== 'production' */\n  debug?: boolean;\n  /** Cache adapter for cachedCall. Defaults to an in-memory adapter */\n  cache?: CacheAdapter;\n  /** HTTP status codes that should fail fast without retrying. Default [400, 401, 403] */\n  nonRetryableStatus?: number[];\n  /** Custom JSON parser. Must return undefined/null on failure. Default: JSON.parse wrapped in try/catch */\n  parseJson?: (content: string) => unknown;\n  /** Called after every successful call with token usage, if the provider reports it */\n  onUsage?: OnUsage;\n  /** Injectable logger. Defaults to a console-based logger gated by `debug` */\n  logger?: import('./logger.js').Logger;\n  /**\n   * Enables a circuit breaker that short-circuits calls after repeated\n   * consecutive failures, instead of continuing to hammer a down provider\n   * Pass `true` for defaults, or an options object to tune threshold/cooldown\n   */\n  circuitBreaker?: boolean | import('./circuitBreaker.js').CircuitBreakerOptions;\n}\n\n/**\n * A provider-native JSON Schema for structured outputs (OpenAI/Groq\n * `response_format: { type: 'json_schema' }`) This is the wire-format\n * schema the model is constrained to generate against, distinct from\n * `schema`, which is a client-side Zod validator run on the parsed result\n * You can use one, both, or neither; using both gets you provider-level\n * constraint plus client-side type inference/validation as a safety net\n */\nexport interface JsonSchemaSpec {\n  name: string;\n  schema: Record<string, unknown>;\n  /** Enforces the schema strictly (OpenAI-specific), default true when supported */\n  strict?: boolean;\n  description?: string;\n}\n\nexport interface CallParams<T = unknown> {\n  systemPrompt: string;\n  userContent: string;\n  temperature?: number;\n  jsonMode?: boolean;\n  maxTokens?: number;\n  requestId?: string;\n  signal?: AbortSignal;\n  /** Overrides the model set on the VernLLM instance for this call only */\n  model?: string;\n  /**\n   * OpenAI-style reasoning effort for reasoning models (o-series, gpt-5, etc).\n   * Passed through as-is, providers/models that don't support it ignore it\n   */\n  reasoningEffort?: 'minimal' | 'low' | 'medium' | 'high';\n  /**\n   * Provider-native JSON Schema structured-output mode. When set, this is sent\n   * as `response_format: { type: 'json_schema', json_schema: ... }` instead of\n   * the looser `json_object` mode, constraining the models output shape at\n   * generation time (not just validating it after the fact). Implies jsonMode: true.\n   */\n  jsonSchema?: JsonSchemaSpec;\n  /**\n   * Optional Zod (or Zod-compatible) schema. When provided, the parsed JSON\n   * is validated against it; on failure an LLMError('validation') is thrown\n   * with `.issues` set to the schema's error object. Implies jsonMode: true.\n   * Can be combined with `jsonSchema` for provider-level constraint + client-side typing.\n   */\n  schema?: SchemaLike<T>;\n}\n\nexport interface CachedCallParams<T> {\n  cacheKey: string;\n  ttl: number;\n  fn: () => Promise<T>;\n  reserveUsage?: ReserveUsage;\n  refundUsage?: RefundUsage;\n}\n","import { LLMError } from './types.js';\n\nexport interface CircuitBreakerOptions {\n  /** Consecutive failures before the circuit opens, default 5 */\n  threshold?: number;\n  /** How long the circuit stays open before allowing a trial request, in ms. Default 30000 */\n  cooldownMs?: number;\n}\n\ntype CircuitState = 'closed' | 'open' | 'half-open';\n\n/**\n * Per retry VernLLM-instance circuit breaker. Tracks consecutive failures across\n * calls. Once the threshold is hit, short-circuits new calls with an\n * LLMError('circuit_open') instead of hitting the provider, until the\n * cooldown elapses and a single trial call is allowed through\n */\nexport class CircuitBreaker {\n  private state: CircuitState = 'closed';\n  private consecutiveFailures = 0;\n  private openedAt = 0;\n  private threshold: number;\n  private cooldownMs: number;\n\n  constructor(options: CircuitBreakerOptions = {}) {\n    this.threshold = options.threshold ?? 5;\n    this.cooldownMs = options.cooldownMs ?? 30_000;\n  }\n\n  /** Throws if the circuit is open and the cooldown hasnt elapsed */\n  assertClosed(): void {\n    if (this.state !== 'open') return;\n\n    const elapsed = Date.now() - this.openedAt;\n    if (elapsed >= this.cooldownMs) {\n      this.state = 'half-open';\n      return;\n    }\n\n    throw new LLMError(\n      `Circuit open — provider has failed ${this.consecutiveFailures} times in a row. Retry in ${Math.ceil((this.cooldownMs - elapsed) / 1000)}s.`,\n      'circuit_open',\n    );\n  }\n\n  recordSuccess(): void {\n    this.consecutiveFailures = 0;\n    this.state = 'closed';\n  }\n\n  recordFailure(): void {\n    this.consecutiveFailures += 1;\n\n    if (this.state === 'half-open') {\n      // Trial call failed: reopen and reset the cooldown window.\n      this.state = 'open';\n      this.openedAt = Date.now();\n      return;\n    }\n\n    if (this.consecutiveFailures >= this.threshold) {\n      this.state = 'open';\n      this.openedAt = Date.now();\n    }\n  }\n\n  getState(): CircuitState {\n    return this.state;\n  }\n}\n","import { LLMError } from '../types.js';\n\nexport function defaultParseJson(content: string): unknown {\n  try {\n    return JSON.parse(content);\n  } catch {\n    return undefined;\n  }\n}\n\n/**\n * Looks inside an unknown error value and pulls out an http status code\n * if one is present. Checks the status field first then the status code\n * field since different client libraries use different names for this.\n * Returns undefined when the error is not an object or carries no status\n */\nexport function extractStatus(err: unknown): number | undefined {\n  if (!err || typeof err !== 'object') return undefined;\n\n  const error = err as {\n    status?: unknown;\n    statusCode?: unknown;\n  };\n\n  if (typeof error.status === 'number') return error.status;\n  if (typeof error.statusCode === 'number') return error.statusCode;\n\n  return undefined;\n}\n\n/**\n * Runs an async function and cancels it if it takes longer than the given\n * timeout. Creates an internal abort controller that fires after the\n * timeout elapses, and combines it with any external signal the caller\n * passed in so either one can cancel the underlying call. If the internal\n * timeout triggers and the underlying operation aborts, the error is\n * converted into an LLMError with type \"timeout\". External cancellations\n * continue to propagate as aborted errors. The internal timer is always\n * cleared afterward, whether the function succeeds, fails, or is aborted,\n * so nothing is left running in the background.\n */\nexport async function withTimeout<T>(\n  fn: (signal: AbortSignal) => Promise<T>,\n  timeoutMs: number,\n  externalSignal?: AbortSignal,\n): Promise<T> {\n  const controller = new AbortController();\n\n  const timer = setTimeout(() => {\n    controller.abort();\n  }, timeoutMs);\n\n  const signal = externalSignal\n    ? AbortSignal.any([externalSignal, controller.signal])\n    : controller.signal;\n\n  try {\n    return await fn(signal);\n  } catch (err) {\n    if (\n      controller.signal.aborted &&\n      !externalSignal?.aborted &&\n      err instanceof DOMException &&\n      err.name === 'AbortError'\n    ) {\n      throw new LLMError('Request timed out', 'timeout');\n    }\n\n    throw err;\n  } finally {\n    clearTimeout(timer);\n  }\n}\n\n/**\n * Exponential backoff with jitter, capped at maxDelayMs.\n * Jitter avoids thundering-herd retries when many callers back off in lockstep,\n * the cap prevents unbounded delays when maxRetries is high\n */\nexport function getBackoffDelay(baseDelayMs: number, attempt: number, maxDelayMs = 10_000): number {\n  const exp = Math.min(baseDelayMs * 2 ** attempt, maxDelayMs);\n  return exp / 2 + Math.random() * (exp / 2);\n}\n\n/**\n * Pauses execution for the given delay before a retry attempt. If an\n * abort signal is provided and it fires while waiting, the pending\n * timer is cancelled immediately and the wait rejects right away with\n * an aborted error instead of continuing to sit idle until the delay\n * would have finished on its own\n */\nexport async function waitForRetry(delay: number, signal?: AbortSignal): Promise<void> {\n  await new Promise<void>((resolve, reject) => {\n    const onAbort = () => {\n      clearTimeout(timer);\n      reject(new LLMError('Operation aborted', 'aborted'));\n    };\n\n    const timer = setTimeout(() => {\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    }, delay);\n\n    signal?.addEventListener('abort', onAbort, { once: true });\n  });\n}\n","export interface Logger {\n  debug(message: string): void;\n  warn(message: string): void;\n  error(message: string, meta?: Record<string, unknown>): void;\n}\n\n/**\n * Default logger. `debug` is gated by the `debug` option on VernLLM\n * warn/error always fire since they indicate real problems (retries, cache failures)\n */\nexport class ConsoleLogger implements Logger {\n  constructor(private debugEnabled: boolean) {}\n\n  debug(message: string): void {\n    if (this.debugEnabled) console.debug(message);\n  }\n\n  warn(message: string): void {\n    console.warn(message);\n  }\n\n  error(message: string, meta?: Record<string, unknown>): void {\n    console.error(message, meta ?? '');\n  }\n}\n","import { randomUUID } from 'crypto';\n\nimport { CircuitBreaker } from './circuitBreaker.js';\nimport {\n  defaultParseJson,\n  extractStatus,\n  withTimeout,\n  getBackoffDelay,\n  waitForRetry,\n} from './internal/vernLLM.utilts.js';\nimport { ConsoleLogger, type Logger } from './logger.js';\nimport {\n  LLMError,\n  InMemoryCacheAdapter,\n  type VernLLMOptions,\n  type CallParams,\n  type CachedCallParams,\n  type CacheAdapter,\n  type LLMClient,\n} from './types.js';\n\n/**\n * A resilient wrapper around an LLM chat completions client, this is VernLLM!\n *\n * Adds retry with exponential backoff and jitter, per-attempt timeouts,\n * an optional circuit breaker, JSON parsing with optional schema\n * validation, usage tracking, and an optional response cache, all\n * configurable, all opt-in beyond sensible defaults\n */\nexport class VernLLM {\n  private readonly client: LLMClient;\n  private readonly model: string;\n\n  private readonly maxRetries: number;\n  private readonly timeoutMs: number;\n  private readonly baseDelayMs: number;\n  private readonly defaultMaxTokens: number;\n\n  private readonly cache: CacheAdapter<unknown>;\n  private readonly nonRetryableStatus: number[];\n\n  private readonly parseJson: (content: string) => unknown;\n  private readonly onUsage?: VernLLMOptions['onUsage'];\n\n  private readonly logger: Logger;\n  private readonly breaker?: CircuitBreaker;\n\n  /**\n   * @param options: Client, model, and all tunables (retries, timeout,\n   * backoff, cache, circuit breaker, logger, etc). See VernLLMOptions in `types.ts`\n   * for individual defaults\n   */\n  constructor(options: VernLLMOptions) {\n    this.client = options.client;\n    this.model = options.model;\n\n    const retryConfig = this.resolveRetryConfig(options);\n    this.maxRetries = retryConfig.maxRetries;\n    this.timeoutMs = retryConfig.timeoutMs;\n    this.baseDelayMs = retryConfig.baseDelayMs;\n    this.defaultMaxTokens = retryConfig.defaultMaxTokens;\n\n    this.cache = options.cache ?? new InMemoryCacheAdapter();\n    this.nonRetryableStatus = options.nonRetryableStatus ?? [400, 401, 403];\n\n    this.parseJson = options.parseJson ?? defaultParseJson;\n    this.onUsage = options.onUsage;\n\n    this.logger = this.resolveLogger(options);\n    this.breaker = this.resolveCircuitBreaker(options);\n  }\n\n  /**\n   * Resolves retry/timeout/token defaults from the given options,\n   * falling back to the librarys built-in defaults for anything unset\n   */\n  private resolveRetryConfig(options: VernLLMOptions) {\n    return {\n      maxRetries: options.maxRetries ?? 1,\n      timeoutMs: options.timeoutMs ?? 25_000,\n      baseDelayMs: options.baseDelayMs ?? 500,\n      defaultMaxTokens: options.defaultMaxTokens ?? 1000,\n    };\n  }\n\n  /**\n   * Returns the caller supplied logger, or a console-based logger whose\n   * debug output is gated by the `debug` option (defaulting to on\n   * outside production)\n   */\n  private resolveLogger(options: VernLLMOptions): Logger {\n    return (\n      options.logger ?? new ConsoleLogger(options.debug ?? process.env.NODE_ENV !== 'production')\n    );\n  }\n\n  /**\n   * Builds a circuit breaker if `circuitBreaker` is truthy on the\n   * options. Passing `true` uses default thresholds, passing an options\n   * object tunes them. Returns undefined when the breaker is disabled\n   */\n  private resolveCircuitBreaker(options: VernLLMOptions): CircuitBreaker | undefined {\n    if (!options.circuitBreaker) return undefined;\n\n    return new CircuitBreaker(options.circuitBreaker === true ? undefined : options.circuitBreaker);\n  }\n\n  /**\n   * Makes a single logical LLM call, transparently retrying on failure\n   * according to the configured retry policy\n   *\n   * Fails fast if the circuit breaker is open or the signal is already\n   * aborted, before any request is dispatched. On exhausting all\n   * retries, records a circuit breaker failure and rejects with a\n   * normalized LLMError\n   *\n   * @param params : System/user content plus per call overrides\n   * (model, temperature, jsonMode, schema, signal, etc)\n   * @returns The parsed and optionally schema-validated response, or\n   * the raw string content when jsonMode is disabled\n   */\n  async call<T = unknown>(params: CallParams<T>): Promise<T> {\n    this.breaker?.assertClosed();\n\n    if (params.signal?.aborted) {\n      throw new LLMError('LLM request aborted', 'aborted');\n    }\n\n    const requestId = params.requestId ?? randomUUID();\n\n    try {\n      const result = await this.retryWithBackoff(\n        () => this.executeCall(params, requestId),\n        requestId,\n        params.signal,\n      );\n\n      return result;\n    } catch (error) {\n      this.breaker?.recordFailure();\n      throw this.normalizeError(error, params.signal);\n    }\n  }\n\n  /**\n   * Runs `fn`, retrying with backoff according to `shouldRetry` policy\n   * Purely mechanical: knows nothing about LLM specifics beyond the retry\n   * predicate, so its testable independent of request/response shaping\n   */\n  private async retryWithBackoff<T>(\n    fn: () => Promise<T>,\n    requestId: string,\n    signal?: AbortSignal,\n  ): Promise<T> {\n    let lastError: unknown;\n\n    for (let attempt = 0; attempt <= this.maxRetries; attempt++) {\n      try {\n        if (attempt > 0) {\n          await this.recoverDelay(requestId, attempt, signal);\n        }\n\n        return await fn();\n      } catch (error) {\n        lastError = error;\n\n        if (!this.shouldRetry(error, signal)) {\n          break;\n        }\n      }\n    }\n\n    throw lastError;\n  }\n\n  /**\n   * Converts any thrown value into a well-typed LLMError for the public\n   * API surface. Preserves an existing LLMError as is, reports aborted\n   * signals as such, classifies errors carrying an http status as\n   * type api, and otherwise falls back to a generic unknown error.\n   */\n  private normalizeError(error: unknown, signal?: AbortSignal): LLMError {\n    if (signal?.aborted) {\n      return new LLMError('LLM request aborted', 'aborted');\n    }\n\n    if (error instanceof LLMError) {\n      return error;\n    }\n\n    const status = extractStatus(error);\n\n    if (status !== undefined) {\n      return new LLMError('LLM request failed', 'api', status);\n    }\n\n    return new LLMError('LLM request failed', 'unknown');\n  }\n\n  /**\n   * Performs a single attempt: builds the request, dispatches it with a\n   * timeout, and shapes the response. Throws on an empty response so\n   * the retry loop treats it like any other transient failure. Records\n   * usage and a circuit breaker success before returning\n   */\n  private async executeCall<T>(params: CallParams<T>, requestId: string): Promise<T> {\n    const { useJson, model, request } = this.buildRequestPayload(params);\n\n    const response = await withTimeout(\n      (attemptSignal) => this.client.chat.completions.create(request, { signal: attemptSignal }),\n      this.timeoutMs,\n      params.signal,\n    );\n\n    const content = response.choices?.[0]?.message?.content?.trim();\n\n    if (!content) {\n      throw new LLMError('Empty LLM response', 'api');\n    }\n\n    this.logger.debug(`[vern:${requestId}] output:\\n${content.slice(0, 800)}`);\n\n    this.recordUsage(response, requestId, model);\n\n    this.breaker?.recordSuccess();\n\n    if (!useJson) {\n      return content as T;\n    }\n\n    return this.parseAndValidate(content, params.schema);\n  }\n\n  /**\n   * Applies per call defaults and shapes the params into the request\n   * object expected by the underlying client, including the resolved\n   * response format. Also returns whether JSON parsing should be\n   * applied to the response and which model was ultimately used\n   */\n  private buildRequestPayload<T>(params: CallParams<T>) {\n    const {\n      systemPrompt,\n      userContent,\n      temperature = 0.2,\n      jsonMode = true,\n      maxTokens = this.defaultMaxTokens,\n      model = this.model,\n      reasoningEffort,\n      jsonSchema,\n    } = params;\n\n    const useJson = jsonMode || Boolean(jsonSchema);\n    const responseFormat = this.buildResponseFormat(jsonSchema, useJson);\n\n    const request = {\n      model,\n      temperature,\n      max_tokens: maxTokens,\n      ...(responseFormat ? { response_format: responseFormat } : {}),\n      ...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}),\n      messages: [\n        { role: 'system' as const, content: systemPrompt },\n        { role: 'user' as const, content: userContent },\n      ],\n    };\n\n    return { useJson, model, request };\n  }\n\n  /**\n   * Chooses the response format to send to the provider. A provider\n   * native json schema takes priority when supplied, constraining\n   * generation directly, otherwise falls back to the looser json\n   * object mode when JSON output is requested, or no format at all\n   * for plain text responses\n   */\n  private buildResponseFormat(jsonSchema: CallParams<unknown>['jsonSchema'], useJson: boolean) {\n    if (jsonSchema) {\n      return {\n        type: 'json_schema' as const,\n        json_schema: {\n          name: jsonSchema.name,\n          schema: jsonSchema.schema,\n          strict: jsonSchema.strict ?? true,\n          description: jsonSchema.description,\n        },\n      };\n    }\n\n    return useJson ? { type: 'json_object' as const } : undefined;\n  }\n\n  /**\n   * Reports token usage to the caller supplied onUsage callback, when\n   * both a callback was configured and the provider actually returned\n   * usage data on this response. A no op otherwise.\n   */\n  private recordUsage(\n    response: Awaited<ReturnType<LLMClient['chat']['completions']['create']>>,\n    requestId: string,\n    model: string,\n  ): void {\n    if (!response.usage || !this.onUsage) return;\n\n    this.onUsage({\n      promptTokens: response.usage.prompt_tokens ?? 0,\n      completionTokens: response.usage.completion_tokens ?? 0,\n      totalTokens: response.usage.total_tokens ?? 0,\n      requestId,\n      model,\n    });\n  }\n\n  /**\n   * Parses the raw response content as JSON and, when a schema is\n   * supplied, validates the parsed value against it. Throws a parse\n   * type LLMError on malformed JSON and a validation type LLMError,\n   * carrying the schemas issues, on a failed validation\n   */\n  private parseAndValidate<T>(content: string, schema?: CallParams<T>['schema']): T {\n    let parsed: unknown;\n\n    try {\n      parsed = this.parseJson(content);\n    } catch {\n      throw new LLMError('Invalid JSON response', 'parse');\n    }\n\n    if (parsed === null || parsed === undefined) {\n      throw new LLMError('Invalid JSON response', 'parse');\n    }\n\n    if (!schema) {\n      return parsed as T;\n    }\n\n    const result = schema.safeParse(parsed);\n\n    if (!result.success) {\n      throw new LLMError('Schema validation failed', 'validation', undefined, result.error);\n    }\n\n    return result.data;\n  }\n\n  /**\n   * Waits out the backoff delay for a given retry attempt, logging the\n   * attempt for observability before the wait begins. Rejects early if\n   * the signal aborts during the wait\n   */\n  private async recoverDelay(requestId: string, attempt: number, signal?: AbortSignal) {\n    const delay = getBackoffDelay(this.baseDelayMs, attempt);\n\n    this.logger.warn(\n      `[vern:${requestId}] recovery attempt ${attempt}/${this.maxRetries}, waiting ${delay}ms`,\n    );\n\n    await waitForRetry(delay, signal);\n  }\n\n  /**\n   * Decides whether a failed attempt is worth retrying. Never retries\n   * once the signal has aborted, never retries a parse or validation\n   * failure since those stem from the response content rather than a\n   * transient fault, and never retries a status code the caller has\n   * marked as non retryable. Retries everything else\n   */\n  private shouldRetry(error: unknown, signal?: AbortSignal): boolean {\n    if (signal?.aborted) {\n      return false;\n    }\n\n    if (error instanceof LLMError && (error.type === 'parse' || error.type === 'validation')) {\n      return false;\n    }\n\n    const status = extractStatus(error);\n\n    if (status !== undefined && this.nonRetryableStatus.includes(status)) {\n      return false;\n    }\n\n    return true;\n  }\n\n  /**\n   * Thin cache wrapper around caller supplied logic. `params.fn` is expected\n   * to be a call that itself invokes `this.call(...)` (see `cachedLLMCall`\n   * below for a convenience wrapper that wires this up automatically),\n   * `cachedCall` does not itself apply retry/timeout policy.\n   */\n  async cachedCall<T>(params: CachedCallParams<T>): Promise<T> {\n    const cached = await this.cache.get(params.cacheKey);\n\n    if (cached !== null) {\n      return cached as T;\n    }\n\n    try {\n      await params.reserveUsage?.();\n\n      const result = await params.fn();\n\n      try {\n        await this.cache.set(params.cacheKey, result, params.ttl);\n      } catch (error) {\n        this.logger.error('[VernLLM] cache write failed', {\n          message: error instanceof Error ? error.message : 'unknown',\n        });\n      }\n\n      return result;\n    } catch (error) {\n      try {\n        await params.refundUsage?.();\n      } catch (refundError) {\n        this.logger.error('[VernLLM] refundUsage failed', {\n          message: refundError instanceof Error ? refundError.message : 'unknown',\n        });\n      }\n\n      throw error;\n    }\n  }\n\n  /**\n   * Convenience wrapper composing `call` + `cachedCall`, so cached LLM calls\n   * automatically get retry/timeout/circuit-breaker behavior without callers\n   * having to remember to wire `fn: () => this.call(...)` themselves\n   */\n  async cachedLLMCall<T>(\n    params: Omit<CachedCallParams<T>, 'fn'> & { call: CallParams<T> },\n  ): Promise<T> {\n    const { call: callParams, ...cacheParams } = params;\n\n    return this.cachedCall({\n      ...cacheParams,\n      fn: () => this.call(callParams),\n    });\n  }\n\n  /**\n   * Returns the current circuit breaker state, or undefined when no\n   * circuit breaker was configured on this instance\n   */\n  getCircuitState() {\n    return this.breaker?.getState();\n  }\n}\n","import type { LLMClient } from '../types.js';\n\n/** Minimal structural type for the Anthropic SDKs `messages.create` */\nexport interface AnthropicClient {\n  messages: {\n    create(\n      params: {\n        model: string;\n        max_tokens: number;\n        temperature?: number;\n        system?: string;\n        messages: Array<{ role: 'user' | 'assistant'; content: string }>;\n      },\n      options: { signal: AbortSignal },\n    ): Promise<{\n      content: Array<{ type: string; text?: string }>;\n      usage?: { input_tokens?: number; output_tokens?: number };\n    }>;\n  };\n}\n\n/**\n * Wraps an Anthropic SDK client so it satisfies the same `LLMClient`\n * interface VernLLM uses for OpenAI/Groq. Anthropics Messages API has no\n * `response_format: json_object` equivalent, so when the caller requests\n * JSON mode, this adapter appends an instruction to the system prompt\n * asking the model to respond with JSON only\n */\nexport function fromAnthropic(anthropicClient: AnthropicClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          // Anthropics Messages API has no `response_format: json_object` or\n          // `json_schema` equivalent, so both are emulated via system prompt\n          // instructions. For json_schema, the schema itself is embedded so the\n          // model has something concrete to conform to (not provider-enforced,\n          // unlike OpenAIs native structured outputs)\n          let jsonInstruction: string | undefined;\n          if (params.response_format?.type === 'json_schema') {\n            const { name, schema } = params.response_format.json_schema;\n            jsonInstruction = `Respond with valid JSON only, no prose or markdown fences. The JSON must conform to this schema (name: \"${name}\"):\\n${JSON.stringify(schema)}`;\n          } else if (params.response_format?.type === 'json_object') {\n            jsonInstruction = 'Respond with valid JSON only, no prose or markdown fences.';\n          }\n\n          // `reasoning_effort` (OpenAI o-series/gpt-5 style) has no direct Anthropic\n          // equivalent, Claudes extended thinking uses a token budget, not a tier\n          // string, so its intentionally dropped here rather than guessed at.\n\n          const system = [systemMessage?.content, jsonInstruction].filter(Boolean).join('\\n\\n');\n\n          const response = await anthropicClient.messages.create(\n            {\n              model: params.model,\n              max_tokens: params.max_tokens,\n              temperature: params.temperature,\n              system: system || undefined,\n              messages: userMessages.map((m) => ({ role: 'user' as const, content: m.content })),\n            },\n            options,\n          );\n\n          const text = response.content.find((block) => block.type === 'text')?.text ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usage?.input_tokens,\n              completion_tokens: response.usage?.output_tokens,\n              total_tokens:\n                (response.usage?.input_tokens ?? 0) + (response.usage?.output_tokens ?? 0),\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Minimal structural type for Geminis `generateContent`, matching both the\n * legacy `@google/generative-ai` SDKs `model.generateContent(...)` and the\n * newer `@google/genai` SDKs `ai.models.generateContent({ model, ... })`\n * closely enough to adapt either — pass whichever `.generateContent` you have.\n */\nexport interface GeminiClient {\n  generateContent(\n    params: {\n      model?: string;\n      contents: Array<{ role: 'user'; parts: Array<{ text: string }> }>;\n      systemInstruction?: { parts: Array<{ text: string }> };\n      generationConfig?: {\n        temperature?: number;\n        maxOutputTokens?: number;\n        responseMimeType?: string;\n        responseSchema?: Record<string, unknown>;\n      };\n    },\n    options: { signal: AbortSignal },\n  ): Promise<{\n    candidates?: Array<{ content?: { parts?: Array<{ text?: string }> } }>;\n    usageMetadata?: {\n      promptTokenCount?: number;\n      candidatesTokenCount?: number;\n      totalTokenCount?: number;\n    };\n  }>;\n}\n\n/**\n * Wraps a Gemini client so it satisfies the `LLMClient` interface VernLLM\n * uses for OpenAI/Groq. Geminis shape differs on nearly every axis: a\n * `contents` array instead of `messages`, a separate `systemInstruction`\n * field instead of a `system` role message, `generationConfig` instead of\n * top-level `temperature`/`max_tokens`, and native JSON Schema support via\n * `responseMimeType: 'application/json'` + `responseSchema` (so `jsonSchema`\n * is provider-enforced here, unlike the Anthropic adapters prompt-embedding\n * fallback). `reasoning_effort` has no equivalent. Geminis thinking models\n * use a token budget, not an effort tier, so its dropped, same as Anthropic.\n */\nexport function fromGemini(geminiClient: GeminiClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          const wantsJson = Boolean(params.response_format);\n          const generationConfig: NonNullable<\n            Parameters<GeminiClient['generateContent']>[0]['generationConfig']\n          > = {\n            temperature: params.temperature,\n            maxOutputTokens: params.max_tokens,\n          };\n\n          if (wantsJson) {\n            generationConfig.responseMimeType = 'application/json';\n          }\n          if (params.response_format?.type === 'json_schema') {\n            generationConfig.responseSchema = params.response_format.json_schema.schema;\n          }\n\n          const response = await geminiClient.generateContent(\n            {\n              model: params.model,\n              contents: userMessages.map((m) => ({\n                role: 'user' as const,\n                parts: [{ text: m.content }],\n              })),\n              systemInstruction: systemMessage\n                ? { parts: [{ text: systemMessage.content }] }\n                : undefined,\n              generationConfig,\n            },\n            options,\n          );\n\n          const text =\n            response.candidates?.[0]?.content?.parts?.map((p) => p.text ?? '').join('') ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usageMetadata?.promptTokenCount,\n              completion_tokens: response.usageMetadata?.candidatesTokenCount,\n              total_tokens: response.usageMetadata?.totalTokenCount,\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Minimal structural type matching AWS Bedrocks Converse API. This is\n * intentionally NOT `BedrockRuntimeClient` itself, the AWS SDK v3 client\n * exposes `.send(command)`, not a direct `.converse()` method, and pulling\n * in `@aws-sdk/client-bedrock-runtime` as a dependency just for its types\n * isn't worth it for a structural adapter. Wrap your client, e.g:\n *\n * ```ts\n * import { BedrockRuntimeClient, ConverseCommand } from '@aws-sdk/client-bedrock-runtime';\n * const client = new BedrockRuntimeClient({ region: 'us-east-1' });\n * const converseClient = {\n *   converse: (params, options) =>\n *     client.send(new ConverseCommand(params), { abortSignal: options.signal }),\n * };\n * ```\n */\nexport interface BedrockConverseClient {\n  converse(\n    params: {\n      modelId: string;\n      messages: Array<{ role: 'user'; content: Array<{ text: string }> }>;\n      system?: Array<{ text: string }>;\n      inferenceConfig?: { temperature?: number; maxTokens?: number };\n    },\n    options: { signal: AbortSignal },\n  ): Promise<{\n    output?: { message?: { content?: Array<{ text?: string }> } };\n    usage?: { inputTokens?: number; outputTokens?: number; totalTokens?: number };\n  }>;\n}\n\n/**\n * Wraps a Bedrock Converse-API client so it satisfies the `LLMClient`\n * interface VernLLM uses for OpenAI/Groq. The Converse API is unified\n * across Bedrocks model families (Anthropic, Titan, Llama, Mistral, etc.),\n * so unlike raw per-model Bedrock invocation, this one adapter works\n * regardless of which underlying model `modelId` points at, as long as\n * that model supports Converse (most current-generation ones do)\n *\n * Theres no uniform native JSON Schema enforcement across families here\n * (some support it via forced tool-use, which varies per model), so\n * `jsonSchema`/`jsonMode` are emulated via a system-prompt instruction, same\n * approach as the Anthropic adapter. `reasoning_effort` has no Converse\n * equivalent and is dropped\n */\nexport function fromBedrock(bedrockClient: BedrockConverseClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          let jsonInstruction: string | undefined;\n          if (params.response_format?.type === 'json_schema') {\n            const { name, schema } = params.response_format.json_schema;\n            jsonInstruction = `Respond with valid JSON only, no prose or markdown fences. The JSON must conform to this schema (name: \"${name}\"):\\n${JSON.stringify(schema)}`;\n          } else if (params.response_format?.type === 'json_object') {\n            jsonInstruction = 'Respond with valid JSON only, no prose or markdown fences.';\n          }\n\n          const systemParts = [systemMessage?.content, jsonInstruction].filter((s): s is string =>\n            Boolean(s),\n          );\n\n          const response = await bedrockClient.converse(\n            {\n              modelId: params.model,\n              messages: userMessages.map((m) => ({\n                role: 'user' as const,\n                content: [{ text: m.content }],\n              })),\n              system: systemParts.length ? systemParts.map((text) => ({ text })) : undefined,\n              inferenceConfig: {\n                temperature: params.temperature,\n                maxTokens: params.max_tokens,\n              },\n            },\n            options,\n          );\n\n          const text = response.output?.message?.content?.map((c) => c.text ?? '').join('') ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usage?.inputTokens,\n              completion_tokens: response.usage?.outputTokens,\n              total_tokens: response.usage?.totalTokens,\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/** The chat-completion-shaped request VernLLM builds internally */\ntype ChatRequest = Parameters<LLMClient['chat']['completions']['create']>[0];\n\nexport interface FetchAdapterConfig {\n  /** Endpoint URL, or a function of the request in case it depends on model/params */\n  url: string | ((params: ChatRequest) => string);\n  /** Static headers, or a function (sync or async) for things like refreshed auth tokens */\n  headers?:\n    | Record<string, string>\n    | (() => Record<string, string> | Promise<Record<string, string>>);\n  /** HTTP method. Default 'POST' */\n  method?: string;\n  /** Maps VernLLMs internal chat-completion request into the providers raw request body */\n  mapRequest: (params: ChatRequest) => unknown;\n  /**\n   * Maps the providers raw JSON response into `{ content, usage? }`\n   * `content` is the assistants text (JSON string when JSON mode was requested)\n   */\n  mapResponse: (json: unknown) => {\n    content: string;\n    usage?: { promptTokens?: number; completionTokens?: number; totalTokens?: number };\n  };\n}\n\n/**\n * A fetch-based escape hatch for providers with no SDK, or where pulling one\n * in isnt worth it. You supply the URL, headers, and two small mapping\n * functions; this handles the HTTP call and slots the result into the same\n * `LLMClient` shape every other adapter produces, so retries, timeouts,\n * the circuit breaker, and JSON/schema handling all still work unmodified\n *\n * Non-2xx responses throw an error with `.status` set to the HTTP status\n * code, so VernLLMs `nonRetryableStatus` handling (e.g. failing fast on\n * 401/403) applies here too\n */\nexport function fromFetch(config: FetchAdapterConfig): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const url = typeof config.url === 'function' ? config.url(params) : config.url;\n          const headers =\n            typeof config.headers === 'function' ? await config.headers() : config.headers;\n\n          const res = await fetch(url, {\n            method: config.method ?? 'POST',\n            headers: { 'Content-Type': 'application/json', ...headers },\n            body: JSON.stringify(config.mapRequest(params)),\n            signal: options.signal,\n          });\n\n          if (!res.ok) {\n            const body = await res.text().catch(() => '');\n            const err = new Error(\n              `Fetch adapter request failed (${res.status}): ${body.slice(0, 500)}`,\n            ) as Error & { status?: number };\n            err.status = res.status;\n            throw err;\n          }\n\n          const json = await res.json();\n          const { content, usage } = config.mapResponse(json);\n\n          return {\n            choices: [{ message: { content } }],\n            usage: usage\n              ? {\n                  prompt_tokens: usage.promptTokens,\n                  completion_tokens: usage.completionTokens,\n                  total_tokens: usage.totalTokens,\n                }\n              : undefined,\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Passthrough adapter for any SDK/client whose `chat.completions.create`\n * already matches the OpenAI wire format 1:1 : this covers most hosted\n * inference providers, since \"OpenAI-compatible\" is a de facto standard for\n * chat completion APIs. No transformation happens here, this exists purely\n * so call sites read clearly (`fromMistral(client)` vs handing a Mistral\n * client to something typed for OpenAI) and so a real transformation could\n * be added later, per-provider, without a breaking change.\n *\n * Not every SDKs own TypeScript types line up exactly with `LLMClient`\n * (extra fields, stricter unions, etc.), so this takes `unknown` and casts:\n * the actual compatibility contract is the JSON each provider sends and\n * receives over the wire, not the SDKs TS types.\n */\nexport function fromOpenAICompatible(client: unknown): LLMClient {\n  return client as LLMClient;\n}\n\n/** Groqs SDK matches the OpenAI wire format */\nexport const fromGroq = fromOpenAICompatible;\n\n/** Mistrals `chat.completions`-shaped client (or their OpenAI-compat endpoint) */\nexport const fromMistral = fromOpenAICompatible;\n\n/** DeepSeeks API is OpenAI-compatible */\nexport const fromDeepSeek = fromOpenAICompatible;\n\n/** Cerebras inference API is OpenAI-compatible */\nexport const fromCerebras = fromOpenAICompatible;\n\n/** Together AIs API is OpenAI-compatible */\nexport const fromTogether = fromOpenAICompatible;\n\n/** Fireworks AIs API is OpenAI-compatible */\nexport const fromFireworks = fromOpenAICompatible;\n\n/**\n * Ollama exposes an OpenAI-compatible endpoint at `/v1/chat/completions`\n * (as opposed to its native `/api/chat` format, which differs). Point an\n * OpenAI SDK instances `baseURL` at your Ollama server and pass it here:\n * this does not talk to Ollamas native API directly.\n */\nexport const fromOllama = fromOpenAICompatible;\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AASA,IAAa,WAAb,cAA8B,MAAM;CAClC,YACEA,SACOC,MACAC,QACAC,QACP;AACA,QAAM,QAAQ;EA2LjB,KA/LU;EA+LT,KA9LS;EA8LR,KA7LQ;AAGP,OAAK,OAAO;CACb;AACF;AAED,SAAgB,WAAWC,KAA+B;AACxD,QAAO,eAAe;AACvB;;;;;AAWD,IAAa,uBAAb,MAA0E;CACxE,AAAQ,QAAQ,IAAI;CAEpB,MAAM,IAAIC,KAAgC;EACxC,MAAM,QAAQ,KAAK,MAAM,IAAI,IAAI;AACjC,OAAK,MAAO,QAAO;AACnB,MAAI,KAAK,KAAK,GAAG,MAAM,WAAW;AAChC,QAAK,MAAM,OAAO,IAAI;AACtB,UAAO;EACR;AACD,SAAO,MAAM;CACd;CAED,MAAM,IAAIA,KAAaC,OAAUC,KAA4B;AAC3D,OAAK,MAAM,IAAI,KAAK;GAAE;GAAO,WAAW,KAAK,KAAK,GAAG,MAAM;EAAM,EAAC;CACnE;AACF;;;;;;;;;;ACjCD,IAAa,iBAAb,MAA4B;CAC1B,AAAQ,QAAsB;CAC9B,AAAQ,sBAAsB;CAC9B,AAAQ,WAAW;CACnB,AAAQ;CACR,AAAQ;CAER,YAAYC,UAAiC,CAAE,GAAE;AAC/C,OAAK,YAAY,QAAQ,aAAa;AACtC,OAAK,aAAa,QAAQ,cAAc;CACzC;;CAGD,eAAqB;AACnB,MAAI,KAAK,UAAU,OAAQ;EAE3B,MAAM,UAAU,KAAK,KAAK,GAAG,KAAK;AAClC,MAAI,WAAW,KAAK,YAAY;AAC9B,QAAK,QAAQ;AACb;EACD;AAED,QAAM,IAAI,UACP,qCAAqC,KAAK,oBAAoB,4BAA4B,KAAK,MAAM,KAAK,aAAa,WAAW,IAAK,CAAC,KACzI;CAEH;CAED,gBAAsB;AACpB,OAAK,sBAAsB;AAC3B,OAAK,QAAQ;CACd;CAED,gBAAsB;AACpB,OAAK,uBAAuB;AAE5B,MAAI,KAAK,UAAU,aAAa;AAE9B,QAAK,QAAQ;AACb,QAAK,WAAW,KAAK,KAAK;AAC1B;EACD;AAED,MAAI,KAAK,uBAAuB,KAAK,WAAW;AAC9C,QAAK,QAAQ;AACb,QAAK,WAAW,KAAK,KAAK;EAC3B;CACF;CAED,WAAyB;AACvB,SAAO,KAAK;CACb;AACF;;;;ACnED,SAAgB,iBAAiBC,SAA0B;AACzD,KAAI;AACF,SAAO,KAAK,MAAM,QAAQ;CAC3B,QAAO;AACN;CACD;AACF;;;;;;;AAQD,SAAgB,cAAcC,KAAkC;AAC9D,MAAK,cAAc,QAAQ,SAAU;CAErC,MAAM,QAAQ;AAKd,YAAW,MAAM,WAAW,SAAU,QAAO,MAAM;AACnD,YAAW,MAAM,eAAe,SAAU,QAAO,MAAM;AAEvD;AACD;;;;;;;;;;;;AAaD,eAAsB,YACpBC,IACAC,WACAC,gBACY;CACZ,MAAM,aAAa,IAAI;CAEvB,MAAM,QAAQ,WAAW,MAAM;AAC7B,aAAW,OAAO;CACnB,GAAE,UAAU;CAEb,MAAM,SAAS,iBACX,YAAY,IAAI,CAAC,gBAAgB,WAAW,MAAO,EAAC,GACpD,WAAW;AAEf,KAAI;AACF,SAAO,MAAM,GAAG,OAAO;CACxB,SAAQ,KAAK;AACZ,MACE,WAAW,OAAO,YACjB,gBAAgB,WACjB,eAAe,gBACf,IAAI,SAAS,aAEb,OAAM,IAAI,SAAS,qBAAqB;AAG1C,QAAM;CACP,UAAS;AACR,eAAa,MAAM;CACpB;AACF;;;;;;AAOD,SAAgB,gBAAgBC,aAAqBC,SAAiB,aAAa,KAAgB;CACjG,MAAM,MAAM,KAAK,IAAI,cAAc,KAAK,SAAS,WAAW;AAC5D,QAAO,MAAM,IAAI,KAAK,QAAQ,IAAI,MAAM;AACzC;;;;;;;;AASD,eAAsB,aAAaC,OAAeC,QAAqC;AACrF,OAAM,IAAI,QAAc,CAAC,SAAS,WAAW;EAC3C,MAAM,UAAU,MAAM;AACpB,gBAAa,MAAM;AACnB,UAAO,IAAI,SAAS,qBAAqB,WAAW;EACrD;EAED,MAAM,QAAQ,WAAW,MAAM;AAC7B,WAAQ,oBAAoB,SAAS,QAAQ;AAC7C,YAAS;EACV,GAAE,MAAM;AAET,UAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAM,EAAC;CAC3D;AACF;;;;;;;;AC/FD,IAAa,gBAAb,MAA6C;CAC3C,YAAoBC,cAAuB;EAc5C,KAdqB;CAAyB;CAE7C,MAAMC,SAAuB;AAC3B,MAAI,KAAK,aAAc,SAAQ,MAAM,QAAQ;CAC9C;CAED,KAAKA,SAAuB;AAC1B,UAAQ,KAAK,QAAQ;CACtB;CAED,MAAMA,SAAiBC,MAAsC;AAC3D,UAAQ,MAAM,SAAS,QAAQ,GAAG;CACnC;AACF;;;;;;;;;;;;ACKD,IAAa,UAAb,MAAqB;CACnB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;;;;;;CAOjB,YAAYC,SAAyB;AACnC,OAAK,SAAS,QAAQ;AACtB,OAAK,QAAQ,QAAQ;EAErB,MAAM,cAAc,KAAK,mBAAmB,QAAQ;AACpD,OAAK,aAAa,YAAY;AAC9B,OAAK,YAAY,YAAY;AAC7B,OAAK,cAAc,YAAY;AAC/B,OAAK,mBAAmB,YAAY;AAEpC,OAAK,QAAQ,QAAQ,SAAS,IAAI;AAClC,OAAK,qBAAqB,QAAQ,sBAAsB;GAAC;GAAK;GAAK;EAAI;AAEvE,OAAK,YAAY,QAAQ,aAAa;AACtC,OAAK,UAAU,QAAQ;AAEvB,OAAK,SAAS,KAAK,cAAc,QAAQ;AACzC,OAAK,UAAU,KAAK,sBAAsB,QAAQ;CACnD;;;;;CAMD,AAAQ,mBAAmBA,SAAyB;AAClD,SAAO;GACL,YAAY,QAAQ,cAAc;GAClC,WAAW,QAAQ,aAAa;GAChC,aAAa,QAAQ,eAAe;GACpC,kBAAkB,QAAQ,oBAAoB;EAC/C;CACF;;;;;;CAOD,AAAQ,cAAcA,SAAiC;AACrD,SACE,QAAQ,UAAU,IAAI,cAAc,QAAQ,SAAS,QAAQ,IAAI,aAAa;CAEjF;;;;;;CAOD,AAAQ,sBAAsBA,SAAqD;AACjF,OAAK,QAAQ,eAAgB;AAE7B,SAAO,IAAI,eAAe,QAAQ,mBAAmB,gBAAmB,QAAQ;CACjF;;;;;;;;;;;;;;;CAgBD,MAAM,KAAkBC,QAAmC;AACzD,OAAK,SAAS,cAAc;AAE5B,MAAI,OAAO,QAAQ,QACjB,OAAM,IAAI,SAAS,uBAAuB;EAG5C,MAAM,YAAY,OAAO,aAAa,wBAAY;AAElD,MAAI;GACF,MAAM,SAAS,MAAM,KAAK,iBACxB,MAAM,KAAK,YAAY,QAAQ,UAAU,EACzC,WACA,OAAO,OACR;AAED,UAAO;EACR,SAAQ,OAAO;AACd,QAAK,SAAS,eAAe;AAC7B,SAAM,KAAK,eAAe,OAAO,OAAO,OAAO;EAChD;CACF;;;;;;CAOD,MAAc,iBACZC,IACAC,WACAC,QACY;EACZ,IAAIC;AAEJ,OAAK,IAAI,UAAU,GAAG,WAAW,KAAK,YAAY,UAChD,KAAI;AACF,OAAI,UAAU,EACZ,OAAM,KAAK,aAAa,WAAW,SAAS,OAAO;AAGrD,UAAO,MAAM,IAAI;EAClB,SAAQ,OAAO;AACd,eAAY;AAEZ,QAAK,KAAK,YAAY,OAAO,OAAO,CAClC;EAEH;AAGH,QAAM;CACP;;;;;;;CAQD,AAAQ,eAAeC,OAAgBF,QAAgC;AACrE,MAAI,QAAQ,QACV,QAAO,IAAI,SAAS,uBAAuB;AAG7C,MAAI,iBAAiB,SACnB,QAAO;EAGT,MAAM,SAAS,cAAc,MAAM;AAEnC,MAAI,kBACF,QAAO,IAAI,SAAS,sBAAsB,OAAO;AAGnD,SAAO,IAAI,SAAS,sBAAsB;CAC3C;;;;;;;CAQD,MAAc,YAAeH,QAAuBE,WAA+B;EACjF,MAAM,EAAE,SAAS,OAAO,SAAS,GAAG,KAAK,oBAAoB,OAAO;EAEpE,MAAM,WAAW,MAAM,YACrB,CAAC,kBAAkB,KAAK,OAAO,KAAK,YAAY,OAAO,SAAS,EAAE,QAAQ,cAAe,EAAC,EAC1F,KAAK,WACL,OAAO,OACR;EAED,MAAM,UAAU,SAAS,UAAU,IAAI,SAAS,SAAS,MAAM;AAE/D,OAAK,QACH,OAAM,IAAI,SAAS,sBAAsB;AAG3C,OAAK,OAAO,OAAO,QAAQ,UAAU,aAAa,QAAQ,MAAM,GAAG,IAAI,CAAC,EAAE;AAE1E,OAAK,YAAY,UAAU,WAAW,MAAM;AAE5C,OAAK,SAAS,eAAe;AAE7B,OAAK,QACH,QAAO;AAGT,SAAO,KAAK,iBAAiB,SAAS,OAAO,OAAO;CACrD;;;;;;;CAQD,AAAQ,oBAAuBF,QAAuB;EACpD,MAAM,EACJ,cACA,aACA,cAAc,IACd,WAAW,MACX,YAAY,KAAK,kBACjB,QAAQ,KAAK,OACb,iBACA,YACD,GAAG;EAEJ,MAAM,UAAU,YAAY,QAAQ,WAAW;EAC/C,MAAM,iBAAiB,KAAK,oBAAoB,YAAY,QAAQ;EAEpE,MAAM,UAAU;GACd;GACA;GACA,YAAY;GACZ,GAAI,iBAAiB,EAAE,iBAAiB,eAAgB,IAAG,CAAE;GAC7D,GAAI,kBAAkB,EAAE,kBAAkB,gBAAiB,IAAG,CAAE;GAChE,UAAU,CACR;IAAE,MAAM;IAAmB,SAAS;GAAc,GAClD;IAAE,MAAM;IAAiB,SAAS;GAAa,CAChD;EACF;AAED,SAAO;GAAE;GAAS;GAAO;EAAS;CACnC;;;;;;;;CASD,AAAQ,oBAAoBM,YAA+CC,SAAkB;AAC3F,MAAI,WACF,QAAO;GACL,MAAM;GACN,aAAa;IACX,MAAM,WAAW;IACjB,QAAQ,WAAW;IACnB,QAAQ,WAAW,UAAU;IAC7B,aAAa,WAAW;GACzB;EACF;AAGH,SAAO,UAAU,EAAE,MAAM,cAAwB;CAClD;;;;;;CAOD,AAAQ,YACNC,UACAN,WACAO,OACM;AACN,OAAK,SAAS,UAAU,KAAK,QAAS;AAEtC,OAAK,QAAQ;GACX,cAAc,SAAS,MAAM,iBAAiB;GAC9C,kBAAkB,SAAS,MAAM,qBAAqB;GACtD,aAAa,SAAS,MAAM,gBAAgB;GAC5C;GACA;EACD,EAAC;CACH;;;;;;;CAQD,AAAQ,iBAAoBC,SAAiBC,QAAqC;EAChF,IAAIC;AAEJ,MAAI;AACF,YAAS,KAAK,UAAU,QAAQ;EACjC,QAAO;AACN,SAAM,IAAI,SAAS,yBAAyB;EAC7C;AAED,MAAI,WAAW,QAAQ,kBACrB,OAAM,IAAI,SAAS,yBAAyB;AAG9C,OAAK,OACH,QAAO;EAGT,MAAM,SAAS,OAAO,UAAU,OAAO;AAEvC,OAAK,OAAO,QACV,OAAM,IAAI,SAAS,4BAA4B,sBAAyB,OAAO;AAGjF,SAAO,OAAO;CACf;;;;;;CAOD,MAAc,aAAaV,WAAmBW,SAAiBV,QAAsB;EACnF,MAAM,QAAQ,gBAAgB,KAAK,aAAa,QAAQ;AAExD,OAAK,OAAO,MACT,QAAQ,UAAU,qBAAqB,QAAQ,GAAG,KAAK,WAAW,YAAY,MAAM,IACtF;AAED,QAAM,aAAa,OAAO,OAAO;CAClC;;;;;;;;CASD,AAAQ,YAAYE,OAAgBF,QAA+B;AACjE,MAAI,QAAQ,QACV,QAAO;AAGT,MAAI,iBAAiB,aAAa,MAAM,SAAS,WAAW,MAAM,SAAS,cACzE,QAAO;EAGT,MAAM,SAAS,cAAc,MAAM;AAEnC,MAAI,qBAAwB,KAAK,mBAAmB,SAAS,OAAO,CAClE,QAAO;AAGT,SAAO;CACR;;;;;;;CAQD,MAAM,WAAcW,QAAyC;EAC3D,MAAM,SAAS,MAAM,KAAK,MAAM,IAAI,OAAO,SAAS;AAEpD,MAAI,WAAW,KACb,QAAO;AAGT,MAAI;AACF,SAAM,OAAO,gBAAgB;GAE7B,MAAM,SAAS,MAAM,OAAO,IAAI;AAEhC,OAAI;AACF,UAAM,KAAK,MAAM,IAAI,OAAO,UAAU,QAAQ,OAAO,IAAI;GAC1D,SAAQ,OAAO;AACd,SAAK,OAAO,MAAM,gCAAgC,EAChD,SAAS,iBAAiB,QAAQ,MAAM,UAAU,UACnD,EAAC;GACH;AAED,UAAO;EACR,SAAQ,OAAO;AACd,OAAI;AACF,UAAM,OAAO,eAAe;GAC7B,SAAQ,aAAa;AACpB,SAAK,OAAO,MAAM,gCAAgC,EAChD,SAAS,uBAAuB,QAAQ,YAAY,UAAU,UAC/D,EAAC;GACH;AAED,SAAM;EACP;CACF;;;;;;CAOD,MAAM,cACJC,QACY;EACZ,MAAM,EAAE,MAAM,WAAY,GAAG,aAAa,GAAG;AAE7C,SAAO,KAAK,WAAW;GACrB,GAAG;GACH,IAAI,MAAM,KAAK,KAAK,WAAW;EAChC,EAAC;CACH;;;;;CAMD,kBAAkB;AAChB,SAAO,KAAK,SAAS,UAAU;CAChC;AACF;;;;;;;;;;;ACpaD,SAAgB,cAAcC,iBAA6C;AACzE,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAOrE,IAAIC;AACJ,MAAI,OAAO,iBAAiB,SAAS,eAAe;GAClD,MAAM,EAAE,MAAM,QAAQ,GAAG,OAAO,gBAAgB;AAChD,sBAAmB,0GAA0G,KAAK,OAAO,KAAK,UAAU,OAAO,CAAC;EACjK,WAAU,OAAO,iBAAiB,SAAS,cAC1C,mBAAkB;EAOpB,MAAM,SAAS,CAAC,eAAe,SAAS,eAAgB,EAAC,OAAO,QAAQ,CAAC,KAAK,OAAO;EAErF,MAAM,WAAW,MAAM,gBAAgB,SAAS,OAC9C;GACE,OAAO,OAAO;GACd,YAAY,OAAO;GACnB,aAAa,OAAO;GACpB,QAAQ;GACR,UAAU,aAAa,IAAI,CAAC,OAAO;IAAE,MAAM;IAAiB,SAAS,EAAE;GAAS,GAAE;EACnF,GACD,QACD;EAED,MAAM,OAAO,SAAS,QAAQ,KAAK,CAAC,UAAU,MAAM,SAAS,OAAO,EAAE,QAAQ;AAE9E,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,OAAO;IAC/B,mBAAmB,SAAS,OAAO;IACnC,eACG,SAAS,OAAO,gBAAgB,MAAM,SAAS,OAAO,iBAAiB;GAC3E;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;ACtCD,SAAgB,WAAWC,cAAuC;AAChE,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAErE,MAAM,YAAY,QAAQ,OAAO,gBAAgB;EACjD,MAAMC,mBAEF;GACF,aAAa,OAAO;GACpB,iBAAiB,OAAO;EACzB;AAED,MAAI,UACF,kBAAiB,mBAAmB;AAEtC,MAAI,OAAO,iBAAiB,SAAS,cACnC,kBAAiB,iBAAiB,OAAO,gBAAgB,YAAY;EAGvE,MAAM,WAAW,MAAM,aAAa,gBAClC;GACE,OAAO,OAAO;GACd,UAAU,aAAa,IAAI,CAAC,OAAO;IACjC,MAAM;IACN,OAAO,CAAC,EAAE,MAAM,EAAE,QAAS,CAAC;GAC7B,GAAE;GACH,mBAAmB,gBACf,EAAE,OAAO,CAAC,EAAE,MAAM,cAAc,QAAS,CAAC,EAAE;GAEhD;EACD,GACD,QACD;EAED,MAAM,OACJ,SAAS,aAAa,IAAI,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,QAAQ,GAAG,CAAC,KAAK,GAAG,IAAI;AAEjF,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,eAAe;IACvC,mBAAmB,SAAS,eAAe;IAC3C,cAAc,SAAS,eAAe;GACvC;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;;;;ACjDD,SAAgB,YAAYC,eAAiD;AAC3E,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAErE,IAAIC;AACJ,MAAI,OAAO,iBAAiB,SAAS,eAAe;GAClD,MAAM,EAAE,MAAM,QAAQ,GAAG,OAAO,gBAAgB;AAChD,sBAAmB,0GAA0G,KAAK,OAAO,KAAK,UAAU,OAAO,CAAC;EACjK,WAAU,OAAO,iBAAiB,SAAS,cAC1C,mBAAkB;EAGpB,MAAM,cAAc,CAAC,eAAe,SAAS,eAAgB,EAAC,OAAO,CAAC,MACpE,QAAQ,EAAE,CACX;EAED,MAAM,WAAW,MAAM,cAAc,SACnC;GACE,SAAS,OAAO;GAChB,UAAU,aAAa,IAAI,CAAC,OAAO;IACjC,MAAM;IACN,SAAS,CAAC,EAAE,MAAM,EAAE,QAAS,CAAC;GAC/B,GAAE;GACH,QAAQ,YAAY,SAAS,YAAY,IAAI,CAACC,YAAU,EAAE,aAAM,GAAE;GAClE,iBAAiB;IACf,aAAa,OAAO;IACpB,WAAW,OAAO;GACnB;EACF,GACD,QACD;EAED,MAAM,OAAO,SAAS,QAAQ,SAAS,SAAS,IAAI,CAAC,MAAM,EAAE,QAAQ,GAAG,CAAC,KAAK,GAAG,IAAI;AAErF,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,OAAO;IAC/B,mBAAmB,SAAS,OAAO;IACnC,cAAc,SAAS,OAAO;GAC/B;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;AC5DD,SAAgB,UAAUC,QAAuC;AAC/D,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,aAAa,OAAO,QAAQ,aAAa,OAAO,IAAI,OAAO,GAAG,OAAO;EAC3E,MAAM,iBACG,OAAO,YAAY,aAAa,MAAM,OAAO,SAAS,GAAG,OAAO;EAEzE,MAAM,MAAM,MAAM,MAAM,KAAK;GAC3B,QAAQ,OAAO,UAAU;GACzB,SAAS;IAAE,gBAAgB;IAAoB,GAAG;GAAS;GAC3D,MAAM,KAAK,UAAU,OAAO,WAAW,OAAO,CAAC;GAC/C,QAAQ,QAAQ;EACjB,EAAC;AAEF,OAAK,IAAI,IAAI;GACX,MAAM,OAAO,MAAM,IAAI,MAAM,CAAC,MAAM,MAAM,GAAG;GAC7C,MAAM,MAAM,IAAI,OACb,gCAAgC,IAAI,OAAO,KAAK,KAAK,MAAM,GAAG,IAAI,CAAC;AAEtE,OAAI,SAAS,IAAI;AACjB,SAAM;EACP;EAED,MAAM,OAAO,MAAM,IAAI,MAAM;EAC7B,MAAM,EAAE,SAAS,OAAO,GAAG,OAAO,YAAY,KAAK;AAEnD,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,QAAS,EAAE,CAAC;GACnC,OAAO,QACH;IACE,eAAe,MAAM;IACrB,mBAAmB,MAAM;IACzB,cAAc,MAAM;GACrB;EAEN;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;;;;AC/DD,SAAgB,qBAAqBC,QAA4B;AAC/D,QAAO;AACR;;AAGD,MAAa,WAAW;;AAGxB,MAAa,cAAc;;AAG3B,MAAa,eAAe;;AAG5B,MAAa,eAAe;;AAG5B,MAAa,eAAe;;AAG5B,MAAa,gBAAgB;;;;;;;AAQ7B,MAAa,aAAa"}

package/dist/index.js

@@ -111,9 +111,12 @@ function extractStatus(err) {
 * Runs an async function and cancels it if it takes longer than the given
 * timeout. Creates an internal abort controller that fires after the
 * timeout elapses, and combines it with any external signal the caller
-* passed in so either one can cancel the underlying call. The internal
-* timer is always cleared afterward, whether the function succeeds,
-* fails, or is aborted, so nothing is left running in the background
+* passed in so either one can cancel the underlying call. If the internal
+* timeout triggers and the underlying operation aborts, the error is
+* converted into an LLMError with type "timeout". External cancellations
+* continue to propagate as aborted errors. The internal timer is always
+* cleared afterward, whether the function succeeds, fails, or is aborted,
+* so nothing is left running in the background.
 */
 async function withTimeout(fn, timeoutMs, externalSignal) {
 	const controller = new AbortController();
@@ -123,6 +126,9 @@ async function withTimeout(fn, timeoutMs, externalSignal) {
 	const signal = externalSignal ? AbortSignal.any([externalSignal, controller.signal]) : controller.signal;
 	try {
 		return await fn(signal);
+	} catch (err) {
+		if (controller.signal.aborted && !externalSignal?.aborted && err instanceof DOMException && err.name === "AbortError") throw new LLMError("Request timed out", "timeout");
+		throw err;
 	} finally {
 		clearTimeout(timer);
 	}

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","names":["message: string","type: LLMErrorType","status?: number","issues?: unknown","err: unknown","key: string","value: T","ttl: number","options: CircuitBreakerOptions","content: string","err: unknown","fn: (signal: AbortSignal) => Promise<T>","timeoutMs: number","externalSignal?: AbortSignal","baseDelayMs: number","attempt: number","delay: number","signal?: AbortSignal","debugEnabled: boolean","message: string","meta?: Record<string, unknown>","options: VernLLMOptions","params: CallParams<T>","fn: () => Promise<T>","requestId: string","signal?: AbortSignal","lastError: unknown","error: unknown","jsonSchema: CallParams<unknown>['jsonSchema']","useJson: boolean","response: Awaited<ReturnType<LLMClient['chat']['completions']['create']>>","model: string","content: string","schema?: CallParams<T>['schema']","parsed: unknown","attempt: number","params: CachedCallParams<T>","params: Omit<CachedCallParams<T>, 'fn'> & { call: CallParams<T> }","anthropicClient: AnthropicClient","jsonInstruction: string | undefined","geminiClient: GeminiClient","generationConfig: NonNullable<\n            Parameters<GeminiClient['generateContent']>[0]['generationConfig']\n          >","bedrockClient: BedrockConverseClient","jsonInstruction: string | undefined","text","config: FetchAdapterConfig","client: unknown"],"sources":["../src/types.ts","../src/circuitBreaker.ts","../src/internal/vernLLM.utilts.ts","../src/logger.ts","../src/vernLLM.ts","../src/adapters/anthropic.ts","../src/adapters/gemini.ts","../src/adapters/bedrock.ts","../src/adapters/fetch.ts","../src/adapters/openaiCompatible.ts"],"sourcesContent":["export type LLMErrorType =\n  | 'timeout'\n  | 'api'\n  | 'parse'\n  | 'validation'\n  | 'circuit_open'\n  | 'unknown'\n  | 'aborted';\n\nexport class LLMError extends Error {\n  constructor(\n    message: string,\n    public type: LLMErrorType,\n    public status?: number,\n    public issues?: unknown,\n  ) {\n    super(message);\n    this.name = 'LLMError';\n  }\n}\n\nexport function isLLMError(err: unknown): err is LLMError {\n  return err instanceof LLMError;\n}\n\nexport interface CacheAdapter<T = unknown> {\n  get(key: string): Promise<T | null>;\n  set(key: string, value: T, ttl: number): Promise<void>;\n}\n\n/**\n * Trivial default so the package works out of the box with no external deps\n * Not shared across processes, swap in Redis/Upstash/etc for production\n */\nexport class InMemoryCacheAdapter<T = unknown> implements CacheAdapter<T> {\n  private store = new Map<string, { value: T; expiresAt: number }>();\n\n  async get(key: string): Promise<T | null> {\n    const entry = this.store.get(key);\n    if (!entry) return null;\n    if (Date.now() > entry.expiresAt) {\n      this.store.delete(key);\n      return null;\n    }\n    return entry.value;\n  }\n\n  async set(key: string, value: T, ttl: number): Promise<void> {\n    this.store.set(key, { value, expiresAt: Date.now() + ttl * 1000 });\n  }\n}\n\n/**\n * Minimal shape compatible with the OpenAI SDKs chat.completions.create,\n * so consumers can pass an OpenAI client directly\n * `response_format.json_schema` and `reasoning_effort` are optional on the wire\n * providers that don't support them will just ignore fields they don't recognize,\n * but not every SDKs TS types accept them, hence this being a structural type\n * rather than importing the SDKs own params type\n */\nexport interface LLMClient {\n  chat: {\n    completions: {\n      create(\n        params: {\n          model: string;\n          temperature: number;\n          max_tokens: number;\n          response_format?:\n            | { type: 'json_object' }\n            | {\n                type: 'json_schema';\n                json_schema: {\n                  name: string;\n                  schema: Record<string, unknown>;\n                  strict?: boolean;\n                  description?: string;\n                };\n              };\n          /** OpenAI reasoning-model param (o-series, gpt-5), ignored by providers that don't support it */\n          reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n          messages: Array<{ role: 'system' | 'user'; content: string }>;\n        },\n        options: { signal: AbortSignal },\n      ): Promise<{\n        choices?: Array<{ message?: { content?: string | null } }>;\n        usage?: {\n          prompt_tokens?: number;\n          completion_tokens?: number;\n          total_tokens?: number;\n        };\n      }>;\n    };\n  };\n}\n\nexport type ReserveUsage = () => Promise<void>;\nexport type RefundUsage = () => Promise<void>;\n\nexport interface TokenUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  requestId: string;\n  model: string;\n}\n\nexport type OnUsage = (usage: TokenUsage) => void;\n\n/**\n * Minimal structural type for a Zod-like schema, so this package doesnt need\n * a hard dependency on a specific Zod major version. Any object exposing\n * `safeParse` (Zod v3/v4, and most Zod-compatible validators) should satisfy this\n */\nexport interface SchemaLike<T> {\n  safeParse(data: unknown): { success: true; data: T } | { success: false; error: unknown };\n}\n\nexport interface VernLLMOptions {\n  client: LLMClient;\n  model: string;\n  /** Max retries after the first attempt. Default 1 (2 attempts total) */\n  maxRetries?: number;\n  /** Per-attempt timeout in ms. Default 25000 */\n  timeoutMs?: number;\n  /** Base delay for exponential backoff in ms. Default 500 */\n  baseDelayMs?: number;\n  /** Default max_tokens for calls that don't override it. Default 1000 */\n  defaultMaxTokens?: number;\n  /** Enables debug logging of raw model output. Default: NODE_ENV !== 'production' */\n  debug?: boolean;\n  /** Cache adapter for cachedCall. Defaults to an in-memory adapter */\n  cache?: CacheAdapter;\n  /** HTTP status codes that should fail fast without retrying. Default [400, 401, 403] */\n  nonRetryableStatus?: number[];\n  /** Custom JSON parser. Must return undefined/null on failure. Default: JSON.parse wrapped in try/catch */\n  parseJson?: (content: string) => unknown;\n  /** Called after every successful call with token usage, if the provider reports it */\n  onUsage?: OnUsage;\n  /** Injectable logger. Defaults to a console-based logger gated by `debug` */\n  logger?: import('./logger.js').Logger;\n  /**\n   * Enables a circuit breaker that short-circuits calls after repeated\n   * consecutive failures, instead of continuing to hammer a down provider\n   * Pass `true` for defaults, or an options object to tune threshold/cooldown\n   */\n  circuitBreaker?: boolean | import('./circuitBreaker.js').CircuitBreakerOptions;\n}\n\n/**\n * A provider-native JSON Schema for structured outputs (OpenAI/Groq\n * `response_format: { type: 'json_schema' }`) This is the wire-format\n * schema the model is constrained to generate against, distinct from\n * `schema`, which is a client-side Zod validator run on the parsed result\n * You can use one, both, or neither; using both gets you provider-level\n * constraint plus client-side type inference/validation as a safety net\n */\nexport interface JsonSchemaSpec {\n  name: string;\n  schema: Record<string, unknown>;\n  /** Enforces the schema strictly (OpenAI-specific), default true when supported */\n  strict?: boolean;\n  description?: string;\n}\n\nexport interface CallParams<T = unknown> {\n  systemPrompt: string;\n  userContent: string;\n  temperature?: number;\n  jsonMode?: boolean;\n  maxTokens?: number;\n  requestId?: string;\n  signal?: AbortSignal;\n  /** Overrides the model set on the VernLLM instance for this call only */\n  model?: string;\n  /**\n   * OpenAI-style reasoning effort for reasoning models (o-series, gpt-5, etc).\n   * Passed through as-is, providers/models that don't support it ignore it\n   */\n  reasoningEffort?: 'minimal' | 'low' | 'medium' | 'high';\n  /**\n   * Provider-native JSON Schema structured-output mode. When set, this is sent\n   * as `response_format: { type: 'json_schema', json_schema: ... }` instead of\n   * the looser `json_object` mode, constraining the models output shape at\n   * generation time (not just validating it after the fact). Implies jsonMode: true.\n   */\n  jsonSchema?: JsonSchemaSpec;\n  /**\n   * Optional Zod (or Zod-compatible) schema. When provided, the parsed JSON\n   * is validated against it; on failure an LLMError('validation') is thrown\n   * with `.issues` set to the schema's error object. Implies jsonMode: true.\n   * Can be combined with `jsonSchema` for provider-level constraint + client-side typing.\n   */\n  schema?: SchemaLike<T>;\n}\n\nexport interface CachedCallParams<T> {\n  cacheKey: string;\n  ttl: number;\n  fn: () => Promise<T>;\n  reserveUsage?: ReserveUsage;\n  refundUsage?: RefundUsage;\n}\n","import { LLMError } from './types.js';\n\nexport interface CircuitBreakerOptions {\n  /** Consecutive failures before the circuit opens, default 5 */\n  threshold?: number;\n  /** How long the circuit stays open before allowing a trial request, in ms. Default 30000 */\n  cooldownMs?: number;\n}\n\ntype CircuitState = 'closed' | 'open' | 'half-open';\n\n/**\n * Per retry VernLLM-instance circuit breaker. Tracks consecutive failures across\n * calls. Once the threshold is hit, short-circuits new calls with an\n * LLMError('circuit_open') instead of hitting the provider, until the\n * cooldown elapses and a single trial call is allowed through\n */\nexport class CircuitBreaker {\n  private state: CircuitState = 'closed';\n  private consecutiveFailures = 0;\n  private openedAt = 0;\n  private threshold: number;\n  private cooldownMs: number;\n\n  constructor(options: CircuitBreakerOptions = {}) {\n    this.threshold = options.threshold ?? 5;\n    this.cooldownMs = options.cooldownMs ?? 30_000;\n  }\n\n  /** Throws if the circuit is open and the cooldown hasnt elapsed */\n  assertClosed(): void {\n    if (this.state !== 'open') return;\n\n    const elapsed = Date.now() - this.openedAt;\n    if (elapsed >= this.cooldownMs) {\n      this.state = 'half-open';\n      return;\n    }\n\n    throw new LLMError(\n      `Circuit open — provider has failed ${this.consecutiveFailures} times in a row. Retry in ${Math.ceil((this.cooldownMs - elapsed) / 1000)}s.`,\n      'circuit_open',\n    );\n  }\n\n  recordSuccess(): void {\n    this.consecutiveFailures = 0;\n    this.state = 'closed';\n  }\n\n  recordFailure(): void {\n    this.consecutiveFailures += 1;\n\n    if (this.state === 'half-open') {\n      // Trial call failed: reopen and reset the cooldown window.\n      this.state = 'open';\n      this.openedAt = Date.now();\n      return;\n    }\n\n    if (this.consecutiveFailures >= this.threshold) {\n      this.state = 'open';\n      this.openedAt = Date.now();\n    }\n  }\n\n  getState(): CircuitState {\n    return this.state;\n  }\n}\n","import { LLMError } from '../types.js';\n\nexport function defaultParseJson(content: string): unknown {\n  try {\n    return JSON.parse(content);\n  } catch {\n    return undefined;\n  }\n}\n\n/**\n * Looks inside an unknown error value and pulls out an http status code\n * if one is present. Checks the status field first then the status code\n * field since different client libraries use different names for this.\n * Returns undefined when the error is not an object or carries no status\n */\nexport function extractStatus(err: unknown): number | undefined {\n  if (!err || typeof err !== 'object') return undefined;\n\n  const error = err as {\n    status?: unknown;\n    statusCode?: unknown;\n  };\n\n  if (typeof error.status === 'number') return error.status;\n  if (typeof error.statusCode === 'number') return error.statusCode;\n\n  return undefined;\n}\n\n/**\n * Runs an async function and cancels it if it takes longer than the given\n * timeout. Creates an internal abort controller that fires after the\n * timeout elapses, and combines it with any external signal the caller\n * passed in so either one can cancel the underlying call. The internal\n * timer is always cleared afterward, whether the function succeeds,\n * fails, or is aborted, so nothing is left running in the background\n */\nexport async function withTimeout<T>(\n  fn: (signal: AbortSignal) => Promise<T>,\n  timeoutMs: number,\n  externalSignal?: AbortSignal,\n): Promise<T> {\n  const controller = new AbortController();\n\n  const timer = setTimeout(() => {\n    controller.abort();\n  }, timeoutMs);\n\n  const signal = externalSignal\n    ? AbortSignal.any([externalSignal, controller.signal])\n    : controller.signal;\n\n  try {\n    return await fn(signal);\n  } finally {\n    clearTimeout(timer);\n  }\n}\n\n/**\n * Exponential backoff with jitter, capped at maxDelayMs.\n * Jitter avoids thundering-herd retries when many callers back off in lockstep,\n * the cap prevents unbounded delays when maxRetries is high\n */\nexport function getBackoffDelay(baseDelayMs: number, attempt: number, maxDelayMs = 10_000): number {\n  const exp = Math.min(baseDelayMs * 2 ** attempt, maxDelayMs);\n  return exp / 2 + Math.random() * (exp / 2);\n}\n\n/**\n * Pauses execution for the given delay before a retry attempt. If an\n * abort signal is provided and it fires while waiting, the pending\n * timer is cancelled immediately and the wait rejects right away with\n * an aborted error instead of continuing to sit idle until the delay\n * would have finished on its own\n */\nexport async function waitForRetry(delay: number, signal?: AbortSignal): Promise<void> {\n  await new Promise<void>((resolve, reject) => {\n    const onAbort = () => {\n      clearTimeout(timer);\n      reject(new LLMError('Operation aborted', 'aborted'));\n    };\n\n    const timer = setTimeout(() => {\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    }, delay);\n\n    signal?.addEventListener('abort', onAbort, { once: true });\n  });\n}\n","export interface Logger {\n  debug(message: string): void;\n  warn(message: string): void;\n  error(message: string, meta?: Record<string, unknown>): void;\n}\n\n/**\n * Default logger. `debug` is gated by the `debug` option on VernLLM\n * warn/error always fire since they indicate real problems (retries, cache failures)\n */\nexport class ConsoleLogger implements Logger {\n  constructor(private debugEnabled: boolean) {}\n\n  debug(message: string): void {\n    if (this.debugEnabled) console.debug(message);\n  }\n\n  warn(message: string): void {\n    console.warn(message);\n  }\n\n  error(message: string, meta?: Record<string, unknown>): void {\n    console.error(message, meta ?? '');\n  }\n}\n","import { randomUUID } from 'crypto';\n\nimport { CircuitBreaker } from './circuitBreaker.js';\nimport {\n  defaultParseJson,\n  extractStatus,\n  withTimeout,\n  getBackoffDelay,\n  waitForRetry,\n} from './internal/vernLLM.utilts.js';\nimport { ConsoleLogger, type Logger } from './logger.js';\nimport {\n  LLMError,\n  InMemoryCacheAdapter,\n  type VernLLMOptions,\n  type CallParams,\n  type CachedCallParams,\n  type CacheAdapter,\n  type LLMClient,\n} from './types.js';\n\n/**\n * A resilient wrapper around an LLM chat completions client, this is VernLLM!\n *\n * Adds retry with exponential backoff and jitter, per-attempt timeouts,\n * an optional circuit breaker, JSON parsing with optional schema\n * validation, usage tracking, and an optional response cache, all\n * configurable, all opt-in beyond sensible defaults\n */\nexport class VernLLM {\n  private readonly client: LLMClient;\n  private readonly model: string;\n\n  private readonly maxRetries: number;\n  private readonly timeoutMs: number;\n  private readonly baseDelayMs: number;\n  private readonly defaultMaxTokens: number;\n\n  private readonly cache: CacheAdapter<unknown>;\n  private readonly nonRetryableStatus: number[];\n\n  private readonly parseJson: (content: string) => unknown;\n  private readonly onUsage?: VernLLMOptions['onUsage'];\n\n  private readonly logger: Logger;\n  private readonly breaker?: CircuitBreaker;\n\n  /**\n   * @param options: Client, model, and all tunables (retries, timeout,\n   * backoff, cache, circuit breaker, logger, etc). See VernLLMOptions in `types.ts`\n   * for individual defaults\n   */\n  constructor(options: VernLLMOptions) {\n    this.client = options.client;\n    this.model = options.model;\n\n    const retryConfig = this.resolveRetryConfig(options);\n    this.maxRetries = retryConfig.maxRetries;\n    this.timeoutMs = retryConfig.timeoutMs;\n    this.baseDelayMs = retryConfig.baseDelayMs;\n    this.defaultMaxTokens = retryConfig.defaultMaxTokens;\n\n    this.cache = options.cache ?? new InMemoryCacheAdapter();\n    this.nonRetryableStatus = options.nonRetryableStatus ?? [400, 401, 403];\n\n    this.parseJson = options.parseJson ?? defaultParseJson;\n    this.onUsage = options.onUsage;\n\n    this.logger = this.resolveLogger(options);\n    this.breaker = this.resolveCircuitBreaker(options);\n  }\n\n  /**\n   * Resolves retry/timeout/token defaults from the given options,\n   * falling back to the librarys built-in defaults for anything unset\n   */\n  private resolveRetryConfig(options: VernLLMOptions) {\n    return {\n      maxRetries: options.maxRetries ?? 1,\n      timeoutMs: options.timeoutMs ?? 25_000,\n      baseDelayMs: options.baseDelayMs ?? 500,\n      defaultMaxTokens: options.defaultMaxTokens ?? 1000,\n    };\n  }\n\n  /**\n   * Returns the caller supplied logger, or a console-based logger whose\n   * debug output is gated by the `debug` option (defaulting to on\n   * outside production)\n   */\n  private resolveLogger(options: VernLLMOptions): Logger {\n    return (\n      options.logger ?? new ConsoleLogger(options.debug ?? process.env.NODE_ENV !== 'production')\n    );\n  }\n\n  /**\n   * Builds a circuit breaker if `circuitBreaker` is truthy on the\n   * options. Passing `true` uses default thresholds, passing an options\n   * object tunes them. Returns undefined when the breaker is disabled\n   */\n  private resolveCircuitBreaker(options: VernLLMOptions): CircuitBreaker | undefined {\n    if (!options.circuitBreaker) return undefined;\n\n    return new CircuitBreaker(options.circuitBreaker === true ? undefined : options.circuitBreaker);\n  }\n\n  /**\n   * Makes a single logical LLM call, transparently retrying on failure\n   * according to the configured retry policy\n   *\n   * Fails fast if the circuit breaker is open or the signal is already\n   * aborted, before any request is dispatched. On exhausting all\n   * retries, records a circuit breaker failure and rejects with a\n   * normalized LLMError\n   *\n   * @param params : System/user content plus per call overrides\n   * (model, temperature, jsonMode, schema, signal, etc)\n   * @returns The parsed and optionally schema-validated response, or\n   * the raw string content when jsonMode is disabled\n   */\n  async call<T = unknown>(params: CallParams<T>): Promise<T> {\n    this.breaker?.assertClosed();\n\n    if (params.signal?.aborted) {\n      throw new LLMError('LLM request aborted', 'aborted');\n    }\n\n    const requestId = params.requestId ?? randomUUID();\n\n    try {\n      const result = await this.retryWithBackoff(\n        () => this.executeCall(params, requestId),\n        requestId,\n        params.signal,\n      );\n\n      return result;\n    } catch (error) {\n      this.breaker?.recordFailure();\n      throw this.normalizeError(error, params.signal);\n    }\n  }\n\n  /**\n   * Runs `fn`, retrying with backoff according to `shouldRetry` policy\n   * Purely mechanical: knows nothing about LLM specifics beyond the retry\n   * predicate, so its testable independent of request/response shaping\n   */\n  private async retryWithBackoff<T>(\n    fn: () => Promise<T>,\n    requestId: string,\n    signal?: AbortSignal,\n  ): Promise<T> {\n    let lastError: unknown;\n\n    for (let attempt = 0; attempt <= this.maxRetries; attempt++) {\n      try {\n        if (attempt > 0) {\n          await this.recoverDelay(requestId, attempt, signal);\n        }\n\n        return await fn();\n      } catch (error) {\n        lastError = error;\n\n        if (!this.shouldRetry(error, signal)) {\n          break;\n        }\n      }\n    }\n\n    throw lastError;\n  }\n\n  /**\n   * Converts any thrown value into a well-typed LLMError for the public\n   * API surface. Preserves an existing LLMError as is, reports aborted\n   * signals as such, classifies errors carrying an http status as\n   * type api, and otherwise falls back to a generic unknown error.\n   */\n  private normalizeError(error: unknown, signal?: AbortSignal): LLMError {\n    if (signal?.aborted) {\n      return new LLMError('LLM request aborted', 'aborted');\n    }\n\n    if (error instanceof LLMError) {\n      return error;\n    }\n\n    const status = extractStatus(error);\n\n    if (status !== undefined) {\n      return new LLMError('LLM request failed', 'api', status);\n    }\n\n    return new LLMError('LLM request failed', 'unknown');\n  }\n\n  /**\n   * Performs a single attempt: builds the request, dispatches it with a\n   * timeout, and shapes the response. Throws on an empty response so\n   * the retry loop treats it like any other transient failure. Records\n   * usage and a circuit breaker success before returning\n   */\n  private async executeCall<T>(params: CallParams<T>, requestId: string): Promise<T> {\n    const { useJson, model, request } = this.buildRequestPayload(params);\n\n    const response = await withTimeout(\n      (attemptSignal) => this.client.chat.completions.create(request, { signal: attemptSignal }),\n      this.timeoutMs,\n      params.signal,\n    );\n\n    const content = response.choices?.[0]?.message?.content?.trim();\n\n    if (!content) {\n      throw new LLMError('Empty LLM response', 'api');\n    }\n\n    this.logger.debug(`[vern:${requestId}] output:\\n${content.slice(0, 800)}`);\n\n    this.recordUsage(response, requestId, model);\n\n    this.breaker?.recordSuccess();\n\n    if (!useJson) {\n      return content as T;\n    }\n\n    return this.parseAndValidate(content, params.schema);\n  }\n\n  /**\n   * Applies per call defaults and shapes the params into the request\n   * object expected by the underlying client, including the resolved\n   * response format. Also returns whether JSON parsing should be\n   * applied to the response and which model was ultimately used\n   */\n  private buildRequestPayload<T>(params: CallParams<T>) {\n    const {\n      systemPrompt,\n      userContent,\n      temperature = 0.2,\n      jsonMode = true,\n      maxTokens = this.defaultMaxTokens,\n      model = this.model,\n      reasoningEffort,\n      jsonSchema,\n    } = params;\n\n    const useJson = jsonMode || Boolean(jsonSchema);\n    const responseFormat = this.buildResponseFormat(jsonSchema, useJson);\n\n    const request = {\n      model,\n      temperature,\n      max_tokens: maxTokens,\n      ...(responseFormat ? { response_format: responseFormat } : {}),\n      ...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}),\n      messages: [\n        { role: 'system' as const, content: systemPrompt },\n        { role: 'user' as const, content: userContent },\n      ],\n    };\n\n    return { useJson, model, request };\n  }\n\n  /**\n   * Chooses the response format to send to the provider. A provider\n   * native json schema takes priority when supplied, constraining\n   * generation directly, otherwise falls back to the looser json\n   * object mode when JSON output is requested, or no format at all\n   * for plain text responses\n   */\n  private buildResponseFormat(jsonSchema: CallParams<unknown>['jsonSchema'], useJson: boolean) {\n    if (jsonSchema) {\n      return {\n        type: 'json_schema' as const,\n        json_schema: {\n          name: jsonSchema.name,\n          schema: jsonSchema.schema,\n          strict: jsonSchema.strict ?? true,\n          description: jsonSchema.description,\n        },\n      };\n    }\n\n    return useJson ? { type: 'json_object' as const } : undefined;\n  }\n\n  /**\n   * Reports token usage to the caller supplied onUsage callback, when\n   * both a callback was configured and the provider actually returned\n   * usage data on this response. A no op otherwise.\n   */\n  private recordUsage(\n    response: Awaited<ReturnType<LLMClient['chat']['completions']['create']>>,\n    requestId: string,\n    model: string,\n  ): void {\n    if (!response.usage || !this.onUsage) return;\n\n    this.onUsage({\n      promptTokens: response.usage.prompt_tokens ?? 0,\n      completionTokens: response.usage.completion_tokens ?? 0,\n      totalTokens: response.usage.total_tokens ?? 0,\n      requestId,\n      model,\n    });\n  }\n\n  /**\n   * Parses the raw response content as JSON and, when a schema is\n   * supplied, validates the parsed value against it. Throws a parse\n   * type LLMError on malformed JSON and a validation type LLMError,\n   * carrying the schemas issues, on a failed validation\n   */\n  private parseAndValidate<T>(content: string, schema?: CallParams<T>['schema']): T {\n    let parsed: unknown;\n\n    try {\n      parsed = this.parseJson(content);\n    } catch {\n      throw new LLMError('Invalid JSON response', 'parse');\n    }\n\n    if (parsed === null || parsed === undefined) {\n      throw new LLMError('Invalid JSON response', 'parse');\n    }\n\n    if (!schema) {\n      return parsed as T;\n    }\n\n    const result = schema.safeParse(parsed);\n\n    if (!result.success) {\n      throw new LLMError('Schema validation failed', 'validation', undefined, result.error);\n    }\n\n    return result.data;\n  }\n\n  /**\n   * Waits out the backoff delay for a given retry attempt, logging the\n   * attempt for observability before the wait begins. Rejects early if\n   * the signal aborts during the wait\n   */\n  private async recoverDelay(requestId: string, attempt: number, signal?: AbortSignal) {\n    const delay = getBackoffDelay(this.baseDelayMs, attempt);\n\n    this.logger.warn(\n      `[vern:${requestId}] recovery attempt ${attempt}/${this.maxRetries}, waiting ${delay}ms`,\n    );\n\n    await waitForRetry(delay, signal);\n  }\n\n  /**\n   * Decides whether a failed attempt is worth retrying. Never retries\n   * once the signal has aborted, never retries a parse or validation\n   * failure since those stem from the response content rather than a\n   * transient fault, and never retries a status code the caller has\n   * marked as non retryable. Retries everything else\n   */\n  private shouldRetry(error: unknown, signal?: AbortSignal): boolean {\n    if (signal?.aborted) {\n      return false;\n    }\n\n    if (error instanceof LLMError && (error.type === 'parse' || error.type === 'validation')) {\n      return false;\n    }\n\n    const status = extractStatus(error);\n\n    if (status !== undefined && this.nonRetryableStatus.includes(status)) {\n      return false;\n    }\n\n    return true;\n  }\n\n  /**\n   * Thin cache wrapper around caller supplied logic. `params.fn` is expected\n   * to be a call that itself invokes `this.call(...)` (see `cachedLLMCall`\n   * below for a convenience wrapper that wires this up automatically),\n   * `cachedCall` does not itself apply retry/timeout policy.\n   */\n  async cachedCall<T>(params: CachedCallParams<T>): Promise<T> {\n    const cached = await this.cache.get(params.cacheKey);\n\n    if (cached !== null) {\n      return cached as T;\n    }\n\n    try {\n      await params.reserveUsage?.();\n\n      const result = await params.fn();\n\n      try {\n        await this.cache.set(params.cacheKey, result, params.ttl);\n      } catch (error) {\n        this.logger.error('[VernLLM] cache write failed', {\n          message: error instanceof Error ? error.message : 'unknown',\n        });\n      }\n\n      return result;\n    } catch (error) {\n      try {\n        await params.refundUsage?.();\n      } catch (refundError) {\n        this.logger.error('[VernLLM] refundUsage failed', {\n          message: refundError instanceof Error ? refundError.message : 'unknown',\n        });\n      }\n\n      throw error;\n    }\n  }\n\n  /**\n   * Convenience wrapper composing `call` + `cachedCall`, so cached LLM calls\n   * automatically get retry/timeout/circuit-breaker behavior without callers\n   * having to remember to wire `fn: () => this.call(...)` themselves\n   */\n  async cachedLLMCall<T>(\n    params: Omit<CachedCallParams<T>, 'fn'> & { call: CallParams<T> },\n  ): Promise<T> {\n    const { call: callParams, ...cacheParams } = params;\n\n    return this.cachedCall({\n      ...cacheParams,\n      fn: () => this.call(callParams),\n    });\n  }\n\n  /**\n   * Returns the current circuit breaker state, or undefined when no\n   * circuit breaker was configured on this instance\n   */\n  getCircuitState() {\n    return this.breaker?.getState();\n  }\n}\n","import type { LLMClient } from '../types.js';\n\n/** Minimal structural type for the Anthropic SDKs `messages.create` */\nexport interface AnthropicClient {\n  messages: {\n    create(\n      params: {\n        model: string;\n        max_tokens: number;\n        temperature?: number;\n        system?: string;\n        messages: Array<{ role: 'user' | 'assistant'; content: string }>;\n      },\n      options: { signal: AbortSignal },\n    ): Promise<{\n      content: Array<{ type: string; text?: string }>;\n      usage?: { input_tokens?: number; output_tokens?: number };\n    }>;\n  };\n}\n\n/**\n * Wraps an Anthropic SDK client so it satisfies the same `LLMClient`\n * interface VernLLM uses for OpenAI/Groq. Anthropics Messages API has no\n * `response_format: json_object` equivalent, so when the caller requests\n * JSON mode, this adapter appends an instruction to the system prompt\n * asking the model to respond with JSON only\n */\nexport function fromAnthropic(anthropicClient: AnthropicClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          // Anthropics Messages API has no `response_format: json_object` or\n          // `json_schema` equivalent, so both are emulated via system prompt\n          // instructions. For json_schema, the schema itself is embedded so the\n          // model has something concrete to conform to (not provider-enforced,\n          // unlike OpenAIs native structured outputs)\n          let jsonInstruction: string | undefined;\n          if (params.response_format?.type === 'json_schema') {\n            const { name, schema } = params.response_format.json_schema;\n            jsonInstruction = `Respond with valid JSON only, no prose or markdown fences. The JSON must conform to this schema (name: \"${name}\"):\\n${JSON.stringify(schema)}`;\n          } else if (params.response_format?.type === 'json_object') {\n            jsonInstruction = 'Respond with valid JSON only, no prose or markdown fences.';\n          }\n\n          // `reasoning_effort` (OpenAI o-series/gpt-5 style) has no direct Anthropic\n          // equivalent, Claudes extended thinking uses a token budget, not a tier\n          // string, so its intentionally dropped here rather than guessed at.\n\n          const system = [systemMessage?.content, jsonInstruction].filter(Boolean).join('\\n\\n');\n\n          const response = await anthropicClient.messages.create(\n            {\n              model: params.model,\n              max_tokens: params.max_tokens,\n              temperature: params.temperature,\n              system: system || undefined,\n              messages: userMessages.map((m) => ({ role: 'user' as const, content: m.content })),\n            },\n            options,\n          );\n\n          const text = response.content.find((block) => block.type === 'text')?.text ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usage?.input_tokens,\n              completion_tokens: response.usage?.output_tokens,\n              total_tokens:\n                (response.usage?.input_tokens ?? 0) + (response.usage?.output_tokens ?? 0),\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Minimal structural type for Geminis `generateContent`, matching both the\n * legacy `@google/generative-ai` SDKs `model.generateContent(...)` and the\n * newer `@google/genai` SDKs `ai.models.generateContent({ model, ... })`\n * closely enough to adapt either — pass whichever `.generateContent` you have.\n */\nexport interface GeminiClient {\n  generateContent(\n    params: {\n      model?: string;\n      contents: Array<{ role: 'user'; parts: Array<{ text: string }> }>;\n      systemInstruction?: { parts: Array<{ text: string }> };\n      generationConfig?: {\n        temperature?: number;\n        maxOutputTokens?: number;\n        responseMimeType?: string;\n        responseSchema?: Record<string, unknown>;\n      };\n    },\n    options: { signal: AbortSignal },\n  ): Promise<{\n    candidates?: Array<{ content?: { parts?: Array<{ text?: string }> } }>;\n    usageMetadata?: {\n      promptTokenCount?: number;\n      candidatesTokenCount?: number;\n      totalTokenCount?: number;\n    };\n  }>;\n}\n\n/**\n * Wraps a Gemini client so it satisfies the `LLMClient` interface VernLLM\n * uses for OpenAI/Groq. Geminis shape differs on nearly every axis: a\n * `contents` array instead of `messages`, a separate `systemInstruction`\n * field instead of a `system` role message, `generationConfig` instead of\n * top-level `temperature`/`max_tokens`, and native JSON Schema support via\n * `responseMimeType: 'application/json'` + `responseSchema` (so `jsonSchema`\n * is provider-enforced here, unlike the Anthropic adapters prompt-embedding\n * fallback). `reasoning_effort` has no equivalent. Geminis thinking models\n * use a token budget, not an effort tier, so its dropped, same as Anthropic.\n */\nexport function fromGemini(geminiClient: GeminiClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          const wantsJson = Boolean(params.response_format);\n          const generationConfig: NonNullable<\n            Parameters<GeminiClient['generateContent']>[0]['generationConfig']\n          > = {\n            temperature: params.temperature,\n            maxOutputTokens: params.max_tokens,\n          };\n\n          if (wantsJson) {\n            generationConfig.responseMimeType = 'application/json';\n          }\n          if (params.response_format?.type === 'json_schema') {\n            generationConfig.responseSchema = params.response_format.json_schema.schema;\n          }\n\n          const response = await geminiClient.generateContent(\n            {\n              model: params.model,\n              contents: userMessages.map((m) => ({\n                role: 'user' as const,\n                parts: [{ text: m.content }],\n              })),\n              systemInstruction: systemMessage\n                ? { parts: [{ text: systemMessage.content }] }\n                : undefined,\n              generationConfig,\n            },\n            options,\n          );\n\n          const text =\n            response.candidates?.[0]?.content?.parts?.map((p) => p.text ?? '').join('') ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usageMetadata?.promptTokenCount,\n              completion_tokens: response.usageMetadata?.candidatesTokenCount,\n              total_tokens: response.usageMetadata?.totalTokenCount,\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Minimal structural type matching AWS Bedrocks Converse API. This is\n * intentionally NOT `BedrockRuntimeClient` itself, the AWS SDK v3 client\n * exposes `.send(command)`, not a direct `.converse()` method, and pulling\n * in `@aws-sdk/client-bedrock-runtime` as a dependency just for its types\n * isn't worth it for a structural adapter. Wrap your client, e.g:\n *\n * ```ts\n * import { BedrockRuntimeClient, ConverseCommand } from '@aws-sdk/client-bedrock-runtime';\n * const client = new BedrockRuntimeClient({ region: 'us-east-1' });\n * const converseClient = {\n *   converse: (params, options) =>\n *     client.send(new ConverseCommand(params), { abortSignal: options.signal }),\n * };\n * ```\n */\nexport interface BedrockConverseClient {\n  converse(\n    params: {\n      modelId: string;\n      messages: Array<{ role: 'user'; content: Array<{ text: string }> }>;\n      system?: Array<{ text: string }>;\n      inferenceConfig?: { temperature?: number; maxTokens?: number };\n    },\n    options: { signal: AbortSignal },\n  ): Promise<{\n    output?: { message?: { content?: Array<{ text?: string }> } };\n    usage?: { inputTokens?: number; outputTokens?: number; totalTokens?: number };\n  }>;\n}\n\n/**\n * Wraps a Bedrock Converse-API client so it satisfies the `LLMClient`\n * interface VernLLM uses for OpenAI/Groq. The Converse API is unified\n * across Bedrocks model families (Anthropic, Titan, Llama, Mistral, etc.),\n * so unlike raw per-model Bedrock invocation, this one adapter works\n * regardless of which underlying model `modelId` points at, as long as\n * that model supports Converse (most current-generation ones do)\n *\n * Theres no uniform native JSON Schema enforcement across families here\n * (some support it via forced tool-use, which varies per model), so\n * `jsonSchema`/`jsonMode` are emulated via a system-prompt instruction, same\n * approach as the Anthropic adapter. `reasoning_effort` has no Converse\n * equivalent and is dropped\n */\nexport function fromBedrock(bedrockClient: BedrockConverseClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          let jsonInstruction: string | undefined;\n          if (params.response_format?.type === 'json_schema') {\n            const { name, schema } = params.response_format.json_schema;\n            jsonInstruction = `Respond with valid JSON only, no prose or markdown fences. The JSON must conform to this schema (name: \"${name}\"):\\n${JSON.stringify(schema)}`;\n          } else if (params.response_format?.type === 'json_object') {\n            jsonInstruction = 'Respond with valid JSON only, no prose or markdown fences.';\n          }\n\n          const systemParts = [systemMessage?.content, jsonInstruction].filter((s): s is string =>\n            Boolean(s),\n          );\n\n          const response = await bedrockClient.converse(\n            {\n              modelId: params.model,\n              messages: userMessages.map((m) => ({\n                role: 'user' as const,\n                content: [{ text: m.content }],\n              })),\n              system: systemParts.length ? systemParts.map((text) => ({ text })) : undefined,\n              inferenceConfig: {\n                temperature: params.temperature,\n                maxTokens: params.max_tokens,\n              },\n            },\n            options,\n          );\n\n          const text = response.output?.message?.content?.map((c) => c.text ?? '').join('') ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usage?.inputTokens,\n              completion_tokens: response.usage?.outputTokens,\n              total_tokens: response.usage?.totalTokens,\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/** The chat-completion-shaped request VernLLM builds internally */\ntype ChatRequest = Parameters<LLMClient['chat']['completions']['create']>[0];\n\nexport interface FetchAdapterConfig {\n  /** Endpoint URL, or a function of the request in case it depends on model/params */\n  url: string | ((params: ChatRequest) => string);\n  /** Static headers, or a function (sync or async) for things like refreshed auth tokens */\n  headers?:\n    | Record<string, string>\n    | (() => Record<string, string> | Promise<Record<string, string>>);\n  /** HTTP method. Default 'POST' */\n  method?: string;\n  /** Maps VernLLMs internal chat-completion request into the providers raw request body */\n  mapRequest: (params: ChatRequest) => unknown;\n  /**\n   * Maps the providers raw JSON response into `{ content, usage? }`\n   * `content` is the assistants text (JSON string when JSON mode was requested)\n   */\n  mapResponse: (json: unknown) => {\n    content: string;\n    usage?: { promptTokens?: number; completionTokens?: number; totalTokens?: number };\n  };\n}\n\n/**\n * A fetch-based escape hatch for providers with no SDK, or where pulling one\n * in isnt worth it. You supply the URL, headers, and two small mapping\n * functions; this handles the HTTP call and slots the result into the same\n * `LLMClient` shape every other adapter produces, so retries, timeouts,\n * the circuit breaker, and JSON/schema handling all still work unmodified\n *\n * Non-2xx responses throw an error with `.status` set to the HTTP status\n * code, so VernLLMs `nonRetryableStatus` handling (e.g. failing fast on\n * 401/403) applies here too\n */\nexport function fromFetch(config: FetchAdapterConfig): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const url = typeof config.url === 'function' ? config.url(params) : config.url;\n          const headers =\n            typeof config.headers === 'function' ? await config.headers() : config.headers;\n\n          const res = await fetch(url, {\n            method: config.method ?? 'POST',\n            headers: { 'Content-Type': 'application/json', ...headers },\n            body: JSON.stringify(config.mapRequest(params)),\n            signal: options.signal,\n          });\n\n          if (!res.ok) {\n            const body = await res.text().catch(() => '');\n            const err = new Error(\n              `Fetch adapter request failed (${res.status}): ${body.slice(0, 500)}`,\n            ) as Error & { status?: number };\n            err.status = res.status;\n            throw err;\n          }\n\n          const json = await res.json();\n          const { content, usage } = config.mapResponse(json);\n\n          return {\n            choices: [{ message: { content } }],\n            usage: usage\n              ? {\n                  prompt_tokens: usage.promptTokens,\n                  completion_tokens: usage.completionTokens,\n                  total_tokens: usage.totalTokens,\n                }\n              : undefined,\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Passthrough adapter for any SDK/client whose `chat.completions.create`\n * already matches the OpenAI wire format 1:1 : this covers most hosted\n * inference providers, since \"OpenAI-compatible\" is a de facto standard for\n * chat completion APIs. No transformation happens here, this exists purely\n * so call sites read clearly (`fromMistral(client)` vs handing a Mistral\n * client to something typed for OpenAI) and so a real transformation could\n * be added later, per-provider, without a breaking change.\n *\n * Not every SDKs own TypeScript types line up exactly with `LLMClient`\n * (extra fields, stricter unions, etc.), so this takes `unknown` and casts:\n * the actual compatibility contract is the JSON each provider sends and\n * receives over the wire, not the SDKs TS types.\n */\nexport function fromOpenAICompatible(client: unknown): LLMClient {\n  return client as LLMClient;\n}\n\n/** Groqs SDK matches the OpenAI wire format */\nexport const fromGroq = fromOpenAICompatible;\n\n/** Mistrals `chat.completions`-shaped client (or their OpenAI-compat endpoint) */\nexport const fromMistral = fromOpenAICompatible;\n\n/** DeepSeeks API is OpenAI-compatible */\nexport const fromDeepSeek = fromOpenAICompatible;\n\n/** Cerebras inference API is OpenAI-compatible */\nexport const fromCerebras = fromOpenAICompatible;\n\n/** Together AIs API is OpenAI-compatible */\nexport const fromTogether = fromOpenAICompatible;\n\n/** Fireworks AIs API is OpenAI-compatible */\nexport const fromFireworks = fromOpenAICompatible;\n\n/**\n * Ollama exposes an OpenAI-compatible endpoint at `/v1/chat/completions`\n * (as opposed to its native `/api/chat` format, which differs). Point an\n * OpenAI SDK instances `baseURL` at your Ollama server and pass it here:\n * this does not talk to Ollamas native API directly.\n */\nexport const fromOllama = fromOpenAICompatible;\n"],"mappings":";;;AASA,IAAa,WAAb,cAA8B,MAAM;CAClC,YACEA,SACOC,MACAC,QACAC,QACP;AACA,QAAM,QAAQ;EA2LjB,KA/LU;EA+LT,KA9LS;EA8LR,KA7LQ;AAGP,OAAK,OAAO;CACb;AACF;AAED,SAAgB,WAAWC,KAA+B;AACxD,QAAO,eAAe;AACvB;;;;;AAWD,IAAa,uBAAb,MAA0E;CACxE,AAAQ,QAAQ,IAAI;CAEpB,MAAM,IAAIC,KAAgC;EACxC,MAAM,QAAQ,KAAK,MAAM,IAAI,IAAI;AACjC,OAAK,MAAO,QAAO;AACnB,MAAI,KAAK,KAAK,GAAG,MAAM,WAAW;AAChC,QAAK,MAAM,OAAO,IAAI;AACtB,UAAO;EACR;AACD,SAAO,MAAM;CACd;CAED,MAAM,IAAIA,KAAaC,OAAUC,KAA4B;AAC3D,OAAK,MAAM,IAAI,KAAK;GAAE;GAAO,WAAW,KAAK,KAAK,GAAG,MAAM;EAAM,EAAC;CACnE;AACF;;;;;;;;;;ACjCD,IAAa,iBAAb,MAA4B;CAC1B,AAAQ,QAAsB;CAC9B,AAAQ,sBAAsB;CAC9B,AAAQ,WAAW;CACnB,AAAQ;CACR,AAAQ;CAER,YAAYC,UAAiC,CAAE,GAAE;AAC/C,OAAK,YAAY,QAAQ,aAAa;AACtC,OAAK,aAAa,QAAQ,cAAc;CACzC;;CAGD,eAAqB;AACnB,MAAI,KAAK,UAAU,OAAQ;EAE3B,MAAM,UAAU,KAAK,KAAK,GAAG,KAAK;AAClC,MAAI,WAAW,KAAK,YAAY;AAC9B,QAAK,QAAQ;AACb;EACD;AAED,QAAM,IAAI,UACP,qCAAqC,KAAK,oBAAoB,4BAA4B,KAAK,MAAM,KAAK,aAAa,WAAW,IAAK,CAAC,KACzI;CAEH;CAED,gBAAsB;AACpB,OAAK,sBAAsB;AAC3B,OAAK,QAAQ;CACd;CAED,gBAAsB;AACpB,OAAK,uBAAuB;AAE5B,MAAI,KAAK,UAAU,aAAa;AAE9B,QAAK,QAAQ;AACb,QAAK,WAAW,KAAK,KAAK;AAC1B;EACD;AAED,MAAI,KAAK,uBAAuB,KAAK,WAAW;AAC9C,QAAK,QAAQ;AACb,QAAK,WAAW,KAAK,KAAK;EAC3B;CACF;CAED,WAAyB;AACvB,SAAO,KAAK;CACb;AACF;;;;ACnED,SAAgB,iBAAiBC,SAA0B;AACzD,KAAI;AACF,SAAO,KAAK,MAAM,QAAQ;CAC3B,QAAO;AACN;CACD;AACF;;;;;;;AAQD,SAAgB,cAAcC,KAAkC;AAC9D,MAAK,cAAc,QAAQ,SAAU;CAErC,MAAM,QAAQ;AAKd,YAAW,MAAM,WAAW,SAAU,QAAO,MAAM;AACnD,YAAW,MAAM,eAAe,SAAU,QAAO,MAAM;AAEvD;AACD;;;;;;;;;AAUD,eAAsB,YACpBC,IACAC,WACAC,gBACY;CACZ,MAAM,aAAa,IAAI;CAEvB,MAAM,QAAQ,WAAW,MAAM;AAC7B,aAAW,OAAO;CACnB,GAAE,UAAU;CAEb,MAAM,SAAS,iBACX,YAAY,IAAI,CAAC,gBAAgB,WAAW,MAAO,EAAC,GACpD,WAAW;AAEf,KAAI;AACF,SAAO,MAAM,GAAG,OAAO;CACxB,UAAS;AACR,eAAa,MAAM;CACpB;AACF;;;;;;AAOD,SAAgB,gBAAgBC,aAAqBC,SAAiB,aAAa,KAAgB;CACjG,MAAM,MAAM,KAAK,IAAI,cAAc,KAAK,SAAS,WAAW;AAC5D,QAAO,MAAM,IAAI,KAAK,QAAQ,IAAI,MAAM;AACzC;;;;;;;;AASD,eAAsB,aAAaC,OAAeC,QAAqC;AACrF,OAAM,IAAI,QAAc,CAAC,SAAS,WAAW;EAC3C,MAAM,UAAU,MAAM;AACpB,gBAAa,MAAM;AACnB,UAAO,IAAI,SAAS,qBAAqB,WAAW;EACrD;EAED,MAAM,QAAQ,WAAW,MAAM;AAC7B,WAAQ,oBAAoB,SAAS,QAAQ;AAC7C,YAAS;EACV,GAAE,MAAM;AAET,UAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAM,EAAC;CAC3D;AACF;;;;;;;;ACjFD,IAAa,gBAAb,MAA6C;CAC3C,YAAoBC,cAAuB;EAc5C,KAdqB;CAAyB;CAE7C,MAAMC,SAAuB;AAC3B,MAAI,KAAK,aAAc,SAAQ,MAAM,QAAQ;CAC9C;CAED,KAAKA,SAAuB;AAC1B,UAAQ,KAAK,QAAQ;CACtB;CAED,MAAMA,SAAiBC,MAAsC;AAC3D,UAAQ,MAAM,SAAS,QAAQ,GAAG;CACnC;AACF;;;;;;;;;;;;ACKD,IAAa,UAAb,MAAqB;CACnB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;;;;;;CAOjB,YAAYC,SAAyB;AACnC,OAAK,SAAS,QAAQ;AACtB,OAAK,QAAQ,QAAQ;EAErB,MAAM,cAAc,KAAK,mBAAmB,QAAQ;AACpD,OAAK,aAAa,YAAY;AAC9B,OAAK,YAAY,YAAY;AAC7B,OAAK,cAAc,YAAY;AAC/B,OAAK,mBAAmB,YAAY;AAEpC,OAAK,QAAQ,QAAQ,SAAS,IAAI;AAClC,OAAK,qBAAqB,QAAQ,sBAAsB;GAAC;GAAK;GAAK;EAAI;AAEvE,OAAK,YAAY,QAAQ,aAAa;AACtC,OAAK,UAAU,QAAQ;AAEvB,OAAK,SAAS,KAAK,cAAc,QAAQ;AACzC,OAAK,UAAU,KAAK,sBAAsB,QAAQ;CACnD;;;;;CAMD,AAAQ,mBAAmBA,SAAyB;AAClD,SAAO;GACL,YAAY,QAAQ,cAAc;GAClC,WAAW,QAAQ,aAAa;GAChC,aAAa,QAAQ,eAAe;GACpC,kBAAkB,QAAQ,oBAAoB;EAC/C;CACF;;;;;;CAOD,AAAQ,cAAcA,SAAiC;AACrD,SACE,QAAQ,UAAU,IAAI,cAAc,QAAQ,SAAS,QAAQ,IAAI,aAAa;CAEjF;;;;;;CAOD,AAAQ,sBAAsBA,SAAqD;AACjF,OAAK,QAAQ,eAAgB;AAE7B,SAAO,IAAI,eAAe,QAAQ,mBAAmB,gBAAmB,QAAQ;CACjF;;;;;;;;;;;;;;;CAgBD,MAAM,KAAkBC,QAAmC;AACzD,OAAK,SAAS,cAAc;AAE5B,MAAI,OAAO,QAAQ,QACjB,OAAM,IAAI,SAAS,uBAAuB;EAG5C,MAAM,YAAY,OAAO,aAAa,YAAY;AAElD,MAAI;GACF,MAAM,SAAS,MAAM,KAAK,iBACxB,MAAM,KAAK,YAAY,QAAQ,UAAU,EACzC,WACA,OAAO,OACR;AAED,UAAO;EACR,SAAQ,OAAO;AACd,QAAK,SAAS,eAAe;AAC7B,SAAM,KAAK,eAAe,OAAO,OAAO,OAAO;EAChD;CACF;;;;;;CAOD,MAAc,iBACZC,IACAC,WACAC,QACY;EACZ,IAAIC;AAEJ,OAAK,IAAI,UAAU,GAAG,WAAW,KAAK,YAAY,UAChD,KAAI;AACF,OAAI,UAAU,EACZ,OAAM,KAAK,aAAa,WAAW,SAAS,OAAO;AAGrD,UAAO,MAAM,IAAI;EAClB,SAAQ,OAAO;AACd,eAAY;AAEZ,QAAK,KAAK,YAAY,OAAO,OAAO,CAClC;EAEH;AAGH,QAAM;CACP;;;;;;;CAQD,AAAQ,eAAeC,OAAgBF,QAAgC;AACrE,MAAI,QAAQ,QACV,QAAO,IAAI,SAAS,uBAAuB;AAG7C,MAAI,iBAAiB,SACnB,QAAO;EAGT,MAAM,SAAS,cAAc,MAAM;AAEnC,MAAI,kBACF,QAAO,IAAI,SAAS,sBAAsB,OAAO;AAGnD,SAAO,IAAI,SAAS,sBAAsB;CAC3C;;;;;;;CAQD,MAAc,YAAeH,QAAuBE,WAA+B;EACjF,MAAM,EAAE,SAAS,OAAO,SAAS,GAAG,KAAK,oBAAoB,OAAO;EAEpE,MAAM,WAAW,MAAM,YACrB,CAAC,kBAAkB,KAAK,OAAO,KAAK,YAAY,OAAO,SAAS,EAAE,QAAQ,cAAe,EAAC,EAC1F,KAAK,WACL,OAAO,OACR;EAED,MAAM,UAAU,SAAS,UAAU,IAAI,SAAS,SAAS,MAAM;AAE/D,OAAK,QACH,OAAM,IAAI,SAAS,sBAAsB;AAG3C,OAAK,OAAO,OAAO,QAAQ,UAAU,aAAa,QAAQ,MAAM,GAAG,IAAI,CAAC,EAAE;AAE1E,OAAK,YAAY,UAAU,WAAW,MAAM;AAE5C,OAAK,SAAS,eAAe;AAE7B,OAAK,QACH,QAAO;AAGT,SAAO,KAAK,iBAAiB,SAAS,OAAO,OAAO;CACrD;;;;;;;CAQD,AAAQ,oBAAuBF,QAAuB;EACpD,MAAM,EACJ,cACA,aACA,cAAc,IACd,WAAW,MACX,YAAY,KAAK,kBACjB,QAAQ,KAAK,OACb,iBACA,YACD,GAAG;EAEJ,MAAM,UAAU,YAAY,QAAQ,WAAW;EAC/C,MAAM,iBAAiB,KAAK,oBAAoB,YAAY,QAAQ;EAEpE,MAAM,UAAU;GACd;GACA;GACA,YAAY;GACZ,GAAI,iBAAiB,EAAE,iBAAiB,eAAgB,IAAG,CAAE;GAC7D,GAAI,kBAAkB,EAAE,kBAAkB,gBAAiB,IAAG,CAAE;GAChE,UAAU,CACR;IAAE,MAAM;IAAmB,SAAS;GAAc,GAClD;IAAE,MAAM;IAAiB,SAAS;GAAa,CAChD;EACF;AAED,SAAO;GAAE;GAAS;GAAO;EAAS;CACnC;;;;;;;;CASD,AAAQ,oBAAoBM,YAA+CC,SAAkB;AAC3F,MAAI,WACF,QAAO;GACL,MAAM;GACN,aAAa;IACX,MAAM,WAAW;IACjB,QAAQ,WAAW;IACnB,QAAQ,WAAW,UAAU;IAC7B,aAAa,WAAW;GACzB;EACF;AAGH,SAAO,UAAU,EAAE,MAAM,cAAwB;CAClD;;;;;;CAOD,AAAQ,YACNC,UACAN,WACAO,OACM;AACN,OAAK,SAAS,UAAU,KAAK,QAAS;AAEtC,OAAK,QAAQ;GACX,cAAc,SAAS,MAAM,iBAAiB;GAC9C,kBAAkB,SAAS,MAAM,qBAAqB;GACtD,aAAa,SAAS,MAAM,gBAAgB;GAC5C;GACA;EACD,EAAC;CACH;;;;;;;CAQD,AAAQ,iBAAoBC,SAAiBC,QAAqC;EAChF,IAAIC;AAEJ,MAAI;AACF,YAAS,KAAK,UAAU,QAAQ;EACjC,QAAO;AACN,SAAM,IAAI,SAAS,yBAAyB;EAC7C;AAED,MAAI,WAAW,QAAQ,kBACrB,OAAM,IAAI,SAAS,yBAAyB;AAG9C,OAAK,OACH,QAAO;EAGT,MAAM,SAAS,OAAO,UAAU,OAAO;AAEvC,OAAK,OAAO,QACV,OAAM,IAAI,SAAS,4BAA4B,sBAAyB,OAAO;AAGjF,SAAO,OAAO;CACf;;;;;;CAOD,MAAc,aAAaV,WAAmBW,SAAiBV,QAAsB;EACnF,MAAM,QAAQ,gBAAgB,KAAK,aAAa,QAAQ;AAExD,OAAK,OAAO,MACT,QAAQ,UAAU,qBAAqB,QAAQ,GAAG,KAAK,WAAW,YAAY,MAAM,IACtF;AAED,QAAM,aAAa,OAAO,OAAO;CAClC;;;;;;;;CASD,AAAQ,YAAYE,OAAgBF,QAA+B;AACjE,MAAI,QAAQ,QACV,QAAO;AAGT,MAAI,iBAAiB,aAAa,MAAM,SAAS,WAAW,MAAM,SAAS,cACzE,QAAO;EAGT,MAAM,SAAS,cAAc,MAAM;AAEnC,MAAI,qBAAwB,KAAK,mBAAmB,SAAS,OAAO,CAClE,QAAO;AAGT,SAAO;CACR;;;;;;;CAQD,MAAM,WAAcW,QAAyC;EAC3D,MAAM,SAAS,MAAM,KAAK,MAAM,IAAI,OAAO,SAAS;AAEpD,MAAI,WAAW,KACb,QAAO;AAGT,MAAI;AACF,SAAM,OAAO,gBAAgB;GAE7B,MAAM,SAAS,MAAM,OAAO,IAAI;AAEhC,OAAI;AACF,UAAM,KAAK,MAAM,IAAI,OAAO,UAAU,QAAQ,OAAO,IAAI;GAC1D,SAAQ,OAAO;AACd,SAAK,OAAO,MAAM,gCAAgC,EAChD,SAAS,iBAAiB,QAAQ,MAAM,UAAU,UACnD,EAAC;GACH;AAED,UAAO;EACR,SAAQ,OAAO;AACd,OAAI;AACF,UAAM,OAAO,eAAe;GAC7B,SAAQ,aAAa;AACpB,SAAK,OAAO,MAAM,gCAAgC,EAChD,SAAS,uBAAuB,QAAQ,YAAY,UAAU,UAC/D,EAAC;GACH;AAED,SAAM;EACP;CACF;;;;;;CAOD,MAAM,cACJC,QACY;EACZ,MAAM,EAAE,MAAM,WAAY,GAAG,aAAa,GAAG;AAE7C,SAAO,KAAK,WAAW;GACrB,GAAG;GACH,IAAI,MAAM,KAAK,KAAK,WAAW;EAChC,EAAC;CACH;;;;;CAMD,kBAAkB;AAChB,SAAO,KAAK,SAAS,UAAU;CAChC;AACF;;;;;;;;;;;ACpaD,SAAgB,cAAcC,iBAA6C;AACzE,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAOrE,IAAIC;AACJ,MAAI,OAAO,iBAAiB,SAAS,eAAe;GAClD,MAAM,EAAE,MAAM,QAAQ,GAAG,OAAO,gBAAgB;AAChD,sBAAmB,0GAA0G,KAAK,OAAO,KAAK,UAAU,OAAO,CAAC;EACjK,WAAU,OAAO,iBAAiB,SAAS,cAC1C,mBAAkB;EAOpB,MAAM,SAAS,CAAC,eAAe,SAAS,eAAgB,EAAC,OAAO,QAAQ,CAAC,KAAK,OAAO;EAErF,MAAM,WAAW,MAAM,gBAAgB,SAAS,OAC9C;GACE,OAAO,OAAO;GACd,YAAY,OAAO;GACnB,aAAa,OAAO;GACpB,QAAQ;GACR,UAAU,aAAa,IAAI,CAAC,OAAO;IAAE,MAAM;IAAiB,SAAS,EAAE;GAAS,GAAE;EACnF,GACD,QACD;EAED,MAAM,OAAO,SAAS,QAAQ,KAAK,CAAC,UAAU,MAAM,SAAS,OAAO,EAAE,QAAQ;AAE9E,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,OAAO;IAC/B,mBAAmB,SAAS,OAAO;IACnC,eACG,SAAS,OAAO,gBAAgB,MAAM,SAAS,OAAO,iBAAiB;GAC3E;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;ACtCD,SAAgB,WAAWC,cAAuC;AAChE,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAErE,MAAM,YAAY,QAAQ,OAAO,gBAAgB;EACjD,MAAMC,mBAEF;GACF,aAAa,OAAO;GACpB,iBAAiB,OAAO;EACzB;AAED,MAAI,UACF,kBAAiB,mBAAmB;AAEtC,MAAI,OAAO,iBAAiB,SAAS,cACnC,kBAAiB,iBAAiB,OAAO,gBAAgB,YAAY;EAGvE,MAAM,WAAW,MAAM,aAAa,gBAClC;GACE,OAAO,OAAO;GACd,UAAU,aAAa,IAAI,CAAC,OAAO;IACjC,MAAM;IACN,OAAO,CAAC,EAAE,MAAM,EAAE,QAAS,CAAC;GAC7B,GAAE;GACH,mBAAmB,gBACf,EAAE,OAAO,CAAC,EAAE,MAAM,cAAc,QAAS,CAAC,EAAE;GAEhD;EACD,GACD,QACD;EAED,MAAM,OACJ,SAAS,aAAa,IAAI,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,QAAQ,GAAG,CAAC,KAAK,GAAG,IAAI;AAEjF,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,eAAe;IACvC,mBAAmB,SAAS,eAAe;IAC3C,cAAc,SAAS,eAAe;GACvC;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;;;;ACjDD,SAAgB,YAAYC,eAAiD;AAC3E,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAErE,IAAIC;AACJ,MAAI,OAAO,iBAAiB,SAAS,eAAe;GAClD,MAAM,EAAE,MAAM,QAAQ,GAAG,OAAO,gBAAgB;AAChD,sBAAmB,0GAA0G,KAAK,OAAO,KAAK,UAAU,OAAO,CAAC;EACjK,WAAU,OAAO,iBAAiB,SAAS,cAC1C,mBAAkB;EAGpB,MAAM,cAAc,CAAC,eAAe,SAAS,eAAgB,EAAC,OAAO,CAAC,MACpE,QAAQ,EAAE,CACX;EAED,MAAM,WAAW,MAAM,cAAc,SACnC;GACE,SAAS,OAAO;GAChB,UAAU,aAAa,IAAI,CAAC,OAAO;IACjC,MAAM;IACN,SAAS,CAAC,EAAE,MAAM,EAAE,QAAS,CAAC;GAC/B,GAAE;GACH,QAAQ,YAAY,SAAS,YAAY,IAAI,CAACC,YAAU,EAAE,aAAM,GAAE;GAClE,iBAAiB;IACf,aAAa,OAAO;IACpB,WAAW,OAAO;GACnB;EACF,GACD,QACD;EAED,MAAM,OAAO,SAAS,QAAQ,SAAS,SAAS,IAAI,CAAC,MAAM,EAAE,QAAQ,GAAG,CAAC,KAAK,GAAG,IAAI;AAErF,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,OAAO;IAC/B,mBAAmB,SAAS,OAAO;IACnC,cAAc,SAAS,OAAO;GAC/B;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;AC5DD,SAAgB,UAAUC,QAAuC;AAC/D,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,aAAa,OAAO,QAAQ,aAAa,OAAO,IAAI,OAAO,GAAG,OAAO;EAC3E,MAAM,iBACG,OAAO,YAAY,aAAa,MAAM,OAAO,SAAS,GAAG,OAAO;EAEzE,MAAM,MAAM,MAAM,MAAM,KAAK;GAC3B,QAAQ,OAAO,UAAU;GACzB,SAAS;IAAE,gBAAgB;IAAoB,GAAG;GAAS;GAC3D,MAAM,KAAK,UAAU,OAAO,WAAW,OAAO,CAAC;GAC/C,QAAQ,QAAQ;EACjB,EAAC;AAEF,OAAK,IAAI,IAAI;GACX,MAAM,OAAO,MAAM,IAAI,MAAM,CAAC,MAAM,MAAM,GAAG;GAC7C,MAAM,MAAM,IAAI,OACb,gCAAgC,IAAI,OAAO,KAAK,KAAK,MAAM,GAAG,IAAI,CAAC;AAEtE,OAAI,SAAS,IAAI;AACjB,SAAM;EACP;EAED,MAAM,OAAO,MAAM,IAAI,MAAM;EAC7B,MAAM,EAAE,SAAS,OAAO,GAAG,OAAO,YAAY,KAAK;AAEnD,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,QAAS,EAAE,CAAC;GACnC,OAAO,QACH;IACE,eAAe,MAAM;IACrB,mBAAmB,MAAM;IACzB,cAAc,MAAM;GACrB;EAEN;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;;;;AC/DD,SAAgB,qBAAqBC,QAA4B;AAC/D,QAAO;AACR;;AAGD,MAAa,WAAW;;AAGxB,MAAa,cAAc;;AAG3B,MAAa,eAAe;;AAG5B,MAAa,eAAe;;AAG5B,MAAa,eAAe;;AAG5B,MAAa,gBAAgB;;;;;;;AAQ7B,MAAa,aAAa"}
+{"version":3,"file":"index.js","names":["message: string","type: LLMErrorType","status?: number","issues?: unknown","err: unknown","key: string","value: T","ttl: number","options: CircuitBreakerOptions","content: string","err: unknown","fn: (signal: AbortSignal) => Promise<T>","timeoutMs: number","externalSignal?: AbortSignal","baseDelayMs: number","attempt: number","delay: number","signal?: AbortSignal","debugEnabled: boolean","message: string","meta?: Record<string, unknown>","options: VernLLMOptions","params: CallParams<T>","fn: () => Promise<T>","requestId: string","signal?: AbortSignal","lastError: unknown","error: unknown","jsonSchema: CallParams<unknown>['jsonSchema']","useJson: boolean","response: Awaited<ReturnType<LLMClient['chat']['completions']['create']>>","model: string","content: string","schema?: CallParams<T>['schema']","parsed: unknown","attempt: number","params: CachedCallParams<T>","params: Omit<CachedCallParams<T>, 'fn'> & { call: CallParams<T> }","anthropicClient: AnthropicClient","jsonInstruction: string | undefined","geminiClient: GeminiClient","generationConfig: NonNullable<\n            Parameters<GeminiClient['generateContent']>[0]['generationConfig']\n          >","bedrockClient: BedrockConverseClient","jsonInstruction: string | undefined","text","config: FetchAdapterConfig","client: unknown"],"sources":["../src/types.ts","../src/circuitBreaker.ts","../src/internal/vernLLM.utilts.ts","../src/logger.ts","../src/vernLLM.ts","../src/adapters/anthropic.ts","../src/adapters/gemini.ts","../src/adapters/bedrock.ts","../src/adapters/fetch.ts","../src/adapters/openaiCompatible.ts"],"sourcesContent":["export type LLMErrorType =\n  | 'timeout'\n  | 'api'\n  | 'parse'\n  | 'validation'\n  | 'circuit_open'\n  | 'unknown'\n  | 'aborted';\n\nexport class LLMError extends Error {\n  constructor(\n    message: string,\n    public type: LLMErrorType,\n    public status?: number,\n    public issues?: unknown,\n  ) {\n    super(message);\n    this.name = 'LLMError';\n  }\n}\n\nexport function isLLMError(err: unknown): err is LLMError {\n  return err instanceof LLMError;\n}\n\nexport interface CacheAdapter<T = unknown> {\n  get(key: string): Promise<T | null>;\n  set(key: string, value: T, ttl: number): Promise<void>;\n}\n\n/**\n * Trivial default so the package works out of the box with no external deps\n * Not shared across processes, swap in Redis/Upstash/etc for production\n */\nexport class InMemoryCacheAdapter<T = unknown> implements CacheAdapter<T> {\n  private store = new Map<string, { value: T; expiresAt: number }>();\n\n  async get(key: string): Promise<T | null> {\n    const entry = this.store.get(key);\n    if (!entry) return null;\n    if (Date.now() > entry.expiresAt) {\n      this.store.delete(key);\n      return null;\n    }\n    return entry.value;\n  }\n\n  async set(key: string, value: T, ttl: number): Promise<void> {\n    this.store.set(key, { value, expiresAt: Date.now() + ttl * 1000 });\n  }\n}\n\n/**\n * Minimal shape compatible with the OpenAI SDKs chat.completions.create,\n * so consumers can pass an OpenAI client directly\n * `response_format.json_schema` and `reasoning_effort` are optional on the wire\n * providers that don't support them will just ignore fields they don't recognize,\n * but not every SDKs TS types accept them, hence this being a structural type\n * rather than importing the SDKs own params type\n */\nexport interface LLMClient {\n  chat: {\n    completions: {\n      create(\n        params: {\n          model: string;\n          temperature: number;\n          max_tokens: number;\n          response_format?:\n            | { type: 'json_object' }\n            | {\n                type: 'json_schema';\n                json_schema: {\n                  name: string;\n                  schema: Record<string, unknown>;\n                  strict?: boolean;\n                  description?: string;\n                };\n              };\n          /** OpenAI reasoning-model param (o-series, gpt-5), ignored by providers that don't support it */\n          reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n          messages: Array<{ role: 'system' | 'user'; content: string }>;\n        },\n        options: { signal: AbortSignal },\n      ): Promise<{\n        choices?: Array<{ message?: { content?: string | null } }>;\n        usage?: {\n          prompt_tokens?: number;\n          completion_tokens?: number;\n          total_tokens?: number;\n        };\n      }>;\n    };\n  };\n}\n\nexport type ReserveUsage = () => Promise<void>;\nexport type RefundUsage = () => Promise<void>;\n\nexport interface TokenUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  requestId: string;\n  model: string;\n}\n\nexport type OnUsage = (usage: TokenUsage) => void;\n\n/**\n * Minimal structural type for a Zod-like schema, so this package doesnt need\n * a hard dependency on a specific Zod major version. Any object exposing\n * `safeParse` (Zod v3/v4, and most Zod-compatible validators) should satisfy this\n */\nexport interface SchemaLike<T> {\n  safeParse(data: unknown): { success: true; data: T } | { success: false; error: unknown };\n}\n\nexport interface VernLLMOptions {\n  client: LLMClient;\n  model: string;\n  /** Max retries after the first attempt. Default 1 (2 attempts total) */\n  maxRetries?: number;\n  /** Per-attempt timeout in ms. Default 25000 */\n  timeoutMs?: number;\n  /** Base delay for exponential backoff in ms. Default 500 */\n  baseDelayMs?: number;\n  /** Default max_tokens for calls that don't override it. Default 1000 */\n  defaultMaxTokens?: number;\n  /** Enables debug logging of raw model output. Default: NODE_ENV !== 'production' */\n  debug?: boolean;\n  /** Cache adapter for cachedCall. Defaults to an in-memory adapter */\n  cache?: CacheAdapter;\n  /** HTTP status codes that should fail fast without retrying. Default [400, 401, 403] */\n  nonRetryableStatus?: number[];\n  /** Custom JSON parser. Must return undefined/null on failure. Default: JSON.parse wrapped in try/catch */\n  parseJson?: (content: string) => unknown;\n  /** Called after every successful call with token usage, if the provider reports it */\n  onUsage?: OnUsage;\n  /** Injectable logger. Defaults to a console-based logger gated by `debug` */\n  logger?: import('./logger.js').Logger;\n  /**\n   * Enables a circuit breaker that short-circuits calls after repeated\n   * consecutive failures, instead of continuing to hammer a down provider\n   * Pass `true` for defaults, or an options object to tune threshold/cooldown\n   */\n  circuitBreaker?: boolean | import('./circuitBreaker.js').CircuitBreakerOptions;\n}\n\n/**\n * A provider-native JSON Schema for structured outputs (OpenAI/Groq\n * `response_format: { type: 'json_schema' }`) This is the wire-format\n * schema the model is constrained to generate against, distinct from\n * `schema`, which is a client-side Zod validator run on the parsed result\n * You can use one, both, or neither; using both gets you provider-level\n * constraint plus client-side type inference/validation as a safety net\n */\nexport interface JsonSchemaSpec {\n  name: string;\n  schema: Record<string, unknown>;\n  /** Enforces the schema strictly (OpenAI-specific), default true when supported */\n  strict?: boolean;\n  description?: string;\n}\n\nexport interface CallParams<T = unknown> {\n  systemPrompt: string;\n  userContent: string;\n  temperature?: number;\n  jsonMode?: boolean;\n  maxTokens?: number;\n  requestId?: string;\n  signal?: AbortSignal;\n  /** Overrides the model set on the VernLLM instance for this call only */\n  model?: string;\n  /**\n   * OpenAI-style reasoning effort for reasoning models (o-series, gpt-5, etc).\n   * Passed through as-is, providers/models that don't support it ignore it\n   */\n  reasoningEffort?: 'minimal' | 'low' | 'medium' | 'high';\n  /**\n   * Provider-native JSON Schema structured-output mode. When set, this is sent\n   * as `response_format: { type: 'json_schema', json_schema: ... }` instead of\n   * the looser `json_object` mode, constraining the models output shape at\n   * generation time (not just validating it after the fact). Implies jsonMode: true.\n   */\n  jsonSchema?: JsonSchemaSpec;\n  /**\n   * Optional Zod (or Zod-compatible) schema. When provided, the parsed JSON\n   * is validated against it; on failure an LLMError('validation') is thrown\n   * with `.issues` set to the schema's error object. Implies jsonMode: true.\n   * Can be combined with `jsonSchema` for provider-level constraint + client-side typing.\n   */\n  schema?: SchemaLike<T>;\n}\n\nexport interface CachedCallParams<T> {\n  cacheKey: string;\n  ttl: number;\n  fn: () => Promise<T>;\n  reserveUsage?: ReserveUsage;\n  refundUsage?: RefundUsage;\n}\n","import { LLMError } from './types.js';\n\nexport interface CircuitBreakerOptions {\n  /** Consecutive failures before the circuit opens, default 5 */\n  threshold?: number;\n  /** How long the circuit stays open before allowing a trial request, in ms. Default 30000 */\n  cooldownMs?: number;\n}\n\ntype CircuitState = 'closed' | 'open' | 'half-open';\n\n/**\n * Per retry VernLLM-instance circuit breaker. Tracks consecutive failures across\n * calls. Once the threshold is hit, short-circuits new calls with an\n * LLMError('circuit_open') instead of hitting the provider, until the\n * cooldown elapses and a single trial call is allowed through\n */\nexport class CircuitBreaker {\n  private state: CircuitState = 'closed';\n  private consecutiveFailures = 0;\n  private openedAt = 0;\n  private threshold: number;\n  private cooldownMs: number;\n\n  constructor(options: CircuitBreakerOptions = {}) {\n    this.threshold = options.threshold ?? 5;\n    this.cooldownMs = options.cooldownMs ?? 30_000;\n  }\n\n  /** Throws if the circuit is open and the cooldown hasnt elapsed */\n  assertClosed(): void {\n    if (this.state !== 'open') return;\n\n    const elapsed = Date.now() - this.openedAt;\n    if (elapsed >= this.cooldownMs) {\n      this.state = 'half-open';\n      return;\n    }\n\n    throw new LLMError(\n      `Circuit open — provider has failed ${this.consecutiveFailures} times in a row. Retry in ${Math.ceil((this.cooldownMs - elapsed) / 1000)}s.`,\n      'circuit_open',\n    );\n  }\n\n  recordSuccess(): void {\n    this.consecutiveFailures = 0;\n    this.state = 'closed';\n  }\n\n  recordFailure(): void {\n    this.consecutiveFailures += 1;\n\n    if (this.state === 'half-open') {\n      // Trial call failed: reopen and reset the cooldown window.\n      this.state = 'open';\n      this.openedAt = Date.now();\n      return;\n    }\n\n    if (this.consecutiveFailures >= this.threshold) {\n      this.state = 'open';\n      this.openedAt = Date.now();\n    }\n  }\n\n  getState(): CircuitState {\n    return this.state;\n  }\n}\n","import { LLMError } from '../types.js';\n\nexport function defaultParseJson(content: string): unknown {\n  try {\n    return JSON.parse(content);\n  } catch {\n    return undefined;\n  }\n}\n\n/**\n * Looks inside an unknown error value and pulls out an http status code\n * if one is present. Checks the status field first then the status code\n * field since different client libraries use different names for this.\n * Returns undefined when the error is not an object or carries no status\n */\nexport function extractStatus(err: unknown): number | undefined {\n  if (!err || typeof err !== 'object') return undefined;\n\n  const error = err as {\n    status?: unknown;\n    statusCode?: unknown;\n  };\n\n  if (typeof error.status === 'number') return error.status;\n  if (typeof error.statusCode === 'number') return error.statusCode;\n\n  return undefined;\n}\n\n/**\n * Runs an async function and cancels it if it takes longer than the given\n * timeout. Creates an internal abort controller that fires after the\n * timeout elapses, and combines it with any external signal the caller\n * passed in so either one can cancel the underlying call. If the internal\n * timeout triggers and the underlying operation aborts, the error is\n * converted into an LLMError with type \"timeout\". External cancellations\n * continue to propagate as aborted errors. The internal timer is always\n * cleared afterward, whether the function succeeds, fails, or is aborted,\n * so nothing is left running in the background.\n */\nexport async function withTimeout<T>(\n  fn: (signal: AbortSignal) => Promise<T>,\n  timeoutMs: number,\n  externalSignal?: AbortSignal,\n): Promise<T> {\n  const controller = new AbortController();\n\n  const timer = setTimeout(() => {\n    controller.abort();\n  }, timeoutMs);\n\n  const signal = externalSignal\n    ? AbortSignal.any([externalSignal, controller.signal])\n    : controller.signal;\n\n  try {\n    return await fn(signal);\n  } catch (err) {\n    if (\n      controller.signal.aborted &&\n      !externalSignal?.aborted &&\n      err instanceof DOMException &&\n      err.name === 'AbortError'\n    ) {\n      throw new LLMError('Request timed out', 'timeout');\n    }\n\n    throw err;\n  } finally {\n    clearTimeout(timer);\n  }\n}\n\n/**\n * Exponential backoff with jitter, capped at maxDelayMs.\n * Jitter avoids thundering-herd retries when many callers back off in lockstep,\n * the cap prevents unbounded delays when maxRetries is high\n */\nexport function getBackoffDelay(baseDelayMs: number, attempt: number, maxDelayMs = 10_000): number {\n  const exp = Math.min(baseDelayMs * 2 ** attempt, maxDelayMs);\n  return exp / 2 + Math.random() * (exp / 2);\n}\n\n/**\n * Pauses execution for the given delay before a retry attempt. If an\n * abort signal is provided and it fires while waiting, the pending\n * timer is cancelled immediately and the wait rejects right away with\n * an aborted error instead of continuing to sit idle until the delay\n * would have finished on its own\n */\nexport async function waitForRetry(delay: number, signal?: AbortSignal): Promise<void> {\n  await new Promise<void>((resolve, reject) => {\n    const onAbort = () => {\n      clearTimeout(timer);\n      reject(new LLMError('Operation aborted', 'aborted'));\n    };\n\n    const timer = setTimeout(() => {\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    }, delay);\n\n    signal?.addEventListener('abort', onAbort, { once: true });\n  });\n}\n","export interface Logger {\n  debug(message: string): void;\n  warn(message: string): void;\n  error(message: string, meta?: Record<string, unknown>): void;\n}\n\n/**\n * Default logger. `debug` is gated by the `debug` option on VernLLM\n * warn/error always fire since they indicate real problems (retries, cache failures)\n */\nexport class ConsoleLogger implements Logger {\n  constructor(private debugEnabled: boolean) {}\n\n  debug(message: string): void {\n    if (this.debugEnabled) console.debug(message);\n  }\n\n  warn(message: string): void {\n    console.warn(message);\n  }\n\n  error(message: string, meta?: Record<string, unknown>): void {\n    console.error(message, meta ?? '');\n  }\n}\n","import { randomUUID } from 'crypto';\n\nimport { CircuitBreaker } from './circuitBreaker.js';\nimport {\n  defaultParseJson,\n  extractStatus,\n  withTimeout,\n  getBackoffDelay,\n  waitForRetry,\n} from './internal/vernLLM.utilts.js';\nimport { ConsoleLogger, type Logger } from './logger.js';\nimport {\n  LLMError,\n  InMemoryCacheAdapter,\n  type VernLLMOptions,\n  type CallParams,\n  type CachedCallParams,\n  type CacheAdapter,\n  type LLMClient,\n} from './types.js';\n\n/**\n * A resilient wrapper around an LLM chat completions client, this is VernLLM!\n *\n * Adds retry with exponential backoff and jitter, per-attempt timeouts,\n * an optional circuit breaker, JSON parsing with optional schema\n * validation, usage tracking, and an optional response cache, all\n * configurable, all opt-in beyond sensible defaults\n */\nexport class VernLLM {\n  private readonly client: LLMClient;\n  private readonly model: string;\n\n  private readonly maxRetries: number;\n  private readonly timeoutMs: number;\n  private readonly baseDelayMs: number;\n  private readonly defaultMaxTokens: number;\n\n  private readonly cache: CacheAdapter<unknown>;\n  private readonly nonRetryableStatus: number[];\n\n  private readonly parseJson: (content: string) => unknown;\n  private readonly onUsage?: VernLLMOptions['onUsage'];\n\n  private readonly logger: Logger;\n  private readonly breaker?: CircuitBreaker;\n\n  /**\n   * @param options: Client, model, and all tunables (retries, timeout,\n   * backoff, cache, circuit breaker, logger, etc). See VernLLMOptions in `types.ts`\n   * for individual defaults\n   */\n  constructor(options: VernLLMOptions) {\n    this.client = options.client;\n    this.model = options.model;\n\n    const retryConfig = this.resolveRetryConfig(options);\n    this.maxRetries = retryConfig.maxRetries;\n    this.timeoutMs = retryConfig.timeoutMs;\n    this.baseDelayMs = retryConfig.baseDelayMs;\n    this.defaultMaxTokens = retryConfig.defaultMaxTokens;\n\n    this.cache = options.cache ?? new InMemoryCacheAdapter();\n    this.nonRetryableStatus = options.nonRetryableStatus ?? [400, 401, 403];\n\n    this.parseJson = options.parseJson ?? defaultParseJson;\n    this.onUsage = options.onUsage;\n\n    this.logger = this.resolveLogger(options);\n    this.breaker = this.resolveCircuitBreaker(options);\n  }\n\n  /**\n   * Resolves retry/timeout/token defaults from the given options,\n   * falling back to the librarys built-in defaults for anything unset\n   */\n  private resolveRetryConfig(options: VernLLMOptions) {\n    return {\n      maxRetries: options.maxRetries ?? 1,\n      timeoutMs: options.timeoutMs ?? 25_000,\n      baseDelayMs: options.baseDelayMs ?? 500,\n      defaultMaxTokens: options.defaultMaxTokens ?? 1000,\n    };\n  }\n\n  /**\n   * Returns the caller supplied logger, or a console-based logger whose\n   * debug output is gated by the `debug` option (defaulting to on\n   * outside production)\n   */\n  private resolveLogger(options: VernLLMOptions): Logger {\n    return (\n      options.logger ?? new ConsoleLogger(options.debug ?? process.env.NODE_ENV !== 'production')\n    );\n  }\n\n  /**\n   * Builds a circuit breaker if `circuitBreaker` is truthy on the\n   * options. Passing `true` uses default thresholds, passing an options\n   * object tunes them. Returns undefined when the breaker is disabled\n   */\n  private resolveCircuitBreaker(options: VernLLMOptions): CircuitBreaker | undefined {\n    if (!options.circuitBreaker) return undefined;\n\n    return new CircuitBreaker(options.circuitBreaker === true ? undefined : options.circuitBreaker);\n  }\n\n  /**\n   * Makes a single logical LLM call, transparently retrying on failure\n   * according to the configured retry policy\n   *\n   * Fails fast if the circuit breaker is open or the signal is already\n   * aborted, before any request is dispatched. On exhausting all\n   * retries, records a circuit breaker failure and rejects with a\n   * normalized LLMError\n   *\n   * @param params : System/user content plus per call overrides\n   * (model, temperature, jsonMode, schema, signal, etc)\n   * @returns The parsed and optionally schema-validated response, or\n   * the raw string content when jsonMode is disabled\n   */\n  async call<T = unknown>(params: CallParams<T>): Promise<T> {\n    this.breaker?.assertClosed();\n\n    if (params.signal?.aborted) {\n      throw new LLMError('LLM request aborted', 'aborted');\n    }\n\n    const requestId = params.requestId ?? randomUUID();\n\n    try {\n      const result = await this.retryWithBackoff(\n        () => this.executeCall(params, requestId),\n        requestId,\n        params.signal,\n      );\n\n      return result;\n    } catch (error) {\n      this.breaker?.recordFailure();\n      throw this.normalizeError(error, params.signal);\n    }\n  }\n\n  /**\n   * Runs `fn`, retrying with backoff according to `shouldRetry` policy\n   * Purely mechanical: knows nothing about LLM specifics beyond the retry\n   * predicate, so its testable independent of request/response shaping\n   */\n  private async retryWithBackoff<T>(\n    fn: () => Promise<T>,\n    requestId: string,\n    signal?: AbortSignal,\n  ): Promise<T> {\n    let lastError: unknown;\n\n    for (let attempt = 0; attempt <= this.maxRetries; attempt++) {\n      try {\n        if (attempt > 0) {\n          await this.recoverDelay(requestId, attempt, signal);\n        }\n\n        return await fn();\n      } catch (error) {\n        lastError = error;\n\n        if (!this.shouldRetry(error, signal)) {\n          break;\n        }\n      }\n    }\n\n    throw lastError;\n  }\n\n  /**\n   * Converts any thrown value into a well-typed LLMError for the public\n   * API surface. Preserves an existing LLMError as is, reports aborted\n   * signals as such, classifies errors carrying an http status as\n   * type api, and otherwise falls back to a generic unknown error.\n   */\n  private normalizeError(error: unknown, signal?: AbortSignal): LLMError {\n    if (signal?.aborted) {\n      return new LLMError('LLM request aborted', 'aborted');\n    }\n\n    if (error instanceof LLMError) {\n      return error;\n    }\n\n    const status = extractStatus(error);\n\n    if (status !== undefined) {\n      return new LLMError('LLM request failed', 'api', status);\n    }\n\n    return new LLMError('LLM request failed', 'unknown');\n  }\n\n  /**\n   * Performs a single attempt: builds the request, dispatches it with a\n   * timeout, and shapes the response. Throws on an empty response so\n   * the retry loop treats it like any other transient failure. Records\n   * usage and a circuit breaker success before returning\n   */\n  private async executeCall<T>(params: CallParams<T>, requestId: string): Promise<T> {\n    const { useJson, model, request } = this.buildRequestPayload(params);\n\n    const response = await withTimeout(\n      (attemptSignal) => this.client.chat.completions.create(request, { signal: attemptSignal }),\n      this.timeoutMs,\n      params.signal,\n    );\n\n    const content = response.choices?.[0]?.message?.content?.trim();\n\n    if (!content) {\n      throw new LLMError('Empty LLM response', 'api');\n    }\n\n    this.logger.debug(`[vern:${requestId}] output:\\n${content.slice(0, 800)}`);\n\n    this.recordUsage(response, requestId, model);\n\n    this.breaker?.recordSuccess();\n\n    if (!useJson) {\n      return content as T;\n    }\n\n    return this.parseAndValidate(content, params.schema);\n  }\n\n  /**\n   * Applies per call defaults and shapes the params into the request\n   * object expected by the underlying client, including the resolved\n   * response format. Also returns whether JSON parsing should be\n   * applied to the response and which model was ultimately used\n   */\n  private buildRequestPayload<T>(params: CallParams<T>) {\n    const {\n      systemPrompt,\n      userContent,\n      temperature = 0.2,\n      jsonMode = true,\n      maxTokens = this.defaultMaxTokens,\n      model = this.model,\n      reasoningEffort,\n      jsonSchema,\n    } = params;\n\n    const useJson = jsonMode || Boolean(jsonSchema);\n    const responseFormat = this.buildResponseFormat(jsonSchema, useJson);\n\n    const request = {\n      model,\n      temperature,\n      max_tokens: maxTokens,\n      ...(responseFormat ? { response_format: responseFormat } : {}),\n      ...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}),\n      messages: [\n        { role: 'system' as const, content: systemPrompt },\n        { role: 'user' as const, content: userContent },\n      ],\n    };\n\n    return { useJson, model, request };\n  }\n\n  /**\n   * Chooses the response format to send to the provider. A provider\n   * native json schema takes priority when supplied, constraining\n   * generation directly, otherwise falls back to the looser json\n   * object mode when JSON output is requested, or no format at all\n   * for plain text responses\n   */\n  private buildResponseFormat(jsonSchema: CallParams<unknown>['jsonSchema'], useJson: boolean) {\n    if (jsonSchema) {\n      return {\n        type: 'json_schema' as const,\n        json_schema: {\n          name: jsonSchema.name,\n          schema: jsonSchema.schema,\n          strict: jsonSchema.strict ?? true,\n          description: jsonSchema.description,\n        },\n      };\n    }\n\n    return useJson ? { type: 'json_object' as const } : undefined;\n  }\n\n  /**\n   * Reports token usage to the caller supplied onUsage callback, when\n   * both a callback was configured and the provider actually returned\n   * usage data on this response. A no op otherwise.\n   */\n  private recordUsage(\n    response: Awaited<ReturnType<LLMClient['chat']['completions']['create']>>,\n    requestId: string,\n    model: string,\n  ): void {\n    if (!response.usage || !this.onUsage) return;\n\n    this.onUsage({\n      promptTokens: response.usage.prompt_tokens ?? 0,\n      completionTokens: response.usage.completion_tokens ?? 0,\n      totalTokens: response.usage.total_tokens ?? 0,\n      requestId,\n      model,\n    });\n  }\n\n  /**\n   * Parses the raw response content as JSON and, when a schema is\n   * supplied, validates the parsed value against it. Throws a parse\n   * type LLMError on malformed JSON and a validation type LLMError,\n   * carrying the schemas issues, on a failed validation\n   */\n  private parseAndValidate<T>(content: string, schema?: CallParams<T>['schema']): T {\n    let parsed: unknown;\n\n    try {\n      parsed = this.parseJson(content);\n    } catch {\n      throw new LLMError('Invalid JSON response', 'parse');\n    }\n\n    if (parsed === null || parsed === undefined) {\n      throw new LLMError('Invalid JSON response', 'parse');\n    }\n\n    if (!schema) {\n      return parsed as T;\n    }\n\n    const result = schema.safeParse(parsed);\n\n    if (!result.success) {\n      throw new LLMError('Schema validation failed', 'validation', undefined, result.error);\n    }\n\n    return result.data;\n  }\n\n  /**\n   * Waits out the backoff delay for a given retry attempt, logging the\n   * attempt for observability before the wait begins. Rejects early if\n   * the signal aborts during the wait\n   */\n  private async recoverDelay(requestId: string, attempt: number, signal?: AbortSignal) {\n    const delay = getBackoffDelay(this.baseDelayMs, attempt);\n\n    this.logger.warn(\n      `[vern:${requestId}] recovery attempt ${attempt}/${this.maxRetries}, waiting ${delay}ms`,\n    );\n\n    await waitForRetry(delay, signal);\n  }\n\n  /**\n   * Decides whether a failed attempt is worth retrying. Never retries\n   * once the signal has aborted, never retries a parse or validation\n   * failure since those stem from the response content rather than a\n   * transient fault, and never retries a status code the caller has\n   * marked as non retryable. Retries everything else\n   */\n  private shouldRetry(error: unknown, signal?: AbortSignal): boolean {\n    if (signal?.aborted) {\n      return false;\n    }\n\n    if (error instanceof LLMError && (error.type === 'parse' || error.type === 'validation')) {\n      return false;\n    }\n\n    const status = extractStatus(error);\n\n    if (status !== undefined && this.nonRetryableStatus.includes(status)) {\n      return false;\n    }\n\n    return true;\n  }\n\n  /**\n   * Thin cache wrapper around caller supplied logic. `params.fn` is expected\n   * to be a call that itself invokes `this.call(...)` (see `cachedLLMCall`\n   * below for a convenience wrapper that wires this up automatically),\n   * `cachedCall` does not itself apply retry/timeout policy.\n   */\n  async cachedCall<T>(params: CachedCallParams<T>): Promise<T> {\n    const cached = await this.cache.get(params.cacheKey);\n\n    if (cached !== null) {\n      return cached as T;\n    }\n\n    try {\n      await params.reserveUsage?.();\n\n      const result = await params.fn();\n\n      try {\n        await this.cache.set(params.cacheKey, result, params.ttl);\n      } catch (error) {\n        this.logger.error('[VernLLM] cache write failed', {\n          message: error instanceof Error ? error.message : 'unknown',\n        });\n      }\n\n      return result;\n    } catch (error) {\n      try {\n        await params.refundUsage?.();\n      } catch (refundError) {\n        this.logger.error('[VernLLM] refundUsage failed', {\n          message: refundError instanceof Error ? refundError.message : 'unknown',\n        });\n      }\n\n      throw error;\n    }\n  }\n\n  /**\n   * Convenience wrapper composing `call` + `cachedCall`, so cached LLM calls\n   * automatically get retry/timeout/circuit-breaker behavior without callers\n   * having to remember to wire `fn: () => this.call(...)` themselves\n   */\n  async cachedLLMCall<T>(\n    params: Omit<CachedCallParams<T>, 'fn'> & { call: CallParams<T> },\n  ): Promise<T> {\n    const { call: callParams, ...cacheParams } = params;\n\n    return this.cachedCall({\n      ...cacheParams,\n      fn: () => this.call(callParams),\n    });\n  }\n\n  /**\n   * Returns the current circuit breaker state, or undefined when no\n   * circuit breaker was configured on this instance\n   */\n  getCircuitState() {\n    return this.breaker?.getState();\n  }\n}\n","import type { LLMClient } from '../types.js';\n\n/** Minimal structural type for the Anthropic SDKs `messages.create` */\nexport interface AnthropicClient {\n  messages: {\n    create(\n      params: {\n        model: string;\n        max_tokens: number;\n        temperature?: number;\n        system?: string;\n        messages: Array<{ role: 'user' | 'assistant'; content: string }>;\n      },\n      options: { signal: AbortSignal },\n    ): Promise<{\n      content: Array<{ type: string; text?: string }>;\n      usage?: { input_tokens?: number; output_tokens?: number };\n    }>;\n  };\n}\n\n/**\n * Wraps an Anthropic SDK client so it satisfies the same `LLMClient`\n * interface VernLLM uses for OpenAI/Groq. Anthropics Messages API has no\n * `response_format: json_object` equivalent, so when the caller requests\n * JSON mode, this adapter appends an instruction to the system prompt\n * asking the model to respond with JSON only\n */\nexport function fromAnthropic(anthropicClient: AnthropicClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          // Anthropics Messages API has no `response_format: json_object` or\n          // `json_schema` equivalent, so both are emulated via system prompt\n          // instructions. For json_schema, the schema itself is embedded so the\n          // model has something concrete to conform to (not provider-enforced,\n          // unlike OpenAIs native structured outputs)\n          let jsonInstruction: string | undefined;\n          if (params.response_format?.type === 'json_schema') {\n            const { name, schema } = params.response_format.json_schema;\n            jsonInstruction = `Respond with valid JSON only, no prose or markdown fences. The JSON must conform to this schema (name: \"${name}\"):\\n${JSON.stringify(schema)}`;\n          } else if (params.response_format?.type === 'json_object') {\n            jsonInstruction = 'Respond with valid JSON only, no prose or markdown fences.';\n          }\n\n          // `reasoning_effort` (OpenAI o-series/gpt-5 style) has no direct Anthropic\n          // equivalent, Claudes extended thinking uses a token budget, not a tier\n          // string, so its intentionally dropped here rather than guessed at.\n\n          const system = [systemMessage?.content, jsonInstruction].filter(Boolean).join('\\n\\n');\n\n          const response = await anthropicClient.messages.create(\n            {\n              model: params.model,\n              max_tokens: params.max_tokens,\n              temperature: params.temperature,\n              system: system || undefined,\n              messages: userMessages.map((m) => ({ role: 'user' as const, content: m.content })),\n            },\n            options,\n          );\n\n          const text = response.content.find((block) => block.type === 'text')?.text ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usage?.input_tokens,\n              completion_tokens: response.usage?.output_tokens,\n              total_tokens:\n                (response.usage?.input_tokens ?? 0) + (response.usage?.output_tokens ?? 0),\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Minimal structural type for Geminis `generateContent`, matching both the\n * legacy `@google/generative-ai` SDKs `model.generateContent(...)` and the\n * newer `@google/genai` SDKs `ai.models.generateContent({ model, ... })`\n * closely enough to adapt either — pass whichever `.generateContent` you have.\n */\nexport interface GeminiClient {\n  generateContent(\n    params: {\n      model?: string;\n      contents: Array<{ role: 'user'; parts: Array<{ text: string }> }>;\n      systemInstruction?: { parts: Array<{ text: string }> };\n      generationConfig?: {\n        temperature?: number;\n        maxOutputTokens?: number;\n        responseMimeType?: string;\n        responseSchema?: Record<string, unknown>;\n      };\n    },\n    options: { signal: AbortSignal },\n  ): Promise<{\n    candidates?: Array<{ content?: { parts?: Array<{ text?: string }> } }>;\n    usageMetadata?: {\n      promptTokenCount?: number;\n      candidatesTokenCount?: number;\n      totalTokenCount?: number;\n    };\n  }>;\n}\n\n/**\n * Wraps a Gemini client so it satisfies the `LLMClient` interface VernLLM\n * uses for OpenAI/Groq. Geminis shape differs on nearly every axis: a\n * `contents` array instead of `messages`, a separate `systemInstruction`\n * field instead of a `system` role message, `generationConfig` instead of\n * top-level `temperature`/`max_tokens`, and native JSON Schema support via\n * `responseMimeType: 'application/json'` + `responseSchema` (so `jsonSchema`\n * is provider-enforced here, unlike the Anthropic adapters prompt-embedding\n * fallback). `reasoning_effort` has no equivalent. Geminis thinking models\n * use a token budget, not an effort tier, so its dropped, same as Anthropic.\n */\nexport function fromGemini(geminiClient: GeminiClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          const wantsJson = Boolean(params.response_format);\n          const generationConfig: NonNullable<\n            Parameters<GeminiClient['generateContent']>[0]['generationConfig']\n          > = {\n            temperature: params.temperature,\n            maxOutputTokens: params.max_tokens,\n          };\n\n          if (wantsJson) {\n            generationConfig.responseMimeType = 'application/json';\n          }\n          if (params.response_format?.type === 'json_schema') {\n            generationConfig.responseSchema = params.response_format.json_schema.schema;\n          }\n\n          const response = await geminiClient.generateContent(\n            {\n              model: params.model,\n              contents: userMessages.map((m) => ({\n                role: 'user' as const,\n                parts: [{ text: m.content }],\n              })),\n              systemInstruction: systemMessage\n                ? { parts: [{ text: systemMessage.content }] }\n                : undefined,\n              generationConfig,\n            },\n            options,\n          );\n\n          const text =\n            response.candidates?.[0]?.content?.parts?.map((p) => p.text ?? '').join('') ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usageMetadata?.promptTokenCount,\n              completion_tokens: response.usageMetadata?.candidatesTokenCount,\n              total_tokens: response.usageMetadata?.totalTokenCount,\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Minimal structural type matching AWS Bedrocks Converse API. This is\n * intentionally NOT `BedrockRuntimeClient` itself, the AWS SDK v3 client\n * exposes `.send(command)`, not a direct `.converse()` method, and pulling\n * in `@aws-sdk/client-bedrock-runtime` as a dependency just for its types\n * isn't worth it for a structural adapter. Wrap your client, e.g:\n *\n * ```ts\n * import { BedrockRuntimeClient, ConverseCommand } from '@aws-sdk/client-bedrock-runtime';\n * const client = new BedrockRuntimeClient({ region: 'us-east-1' });\n * const converseClient = {\n *   converse: (params, options) =>\n *     client.send(new ConverseCommand(params), { abortSignal: options.signal }),\n * };\n * ```\n */\nexport interface BedrockConverseClient {\n  converse(\n    params: {\n      modelId: string;\n      messages: Array<{ role: 'user'; content: Array<{ text: string }> }>;\n      system?: Array<{ text: string }>;\n      inferenceConfig?: { temperature?: number; maxTokens?: number };\n    },\n    options: { signal: AbortSignal },\n  ): Promise<{\n    output?: { message?: { content?: Array<{ text?: string }> } };\n    usage?: { inputTokens?: number; outputTokens?: number; totalTokens?: number };\n  }>;\n}\n\n/**\n * Wraps a Bedrock Converse-API client so it satisfies the `LLMClient`\n * interface VernLLM uses for OpenAI/Groq. The Converse API is unified\n * across Bedrocks model families (Anthropic, Titan, Llama, Mistral, etc.),\n * so unlike raw per-model Bedrock invocation, this one adapter works\n * regardless of which underlying model `modelId` points at, as long as\n * that model supports Converse (most current-generation ones do)\n *\n * Theres no uniform native JSON Schema enforcement across families here\n * (some support it via forced tool-use, which varies per model), so\n * `jsonSchema`/`jsonMode` are emulated via a system-prompt instruction, same\n * approach as the Anthropic adapter. `reasoning_effort` has no Converse\n * equivalent and is dropped\n */\nexport function fromBedrock(bedrockClient: BedrockConverseClient): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const systemMessage = params.messages.find((m) => m.role === 'system');\n          const userMessages = params.messages.filter((m) => m.role === 'user');\n\n          let jsonInstruction: string | undefined;\n          if (params.response_format?.type === 'json_schema') {\n            const { name, schema } = params.response_format.json_schema;\n            jsonInstruction = `Respond with valid JSON only, no prose or markdown fences. The JSON must conform to this schema (name: \"${name}\"):\\n${JSON.stringify(schema)}`;\n          } else if (params.response_format?.type === 'json_object') {\n            jsonInstruction = 'Respond with valid JSON only, no prose or markdown fences.';\n          }\n\n          const systemParts = [systemMessage?.content, jsonInstruction].filter((s): s is string =>\n            Boolean(s),\n          );\n\n          const response = await bedrockClient.converse(\n            {\n              modelId: params.model,\n              messages: userMessages.map((m) => ({\n                role: 'user' as const,\n                content: [{ text: m.content }],\n              })),\n              system: systemParts.length ? systemParts.map((text) => ({ text })) : undefined,\n              inferenceConfig: {\n                temperature: params.temperature,\n                maxTokens: params.max_tokens,\n              },\n            },\n            options,\n          );\n\n          const text = response.output?.message?.content?.map((c) => c.text ?? '').join('') ?? '';\n\n          return {\n            choices: [{ message: { content: text } }],\n            usage: {\n              prompt_tokens: response.usage?.inputTokens,\n              completion_tokens: response.usage?.outputTokens,\n              total_tokens: response.usage?.totalTokens,\n            },\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/** The chat-completion-shaped request VernLLM builds internally */\ntype ChatRequest = Parameters<LLMClient['chat']['completions']['create']>[0];\n\nexport interface FetchAdapterConfig {\n  /** Endpoint URL, or a function of the request in case it depends on model/params */\n  url: string | ((params: ChatRequest) => string);\n  /** Static headers, or a function (sync or async) for things like refreshed auth tokens */\n  headers?:\n    | Record<string, string>\n    | (() => Record<string, string> | Promise<Record<string, string>>);\n  /** HTTP method. Default 'POST' */\n  method?: string;\n  /** Maps VernLLMs internal chat-completion request into the providers raw request body */\n  mapRequest: (params: ChatRequest) => unknown;\n  /**\n   * Maps the providers raw JSON response into `{ content, usage? }`\n   * `content` is the assistants text (JSON string when JSON mode was requested)\n   */\n  mapResponse: (json: unknown) => {\n    content: string;\n    usage?: { promptTokens?: number; completionTokens?: number; totalTokens?: number };\n  };\n}\n\n/**\n * A fetch-based escape hatch for providers with no SDK, or where pulling one\n * in isnt worth it. You supply the URL, headers, and two small mapping\n * functions; this handles the HTTP call and slots the result into the same\n * `LLMClient` shape every other adapter produces, so retries, timeouts,\n * the circuit breaker, and JSON/schema handling all still work unmodified\n *\n * Non-2xx responses throw an error with `.status` set to the HTTP status\n * code, so VernLLMs `nonRetryableStatus` handling (e.g. failing fast on\n * 401/403) applies here too\n */\nexport function fromFetch(config: FetchAdapterConfig): LLMClient {\n  return {\n    chat: {\n      completions: {\n        async create(params, options) {\n          const url = typeof config.url === 'function' ? config.url(params) : config.url;\n          const headers =\n            typeof config.headers === 'function' ? await config.headers() : config.headers;\n\n          const res = await fetch(url, {\n            method: config.method ?? 'POST',\n            headers: { 'Content-Type': 'application/json', ...headers },\n            body: JSON.stringify(config.mapRequest(params)),\n            signal: options.signal,\n          });\n\n          if (!res.ok) {\n            const body = await res.text().catch(() => '');\n            const err = new Error(\n              `Fetch adapter request failed (${res.status}): ${body.slice(0, 500)}`,\n            ) as Error & { status?: number };\n            err.status = res.status;\n            throw err;\n          }\n\n          const json = await res.json();\n          const { content, usage } = config.mapResponse(json);\n\n          return {\n            choices: [{ message: { content } }],\n            usage: usage\n              ? {\n                  prompt_tokens: usage.promptTokens,\n                  completion_tokens: usage.completionTokens,\n                  total_tokens: usage.totalTokens,\n                }\n              : undefined,\n          };\n        },\n      },\n    },\n  };\n}\n","import type { LLMClient } from '../types.js';\n\n/**\n * Passthrough adapter for any SDK/client whose `chat.completions.create`\n * already matches the OpenAI wire format 1:1 : this covers most hosted\n * inference providers, since \"OpenAI-compatible\" is a de facto standard for\n * chat completion APIs. No transformation happens here, this exists purely\n * so call sites read clearly (`fromMistral(client)` vs handing a Mistral\n * client to something typed for OpenAI) and so a real transformation could\n * be added later, per-provider, without a breaking change.\n *\n * Not every SDKs own TypeScript types line up exactly with `LLMClient`\n * (extra fields, stricter unions, etc.), so this takes `unknown` and casts:\n * the actual compatibility contract is the JSON each provider sends and\n * receives over the wire, not the SDKs TS types.\n */\nexport function fromOpenAICompatible(client: unknown): LLMClient {\n  return client as LLMClient;\n}\n\n/** Groqs SDK matches the OpenAI wire format */\nexport const fromGroq = fromOpenAICompatible;\n\n/** Mistrals `chat.completions`-shaped client (or their OpenAI-compat endpoint) */\nexport const fromMistral = fromOpenAICompatible;\n\n/** DeepSeeks API is OpenAI-compatible */\nexport const fromDeepSeek = fromOpenAICompatible;\n\n/** Cerebras inference API is OpenAI-compatible */\nexport const fromCerebras = fromOpenAICompatible;\n\n/** Together AIs API is OpenAI-compatible */\nexport const fromTogether = fromOpenAICompatible;\n\n/** Fireworks AIs API is OpenAI-compatible */\nexport const fromFireworks = fromOpenAICompatible;\n\n/**\n * Ollama exposes an OpenAI-compatible endpoint at `/v1/chat/completions`\n * (as opposed to its native `/api/chat` format, which differs). Point an\n * OpenAI SDK instances `baseURL` at your Ollama server and pass it here:\n * this does not talk to Ollamas native API directly.\n */\nexport const fromOllama = fromOpenAICompatible;\n"],"mappings":";;;AASA,IAAa,WAAb,cAA8B,MAAM;CAClC,YACEA,SACOC,MACAC,QACAC,QACP;AACA,QAAM,QAAQ;EA2LjB,KA/LU;EA+LT,KA9LS;EA8LR,KA7LQ;AAGP,OAAK,OAAO;CACb;AACF;AAED,SAAgB,WAAWC,KAA+B;AACxD,QAAO,eAAe;AACvB;;;;;AAWD,IAAa,uBAAb,MAA0E;CACxE,AAAQ,QAAQ,IAAI;CAEpB,MAAM,IAAIC,KAAgC;EACxC,MAAM,QAAQ,KAAK,MAAM,IAAI,IAAI;AACjC,OAAK,MAAO,QAAO;AACnB,MAAI,KAAK,KAAK,GAAG,MAAM,WAAW;AAChC,QAAK,MAAM,OAAO,IAAI;AACtB,UAAO;EACR;AACD,SAAO,MAAM;CACd;CAED,MAAM,IAAIA,KAAaC,OAAUC,KAA4B;AAC3D,OAAK,MAAM,IAAI,KAAK;GAAE;GAAO,WAAW,KAAK,KAAK,GAAG,MAAM;EAAM,EAAC;CACnE;AACF;;;;;;;;;;ACjCD,IAAa,iBAAb,MAA4B;CAC1B,AAAQ,QAAsB;CAC9B,AAAQ,sBAAsB;CAC9B,AAAQ,WAAW;CACnB,AAAQ;CACR,AAAQ;CAER,YAAYC,UAAiC,CAAE,GAAE;AAC/C,OAAK,YAAY,QAAQ,aAAa;AACtC,OAAK,aAAa,QAAQ,cAAc;CACzC;;CAGD,eAAqB;AACnB,MAAI,KAAK,UAAU,OAAQ;EAE3B,MAAM,UAAU,KAAK,KAAK,GAAG,KAAK;AAClC,MAAI,WAAW,KAAK,YAAY;AAC9B,QAAK,QAAQ;AACb;EACD;AAED,QAAM,IAAI,UACP,qCAAqC,KAAK,oBAAoB,4BAA4B,KAAK,MAAM,KAAK,aAAa,WAAW,IAAK,CAAC,KACzI;CAEH;CAED,gBAAsB;AACpB,OAAK,sBAAsB;AAC3B,OAAK,QAAQ;CACd;CAED,gBAAsB;AACpB,OAAK,uBAAuB;AAE5B,MAAI,KAAK,UAAU,aAAa;AAE9B,QAAK,QAAQ;AACb,QAAK,WAAW,KAAK,KAAK;AAC1B;EACD;AAED,MAAI,KAAK,uBAAuB,KAAK,WAAW;AAC9C,QAAK,QAAQ;AACb,QAAK,WAAW,KAAK,KAAK;EAC3B;CACF;CAED,WAAyB;AACvB,SAAO,KAAK;CACb;AACF;;;;ACnED,SAAgB,iBAAiBC,SAA0B;AACzD,KAAI;AACF,SAAO,KAAK,MAAM,QAAQ;CAC3B,QAAO;AACN;CACD;AACF;;;;;;;AAQD,SAAgB,cAAcC,KAAkC;AAC9D,MAAK,cAAc,QAAQ,SAAU;CAErC,MAAM,QAAQ;AAKd,YAAW,MAAM,WAAW,SAAU,QAAO,MAAM;AACnD,YAAW,MAAM,eAAe,SAAU,QAAO,MAAM;AAEvD;AACD;;;;;;;;;;;;AAaD,eAAsB,YACpBC,IACAC,WACAC,gBACY;CACZ,MAAM,aAAa,IAAI;CAEvB,MAAM,QAAQ,WAAW,MAAM;AAC7B,aAAW,OAAO;CACnB,GAAE,UAAU;CAEb,MAAM,SAAS,iBACX,YAAY,IAAI,CAAC,gBAAgB,WAAW,MAAO,EAAC,GACpD,WAAW;AAEf,KAAI;AACF,SAAO,MAAM,GAAG,OAAO;CACxB,SAAQ,KAAK;AACZ,MACE,WAAW,OAAO,YACjB,gBAAgB,WACjB,eAAe,gBACf,IAAI,SAAS,aAEb,OAAM,IAAI,SAAS,qBAAqB;AAG1C,QAAM;CACP,UAAS;AACR,eAAa,MAAM;CACpB;AACF;;;;;;AAOD,SAAgB,gBAAgBC,aAAqBC,SAAiB,aAAa,KAAgB;CACjG,MAAM,MAAM,KAAK,IAAI,cAAc,KAAK,SAAS,WAAW;AAC5D,QAAO,MAAM,IAAI,KAAK,QAAQ,IAAI,MAAM;AACzC;;;;;;;;AASD,eAAsB,aAAaC,OAAeC,QAAqC;AACrF,OAAM,IAAI,QAAc,CAAC,SAAS,WAAW;EAC3C,MAAM,UAAU,MAAM;AACpB,gBAAa,MAAM;AACnB,UAAO,IAAI,SAAS,qBAAqB,WAAW;EACrD;EAED,MAAM,QAAQ,WAAW,MAAM;AAC7B,WAAQ,oBAAoB,SAAS,QAAQ;AAC7C,YAAS;EACV,GAAE,MAAM;AAET,UAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAM,EAAC;CAC3D;AACF;;;;;;;;AC/FD,IAAa,gBAAb,MAA6C;CAC3C,YAAoBC,cAAuB;EAc5C,KAdqB;CAAyB;CAE7C,MAAMC,SAAuB;AAC3B,MAAI,KAAK,aAAc,SAAQ,MAAM,QAAQ;CAC9C;CAED,KAAKA,SAAuB;AAC1B,UAAQ,KAAK,QAAQ;CACtB;CAED,MAAMA,SAAiBC,MAAsC;AAC3D,UAAQ,MAAM,SAAS,QAAQ,GAAG;CACnC;AACF;;;;;;;;;;;;ACKD,IAAa,UAAb,MAAqB;CACnB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;CAEjB,AAAiB;CACjB,AAAiB;;;;;;CAOjB,YAAYC,SAAyB;AACnC,OAAK,SAAS,QAAQ;AACtB,OAAK,QAAQ,QAAQ;EAErB,MAAM,cAAc,KAAK,mBAAmB,QAAQ;AACpD,OAAK,aAAa,YAAY;AAC9B,OAAK,YAAY,YAAY;AAC7B,OAAK,cAAc,YAAY;AAC/B,OAAK,mBAAmB,YAAY;AAEpC,OAAK,QAAQ,QAAQ,SAAS,IAAI;AAClC,OAAK,qBAAqB,QAAQ,sBAAsB;GAAC;GAAK;GAAK;EAAI;AAEvE,OAAK,YAAY,QAAQ,aAAa;AACtC,OAAK,UAAU,QAAQ;AAEvB,OAAK,SAAS,KAAK,cAAc,QAAQ;AACzC,OAAK,UAAU,KAAK,sBAAsB,QAAQ;CACnD;;;;;CAMD,AAAQ,mBAAmBA,SAAyB;AAClD,SAAO;GACL,YAAY,QAAQ,cAAc;GAClC,WAAW,QAAQ,aAAa;GAChC,aAAa,QAAQ,eAAe;GACpC,kBAAkB,QAAQ,oBAAoB;EAC/C;CACF;;;;;;CAOD,AAAQ,cAAcA,SAAiC;AACrD,SACE,QAAQ,UAAU,IAAI,cAAc,QAAQ,SAAS,QAAQ,IAAI,aAAa;CAEjF;;;;;;CAOD,AAAQ,sBAAsBA,SAAqD;AACjF,OAAK,QAAQ,eAAgB;AAE7B,SAAO,IAAI,eAAe,QAAQ,mBAAmB,gBAAmB,QAAQ;CACjF;;;;;;;;;;;;;;;CAgBD,MAAM,KAAkBC,QAAmC;AACzD,OAAK,SAAS,cAAc;AAE5B,MAAI,OAAO,QAAQ,QACjB,OAAM,IAAI,SAAS,uBAAuB;EAG5C,MAAM,YAAY,OAAO,aAAa,YAAY;AAElD,MAAI;GACF,MAAM,SAAS,MAAM,KAAK,iBACxB,MAAM,KAAK,YAAY,QAAQ,UAAU,EACzC,WACA,OAAO,OACR;AAED,UAAO;EACR,SAAQ,OAAO;AACd,QAAK,SAAS,eAAe;AAC7B,SAAM,KAAK,eAAe,OAAO,OAAO,OAAO;EAChD;CACF;;;;;;CAOD,MAAc,iBACZC,IACAC,WACAC,QACY;EACZ,IAAIC;AAEJ,OAAK,IAAI,UAAU,GAAG,WAAW,KAAK,YAAY,UAChD,KAAI;AACF,OAAI,UAAU,EACZ,OAAM,KAAK,aAAa,WAAW,SAAS,OAAO;AAGrD,UAAO,MAAM,IAAI;EAClB,SAAQ,OAAO;AACd,eAAY;AAEZ,QAAK,KAAK,YAAY,OAAO,OAAO,CAClC;EAEH;AAGH,QAAM;CACP;;;;;;;CAQD,AAAQ,eAAeC,OAAgBF,QAAgC;AACrE,MAAI,QAAQ,QACV,QAAO,IAAI,SAAS,uBAAuB;AAG7C,MAAI,iBAAiB,SACnB,QAAO;EAGT,MAAM,SAAS,cAAc,MAAM;AAEnC,MAAI,kBACF,QAAO,IAAI,SAAS,sBAAsB,OAAO;AAGnD,SAAO,IAAI,SAAS,sBAAsB;CAC3C;;;;;;;CAQD,MAAc,YAAeH,QAAuBE,WAA+B;EACjF,MAAM,EAAE,SAAS,OAAO,SAAS,GAAG,KAAK,oBAAoB,OAAO;EAEpE,MAAM,WAAW,MAAM,YACrB,CAAC,kBAAkB,KAAK,OAAO,KAAK,YAAY,OAAO,SAAS,EAAE,QAAQ,cAAe,EAAC,EAC1F,KAAK,WACL,OAAO,OACR;EAED,MAAM,UAAU,SAAS,UAAU,IAAI,SAAS,SAAS,MAAM;AAE/D,OAAK,QACH,OAAM,IAAI,SAAS,sBAAsB;AAG3C,OAAK,OAAO,OAAO,QAAQ,UAAU,aAAa,QAAQ,MAAM,GAAG,IAAI,CAAC,EAAE;AAE1E,OAAK,YAAY,UAAU,WAAW,MAAM;AAE5C,OAAK,SAAS,eAAe;AAE7B,OAAK,QACH,QAAO;AAGT,SAAO,KAAK,iBAAiB,SAAS,OAAO,OAAO;CACrD;;;;;;;CAQD,AAAQ,oBAAuBF,QAAuB;EACpD,MAAM,EACJ,cACA,aACA,cAAc,IACd,WAAW,MACX,YAAY,KAAK,kBACjB,QAAQ,KAAK,OACb,iBACA,YACD,GAAG;EAEJ,MAAM,UAAU,YAAY,QAAQ,WAAW;EAC/C,MAAM,iBAAiB,KAAK,oBAAoB,YAAY,QAAQ;EAEpE,MAAM,UAAU;GACd;GACA;GACA,YAAY;GACZ,GAAI,iBAAiB,EAAE,iBAAiB,eAAgB,IAAG,CAAE;GAC7D,GAAI,kBAAkB,EAAE,kBAAkB,gBAAiB,IAAG,CAAE;GAChE,UAAU,CACR;IAAE,MAAM;IAAmB,SAAS;GAAc,GAClD;IAAE,MAAM;IAAiB,SAAS;GAAa,CAChD;EACF;AAED,SAAO;GAAE;GAAS;GAAO;EAAS;CACnC;;;;;;;;CASD,AAAQ,oBAAoBM,YAA+CC,SAAkB;AAC3F,MAAI,WACF,QAAO;GACL,MAAM;GACN,aAAa;IACX,MAAM,WAAW;IACjB,QAAQ,WAAW;IACnB,QAAQ,WAAW,UAAU;IAC7B,aAAa,WAAW;GACzB;EACF;AAGH,SAAO,UAAU,EAAE,MAAM,cAAwB;CAClD;;;;;;CAOD,AAAQ,YACNC,UACAN,WACAO,OACM;AACN,OAAK,SAAS,UAAU,KAAK,QAAS;AAEtC,OAAK,QAAQ;GACX,cAAc,SAAS,MAAM,iBAAiB;GAC9C,kBAAkB,SAAS,MAAM,qBAAqB;GACtD,aAAa,SAAS,MAAM,gBAAgB;GAC5C;GACA;EACD,EAAC;CACH;;;;;;;CAQD,AAAQ,iBAAoBC,SAAiBC,QAAqC;EAChF,IAAIC;AAEJ,MAAI;AACF,YAAS,KAAK,UAAU,QAAQ;EACjC,QAAO;AACN,SAAM,IAAI,SAAS,yBAAyB;EAC7C;AAED,MAAI,WAAW,QAAQ,kBACrB,OAAM,IAAI,SAAS,yBAAyB;AAG9C,OAAK,OACH,QAAO;EAGT,MAAM,SAAS,OAAO,UAAU,OAAO;AAEvC,OAAK,OAAO,QACV,OAAM,IAAI,SAAS,4BAA4B,sBAAyB,OAAO;AAGjF,SAAO,OAAO;CACf;;;;;;CAOD,MAAc,aAAaV,WAAmBW,SAAiBV,QAAsB;EACnF,MAAM,QAAQ,gBAAgB,KAAK,aAAa,QAAQ;AAExD,OAAK,OAAO,MACT,QAAQ,UAAU,qBAAqB,QAAQ,GAAG,KAAK,WAAW,YAAY,MAAM,IACtF;AAED,QAAM,aAAa,OAAO,OAAO;CAClC;;;;;;;;CASD,AAAQ,YAAYE,OAAgBF,QAA+B;AACjE,MAAI,QAAQ,QACV,QAAO;AAGT,MAAI,iBAAiB,aAAa,MAAM,SAAS,WAAW,MAAM,SAAS,cACzE,QAAO;EAGT,MAAM,SAAS,cAAc,MAAM;AAEnC,MAAI,qBAAwB,KAAK,mBAAmB,SAAS,OAAO,CAClE,QAAO;AAGT,SAAO;CACR;;;;;;;CAQD,MAAM,WAAcW,QAAyC;EAC3D,MAAM,SAAS,MAAM,KAAK,MAAM,IAAI,OAAO,SAAS;AAEpD,MAAI,WAAW,KACb,QAAO;AAGT,MAAI;AACF,SAAM,OAAO,gBAAgB;GAE7B,MAAM,SAAS,MAAM,OAAO,IAAI;AAEhC,OAAI;AACF,UAAM,KAAK,MAAM,IAAI,OAAO,UAAU,QAAQ,OAAO,IAAI;GAC1D,SAAQ,OAAO;AACd,SAAK,OAAO,MAAM,gCAAgC,EAChD,SAAS,iBAAiB,QAAQ,MAAM,UAAU,UACnD,EAAC;GACH;AAED,UAAO;EACR,SAAQ,OAAO;AACd,OAAI;AACF,UAAM,OAAO,eAAe;GAC7B,SAAQ,aAAa;AACpB,SAAK,OAAO,MAAM,gCAAgC,EAChD,SAAS,uBAAuB,QAAQ,YAAY,UAAU,UAC/D,EAAC;GACH;AAED,SAAM;EACP;CACF;;;;;;CAOD,MAAM,cACJC,QACY;EACZ,MAAM,EAAE,MAAM,WAAY,GAAG,aAAa,GAAG;AAE7C,SAAO,KAAK,WAAW;GACrB,GAAG;GACH,IAAI,MAAM,KAAK,KAAK,WAAW;EAChC,EAAC;CACH;;;;;CAMD,kBAAkB;AAChB,SAAO,KAAK,SAAS,UAAU;CAChC;AACF;;;;;;;;;;;ACpaD,SAAgB,cAAcC,iBAA6C;AACzE,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAOrE,IAAIC;AACJ,MAAI,OAAO,iBAAiB,SAAS,eAAe;GAClD,MAAM,EAAE,MAAM,QAAQ,GAAG,OAAO,gBAAgB;AAChD,sBAAmB,0GAA0G,KAAK,OAAO,KAAK,UAAU,OAAO,CAAC;EACjK,WAAU,OAAO,iBAAiB,SAAS,cAC1C,mBAAkB;EAOpB,MAAM,SAAS,CAAC,eAAe,SAAS,eAAgB,EAAC,OAAO,QAAQ,CAAC,KAAK,OAAO;EAErF,MAAM,WAAW,MAAM,gBAAgB,SAAS,OAC9C;GACE,OAAO,OAAO;GACd,YAAY,OAAO;GACnB,aAAa,OAAO;GACpB,QAAQ;GACR,UAAU,aAAa,IAAI,CAAC,OAAO;IAAE,MAAM;IAAiB,SAAS,EAAE;GAAS,GAAE;EACnF,GACD,QACD;EAED,MAAM,OAAO,SAAS,QAAQ,KAAK,CAAC,UAAU,MAAM,SAAS,OAAO,EAAE,QAAQ;AAE9E,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,OAAO;IAC/B,mBAAmB,SAAS,OAAO;IACnC,eACG,SAAS,OAAO,gBAAgB,MAAM,SAAS,OAAO,iBAAiB;GAC3E;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;ACtCD,SAAgB,WAAWC,cAAuC;AAChE,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAErE,MAAM,YAAY,QAAQ,OAAO,gBAAgB;EACjD,MAAMC,mBAEF;GACF,aAAa,OAAO;GACpB,iBAAiB,OAAO;EACzB;AAED,MAAI,UACF,kBAAiB,mBAAmB;AAEtC,MAAI,OAAO,iBAAiB,SAAS,cACnC,kBAAiB,iBAAiB,OAAO,gBAAgB,YAAY;EAGvE,MAAM,WAAW,MAAM,aAAa,gBAClC;GACE,OAAO,OAAO;GACd,UAAU,aAAa,IAAI,CAAC,OAAO;IACjC,MAAM;IACN,OAAO,CAAC,EAAE,MAAM,EAAE,QAAS,CAAC;GAC7B,GAAE;GACH,mBAAmB,gBACf,EAAE,OAAO,CAAC,EAAE,MAAM,cAAc,QAAS,CAAC,EAAE;GAEhD;EACD,GACD,QACD;EAED,MAAM,OACJ,SAAS,aAAa,IAAI,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,QAAQ,GAAG,CAAC,KAAK,GAAG,IAAI;AAEjF,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,eAAe;IACvC,mBAAmB,SAAS,eAAe;IAC3C,cAAc,SAAS,eAAe;GACvC;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;;;;ACjDD,SAAgB,YAAYC,eAAiD;AAC3E,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,gBAAgB,OAAO,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS;EACtE,MAAM,eAAe,OAAO,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO;EAErE,IAAIC;AACJ,MAAI,OAAO,iBAAiB,SAAS,eAAe;GAClD,MAAM,EAAE,MAAM,QAAQ,GAAG,OAAO,gBAAgB;AAChD,sBAAmB,0GAA0G,KAAK,OAAO,KAAK,UAAU,OAAO,CAAC;EACjK,WAAU,OAAO,iBAAiB,SAAS,cAC1C,mBAAkB;EAGpB,MAAM,cAAc,CAAC,eAAe,SAAS,eAAgB,EAAC,OAAO,CAAC,MACpE,QAAQ,EAAE,CACX;EAED,MAAM,WAAW,MAAM,cAAc,SACnC;GACE,SAAS,OAAO;GAChB,UAAU,aAAa,IAAI,CAAC,OAAO;IACjC,MAAM;IACN,SAAS,CAAC,EAAE,MAAM,EAAE,QAAS,CAAC;GAC/B,GAAE;GACH,QAAQ,YAAY,SAAS,YAAY,IAAI,CAACC,YAAU,EAAE,aAAM,GAAE;GAClE,iBAAiB;IACf,aAAa,OAAO;IACpB,WAAW,OAAO;GACnB;EACF,GACD,QACD;EAED,MAAM,OAAO,SAAS,QAAQ,SAAS,SAAS,IAAI,CAAC,MAAM,EAAE,QAAQ,GAAG,CAAC,KAAK,GAAG,IAAI;AAErF,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,SAAS,KAAM,EAAE,CAAC;GACzC,OAAO;IACL,eAAe,SAAS,OAAO;IAC/B,mBAAmB,SAAS,OAAO;IACnC,cAAc,SAAS,OAAO;GAC/B;EACF;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;AC5DD,SAAgB,UAAUC,QAAuC;AAC/D,QAAO,EACL,MAAM,EACJ,aAAa,EACX,MAAM,OAAO,QAAQ,SAAS;EAC5B,MAAM,aAAa,OAAO,QAAQ,aAAa,OAAO,IAAI,OAAO,GAAG,OAAO;EAC3E,MAAM,iBACG,OAAO,YAAY,aAAa,MAAM,OAAO,SAAS,GAAG,OAAO;EAEzE,MAAM,MAAM,MAAM,MAAM,KAAK;GAC3B,QAAQ,OAAO,UAAU;GACzB,SAAS;IAAE,gBAAgB;IAAoB,GAAG;GAAS;GAC3D,MAAM,KAAK,UAAU,OAAO,WAAW,OAAO,CAAC;GAC/C,QAAQ,QAAQ;EACjB,EAAC;AAEF,OAAK,IAAI,IAAI;GACX,MAAM,OAAO,MAAM,IAAI,MAAM,CAAC,MAAM,MAAM,GAAG;GAC7C,MAAM,MAAM,IAAI,OACb,gCAAgC,IAAI,OAAO,KAAK,KAAK,MAAM,GAAG,IAAI,CAAC;AAEtE,OAAI,SAAS,IAAI;AACjB,SAAM;EACP;EAED,MAAM,OAAO,MAAM,IAAI,MAAM;EAC7B,MAAM,EAAE,SAAS,OAAO,GAAG,OAAO,YAAY,KAAK;AAEnD,SAAO;GACL,SAAS,CAAC,EAAE,SAAS,EAAE,QAAS,EAAE,CAAC;GACnC,OAAO,QACH;IACE,eAAe,MAAM;IACrB,mBAAmB,MAAM;IACzB,cAAc,MAAM;GACrB;EAEN;CACF,EACF,EACF,EACF;AACF;;;;;;;;;;;;;;;;;;AC/DD,SAAgB,qBAAqBC,QAA4B;AAC/D,QAAO;AACR;;AAGD,MAAa,WAAW;;AAGxB,MAAa,cAAc;;AAG3B,MAAa,eAAe;;AAG5B,MAAa,eAAe;;AAG5B,MAAa,eAAe;;AAG5B,MAAa,gBAAgB;;;;;;;AAQ7B,MAAa,aAAa"}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "vern-llm",
-  "version": "0.2.1",
-  "description": "Retry + timeout + cache wrapper for OpenAI-compatible LLM chat completion calls (OpenAI, Groq, etc).",
+  "version": "0.3.0",
+  "description": "Retry, timeout, caching, and circuit breaker for OpenAI-compatible LLM calls, Groq, Anthropic, Gemini, Bedrock, and more.",
   "type": "module",
   "main": "./dist/index.cjs",
   "module": "./dist/index.js",
@@ -39,7 +39,8 @@
   },
   "repository": {
     "type": "git",
-    "url": "https://github.com/LakBud/vernLLM"
+    "url": "https://github.com/LakBud/vernLLM",
+    "directory": "packages/vern-llm"
   },
   "license": "MIT",
   "publishConfig": {