import type { Brain } from "../core/types.js";
import { type StreamEngineConfig } from "./stream-engine.js";
export interface OpenAIBrainConfig extends StreamEngineConfig {
    /** Base URL of your OpenAI-compatible gateway (without /chat/completions). Overridden by model.baseUrl if set. */
    baseUrl?: string;
    /** API key. Overridden by per-call options.apiKey / getApiKeyAndHeaders. */
    apiKey?: string;
    /** Extra headers merged into every request. */
    headers?: Record<string, string>;
    /** Inject a custom fetch (tests / proxies). Defaults to global fetch. */
    fetchImpl?: typeof fetch;
    /**
     * For weaker models that emit tool calls as TEXT (`<tool_call>{...}</tool_call>` / fenced json)
     * instead of native tool calls: promote them to real tool calls. Off by default. Native models
     * don't need this.
     */
    repairTextToolCalls?: boolean;
    /**
     * Replay assistant `reasoning_content` back to the provider on follow-up requests (design/46).
     * **On by default**, and **conditional**: thinking is replayed only on prior assistant **tool-call**
     * turns — DeepSeek V4 **requires** it there (else it errors/degrades), while non-tool turns ignore it,
     * so replaying there only bloats the context window (prefix-cache cuts the *cost* of that bloat but not
     * the *window occupancy* → earlier compaction). Replaying just the tool-call turns is correctness-
     * equivalent and leaner. Set `false` to never replay (e.g. a provider that rejects the field).
     * Anthropic uses `createAnthropicBrain`, which always replays thinking blocks with their signature.
     */
    replayThinking?: boolean;
    /**
     * Detect **degenerate repetition** (the model looping on the same char/phrase forever) mid-stream
     * and cut it off — the turn ends `error` ("degenerate repetition") so the loop stops paying and a
     * team member's looped output is treated as a failed turn, not a real statement. Default true.
     */
    detectRepetition?: boolean;
}
/**
 * Create a Brain that talks to an OpenAI-compatible chat-completions endpoint (your model gateway).
 * Translates the unified Context into an OpenAI request and the streamed response back into the unified
 * AssistantMessageEvent protocol, including native tool calls. The connect/retry/timeout/read-loop
 * machinery lives in the shared {@link runStreamingBrain} engine (design/32); this adapter owns the
 * OpenAI request shaping + chat-completions SSE parsing + finalization.
 */
export declare function createOpenAIBrain(config?: OpenAIBrainConfig): Brain;
//# sourceMappingURL=openai.d.ts.map