OmniRoute/open-sse/executors/xai.ts

import { BaseExecutor, type ProviderCredentials } from "./base.ts";
import { PROVIDERS } from "../config/constants.ts";
import { getModelTargetFormat } from "../config/providerModels.ts";

type JsonRecord = Record<string, unknown>;

/**
 * xAI/Grok model ids (open-sse/config/providers/registry/xai/index.ts) that accept
 * a graduated `reasoning_effort`. Kept narrow and reconciled against the REAL
 * catalog rather than upstream's example ids (grok-4/grok-3 do not exist here):
 *   - grok-4.3                  — current-generation flagship, reasoning-capable.
 *   - grok-4.20-0309-reasoning  — explicit reasoning variant.
 *
 * grok-4.20-multi-agent-0309 is intentionally left unclassified (neither allow
 * nor deny): its reasoning support is not documented in the local catalog, so
 * we pass it through unchanged rather than guess.
 */
const REASONING_ALLOWED = ["grok-4.3", "grok-4.20-0309-reasoning"];

/**
 * Model ids that reject `reasoning_effort` outright:
 *   - grok-build-0.1                — build/tool-oriented model, no reasoning mode.
 *   - grok-4.20-0309-non-reasoning   — already encodes "no reasoning" in the id;
 *     forwarding reasoning_effort here would be redundant/rejected upstream.
 */
const REASONING_DENIED = ["grok-build-0.1", "grok-4.20-0309-non-reasoning"];

/** `-{level}` suffixes some clients append to a model id to select reasoning intensity. */
const EFFORT_SUFFIXES = ["low", "medium", "high", "xhigh"] as const;

function asRecord(value: unknown): JsonRecord | null {
  return value && typeof value === "object" && !Array.isArray(value) ? (value as JsonRecord) : null;
}

/**
 * xAI/Grok executor (port of decolua/9router#2147).
 *
 * Some Grok clients select reasoning intensity via a `-{low,medium,high,xhigh}`
 * suffix on the model id (e.g. `grok-4.3-high`) rather than a native
 * `reasoning_effort` field — xAI itself does not recognize the suffixed id.
 * This executor:
 *   1. Parses and strips that suffix off the model id before the request
 *      reaches xAI, mapping it to `reasoning_effort` for allow-listed models.
 *   2. Strips any `reasoning_effort` for deny-listed models — including ids
 *      that already encode their reasoning state in the name (`-reasoning` /
 *      `-non-reasoning`), which must not be double-mutated by also stacking a
 *      `reasoning_effort` field on top of what the id already declares.
 *   3. Leaves unclassified models and bodies untouched otherwise.
 */
export class XaiExecutor extends BaseExecutor {
  constructor() {
    super("xai", PROVIDERS.xai);
  }

  /**
   * Port of decolua/9router#2439 (author: @ryanngit): xAI ships a native
   * `/v1/responses` endpoint alongside `/v1/chat/completions`. Models tagged
   * `targetFormat: "openai-responses"` in the registry (currently
   * grok-4.20-multi-agent-0309, per upstream) resolve to that endpoint instead
   * of the default chat-completions bridge. The per-model registry tag is the
   * single source of truth — it also drives chatCore's body translation — so
   * the URL stays in lockstep with the translated body, mirroring the gh
   * executor's targetFormat-driven routing (9router#102) and the "openai"
   * -pro heuristic in open-sse/executors/default.ts.
   */
  buildUrl(model: string, _stream: boolean, _urlIndex = 0) {
    if (getModelTargetFormat("xai", model) === "openai-responses") {
      return this.config.responsesBaseUrl || this.config.baseUrl;
    }
    return this.config.baseUrl;
  }

  transformRequest(
    model: string,
    body: unknown,
    stream: boolean,
    credentials: ProviderCredentials
  ): unknown {
    const cleaned = super.transformRequest(model, body, stream, credentials);
    const record = asRecord(cleaned);
    if (!record) return cleaned;

    const out: JsonRecord = { ...record };
    let modelId = typeof out.model === "string" ? out.model : model;

    let suffixEffort: string | null = null;
    for (const level of EFFORT_SUFFIXES) {
      const suffix = `-${level}`;
      if (modelId.endsWith(suffix)) {
        suffixEffort = level;
        modelId = modelId.slice(0, -suffix.length);
        break;
      }
    }
    if (suffixEffort && typeof out.model === "string") {
      out.model = modelId;
    }

    const isDenied = REASONING_DENIED.some((id) => modelId.includes(id));
    const isAllowed = REASONING_ALLOWED.some((id) => modelId.includes(id));

    if (isDenied) {
      delete out.reasoning_effort;
    } else if (isAllowed) {
      const effort = suffixEffort || out.reasoning_effort;
      if (effort) out.reasoning_effort = effort;
    }

    return out;
  }
}

export default XaiExecutor;