granthi-review/src/lib/router.js

// shre-router client. POST /v1/chat with stream:false.
// Real model answer lives in .message.content; top-level .content can carry an
// upstream-down apology — always prefer .message.content.

export async function chat({ routerUrl, model, messages, agentId, sessionId, tenantId, timeoutMs }) {
  const res = await fetch(`${routerUrl.replace(/\/$/, '')}/v1/chat`, {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    // tools:false + raw:true — reviews need pure model reasoning. raw:true is
    // the router's sanctioned _bypassToolRouting flag: without it, keyword
    // fast-paths (current-info briefing, store-resolver, retail dispatch)
    // hijack prompts whose DIFF CONTENT mentions github/sales/store-ish terms
    // and return non-review text. The domain preflight (agent_mismatch 400)
    // still applies per agent tier; the pipeline's attempt-2 fallback agent
    // covers that.
    body: JSON.stringify({ model, messages, agentId, sessionId, tenantId, stream: false, tools: false, raw: true }),
    signal: AbortSignal.timeout(timeoutMs)
  });
  if (!res.ok) {
    const text = await res.text().catch(() => '');
    throw new Error(`router -> ${res.status}: ${text.slice(0, 300)}`);
  }
  const data = await res.json();
  // Response shape varies by upstream: prefer .message.content, then
  // OpenAI-style choices, then top-level .content (which can carry an
  // upstream-down apology — hence last).
  const content = data?.message?.content
    ?? data?.choices?.[0]?.message?.content
    ?? data?.content;
  if (typeof content !== 'string' || !content.trim()) {
    throw new Error('router returned empty content');
  }
  return { content, servedModel: data?._shre?.model ?? null };
}