SkateSpots/src/lib/sasquatchBrain.ts

/**
 * Sasquatch brain — the shared AI caller used by both the global Sasquatch
 * guide and the per-crew Crew Agent. Kept framework-agnostic (no React) so any
 * surface can reuse the same model list + fallback behaviour.
 */

import { sasquatchAiBase } from '@/lib/sasquatchAi';

// ─── AI models, via our own server-side proxy ────────────────────────────────
// SECURITY: no API key is ever shipped to the client. Requests go to our proxy
// (`/api/ai`), which holds the Routstr key server-side and only allows a curated
// model list. Where that proxy lives is decided in one place — lib/sasquatchAi.ts
// (the Netlify Function by default, or our own ai-proxy/ service on the CVM host
// via VITE_SASQUATCH_AI_BASE). If it's unavailable we fall back to the keyless
// Pollinations provider below, so Sasquatch still works with zero secrets shipped.
const AI_PROXY_URL = sasquatchAiBase();

// Models Sasquatch requests through the proxy, in preference order. These are
// the slugs served by our upstream (a Routstr node — see AGENTS.md), which are
// OpenAI-style ids rather than OpenRouter's. If one is rate-limited (429) we back
// off briefly then try the next; Pollinations is the keyless fallback.
// The server is authoritative: keep this list in sync with `SASQUATCH_MODELS`
// (and ALLOWED_MODELS in netlify/functions/ai.mjs) — but it's not fatal if they
// drift, because the proxy substitutes its first allowed model.
export const PROXY_MODELS = [
  'gemini-3-6-flash',
  'z-ai-glm-5-3',
  'gemini-3-7-flash',
];

export type ChatTurn = { role: 'system' | 'user' | 'assistant'; content: string };

const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));

// Hard caps so a hung provider can never wedge the chat: every HTTP request is
// aborted after PER_REQUEST_TIMEOUT_MS, and the whole multi-provider attempt
// gives up once TOTAL_TIMEOUT_MS is spent (the caller then falls back to the
// static FAQ / glossary / local data paths).
const PER_REQUEST_TIMEOUT_MS = 15000;
const TOTAL_TIMEOUT_MS = 35000;

// ─── Backup provider — Pollinations (keyless, OpenAI-compatible) ─────────────
// Independent of the proxy, so if the proxy (or the key behind it) is down or
// rate-limited we still have a second brain to fall back to. No API key needed.
const POLLINATIONS_URL = 'https://text.pollinations.ai/openai';
// `mistral` was removed upstream (404 on the legacy endpoint), so keep only
// live slugs. `openai` currently maps to gpt-oss-20b server-side.
const POLLINATIONS_MODELS = ['openai'];

// Only try the top few models before failing over to the keyless backup.
// Grinding through the whole list is the main cause of a "why is this so slow"
// wait; the first couple answer the vast majority of the time.
const MAX_PROXY_MODELS = 3;

/** Try our AI proxy (OpenRouter free models), respecting 429 backoff. */
async function tryProxy(
  messages: ChatTurn[],
  maxTokens: number,
  remainingMs: () => number,
  authHeaders: Record<string, string> = {},
): Promise<string> {
  const models = PROXY_MODELS.slice(0, MAX_PROXY_MODELS);
  for (const model of models) {
    if (remainingMs() <= 1000) return '';
    // Single 429 retry per model with a short, capped backoff — a rate-limited
    // model rarely frees up quickly, so we'd rather move on than wait it out.
    for (let attempt = 0; attempt < 2; attempt++) {
      try {
        const left = remainingMs();
        if (left <= 1000) return '';
        const response = await fetch(AI_PROXY_URL, {
          method: 'POST',
          headers: { 'Content-Type': 'application/json', ...authHeaders },
          body: JSON.stringify({ model, messages, max_tokens: maxTokens }),
          signal: AbortSignal.timeout(Math.min(PER_REQUEST_TIMEOUT_MS, left)),
        });

        if (response.status === 429 && attempt === 0) {
          const retryAfter = Number(response.headers.get('retry-after'));
          const waitMs = Number.isFinite(retryAfter) && retryAfter > 0
            ? Math.min(retryAfter * 1000, 800)
            : 400;
          console.warn(`[Sasquatch] ${model} rate-limited, retrying in ${waitMs}ms`);
          await sleep(waitMs);
          continue;
        }

        if (!response.ok) {
          console.warn(`[Sasquatch] AI proxy ${model} failed ${response.status}`);
          break;
        }

        const data = await response.json();
        const content = data.choices?.[0]?.message?.content ?? '';
        if (content) return content;
        break;
      } catch (err) {
        console.warn(`[Sasquatch] AI proxy ${model} threw:`, err);
        break;
      }
    }
  }
  return '';
}

/** Try the keyless Pollinations backup. Returns text or ''. */
async function tryPollinations(
  messages: ChatTurn[],
  maxTokens: number,
  remainingMs: () => number,
): Promise<string> {
  for (const model of POLLINATIONS_MODELS) {
    try {
      const left = remainingMs();
      if (left <= 1000) return '';
      const response = await fetch(POLLINATIONS_URL, {
        method: 'POST',
        headers: { 'Content-Type': 'application/json' },
        body: JSON.stringify({ model, messages, max_tokens: maxTokens }),
        signal: AbortSignal.timeout(Math.min(PER_REQUEST_TIMEOUT_MS, left)),
      });
      if (!response.ok) {
        console.warn(`[Sasquatch] Pollinations ${model} failed ${response.status}`);
        continue;
      }
      const data = await response.json();
      const content =
        data.choices?.[0]?.message?.content ??
        (typeof data === 'string' ? data : '');
      if (content) return content;
    } catch (err) {
      console.warn(`[Sasquatch] Pollinations ${model} threw:`, err);
    }
  }
  return '';
}

/**
 * Call the AI with multi-provider fallback + rate-limit backoff.
 *
 * Order: our AI proxy (OpenRouter free models, 429 retry) → keyless
 * Pollinations backup. Throws only if all available providers fail, so the
 * caller can fall back to local knowledge (static FAQ / glossary).
 */
export async function callSasquatchAI(
  messages: ChatTurn[],
  maxTokens = 400,
  totalTimeoutMs = TOTAL_TIMEOUT_MS,
  authHeaders: Record<string, string> = {},
): Promise<string> {
  const deadline = Date.now() + totalTimeoutMs;
  const remainingMs = () => deadline - Date.now();

  const primary = await tryProxy(messages, maxTokens, remainingMs, authHeaders);
  if (primary) return primary;

  console.warn('[Sasquatch] AI proxy exhausted — trying backup provider');
  const backup = await tryPollinations(messages, maxTokens, remainingMs);
  if (backup) return backup;

  throw new Error('All AI providers failed (proxy + backup).');
}

// ─── Streaming variant ───────────────────────────────────────────────────────
// Same provider order + budget as callSasquatchAI, but reads the OpenAI-style
// SSE stream and reports the growing text via `onDelta`. Used for the first
// pass so answers appear as they're written. Falls back to the non-streaming
// caller only if nothing at all was delivered.

/** Result of one provider's streaming attempt. */
interface StreamAttempt {
  text: string;
  /** True once at least one delta reached `onDelta`. */
  delivered: boolean;
}

async function streamOpenAiCompatible(
  url: string,
  headers: Record<string, string>,
  body: Record<string, unknown>,
  onDelta: ((cumulative: string) => void) | undefined,
  remainingMs: () => number,
): Promise<StreamAttempt> {
  const left = remainingMs();
  if (left <= 1000) return { text: '', delivered: false };

  const response = await fetch(url, {
    method: 'POST',
    headers: { 'Content-Type': 'application/json', ...headers },
    body: JSON.stringify({ ...body, stream: true }),
    signal: AbortSignal.timeout(Math.min(PER_REQUEST_TIMEOUT_MS, left)),
  });
  if (!response.ok || !response.body) return { text: '', delivered: false };

  const reader = response.body.getReader();
  const decoder = new TextDecoder();
  let sseBuffer = '';
  let text = '';
  let delivered = false;

  for (;;) {
    const { done, value } = await reader.read();
    if (done) break;
    sseBuffer += decoder.decode(value, { stream: true });

    let newline = sseBuffer.indexOf('\n');
    while (newline !== -1) {
      const line = sseBuffer.slice(0, newline).trim();
      sseBuffer = sseBuffer.slice(newline + 1);
      if (line.startsWith('data:')) {
        const payload = line.slice(5).trim();
        if (payload && payload !== '[DONE]') {
          try {
            const json = JSON.parse(payload);
            const delta: string = json.choices?.[0]?.delta?.content ?? '';
            if (delta) {
              text += delta;
              delivered = true;
              onDelta?.(text);
            }
          } catch {
            // Keep-alive / partial line — ignore.
          }
        }
      }
      newline = sseBuffer.indexOf('\n');
    }
  }

  return { text, delivered };
}

async function streamProxy(
  messages: ChatTurn[],
  maxTokens: number,
  onDelta: ((cumulative: string) => void) | undefined,
  remainingMs: () => number,
  authHeaders: Record<string, string> = {},
): Promise<StreamAttempt> {
  const models = PROXY_MODELS.slice(0, MAX_PROXY_MODELS);
  for (const model of models) {
    if (remainingMs() <= 1000) break;
    try {
      const attempt = await streamOpenAiCompatible(
        AI_PROXY_URL,
        authHeaders,
        { model, messages, max_tokens: maxTokens },
        onDelta,
        remainingMs,
      );
      if (attempt.text || attempt.delivered) return attempt;
    } catch (err) {
      console.warn(`[Sasquatch] AI proxy stream ${model} threw:`, err);
    }
  }
  return { text: '', delivered: false };
}

async function streamPollinations(
  messages: ChatTurn[],
  maxTokens: number,
  onDelta: ((cumulative: string) => void) | undefined,
  remainingMs: () => number,
): Promise<StreamAttempt> {
  for (const model of POLLINATIONS_MODELS) {
    if (remainingMs() <= 1000) break;
    try {
      const attempt = await streamOpenAiCompatible(
        POLLINATIONS_URL,
        {},
        { model, messages, max_tokens: maxTokens },
        onDelta,
        remainingMs,
      );
      if (attempt.text || attempt.delivered) return attempt;
    } catch (err) {
      console.warn(`[Sasquatch] Pollinations stream ${model} threw:`, err);
    }
  }
  return { text: '', delivered: false };
}

/**
 * Streaming version of {@link callSasquatchAI}. Reports the growing response
 * via `onDelta` (always the *cumulative* text, so the caller can just replace
 * the bubble) and resolves with the full text. Only if nothing was streamed at
 * all does it fall back to the non-streaming path, so the user still gets an
 * answer on environments that block streaming.
 */
export async function streamSasquatchAI(
  messages: ChatTurn[],
  maxTokens = 400,
  handlers: { onDelta?: (cumulative: string) => void; authHeaders?: Record<string, string> } = {},
): Promise<string> {
  const deadline = Date.now() + TOTAL_TIMEOUT_MS;
  const remainingMs = () => deadline - Date.now();

  const primary = await streamProxy(
    messages,
    maxTokens,
    handlers.onDelta,
    remainingMs,
    handlers.authHeaders,
  );
  if (primary.text) return primary.text;

  const backup = await streamPollinations(messages, maxTokens, handlers.onDelta, remainingMs);
  if (backup.text) return backup.text;

  // A partial answer was already shown — don't restart with a fresh call.
  if (primary.delivered || backup.delivered) return primary.text || backup.text;

  console.warn('[Sasquatch] Streaming unavailable — falling back to non-streaming call');
  return callSasquatchAI(messages, maxTokens, undefined, handlers.authHeaders);
}

/** Parse a `{"tool":"name","args":{…}}` tool call out of an AI response, if any. */
export function parseToolCall(aiText: string): { tool: string; args: Record<string, unknown> } | null {
  const jsonMatch = aiText.match(/\{[\s\S]*"tool"\s*:\s*"(\w+)"[\s\S]*\}/);
  if (!jsonMatch) return null;
  try {
    return JSON.parse(jsonMatch[0]);
  } catch {
    return null;
  }
}