n9router 0.4.43 → 0.4.45

package/app/open-sse/services/combo.js

@@ -4,6 +4,82 @@
 
 import { checkFallbackError, formatRetryAfter } from "./accountFallback.js";
 import { unavailableResponse } from "../utils/error.js";
+import { getCapabilitiesForModel } from "../providers/capabilities.js";
+import { extractTextContent } from "../translator/formats/gemini.js";
+
+// Hard capabilities = input modalities; missing one drops request data (e.g. image
+// stripped). Must be prioritized. Soft (e.g. search) only degrades a feature.
+const HARD_CAPS = new Set(["vision", "pdf", "audioInput", "videoInput"]);
+
+// Prefixes used when flattening tool turns into plain prose for panel models.
+const TOOL_CALL_PREFIX = "[Called tools: ";
+const TOOL_RESULT_PREFIX = "[Tool result: ";
+
+// Flatten tool turns into prose so panel models keep the context but can't loop
+// on tools: drop the request's tools, turn tool/function results into assistant
+// text, and inline assistant tool_calls names instead of the structured field.
+function flattenToolHistory(messages) {
+  return messages
+    .filter((msg) => msg)
+    .map((msg) => {
+      if (msg.role === "tool" || msg.role === "function") {
+        return { role: "assistant", content: `${TOOL_RESULT_PREFIX}${extractTextContent(msg.content) || String(msg.content ?? "")}]` };
+      }
+      if (msg.role === "assistant" && Array.isArray(msg.tool_calls)) {
+        const { tool_calls, ...rest } = msg;
+        const names = tool_calls.map((c) => c?.function?.name || c?.name || "tool").join(", ");
+        const base = extractTextContent(rest.content) || (typeof rest.content === "string" ? rest.content : "");
+        return { ...rest, content: `${base}${base ? "\n" : ""}${TOOL_CALL_PREFIX}${names}]` };
+      }
+      if (Array.isArray(msg.content)) {
+        const hasToolUse = msg.content.some((c) => c.type === "tool_use");
+        const hasToolResult = msg.content.some((c) => c.type === "tool_result");
+        if (hasToolUse || hasToolResult) {
+          const textParts = [];
+          const toolNames = [];
+          const toolResults = [];
+          for (const block of msg.content) {
+            if (block.type === "text" && block.text) textParts.push(block.text);
+            if (block.type === "tool_use") toolNames.push(block.name || "tool");
+            if (block.type === "tool_result") toolResults.push(extractTextContent(block.content) || String(block.content ?? ""));
+          }
+          const { ...rest } = msg;
+          let newContent = textParts.join("\n");
+          if (toolNames.length > 0) {
+            newContent = `${newContent}${newContent ? "\n" : ""}${TOOL_CALL_PREFIX}${toolNames.join(", ")}]`;
+          }
+          if (toolResults.length > 0) {
+            newContent = `${newContent}${newContent ? "\n" : ""}${TOOL_RESULT_PREFIX}${toolResults.join("\n")}]`;
+          }
+          return { ...rest, content: newContent };
+        }
+      }
+      return msg;
+    });
+}
+
+// Reorder combo models by capability fit. Stable; never drops a model (fallback intact).
+// Tier 0: satisfies all hard + all soft. Tier 1: all hard only. Tier 2: rest.
+export function reorderByCapabilities(models, required) {
+  if (!required || required.size === 0 || !Array.isArray(models) || models.length <= 1) return models;
+  const hard = [...required].filter((c) => HARD_CAPS.has(c));
+  const soft = [...required].filter((c) => !HARD_CAPS.has(c));
+
+  const tierOf = (m) => {
+    const slash = typeof m === "string" ? m.indexOf("/") : -1;
+    const provider = slash > 0 ? m.slice(0, slash) : "";
+    const model = slash > 0 ? m.slice(slash + 1) : m;
+    const caps = getCapabilitiesForModel(provider, model);
+    if (!hard.every((c) => caps[c] === true)) return 2;
+    return soft.every((c) => caps[c] === true) ? 0 : 1;
+  };
+
+  // Stable sort by tier (Array.prototype.sort is stable in modern engines).
+  return models
+    .map((m, i) => ({ m, i, t: tierOf(m) }))
+    .sort((a, b) => a.t - b.t || a.i - b.i)
+    .map((x) => x.m);
+}
 
 /**
  * Track rotation state per combo (for round-robin strategy)
@@ -11,6 +87,51 @@ import { unavailableResponse } from "../utils/error.js";
  */
 const comboRotationState = new Map();
 
+// Trailing run of items after the last assistant/model turn = the current user
+// turn. It may span several messages (e.g. text + image split across blocks),
+// so we return all of them. History media (older turns) must not pin the combo
+// to a vision model — those get stripped + placeholdered downstream instead.
+function trailingUserItems(arr) {
+  if (!Array.isArray(arr) || arr.length === 0) return [];
+  const isAssistant = (r) => r === "assistant" || r === "model";
+  let i = arr.length - 1;
+  while (i >= 0 && !isAssistant(arr[i]?.role)) i--;
+  return arr.slice(i + 1);
+}
+
+// Detect which capabilities a request needs. Modalities (vision/pdf) are scanned
+// only on the current user turn; "search" is request-wide (lives in tools).
+// Returns a Set of: "vision" | "pdf" | "search".
+export function detectRequiredCapabilities(body) {
+  const required = new Set();
+  if (!body || typeof body !== "object") return required;
+
+  const scanBlock = (b) => {
+    if (!b || typeof b !== "object") return;
+    const t = b.type;
+    if (t === "image_url" || t === "image" || t === "input_image") required.add("vision");
+    if (t === "file" || t === "document" || t === "input_file") required.add("pdf");
+    // gemini parts: inlineData/fileData carry a mime
+    const mime = b.inlineData?.mimeType || b.fileData?.mimeType;
+    if (typeof mime === "string" && mime.startsWith("image/")) required.add("vision");
+    if (mime === "application/pdf") required.add("pdf");
+  };
+
+  const scanContent = (content) => {
+    if (Array.isArray(content)) for (const b of content) scanBlock(b);
+  };
+
+  // Modalities: current user turn only (trailing user run across each known shape).
+  for (const m of trailingUserItems(body.messages)) scanContent(m.content);      // openai / claude
+  for (const it of trailingUserItems(body.input)) scanContent(it.content);       // responses
+  const contents = body.contents || body.request?.contents;                      // gemini / antigravity
+  for (const c of trailingUserItems(contents)) scanContent(c.parts);
+
+  // search: temporarily disabled in auto-switch (feature not wired yet).
+
+  return required;
+}
+
 function normalizeStickyLimit(stickyLimit) {
   const parsed = Number.parseInt(stickyLimit, 10);
   return Number.isFinite(parsed) && parsed > 0 ? parsed : 1;
@@ -105,9 +226,21 @@ export function getComboModelsFromData(modelStr, combosData) {
  * @param {number|string} [options.comboStickyLimit=1] - Requests per combo model before switching
  * @returns {Promise<Response>}
  */
-export async function handleComboChat({ body, models, handleSingleModel, log, comboName, comboStrategy, comboStickyLimit = 1 }) {
+export async function handleComboChat({ body, models, handleSingleModel, log, comboName, comboStrategy, comboStickyLimit = 1, autoSwitch = true }) {
   // Apply rotation strategy if enabled
-  const rotatedModels = getRotatedModels(models, comboName, comboStrategy, comboStickyLimit);
+  let rotatedModels = getRotatedModels(models, comboName, comboStrategy, comboStickyLimit);
+
+  // Auto-switch: float models that satisfy the request's required capabilities to the front.
+  if (autoSwitch) {
+    const required = detectRequiredCapabilities(body);
+    if (required.size > 0) {
+      const reordered = reorderByCapabilities(rotatedModels, required);
+      if (reordered[0] !== rotatedModels[0]) {
+        log.info("COMBO", `auto-switch for [${[...required].join(",")}] → ${reordered[0]}`);
+      }
+      rotatedModels = reordered;
+    }
+  }
   
   let lastError = null;
   let earliestRetryAfter = null;
@@ -196,3 +329,243 @@ export async function handleComboChat({ body, models, handleSingleModel, log, co
     { status, headers: { "Content-Type": "application/json" } }
   );
 }
+
+/**
+ * Extract assistant text from a non-stream completion across formats
+ * (OpenAI chat, Claude messages, Gemini, OpenAI Responses). Returns "" if none.
+ * Panel responses are already translated to the client format by chatCore, so the
+ * leaf content→string step reuses the translator's own extractTextContent.
+ */
+function extractPanelText(json) {
+  if (!json || typeof json !== "object") return "";
+
+  // OpenAI chat completion
+  const choice = json.choices?.[0];
+  if (choice) {
+    const msg = choice.message ?? choice.delta ?? {};
+    const t = extractTextContent(msg.content);
+    if (t.trim()) return t;
+    if (typeof choice.text === "string" && choice.text.trim()) return choice.text;
+  }
+
+  // Claude messages (text blocks share OpenAI's {type:"text"} shape)
+  const claudeText = extractTextContent(json.content);
+  if (claudeText.trim()) return claudeText;
+
+  // Gemini (parts carry .text without a type discriminator)
+  const parts = json.candidates?.[0]?.content?.parts;
+  if (Array.isArray(parts)) {
+    const t = parts.map((p) => p?.text || "").join("");
+    if (t.trim()) return t;
+  }
+
+  // OpenAI Responses API
+  if (Array.isArray(json.output)) {
+    const t = json.output
+      .flatMap((o) => (Array.isArray(o.content) ? o.content.map((c) => c?.text || "") : []))
+      .join("");
+    if (t.trim()) return t;
+  }
+
+  return "";
+}
+
+/**
+ * Append a synthesized user turn to whichever message array the request format uses.
+ * Preserves the original conversation + system prompt so the judge has full context.
+ */
+function appendUserTurn(body, text) {
+  const next = { ...body };
+  if (Array.isArray(body.messages)) {
+    next.messages = [...body.messages, { role: "user", content: text }];
+  } else if (Array.isArray(body.input)) {
+    next.input = [...body.input, { role: "user", content: text }];
+  } else if (Array.isArray(body.contents)) {
+    next.contents = [...body.contents, { role: "user", parts: [{ text }] }];
+  } else {
+    next.messages = [{ role: "user", content: text }];
+  }
+  return next;
+}
+
+/**
+ * Build the judge directive. Per OpenRouter's Fusion design, the judge does NOT
+ * merge — it analyzes (consensus / contradictions / partial coverage / unique
+ * insights / blind spots) then writes one answer grounded in that analysis.
+ * ~3/4 of fusion's quality lift comes from this synthesis step.
+ *
+ * Sources are anonymized ("Source N") so the judge weighs substance, not the
+ * reputation of a model brand.
+ */
+function buildJudgePrompt(answers) {
+  const panel = answers
+    .map((a, i) => `[Source ${i + 1}]\n${a.text}`)
+    .join("\n\n");
+
+  return [
+    `You are the JUDGE in a model-fusion panel. ${answers.length} expert models independently answered the user's most recent request. Their responses are below, anonymized by source.`,
+    "",
+    "Do NOT mention that multiple models were used, and do NOT refer to the sources. Produce ONE authoritative final answer addressed directly to the user.",
+    "",
+    "First, internally analyze the panel along these dimensions: consensus (points most sources agree on — treat as higher-confidence), contradictions (where they disagree — resolve with your own judgment), partial coverage, unique insights only one source surfaced, and blind spots every source missed. Then write the best possible final answer grounded in that analysis — more complete and correct than any single response, with no filler.",
+    "",
+    "=== PANEL RESPONSES ===",
+    panel,
+    "=== END PANEL RESPONSES ===",
+    "",
+    "Now write the final answer to the user's original request.",
+  ].join("\n");
+}
+
+// Fusion tuning. Overridable per-combo via settings.comboStrategies[name].
+const FUSION_DEFAULTS = {
+  minPanel: 2,             // answers needed before stragglers get a grace window
+  stragglerGraceMs: 8000,  // wait this long for laggards once quorum is reached
+  panelHardTimeoutMs: 90000, // absolute cap so one hung model can't stall forever
+};
+
+// Resolve a Response (or {__error}) within ms; the loser keeps running but is ignored.
+function withTimeout(promise, ms) {
+  return new Promise((resolve) => {
+    const t = setTimeout(() => resolve({ __timeout: true }), ms);
+    Promise.resolve(promise)
+      .then((v) => { clearTimeout(t); resolve(v); })
+      .catch((e) => { clearTimeout(t); resolve({ __error: e }); });
+  });
+}
+
+/**
+ * Collect panel responses with quorum-grace: as soon as `minPanel` calls succeed,
+ * start a short grace timer for the rest, then proceed with whatever arrived. This
+ * caps the straggler penalty (the slowest model otherwise dominates wall time) while
+ * still preferring a full panel when everyone is fast. Bounded by a hard timeout.
+ * Returns a sparse array aligned to `calls` (undefined = not yet / dropped).
+ */
+function collectPanel(calls, { minPanel, stragglerGraceMs, panelHardTimeoutMs }) {
+  return new Promise((resolve) => {
+    const out = new Array(calls.length);
+    let settled = 0;
+    let ok = 0;
+    let finished = false;
+    let graceTimer = null;
+    const finish = () => {
+      if (finished) return;
+      finished = true;
+      clearTimeout(hardTimer);
+      if (graceTimer) clearTimeout(graceTimer);
+      resolve(out);
+    };
+    const hardTimer = setTimeout(finish, panelHardTimeoutMs);
+    calls.forEach((p, i) => {
+      Promise.resolve(p)
+        .then((v) => { out[i] = v; })
+        .catch((e) => { out[i] = { __error: e }; })
+        .finally(() => {
+          settled++;
+          if (out[i] && out[i].ok) ok++;
+          if (settled === calls.length) return finish();
+          if (ok >= minPanel && !graceTimer) graceTimer = setTimeout(finish, stragglerGraceMs);
+        });
+    });
+  });
+}
+
+/**
+ * Handle a fusion combo: fan the prompt out to every panel model in parallel,
+ * then a judge model synthesizes one final answer from all panel responses.
+ *
+ * Panel calls are forced non-streaming with tools stripped (the judge needs
+ * complete prose to synthesize). The judge call keeps the client's original
+ * stream flag + tools, so streaming and downstream tool use still work.
+ *
+ * Speed: quorum-grace collection caps the straggler penalty. Quality: the judge
+ * runs the consensus/contradiction/blind-spot analysis before writing.
+ *
+ * Degrades gracefully: 0 panel answers -> 503, exactly 1 -> return it directly.
+ *
+ * @param {Object} options
+ * @param {Object} options.body - Request body (client format)
+ * @param {string[]} options.models - Panel model strings
+ * @param {Function} options.handleSingleModel - (body, modelStr) => Promise<Response>
+ * @param {Object} options.log - Logger
+ * @param {string} [options.comboName] - Combo name (logging)
+ * @param {string} [options.judgeModel] - Judge model; falls back to panel[0]
+ * @param {Object} [options.tuning] - Override FUSION_DEFAULTS (minPanel, grace, timeout)
+ * @returns {Promise<Response>}
+ */
+export async function handleFusionChat({ body, models, handleSingleModel, log, comboName, judgeModel, tuning }) {
+  const panel = Array.isArray(models) ? models.filter(Boolean) : [];
+  if (panel.length === 0) {
+    return new Response(
+      JSON.stringify({ error: { message: "Fusion combo has no models" } }),
+      { status: 400, headers: { "Content-Type": "application/json" } }
+    );
+  }
+
+  // A single-model fusion has nothing to fuse — just answer directly.
+  if (panel.length === 1) {
+    return handleSingleModel(body, panel[0]);
+  }
+
+  const cfg = { ...FUSION_DEFAULTS, ...(tuning || {}) };
+  const minPanel = Math.min(Math.max(2, cfg.minPanel), panel.length);
+  const judge = judgeModel && judgeModel.trim() ? judgeModel.trim() : panel[0];
+  log.info("FUSION", `Combo "${comboName}" | panel=${panel.length} [${panel.join(", ")}] | judge=${judge} | quorum=${minPanel}`);
+
+  // 1. Fan out to the panel in parallel: non-streaming, tools stripped (we want prose).
+  const { tools, tool_choice, ...rest } = body;
+  const panelBody = { ...rest, stream: false };
+
+  // Flatten tool turns to prose so panel models keep context without emitting tool_calls.
+  if (Array.isArray(panelBody.messages)) {
+    panelBody.messages = flattenToolHistory(panelBody.messages);
+  } else if (Array.isArray(panelBody.input)) {
+    panelBody.input = flattenToolHistory(panelBody.input);
+  }
+
+  const t0 = Date.now();
+  const calls = panel.map((m) => withTimeout(handleSingleModel(panelBody, m, true), cfg.panelHardTimeoutMs));
+  const settled = await collectPanel(calls, { ...cfg, minPanel });
+  log.info("FUSION", `fan-out collected in ${Date.now() - t0}ms`);
+
+  // 2. Collect successful answers.
+  const answers = [];
+  for (let i = 0; i < settled.length; i++) {
+    const res = settled[i];
+    const model = panel[i];
+    if (!res) { log.warn("FUSION", `Panel ${model} dropped (straggler/timeout)`); continue; }
+    if (res.__timeout) { log.warn("FUSION", `Panel ${model} timed out`); continue; }
+    if (res.__error) { log.warn("FUSION", `Panel ${model} threw`, { error: res.__error?.message || String(res.__error) }); continue; }
+    if (!res.ok) { log.warn("FUSION", `Panel ${model} failed`, { status: res.status }); continue; }
+    try {
+      const json = await res.clone().json();
+      const text = extractPanelText(json);
+      if (text) {
+        answers.push({ model, text });
+        log.info("FUSION", `Panel ${model} ok (${text.length} chars)`);
+      } else {
+        log.warn("FUSION", `Panel ${model} returned empty content`);
+      }
+    } catch (e) {
+      log.warn("FUSION", `Panel ${model} unparseable`, { error: e.message || String(e) });
+    }
+  }
+
+  // 3. Degrade gracefully when the panel is too thin to fuse.
+  if (answers.length === 0) {
+    log.warn("FUSION", "All panel models failed");
+    return new Response(
+      JSON.stringify({ error: { message: "All fusion panel models failed" } }),
+      { status: 503, headers: { "Content-Type": "application/json" } }
+    );
+  }
+  if (answers.length === 1) {
+    log.info("FUSION", `Only ${answers[0].model} succeeded — answering directly (no fusion)`);
+    return handleSingleModel(body, answers[0].model);
+  }
+
+  // 4. Judge analyzes + writes one final answer (streams to client if requested).
+  const judgeBody = appendUserTurn(body, buildJudgePrompt(answers));
+  log.info("FUSION", `Judging ${answers.length} answers with ${judge}`);
+  return handleSingleModel(judgeBody, judge);
+}

package/app/open-sse/services/copilotModels.js

@@ -0,0 +1,155 @@
+/**
+ * GitHub Copilot model catalog fetcher.
+ *
+ * Calls Copilot's `GET /models` endpoint to get the live catalog for an
+ * authenticated account, so `/v1/models` reflects what the account can
+ * actually use (e.g. newly shipped `claude-opus-4.8`, `gpt-5.5`) instead of
+ * the hand-maintained static registry, which inevitably lags behind.
+ *
+ * Returns chat-capable models the account's policy allows. Embeddings and
+ * disabled models are filtered out.
+ */
+
+import { proxyAwareFetch } from "../utils/proxyFetch.js";
+import { GITHUB_COPILOT } from "../config/appConstants.js";
+import { refreshCopilotToken } from "./tokenRefresh.js";
+
+const MODELS_URL = "https://api.githubcopilot.com/models";
+const FETCH_TIMEOUT_MS = 10_000;
+const CACHE_TTL_MS = 5 * 60 * 1000; // 5 minutes per credential
+
+/** @type {Map<string, { expiresAt: number, models: any[] }>} */
+const catalogCache = new Map();
+
+function cacheKey(credentials) {
+  return credentials?.providerSpecificData?.copilotToken
+    || credentials?.accessToken
+    || "copilot-anonymous";
+}
+
+function buildHeaders(token) {
+  return {
+    "Authorization": `Bearer ${token}`,
+    "Content-Type": "application/json",
+    "Copilot-Integration-Id": "vscode-chat",
+    "editor-version": `vscode/${GITHUB_COPILOT.VSCODE_VERSION}`,
+    "editor-plugin-version": `copilot-chat/${GITHUB_COPILOT.COPILOT_CHAT_VERSION}`,
+    "user-agent": GITHUB_COPILOT.USER_AGENT,
+    "x-github-api-version": GITHUB_COPILOT.API_VERSION,
+  };
+}
+
+async function fetchCatalogRaw(token, signal) {
+  const controller = new AbortController();
+  const timeoutId = setTimeout(() => controller.abort(), FETCH_TIMEOUT_MS);
+  try {
+    const response = await proxyAwareFetch(MODELS_URL, {
+      method: "GET",
+      headers: buildHeaders(token),
+      cache: "no-store",
+      signal: signal || controller.signal,
+    });
+    if (!response.ok) {
+      const err = new Error(`Copilot /models returned ${response.status}`);
+      err.status = response.status;
+      throw err;
+    }
+    const data = await response.json();
+    return Array.isArray(data?.data) ? data.data : [];
+  } finally {
+    clearTimeout(timeoutId);
+  }
+}
+
+// Keep only chat models the account is allowed to use. The static registry
+// surfaced disabled/embedding entries inconsistently; here we trust upstream.
+function expandCatalog(raw) {
+  const seen = new Set();
+  const models = [];
+  for (const m of raw) {
+    if (!m || typeof m !== "object") continue;
+    if (m.capabilities?.type !== "chat") continue;
+    if (m.policy && m.policy.state !== "enabled") continue;
+    const id = m.id;
+    if (!id || seen.has(id)) continue;
+    seen.add(id);
+    models.push({ id, name: m.name || id });
+  }
+  return models;
+}
+
+/**
+ * Resolve the live Copilot model catalog for a connection.
+ *
+ * @param {object} credentials Connection record (accessToken, refreshToken,
+ *   providerSpecificData {copilotToken, copilotTokenExpiresAt}).
+ * @param {object} [options]
+ * @param {boolean} [options.forceRefresh] Bypass the per-credential cache.
+ * @param {object}  [options.log] Logger.
+ * @param {function} [options.onCredentialsRefreshed] Persist a refreshed
+ *   Copilot token back to your store. Called with `{ copilotToken,
+ *   copilotTokenExpiresAt }` whenever a 401 triggers a refresh.
+ * @returns {Promise<{ models: object[] } | null>}
+ */
+export async function resolveCopilotModels(credentials, options = {}) {
+  const token = credentials?.providerSpecificData?.copilotToken || credentials?.accessToken;
+  if (!token) {
+    options.log?.debug?.("COPILOT_MODELS", "No copilotToken/accessToken; skipping live fetch");
+    return null;
+  }
+
+  const key = cacheKey(credentials);
+  const now = Date.now();
+  if (!options.forceRefresh) {
+    const cached = catalogCache.get(key);
+    if (cached && cached.expiresAt > now) {
+      return { models: cached.models };
+    }
+  }
+
+  let raw;
+  try {
+    raw = await fetchCatalogRaw(token, options.signal);
+  } catch (err) {
+    // A 401/403 means the Copilot token is stale — refresh from the GitHub
+    // access token and retry once.
+    if (err && (err.status === 401 || err.status === 403) && credentials.accessToken) {
+      options.log?.info?.("COPILOT_MODELS", `Got ${err.status}; refreshing Copilot token`);
+      const refreshed = await refreshCopilotToken(credentials.accessToken);
+      if (refreshed?.token) {
+        if (typeof options.onCredentialsRefreshed === "function") {
+          try {
+            await options.onCredentialsRefreshed({
+              copilotToken: refreshed.token,
+              copilotTokenExpiresAt: refreshed.expiresAt,
+            });
+          } catch (e) {
+            options.log?.warn?.("COPILOT_MODELS", `onCredentialsRefreshed failed: ${e?.message || e}`);
+          }
+        }
+        try {
+          raw = await fetchCatalogRaw(refreshed.token, options.signal);
+        } catch (err2) {
+          options.log?.warn?.("COPILOT_MODELS", `Retry after refresh failed: ${err2?.message || err2}`);
+          return null;
+        }
+      } else {
+        options.log?.warn?.("COPILOT_MODELS", "Token refresh did not return a token");
+        return null;
+      }
+    } else {
+      options.log?.warn?.("COPILOT_MODELS", `Live model fetch failed: ${err?.message || err}`);
+      return null;
+    }
+  }
+
+  const models = expandCatalog(raw);
+  if (!models.length) return null;
+
+  catalogCache.set(key, { expiresAt: now + CACHE_TTL_MS, models });
+  return { models };
+}
+
+export function clearCopilotModelCache() {
+  catalogCache.clear();
+}