SRCmail/lib/ai/opencode.ts

// Shared server-side helpers for the OpenCode AI class.
//
// OpenCode (github.com/sst/opencode) runs as a local headless server
// (`opencode serve`) — the same runtime Paperclip drives as an agent adapter.
// Here it is used only as a one-shot chat backend for the mail assistant, so
// its HTTP surface is enough and no subprocess needs spawning from this app.
//
// IT IS NOT OpenAI-COMPATIBLE, despite `/v1/models` and `/v1/chat/completions`
// both answering 200: opencode serves a web UI from the same port with a
// catch-all route, so ANY unknown path returns the SPA's index.html with a 200.
// Checking `res.ok` alone therefore "verifies" endpoints that do not exist —
// verified the hard way, by believing exactly that before reading a body.
// Every probe here validates the parsed SHAPE, never the status code alone.
//
// The real API (from the server's own /doc OpenAPI spec):
//   GET  /api/model                  -> { data: [{ id, providerID, name, ... }] }
//   POST /session                    -> { id: "ses_..." }
//   POST /session/{id}/message       -> { info, parts: [{ type: 'text', text }, ...] }
//
// Address resolution is deliberately narrow: loopback only. This class exists
// to reach a runtime on the user's OWN machine — pointing it at a remote host
// would silently turn "local, no keys, nothing leaves the device" into the
// opposite, so a non-loopback OPENCODE_BASE_URL is refused rather than honoured.

const DEFAULT_BASE_URLS = ['http://127.0.0.1:4096', 'http://localhost:4096'];
const PROBE_TIMEOUT_MS = 2500;
const PROMPT_TIMEOUT_MS = 120_000;

function isLoopback(raw: string): boolean {
  try {
    const url = new URL(raw);
    return url.hostname === '127.0.0.1' || url.hostname === 'localhost' || url.hostname === '::1';
  } catch {
    return false;
  }
}

/** Candidate addresses, honouring an explicit OPENCODE_BASE_URL when it is
 *  loopback. `opencode serve` defaults to a RANDOM port (`--port 0`), so the
 *  conventional 4096 only finds a server deliberately started there; the env
 *  var is how someone on another port points us at it. */
export function opencodeBaseUrls(): string[] {
  const configured = process.env.OPENCODE_BASE_URL?.trim();
  if (configured) {
    if (!isLoopback(configured)) {
      console.error('[opencode] ignoring non-loopback OPENCODE_BASE_URL:', configured);
      return DEFAULT_BASE_URLS;
    }
    return [configured.replace(/\/+$/, ''), ...DEFAULT_BASE_URLS];
  }
  return DEFAULT_BASE_URLS;
}

interface OpencodeModelListResponse {
  data?: Array<{ id?: string; providerID?: string; name?: string }>;
}

export interface OpencodeModel {
  /** "providerID/modelID" — the reference shown in the picker and stored in
   *  settings, matching how opencode itself names models on the CLI. */
  ref: string;
  providerID: string;
  modelID: string;
  label: string;
}

/** Splits the stored "providerID/modelID" reference back into the pair the
 *  message API wants. Returns null for anything malformed rather than
 *  guessing, so a corrupted setting surfaces as a clear error. */
export function parseModelRef(ref: string): { providerID: string; modelID: string } | null {
  const slash = ref.indexOf('/');
  if (slash <= 0 || slash === ref.length - 1) return null;
  return { providerID: ref.slice(0, slash), modelID: ref.slice(slash + 1) };
}

/**
 * Auth header for a password-protected server.
 *
 * HTTP Basic with the username EXACTLY `opencode` — verified against 1.18.14:
 * an empty username, an arbitrary one, a Bearer token and every plausible
 * custom header all 401 with the correct password. Its own OpenAPI spec
 * declares no securitySchemes at all, so this is only knowable by trying it.
 * Absent password = an unsecured server (the desktop shell always sets one;
 * a hand-started `opencode serve` typically has none).
 */
function authHeaders(): Record<string, string> {
  const password = process.env.OPENCODE_SERVER_PASSWORD;
  if (!password) return {};
  return { Authorization: `Basic ${Buffer.from(`opencode:${password}`).toString('base64')}` };
}

async function fetchJson(url: string, init: RequestInit, timeoutMs: number): Promise<unknown | null> {
  const controller = new AbortController();
  const timer = setTimeout(() => controller.abort(), timeoutMs);
  try {
    const res = await fetch(url, {
      ...init,
      headers: { ...authHeaders(), ...(init.headers as Record<string, string> | undefined) },
      signal: controller.signal,
    });
    if (!res.ok) return null;
    // The SPA catch-all returns HTML with a 200 for unknown paths — see the
    // module header. Content-type is what actually distinguishes a real API
    // response from the web UI.
    const contentType = res.headers.get('content-type') ?? '';
    if (!contentType.includes('application/json')) return null;
    return await res.json();
  } catch {
    return null;
  } finally {
    clearTimeout(timer);
  }
}

/** First reachable candidate that answers /api/model with a real model list. */
export async function findOpencodeServer(): Promise<{ baseUrl: string; models: OpencodeModel[] } | null> {
  for (const baseUrl of opencodeBaseUrls()) {
    const body = (await fetchJson(`${baseUrl}/api/model`, {}, PROBE_TIMEOUT_MS)) as OpencodeModelListResponse | null;
    if (!body || !Array.isArray(body.data)) continue;
    const models: OpencodeModel[] = body.data
      .filter((m): m is { id: string; providerID: string; name?: string } =>
        typeof m?.id === 'string' && !!m.id && typeof m?.providerID === 'string' && !!m.providerID)
      .map((m) => ({
        ref: `${m.providerID}/${m.id}`,
        providerID: m.providerID,
        modelID: m.id,
        label: m.name ? `${m.name} (${m.providerID})` : `${m.providerID}/${m.id}`,
      }));
    if (models.length > 0) return { baseUrl, models };
  }
  return null;
}

interface OpencodeMessageResponse {
  parts?: Array<{ type?: string; text?: string }>;
}

/**
 * One prompt, one answer. Creates a throwaway session per question — this is
 * a stateless "ask about my mail" box, not a running conversation, and a fresh
 * session keeps one question's context from leaking into the next.
 *
 * `system` is passed as opencode's own system field rather than as a message
 * part, so the retrieved-mail prompt keeps the same shape it has for every
 * other provider class (see buildPrompt in lib/ai/local-client.ts).
 */
export async function opencodePrompt(
  baseUrl: string,
  model: { providerID: string; modelID: string },
  system: string | undefined,
  userText: string,
): Promise<{ ok: true; answer: string } | { ok: false; error: string }> {
  const session = (await fetchJson(
    `${baseUrl}/session`,
    { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: '{}' },
    PROBE_TIMEOUT_MS,
  )) as { id?: string } | null;
  if (!session?.id) return { ok: false, error: 'OpenCode would not start a session' };

  const body = (await fetchJson(
    `${baseUrl}/session/${encodeURIComponent(session.id)}/message`,
    {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({
        model,
        ...(system ? { system } : {}),
        parts: [{ type: 'text', text: userText }],
      }),
    },
    PROMPT_TIMEOUT_MS,
  )) as OpencodeMessageResponse | null;

  if (!body) return { ok: false, error: 'OpenCode returned no usable response' };
  // A reply carries several parts (step-start / reasoning / text / step-finish).
  // Only the `text` parts are the answer; `reasoning` is the model's private
  // chain of thought and must not be shown as the reply.
  const answer = (body.parts ?? [])
    .filter((p) => p.type === 'text' && typeof p.text === 'string' && p.text.trim())
    .map((p) => (p.text as string).trim())
    .join('\n\n');
  if (!answer) return { ok: false, error: 'OpenCode returned no message content' };
  return { ok: true, answer };
}