fix: resolve model-specific OpenCode inference routes

.changeset/tiny-days-retire.md

@@ -0,0 +1,5 @@
+---
+"opencode-bridge-for-copilot-chat": patch
+---
+
+Resolve OpenCode gateway URLs by model protocol, including Gemini on Go and metadata fallback. Derive discovery and inference URLs from shared canonical OpenCode constants while supporting legacy cached gateway bases.

docs/models.md

@@ -3,10 +3,9 @@
 ## Live metadata
 
 The extension discovers the catalog available to each configured provider entry
-from that gateway's live `models` endpoint: `https://opencode.ai/zen/v1/models`
-for Console and `https://opencode.ai/zen/go/v1/models` for Go. (The paths retain
-the historical `zen` segment upstream; the former Zen gateway is now Console's
-pay-as-you-go inference service.) Live responses
+from that gateway's live `models` endpoint: `https://opencode.ai/inference/v1/models`
+for Console and `https://opencode.ai/inference/go/v1/models` for Go. The former
+Zen gateway is now Console's pay-as-you-go inference service. Live responses
 provide names, context windows, maximum output lengths, reasoning support and
 effort options, image input, and tool-calling support used by Copilot Chat.
 Console free-tier entries can be filtered to cost-free models only.
@@ -16,6 +15,16 @@ organization's configuration on the Console server and never show public or
 other-organization models. Console service-account key entries use the public
 discovery catalog like Go.
 
+Requests resolve each model's SDK metadata to its wire protocol, with model-family
+fallbacks when metadata is unavailable. Generic OpenCode gateway bases are mapped
+to the canonical `/inference` routes: `openai/v1` for Chat Completions and Responses,
+`anthropic/v1` for Messages, and `google/v1beta/models` for Gemini, with `/go` after
+`/inference` for Go models. This includes Gemini contributor models on Go, whose
+Google route is absent from the legacy `/zen/go/v1` gateway. Custom endpoints and
+explicit protocol-specific bases are preserved.
+The Console server, discovery URLs, and inference URLs share one OpenCode origin
+constant. Legacy `/zen` bases in cached descriptors are normalized on use.
+
 Fields the live responses omit are enriched from the canonical `opencode` and
 `opencode-go` providers in a six-hour models.dev snapshot
 (`https://models.dev/api.json`) stored in VS Code `globalState`. Stale metadata

src/autocomplete/engine.test.ts

@@ -24,7 +24,7 @@ test("parses visible content deltas and ignores reasoning-only events", () => {
 test("streams a completion and sends the thinking-off payload", async () => {
   const requests: Array<{ url: string; init: RequestInit }> = [];
   const engine = new ChatCompletionEngine({
-    url: "https://opencode.ai/zen/go/v1/chat/completions",
+    url: "https://opencode.ai/inference/go/openai/v1/chat/completions",
     headers: { Authorization: "Bearer test-key" },
     timeoutMs: 1_000,
     fetcher: async (url, init) => {
@@ -50,7 +50,7 @@ test("streams a completion and sends the thinking-off payload", async () => {
     max_tokens: number;
     temperature: number;
   };
-  assert.equal(requests[0]?.url, "https://opencode.ai/zen/go/v1/chat/completions");
+  assert.equal(requests[0]?.url, "https://opencode.ai/inference/go/openai/v1/chat/completions");
   assert.equal(body.model, "qwen3.7-plus");
   assert.equal(body.enable_thinking, false);
   assert.equal(body.stream, true);
@@ -64,7 +64,7 @@ test("streams a completion and sends the thinking-off payload", async () => {
 
 test("never echoes upstream error bodies", async () => {
   const engine = new ChatCompletionEngine({
-    url: "https://opencode.ai/zen/go/v1/chat/completions",
+    url: "https://opencode.ai/inference/go/openai/v1/chat/completions",
     headers: {},
     timeoutMs: 1_000,
     fetcher: async () => new Response('{"error":"prompt context leak"}', { status: 401 }),
@@ -78,7 +78,7 @@ test("never echoes upstream error bodies", async () => {
 test("aborted caller signal surfaces as a quiet no-result", async () => {
   const controller = new AbortController();
   const engine = new ChatCompletionEngine({
-    url: "https://opencode.ai/zen/go/v1/chat/completions",
+    url: "https://opencode.ai/inference/go/openai/v1/chat/completions",
     headers: {},
     timeoutMs: 1_000,
     fetcher: async (_url, init) => {
@@ -96,7 +96,7 @@ test("aborted caller signal surfaces as a quiet no-result", async () => {
 
 test("times out long requests via the request timeout", async () => {
   const engine = new ChatCompletionEngine({
-    url: "https://opencode.ai/zen/go/v1/chat/completions",
+    url: "https://opencode.ai/inference/go/openai/v1/chat/completions",
     headers: {},
     timeoutMs: 30,
     fetcher: (async (_url: string | URL | Request, init?: RequestInit) => {

src/models/catalog.test.ts

@@ -1,6 +1,22 @@
 import assert from "node:assert/strict";
 import test from "node:test";
 import { catalogScope, ModelCatalog, modelsFromProvider } from "./catalog";
+import { endpointUrl } from "../transport/protocol";
+
+test("routes Go Gemini discovery with fresh, missing, or unavailable metadata", async () => {
+  for (const metadata of [
+    { "opencode-go": { npm: "@ai-sdk/openai-compatible", api: "https://opencode.ai/inference/go/v1", models: { "gemini-4-argon-contributor": { provider: { npm: "@ai-sdk/google" } } } } },
+    { "opencode-go": { models: {} } },
+    undefined,
+  ]) {
+    const catalog = new ModelCatalog(async (input) => String(input).endsWith("/models")
+      ? Response.json({ data: [{ id: "gemini-4-argon-contributor" }] })
+      : metadata ? Response.json(metadata) : new Response("unavailable", { status: 503 }));
+    const [model] = await catalog.refresh("go", undefined, false);
+    assert.equal(model.endpoint, "google");
+    assert.equal(endpointUrl(model.baseUrl, model.endpoint, model.rawModelId), "https://opencode.ai/inference/go/google/v1beta/models/gemini-4-argon-contributor:streamGenerateContent?alt=sse");
+  }
+});
 
 test("filters deprecated and paid Console models when free-only is enabled", () => {
   const models = modelsFromProvider("console", "opencode", { api: "https://example.test/v1" }, {
@@ -60,8 +76,10 @@ test("uses authenticated live models and enriches fields from models.dev", async
 
 test("uses the live public catalog without credentials and excludes stale metadata-only models", async () => {
   let authorization: string | null | undefined;
+  let discoveryUrl: string | undefined;
   const catalog = new ModelCatalog(async (input, init) => {
     if (String(input).endsWith("/models")) {
+      discoveryUrl = String(input);
       authorization = new Headers(init?.headers).get("authorization");
       return Response.json({ data: [{ id: "current" }] });
     }
@@ -74,6 +92,7 @@ test("uses the live public catalog without credentials and excludes stale metada
   assert.deepEqual(models.map((model) => model.id), ["current"]);
   assert.equal(models[0].name, "Current");
   assert.equal(authorization, null);
+  assert.equal(discoveryUrl, "https://opencode.ai/inference/v1/models");
 });
 
 test("enriches discovery-only Go models with supplemental metadata", async () => {

src/transport/protocol.test.ts

@@ -7,8 +7,8 @@ test("maps legacy zen mode values onto the Console gateway", () => {
   assert.equal(normalizeMode("console"), "console");
   assert.equal(normalizeMode("go"), "go");
   assert.equal(normalizeMode(undefined), "console");
-  assert.equal(apiBaseForMode("console"), "https://opencode.ai/zen/v1");
-  assert.equal(apiBaseForMode("go"), "https://opencode.ai/zen/go/v1");
+  assert.equal(apiBaseForMode("console"), "https://opencode.ai/inference/v1");
+  assert.equal(apiBaseForMode("go"), "https://opencode.ai/inference/go/v1");
 });
 
 test("resolves Console device verification URLs onto the /console subpath", () => {
@@ -49,6 +49,7 @@ test("routes known OpenCode model families to their native gateway shape", () =>
   assert.equal(resolveEndpointKind("gpt-6-astra", "console"), "responses");
   assert.equal(resolveEndpointKind("claude-fable-5", "console"), "messages");
   assert.equal(resolveEndpointKind("gemini-3.5-flash", "console"), "google");
+  assert.equal(resolveEndpointKind("gemini-4-argon-contributor", "go"), "google");
   assert.equal(resolveEndpointKind("grok-4.6", "go"), "responses");
   assert.equal(resolveEndpointKind("grok-build-0.1", "console"), "responses");
   assert.equal(resolveEndpointKind("muse-spark-1.3", "console"), "responses");
@@ -88,7 +89,6 @@ test("keeps undiscovered families on the shared chat completions shape", () => {
   assert.equal(resolveEndpointKind("omen-alpha", "go"), "chat-completions");
   assert.equal(resolveEndpointKind("big-pickle", "console"), "chat-completions");
   assert.equal(resolveEndpointKind("nemotron-3-ultra-free", "console"), "chat-completions");
-  assert.equal(resolveEndpointKind("gemini-3.5-flash", "go"), "chat-completions");
 });
 
 test("lets catalog package names override family heuristics", () => {
@@ -97,3 +97,30 @@ test("lets catalog package names override family heuristics", () => {
   assert.equal(resolveEndpointKind("custom-model", "go", "@ai-sdk/openai"), "responses");
   assert.equal(resolveEndpointKind("grok-4.6", "go", "@ai-sdk/openai-compatible"), "responses");
 });
+
+test("keeps MiniMax fallbacks gateway-specific while honoring model SDK metadata", () => {
+  assert.equal(resolveEndpointKind("minimax-m3", "console"), "chat-completions");
+  assert.equal(resolveEndpointKind("minimax-m3", "go"), "messages");
+  assert.equal(resolveEndpointKind("minimax-m3-free", "console", "@ai-sdk/anthropic"), "messages");
+});
+
+test("resolves generic OpenCode gateway bases to protocol-specific inference routes", () => {
+  for (const mode of ["console", "go"] as const) {
+    const scope = mode === "go" ? "/go" : "";
+    for (const base of [apiBaseForMode(mode), `https://opencode.ai/zen${scope}/v1/`]) {
+      assert.equal(endpointUrl(base, "chat-completions", "kimi-k3"), `https://opencode.ai/inference${scope}/openai/v1/chat/completions`);
+      assert.equal(endpointUrl(base, "responses", "gpt-6-luna"), `https://opencode.ai/inference${scope}/openai/v1/responses`);
+      assert.equal(endpointUrl(base, "messages", "qwen3.8-max"), `https://opencode.ai/inference${scope}/anthropic/v1/messages`);
+      assert.equal(endpointUrl(base, "google", "gemini-4-argon-contributor"), `https://opencode.ai/inference${scope}/google/v1beta/models/gemini-4-argon-contributor:streamGenerateContent?alt=sse`);
+    }
+  }
+});
+
+test("preserves custom and explicit protocol endpoints and encodes Google model ids", () => {
+  assert.equal(endpointUrl("https://example.test/v1/", "messages", "custom"), "https://example.test/v1/messages");
+  for (const base of ["https://example.test/v1beta", "https://example.test/v1beta/models/"]) {
+    assert.equal(endpointUrl(base, "google", "model/id"), "https://example.test/v1beta/models/model%2Fid:streamGenerateContent?alt=sse");
+  }
+  assert.equal(endpointUrl("https://opencode.ai/inference/go/google/v1beta", "google", "gemini-4-argon-contributor"), "https://opencode.ai/inference/go/google/v1beta/models/gemini-4-argon-contributor:streamGenerateContent?alt=sse");
+  assert.equal(endpointUrl("https://opencode.ai.example.test/inference/v1", "responses", "custom"), "https://opencode.ai.example.test/inference/v1/responses");
+});

src/transport/protocol.ts

@@ -1,6 +1,8 @@
-export const DEFAULT_CONSOLE_SERVER = "https://opencode.ai/console";
-export const CONSOLE_API_BASE_URL = "https://opencode.ai/zen/v1";
-export const GO_API_BASE_URL = "https://opencode.ai/zen/go/v1";
+export const OPENCODE_ORIGIN = "https://opencode.ai";
+export const OPENCODE_INFERENCE_BASE = `${OPENCODE_ORIGIN}/inference`;
+export const DEFAULT_CONSOLE_SERVER = `${OPENCODE_ORIGIN}/console`;
+export const CONSOLE_API_BASE_URL = `${OPENCODE_INFERENCE_BASE}/v1`;
+export const GO_API_BASE_URL = `${OPENCODE_INFERENCE_BASE}/go/v1`;
 export const OPENCODE_CLIENT_ID = "opencode-cli";
 export const OPENCODE_CLIENT = "opencode-copilot-chat";
 
@@ -10,6 +12,13 @@ export const LEGACY_ZEN_MODE = "zen";
 export type OpenCodeMode = "console" | "go";
 export type EndpointKind = "chat-completions" | "messages" | "responses" | "google";
 
+const PROTOCOL_BASE_PATHS: Readonly<Record<EndpointKind, string>> = {
+  "chat-completions": "openai/v1",
+  messages: "anthropic/v1",
+  responses: "openai/v1",
+  google: "google/v1beta",
+};
+
 /** Modes that can appear in persisted settings or state written by older versions. */
 export type LegacyOpenCodeMode = OpenCodeMode | "zen";
 
@@ -73,25 +82,37 @@ export function resolveEndpointKind(
   if (npm === "@ai-sdk/openai" || npm.endsWith("/openai")) return "responses";
   if (/^gpt-/i.test(modelId)) return "responses";
   if (/^claude-/i.test(modelId)) return "messages";
-  // Family-wide rather than version-pinned: every live grok* and qwen* model
-  // on both gateways is responses or messages, and models.dev omits `npm` for
-  // all qwen ids, which is the only signal the checks above have.
+  // Family fallbacks cover models whose SDK metadata is unavailable.
   if (/^grok(?:-|$)/i.test(modelId)) return "responses";
   if (/^muse-spark-/i.test(modelId)) return "responses";
   // qwen3.8-max is the only live qwen whose documented and live-verified shape
   // on the zen gateway is chat completions (the gateway rejects it on
   // /messages with a ModelProtocolUnsupported error); Go keeps messages.
   if (/^qwen-?3\.8-max$/i.test(modelId)) return mode === "go" ? "messages" : "chat-completions";
   if (/^qwen/i.test(modelId)) return "messages";
+  // MiniMax defaults to Messages on Go and Chat Completions on Console.
+  // Explicit model SDK metadata above can override either fallback.
   if (mode === "go" && /^minimax-/i.test(modelId)) return "messages";
-  if (mode === "console" && /^gemini-/i.test(modelId)) return "google";
+  if (/^gemini-/i.test(modelId)) return "google";
   return "chat-completions";
 }
 
 export function endpointUrl(baseUrl: string, endpoint: EndpointKind, modelId: string): string {
-  const base = baseUrl.replace(/\/+$/, "");
+  const base = gatewayBaseUrl(baseUrl, endpoint);
   if (endpoint === "messages") return `${base}/messages`;
   if (endpoint === "responses") return `${base}/responses`;
-  if (endpoint === "google") return `${base}/${encodeURIComponent(modelId)}:streamGenerateContent?alt=sse`;
+  if (endpoint === "google") return `${base.endsWith("/models") ? base : `${base}/models`}/${encodeURIComponent(modelId)}:streamGenerateContent?alt=sse`;
   return `${base}/chat/completions`;
 }
+
+/** Generic catalog bases are not inference roots for every protocol. */
+function gatewayBaseUrl(baseUrl: string, endpoint: EndpointKind): string {
+  const base = baseUrl.replace(/\/+$/, "");
+  let url: URL;
+  try { url = new URL(base); }
+  catch { return base; }
+  if (url.origin !== OPENCODE_ORIGIN || url.username || url.password || url.search || url.hash) return base;
+  const gateway = /^\/(?:zen|inference)(\/go)?\/v1$/.exec(url.pathname);
+  if (!gateway) return base;
+  return `${OPENCODE_INFERENCE_BASE}${gateway[1] ?? ""}/${PROTOCOL_BASE_PATHS[endpoint]}`;
+}