ada-agent 0.1.0 → 0.2.0

package/src/client/models-dev.ts

@@ -1,52 +1,106 @@
-// models.dev catalog — model metadata (context limits, pricing, capabilities). Prefetched once at
-// startup and cached for an hour; reads are synchronous from the in-memory cache. Offline-safe:
-// if the fetch fails, the cache stays empty and callers fall back to their own tables.
-
-interface Info {
-  context?: number;
-  output?: number;
-  inputCost?: number; // $ per 1M input tokens
-  outputCost?: number; // $ per 1M output tokens
-  reasoning?: boolean;
-}
-
-const cache = new Map<string, Info>();
-let fetchedAt = 0;
-
-/** Fetch and cache the models.dev catalog (no-op if fetched within the last hour). */
-export async function prefetch(): Promise<void> {
-  if (cache.size && Date.now() - fetchedAt < 3_600_000) return;
-  try {
-    const res = await fetch("https://models.dev/api.json", { signal: AbortSignal.timeout(10_000) });
-    if (!res.ok) return;
-    const data = (await res.json()) as Record<string, { models?: Record<string, { limit?: { context?: number; output?: number }; cost?: { input?: number; output?: number }; reasoning?: boolean }> }>;
-    cache.clear();
-    for (const prov of Object.values(data)) {
-      for (const [id, m] of Object.entries(prov.models ?? {})) {
-        cache.set(id, { context: m.limit?.context, output: m.limit?.output, inputCost: m.cost?.input, outputCost: m.cost?.output, reasoning: m.reasoning });
-      }
-    }
-    fetchedAt = Date.now();
-  } catch {
-    /* offline — keep whatever's cached */
-  }
-}
-
-function lookup(modelId: string): Info | null {
-  return cache.get(modelId) ?? cache.get(modelId.split("/").pop() ?? "") ?? cache.get(modelId.split(":")[0] ?? "") ?? null;
-}
-
-/** [inputCostPer1M, outputCostPer1M] from models.dev, or null. */
-export function priceOf(modelId: string): [number, number] | null {
-  const i = lookup(modelId);
-  return i && i.inputCost != null && i.outputCost != null ? [i.inputCost, i.outputCost] : null;
-}
-
-/** Context-window limit (tokens) from models.dev, or null. */
-export function contextOf(modelId: string): number | null {
-  return lookup(modelId)?.context ?? null;
-}
-
-export function catalogSize(): number {
-  return cache.size;
-}
+// models.dev catalog — model metadata (context limits, pricing, capabilities). The cache is seeded
+// at load from a baked, curated subset (catalog.json — popular providers, generated by
+// `npm run catalog:refresh`) so pricing/limits work offline; a live prefetch() then overlays the
+// full models.dev catalog. Reads are synchronous from the in-memory cache.
+
+import { readFileSync } from "node:fs";
+
+interface Info {
+  context?: number;
+  output?: number;
+  inputCost?: number; // $ per 1M input tokens
+  outputCost?: number; // $ per 1M output tokens
+  reasoning?: boolean;
+}
+
+interface CatalogModel {
+  name: string;
+  context: number | null;
+  output: number | null;
+  in: number | null;
+  out: number | null;
+  reasoning?: boolean;
+  cacheRead?: number;
+  toolCall?: boolean;
+}
+interface Catalog {
+  providers: Record<string, { name: string; npm?: string; models: Record<string, CatalogModel> }>;
+}
+
+const cache = new Map<string, Info>();
+let fetchedAt = 0;
+
+// The baked offline catalog (curated popular providers). Seeds the cache; live prefetch overlays it.
+const CATALOG: Catalog = (() => {
+  try {
+    return JSON.parse(readFileSync(new URL("./catalog.json", import.meta.url), "utf8")) as Catalog;
+  } catch {
+    return { providers: {} };
+  }
+})();
+for (const prov of Object.values(CATALOG.providers)) {
+  for (const [id, m] of Object.entries(prov.models)) {
+    cache.set(id, { context: m.context ?? undefined, output: m.output ?? undefined, inputCost: m.in ?? undefined, outputCost: m.out ?? undefined, reasoning: m.reasoning });
+  }
+}
+
+/** Fetch and cache the models.dev catalog (no-op if fetched within the last hour). */
+export async function prefetch(): Promise<void> {
+  if (cache.size && Date.now() - fetchedAt < 3_600_000) return;
+  try {
+    const res = await fetch("https://models.dev/api.json", { signal: AbortSignal.timeout(10_000) });
+    if (!res.ok) return;
+    const data = (await res.json()) as Record<string, { models?: Record<string, { limit?: { context?: number; output?: number }; cost?: { input?: number; output?: number }; reasoning?: boolean }> }>;
+    cache.clear();
+    for (const prov of Object.values(data)) {
+      for (const [id, m] of Object.entries(prov.models ?? {})) {
+        cache.set(id, { context: m.limit?.context, output: m.limit?.output, inputCost: m.cost?.input, outputCost: m.cost?.output, reasoning: m.reasoning });
+      }
+    }
+    fetchedAt = Date.now();
+  } catch {
+    /* offline — keep whatever's cached */
+  }
+}
+
+function lookup(modelId: string): Info | null {
+  return cache.get(modelId) ?? cache.get(modelId.split("/").pop() ?? "") ?? cache.get(modelId.split(":")[0] ?? "") ?? null;
+}
+
+/** [inputCostPer1M, outputCostPer1M] from models.dev, or null. */
+export function priceOf(modelId: string): [number, number] | null {
+  const i = lookup(modelId);
+  return i && i.inputCost != null && i.outputCost != null ? [i.inputCost, i.outputCost] : null;
+}
+
+/** Context-window limit (tokens) from models.dev, or null. */
+export function contextOf(modelId: string): number | null {
+  return lookup(modelId)?.context ?? null;
+}
+
+export function catalogSize(): number {
+  return cache.size;
+}
+
+/** Human-readable listing of the baked offline catalog. No filter → provider summary; a filter
+ *  (provider id/name substring) → that provider's models with context + price. */
+export function catalogText(filter?: string): string {
+  const f = filter?.toLowerCase();
+  const out: string[] = [];
+  for (const [pid, prov] of Object.entries(CATALOG.providers)) {
+    const models = Object.entries(prov.models);
+    if (!f) {
+      out.push(`${pid.padEnd(24)} ${String(models.length).padStart(3)} models  \x1b[2m${prov.name}\x1b[0m`);
+      continue;
+    }
+    if (!pid.toLowerCase().includes(f) && !prov.name.toLowerCase().includes(f)) continue;
+    out.push(`\n\x1b[1m${prov.name}\x1b[0m \x1b[2m(${pid})\x1b[0m`);
+    for (const [id, m] of models) {
+      const price = m.in != null && m.out != null ? `$${m.in}/$${m.out}` : "—";
+      const ctx = m.context ? `${Math.round(m.context / 1000)}k` : "—";
+      out.push(`  ${id.padEnd(40)} ${ctx.padStart(6)} ctx · ${price}/1M${m.reasoning ? " · reasoning" : ""}`);
+    }
+  }
+  if (!out.length) return `no providers match "${filter}". Try /catalog with no argument for the list.`;
+  return f ? out.join("\n") : `${out.join("\n")}\n\x1b[2m/catalog <provider> for models · npm run catalog:refresh to update\x1b[0m`;
+}

package/src/selfcheck.ts

@@ -255,6 +255,32 @@ async function main(): Promise<void> {
   assert.equal(permPhrase("write_file", false), "create or modify files on disk", "write phrase");
   assert.ok(permPhrase("merchant__x", false).includes("connector"), "MCP phrase mentions the connector");
 
+  // --- baked offline catalog seeds pricing/limits (no network) ---
+  {
+    const { priceOf, contextOf, catalogSize, catalogText } = await import("./client/models-dev.ts");
+    assert.ok(catalogSize() > 100, `catalog seeded from catalog.json (${catalogSize()} models)`);
+    const op = priceOf("claude-opus-4-8");
+    assert.ok(op && op[0] > 0 && op[1] > 0, "priceOf resolves a baked model offline");
+    assert.ok((contextOf("claude-opus-4-8") ?? 0) >= 200000, "contextOf resolves a baked model offline");
+    assert.ok(/anthropic/.test(catalogText()) && /openai/.test(catalogText()) && /cloudflare/.test(catalogText()), "catalogText lists the popular providers");
+    assert.ok(/claude-opus-4-8/.test(catalogText("anthropic")), "catalogText <provider> lists its models");
+  }
+
+  // --- provider routing (incl. the new cloudflare + groq/together disambiguation) ---
+  {
+    const { route } = await import("./server/router.ts");
+    const { PROVIDERS } = await import("./server/config.ts");
+    assert.ok("cloudflare" in PROVIDERS, "cloudflare provider is registered");
+    assert.equal(route("@cf/moonshotai/kimi-k2.7-code"), "cloudflare", "@cf/ → cloudflare");
+    assert.equal(route("groq/llama-3.3-70b"), "groq", "groq/ → groq");
+    assert.equal(route("together/x"), "together", "together/ → together");
+    assert.equal(route("claude-opus-4-8"), "anthropic", "claude → anthropic");
+    assert.equal(route("gpt-5"), "openai", "gpt → openai");
+    assert.equal(route("gemini-3-pro"), "google", "gemini → google");
+    assert.equal(route("qwen3-coder"), "dashscope", "qwen → dashscope");
+    assert.equal(route("anything-else"), "openrouter", "unmatched → openrouter");
+  }
+
   // --- background job runs and reports ---
   const jid = startJob("selfcheck job", async () => "job-done-ok");
   await new Promise((r) => setTimeout(r, 30));

package/src/server/config.ts

@@ -1,58 +1,65 @@
-// Backend configuration: provider upstreams, keys, client-key auth, port.
-// Everything is env-driven. The backend is the only place provider keys live.
-
-import { getCredential } from "./credentials.ts";
-import type { ProviderName } from "../shared/types.ts";
-
-export interface ProviderDef {
-  baseURL: string; // OpenAI-compatible base (…/v1) — every provider is proxied as-is
-  keyEnv: string; // env var holding this provider's key ("" = keyless, e.g. local Ollama)
-}
-
-export const PROVIDERS: Record<ProviderName, ProviderDef> = {
-  openai: { baseURL: "https://api.openai.com/v1", keyEnv: "OPENAI_API_KEY" },
-  anthropic: { baseURL: "https://api.anthropic.com/v1", keyEnv: "ANTHROPIC_API_KEY" },
-  google: { baseURL: "https://generativelanguage.googleapis.com/v1beta/openai", keyEnv: "GEMINI_API_KEY" },
-  mistral: { baseURL: "https://api.mistral.ai/v1", keyEnv: "MISTRAL_API_KEY" },
-  openrouter: { baseURL: "https://openrouter.ai/api/v1", keyEnv: "OPENROUTER_API_KEY" },
-  groq: { baseURL: "https://api.groq.com/openai/v1", keyEnv: "GROQ_API_KEY" },
-  deepseek: { baseURL: "https://api.deepseek.com", keyEnv: "DEEPSEEK_API_KEY" },
-  together: { baseURL: "https://api.together.xyz/v1", keyEnv: "TOGETHER_API_KEY" },
-  xai: { baseURL: "https://api.x.ai/v1", keyEnv: "XAI_API_KEY" },
-  dashscope: {
-    baseURL: process.env.DASHSCOPE_BASE_URL ?? "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
-    keyEnv: "DASHSCOPE_API_KEY",
-  },
-  // GitHub Copilot — OpenAI-compatible chat endpoint. COPILOT_API_KEY must be a Copilot *bearer*
-  // token (exchanged from a GitHub OAuth token at /copilot_internal/v2/token — that exchange is not
-  // implemented here; it needs a Copilot subscription). Required headers are added in the adapter.
-  copilot: { baseURL: process.env.COPILOT_BASE_URL ?? "https://api.githubcopilot.com", keyEnv: "COPILOT_API_KEY" },
-  ollama: { baseURL: process.env.OLLAMA_BASE_URL ?? "http://localhost:11434/v1", keyEnv: "" },
-};
-
-export const PORT = Number(process.env.ADA_PORT) || 8787;
-
-/** The ada client keys allowed to use this backend. null = auth disabled (dev mode). */
-export function clientKeys(): string[] | null {
-  const v = process.env.ADA_CLIENT_KEYS;
-  if (!v) return null;
-  return v.split(",").map((s) => s.trim()).filter(Boolean);
-}
-
-/** The upstream provider key: env var first, then a stored credential (API key or OAuth token). */
-export function providerKey(p: ProviderName): string | undefined {
-  const env = PROVIDERS[p].keyEnv;
-  if (env && process.env[env]) return process.env[env];
-  const cred = getCredential(p);
-  if (cred) return cred.type === "oauth" ? cred.access : cred.key;
-  return undefined; // keyless provider (Ollama) or unconfigured
-}
-
-/** A provider is usable if it's keyless, its key env var is set, or a credential is stored. */
-export function isConfigured(p: ProviderName): boolean {
-  return PROVIDERS[p].keyEnv === "" || !!process.env[PROVIDERS[p].keyEnv] || !!getCredential(p);
-}
-
-export function configuredProviders(): ProviderName[] {
-  return (Object.keys(PROVIDERS) as ProviderName[]).filter(isConfigured);
-}
+// Backend configuration: provider upstreams, keys, client-key auth, port.
+// Everything is env-driven. The backend is the only place provider keys live.
+
+import { getCredential } from "./credentials.ts";
+import type { ProviderName } from "../shared/types.ts";
+
+export interface ProviderDef {
+  baseURL: string; // OpenAI-compatible base (…/v1) — every provider is proxied as-is
+  keyEnv: string; // env var holding this provider's key ("" = keyless, e.g. local Ollama)
+}
+
+export const PROVIDERS: Record<ProviderName, ProviderDef> = {
+  openai: { baseURL: "https://api.openai.com/v1", keyEnv: "OPENAI_API_KEY" },
+  anthropic: { baseURL: "https://api.anthropic.com/v1", keyEnv: "ANTHROPIC_API_KEY" },
+  google: { baseURL: "https://generativelanguage.googleapis.com/v1beta/openai", keyEnv: "GEMINI_API_KEY" },
+  mistral: { baseURL: "https://api.mistral.ai/v1", keyEnv: "MISTRAL_API_KEY" },
+  openrouter: { baseURL: "https://openrouter.ai/api/v1", keyEnv: "OPENROUTER_API_KEY" },
+  groq: { baseURL: "https://api.groq.com/openai/v1", keyEnv: "GROQ_API_KEY" },
+  deepseek: { baseURL: "https://api.deepseek.com", keyEnv: "DEEPSEEK_API_KEY" },
+  together: { baseURL: "https://api.together.xyz/v1", keyEnv: "TOGETHER_API_KEY" },
+  xai: { baseURL: "https://api.x.ai/v1", keyEnv: "XAI_API_KEY" },
+  dashscope: {
+    baseURL: process.env.DASHSCOPE_BASE_URL ?? "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
+    keyEnv: "DASHSCOPE_API_KEY",
+  },
+  // GitHub Copilot — OpenAI-compatible chat endpoint. COPILOT_API_KEY must be a Copilot *bearer*
+  // token (exchanged from a GitHub OAuth token at /copilot_internal/v2/token — that exchange is not
+  // implemented here; it needs a Copilot subscription). Required headers are added in the adapter.
+  copilot: { baseURL: process.env.COPILOT_BASE_URL ?? "https://api.githubcopilot.com", keyEnv: "COPILOT_API_KEY" },
+  // Cloudflare Workers AI / AI Gateway — OpenAI-compatible. Workers AI: set CLOUDFLARE_ACCOUNT_ID +
+  // CLOUDFLARE_API_TOKEN (default URL). AI Gateway: point CLOUDFLARE_BASE_URL at the gateway URL.
+  // Model ids are `@cf/<vendor>/<model>` (e.g. @cf/moonshotai/kimi-k2.7-code) — sent through as-is.
+  cloudflare: {
+    baseURL: process.env.CLOUDFLARE_BASE_URL ?? `https://api.cloudflare.com/client/v4/accounts/${process.env.CLOUDFLARE_ACCOUNT_ID ?? ""}/ai/v1`,
+    keyEnv: "CLOUDFLARE_API_TOKEN",
+  },
+  ollama: { baseURL: process.env.OLLAMA_BASE_URL ?? "http://localhost:11434/v1", keyEnv: "" },
+};
+
+export const PORT = Number(process.env.ADA_PORT) || 8787;
+
+/** The ada client keys allowed to use this backend. null = auth disabled (dev mode). */
+export function clientKeys(): string[] | null {
+  const v = process.env.ADA_CLIENT_KEYS;
+  if (!v) return null;
+  return v.split(",").map((s) => s.trim()).filter(Boolean);
+}
+
+/** The upstream provider key: env var first, then a stored credential (API key or OAuth token). */
+export function providerKey(p: ProviderName): string | undefined {
+  const env = PROVIDERS[p].keyEnv;
+  if (env && process.env[env]) return process.env[env];
+  const cred = getCredential(p);
+  if (cred) return cred.type === "oauth" ? cred.access : cred.key;
+  return undefined; // keyless provider (Ollama) or unconfigured
+}
+
+/** A provider is usable if it's keyless, its key env var is set, or a credential is stored. */
+export function isConfigured(p: ProviderName): boolean {
+  return PROVIDERS[p].keyEnv === "" || !!process.env[PROVIDERS[p].keyEnv] || !!getCredential(p);
+}
+
+export function configuredProviders(): ProviderName[] {
+  return (Object.keys(PROVIDERS) as ProviderName[]).filter(isConfigured);
+}

package/src/server/providers/openai-compat.ts

@@ -1,76 +1,78 @@
-// OpenAI-compatible adapter. Covers every provider that speaks the OpenAI Chat
-// Completions format: OpenAI, Mistral, Groq, DeepSeek, xAI, OpenRouter, Together, Ollama,
-// and Gemini (via Google's OpenAI-compatible endpoint). Because the client also speaks
-// that format, this adapter just swaps in the upstream base URL + key and streams the
-// response straight back — no translation needed.
-
-import type { ProviderName } from "../../shared/types.ts";
-import { PROVIDERS, providerKey } from "../config.ts";
-import { SSE_HEADERS } from "../sse.ts";
-import type { Adapter, ChatRequest } from "./adapter.ts";
-
-function authHeaders(provider: ProviderName): Record<string, string> {
-  const key = providerKey(provider);
-  const base: Record<string, string> = key ? { authorization: `Bearer ${key}` } : {};
-  // GitHub Copilot's endpoint requires these editor-identification headers.
-  if (provider === "copilot") return { ...base, "Copilot-Integration-Id": "vscode-chat", "Editor-Version": "ada/0.0.1", "Editor-Plugin-Version": "ada/0.0.1" };
-  return base;
-}
-
-export const openAICompatAdapter: Adapter = {
-  async chat({ provider, body, res }: ChatRequest): Promise<void> {
-    const def = PROVIDERS[provider];
-    // Copilot is addressed as "copilot/<model>" but the endpoint wants the bare model id.
-    const outBody = provider === "copilot" && typeof body.model === "string" && body.model.startsWith("copilot/") ? { ...body, model: body.model.slice("copilot/".length) } : body;
-    let upstream: Awaited<ReturnType<typeof fetch>>;
-    try {
-      upstream = await fetch(`${def.baseURL}/chat/completions`, {
-        method: "POST",
-        headers: { "content-type": "application/json", ...authHeaders(provider) },
-        body: JSON.stringify(outBody),
-      });
-    } catch (e) {
-      res.writeHead(502, { "content-type": "application/json" });
-      res.end(
-        JSON.stringify({
-          error: { message: `could not reach ${provider} upstream at ${def.baseURL}: ${e instanceof Error ? e.message : String(e)}` },
-        }),
-      );
-      return;
-    }
-
-    if (!upstream.ok || !upstream.body) {
-      const text = await upstream.text().catch(() => "");
-      res.writeHead(upstream.status || 502, { "content-type": "application/json" });
-      res.end(text || JSON.stringify({ error: { message: `upstream error ${upstream.status}` } }));
-      return;
-    }
-
-    if (body.stream) {
-      res.writeHead(200, SSE_HEADERS);
-      const reader = upstream.body.getReader();
-      for (;;) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        if (value) res.write(Buffer.from(value));
-      }
-      res.end();
-    } else {
-      const text = await upstream.text();
-      res.writeHead(upstream.status, { "content-type": upstream.headers.get("content-type") ?? "application/json" });
-      res.end(text);
-    }
-  },
-
-  async listModels(provider: ProviderName): Promise<string[]> {
-    const def = PROVIDERS[provider];
-    try {
-      const r = await fetch(`${def.baseURL}/models`, { headers: authHeaders(provider) });
-      if (!r.ok) return [];
-      const j = (await r.json()) as { data?: Array<{ id?: unknown }> };
-      return (j.data ?? []).map((m) => m.id).filter((x): x is string => typeof x === "string");
-    } catch {
-      return [];
-    }
-  },
-};
+// OpenAI-compatible adapter. Covers every provider that speaks the OpenAI Chat
+// Completions format: OpenAI, Mistral, Groq, DeepSeek, xAI, OpenRouter, Together, Ollama,
+// and Gemini (via Google's OpenAI-compatible endpoint). Because the client also speaks
+// that format, this adapter just swaps in the upstream base URL + key and streams the
+// response straight back — no translation needed.
+
+import type { ProviderName } from "../../shared/types.ts";
+import { PROVIDERS, providerKey } from "../config.ts";
+import { SSE_HEADERS } from "../sse.ts";
+import type { Adapter, ChatRequest } from "./adapter.ts";
+
+function authHeaders(provider: ProviderName): Record<string, string> {
+  const key = providerKey(provider);
+  const base: Record<string, string> = key ? { authorization: `Bearer ${key}` } : {};
+  // GitHub Copilot's endpoint requires these editor-identification headers.
+  if (provider === "copilot") return { ...base, "Copilot-Integration-Id": "vscode-chat", "Editor-Version": "ada/0.0.1", "Editor-Plugin-Version": "ada/0.0.1" };
+  return base;
+}
+
+export const openAICompatAdapter: Adapter = {
+  async chat({ provider, body, res }: ChatRequest): Promise<void> {
+    const def = PROVIDERS[provider];
+    // Strip a leading "<provider>/" the router used only to disambiguate (copilot/groq/together) — the
+    // endpoint wants the bare id. (Cloudflare's "@cf/…" ids aren't "cloudflare/…", so they pass through.)
+    const prefix = `${provider}/`;
+    const outBody = typeof body.model === "string" && body.model.startsWith(prefix) ? { ...body, model: body.model.slice(prefix.length) } : body;
+    let upstream: Awaited<ReturnType<typeof fetch>>;
+    try {
+      upstream = await fetch(`${def.baseURL}/chat/completions`, {
+        method: "POST",
+        headers: { "content-type": "application/json", ...authHeaders(provider) },
+        body: JSON.stringify(outBody),
+      });
+    } catch (e) {
+      res.writeHead(502, { "content-type": "application/json" });
+      res.end(
+        JSON.stringify({
+          error: { message: `could not reach ${provider} upstream at ${def.baseURL}: ${e instanceof Error ? e.message : String(e)}` },
+        }),
+      );
+      return;
+    }
+
+    if (!upstream.ok || !upstream.body) {
+      const text = await upstream.text().catch(() => "");
+      res.writeHead(upstream.status || 502, { "content-type": "application/json" });
+      res.end(text || JSON.stringify({ error: { message: `upstream error ${upstream.status}` } }));
+      return;
+    }
+
+    if (body.stream) {
+      res.writeHead(200, SSE_HEADERS);
+      const reader = upstream.body.getReader();
+      for (;;) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        if (value) res.write(Buffer.from(value));
+      }
+      res.end();
+    } else {
+      const text = await upstream.text();
+      res.writeHead(upstream.status, { "content-type": upstream.headers.get("content-type") ?? "application/json" });
+      res.end(text);
+    }
+  },
+
+  async listModels(provider: ProviderName): Promise<string[]> {
+    const def = PROVIDERS[provider];
+    try {
+      const r = await fetch(`${def.baseURL}/models`, { headers: authHeaders(provider) });
+      if (!r.ok) return [];
+      const j = (await r.json()) as { data?: Array<{ id?: unknown }> };
+      return (j.data ?? []).map((m) => m.id).filter((x): x is string => typeof x === "string");
+    } catch {
+      return [];
+    }
+  },
+};

package/src/server/providers/registry.ts

@@ -1,31 +1,32 @@
-// Provider → adapter map. This table is the whole routing story at a glance:
-// who shares the OpenAI-compatible adapter, and who has a dedicated one.
-//
-// Adding support is obvious from here:
-//   - new model on an existing provider      → nothing to change
-//   - new OpenAI-compatible provider          → add it in config.ts + a line below
-//   - new provider with a divergent format    → write an adapter, map it below
-
-import type { ProviderName } from "../../shared/types.ts";
-import type { Adapter } from "./adapter.ts";
-import { anthropicAdapter } from "./anthropic.ts";
-import { openAICompatAdapter } from "./openai-compat.ts";
-
-const ADAPTERS: Record<ProviderName, Adapter> = {
-  anthropic: anthropicAdapter, // native: Anthropic Messages API
-  openai: openAICompatAdapter,
-  google: openAICompatAdapter, // via Google's OpenAI-compatible endpoint
-  mistral: openAICompatAdapter,
-  openrouter: openAICompatAdapter,
-  groq: openAICompatAdapter,
-  deepseek: openAICompatAdapter,
-  together: openAICompatAdapter,
-  xai: openAICompatAdapter,
-  dashscope: openAICompatAdapter, // Alibaba Qwen via DashScope's OpenAI-compatible endpoint
-  copilot: openAICompatAdapter, // GitHub Copilot's OpenAI-compatible endpoint (+ custom headers in the adapter)
-  ollama: openAICompatAdapter,
-};
-
-export function adapterFor(provider: ProviderName): Adapter {
-  return ADAPTERS[provider];
-}
+// Provider → adapter map. This table is the whole routing story at a glance:
+// who shares the OpenAI-compatible adapter, and who has a dedicated one.
+//
+// Adding support is obvious from here:
+//   - new model on an existing provider      → nothing to change
+//   - new OpenAI-compatible provider          → add it in config.ts + a line below
+//   - new provider with a divergent format    → write an adapter, map it below
+
+import type { ProviderName } from "../../shared/types.ts";
+import type { Adapter } from "./adapter.ts";
+import { anthropicAdapter } from "./anthropic.ts";
+import { openAICompatAdapter } from "./openai-compat.ts";
+
+const ADAPTERS: Record<ProviderName, Adapter> = {
+  anthropic: anthropicAdapter, // native: Anthropic Messages API
+  openai: openAICompatAdapter,
+  google: openAICompatAdapter, // via Google's OpenAI-compatible endpoint
+  mistral: openAICompatAdapter,
+  openrouter: openAICompatAdapter,
+  groq: openAICompatAdapter,
+  deepseek: openAICompatAdapter,
+  together: openAICompatAdapter,
+  xai: openAICompatAdapter,
+  dashscope: openAICompatAdapter, // Alibaba Qwen via DashScope's OpenAI-compatible endpoint
+  copilot: openAICompatAdapter, // GitHub Copilot's OpenAI-compatible endpoint (+ custom headers in the adapter)
+  cloudflare: openAICompatAdapter, // Cloudflare Workers AI / AI Gateway (OpenAI-compatible)
+  ollama: openAICompatAdapter,
+};
+
+export function adapterFor(provider: ProviderName): Adapter {
+  return ADAPTERS[provider];
+}

package/src/server/router.ts

@@ -1,29 +1,33 @@
-// Map a model id (and optional explicit provider) to a provider.
-// Order matters: explicit wins; then the shape of the id (namespaced / local); then base-name prefixes.
-
-import type { ProviderName } from "../shared/types.ts";
-import { PROVIDERS } from "./config.ts";
-
-export function route(model: string, explicit?: string): ProviderName {
-  if (explicit && explicit in PROVIDERS) return explicit as ProviderName;
-
-  const m = model.toLowerCase();
-
-  // "vendor/model" → OpenRouter's namespacing convention. Checked before base-name prefixes
-  // so e.g. "mistralai/…" routes to OpenRouter, not the Mistral API.
-  // "copilot/<model>" → GitHub Copilot (checked before the OpenRouter "/" rule).
-  if (m.startsWith("copilot/")) return "copilot";
-  if (m.includes("/")) return "openrouter";
-  // "model:tag" → a local Ollama model (e.g. gemma4:latest).
-  if (m.includes(":")) return "ollama";
-
-  if (/^(gpt|o1|o3|o4|chatgpt|text-|davinci)/.test(m)) return "openai";
-  if (m.startsWith("claude")) return "anthropic";
-  if (m.startsWith("gemini") || m.startsWith("gemma")) return "google";
-  if (/^(mistral|codestral|magistral|ministral|devstral|pixtral|open-mi)/.test(m)) return "mistral";
-  if (m.startsWith("grok")) return "xai";
-  if (m.startsWith("deepseek")) return "deepseek";
-  if (m.startsWith("qwen") || m.startsWith("qwq")) return "dashscope";
-
-  return "openrouter"; // default: one key, every model
-}
+// Map a model id (and optional explicit provider) to a provider.
+// Order matters: explicit wins; then the shape of the id (namespaced / local); then base-name prefixes.
+
+import type { ProviderName } from "../shared/types.ts";
+import { PROVIDERS } from "./config.ts";
+
+export function route(model: string, explicit?: string): ProviderName {
+  if (explicit && explicit in PROVIDERS) return explicit as ProviderName;
+
+  const m = model.toLowerCase();
+
+  // "vendor/model" → OpenRouter's namespacing convention. Checked before base-name prefixes
+  // so e.g. "mistralai/…" routes to OpenRouter, not the Mistral API.
+  // Prefixed ids that must beat the OpenRouter "/" rule below:
+  if (m.startsWith("@cf/")) return "cloudflare"; // Cloudflare Workers AI model ids
+  if (m.startsWith("copilot/")) return "copilot";
+  // `groq/…` / `together/…` disambiguate shared model names (llama-3.3, gemma2…) that no prefix can.
+  if (m.startsWith("groq/")) return "groq";
+  if (m.startsWith("together/")) return "together";
+  if (m.includes("/")) return "openrouter";
+  // "model:tag" → a local Ollama model (e.g. gemma4:latest).
+  if (m.includes(":")) return "ollama";
+
+  if (/^(gpt|o1|o3|o4|chatgpt|text-|davinci)/.test(m)) return "openai";
+  if (m.startsWith("claude")) return "anthropic";
+  if (m.startsWith("gemini") || m.startsWith("gemma")) return "google";
+  if (/^(mistral|codestral|magistral|ministral|devstral|pixtral|open-mi)/.test(m)) return "mistral";
+  if (m.startsWith("grok")) return "xai";
+  if (m.startsWith("deepseek")) return "deepseek";
+  if (m.startsWith("qwen") || m.startsWith("qwq")) return "dashscope";
+
+  return "openrouter"; // default: one key, every model
+}