feat(plugin): add Microsoft Foundry as a built-in auth provider

packages/opencode/src/plugin/foundry.ts

@@ -0,0 +1,79 @@
+import type { Hooks, PluginInput } from "@opencode-ai/plugin"
+
+// Microsoft Foundry auth plugin.
+//
+// Foundry exposes models behind project-scoped URLs of the form:
+//   https://<resource>.services.ai.azure.com/api/projects/<project>/openai/v1
+// authenticated via either `Authorization: Bearer <key>` or `api-key: <key>`.
+//
+// Deployments are user-defined (e.g. `gpt-5.5`, `gpt-4o-mini`) and not present in
+// the models.dev catalog, so the user supplies them as part of `/login` and then
+// references them in their own `opencode.json` provider entry. Example:
+//
+//   {
+//     "provider": {
+//       "foundry": {
+//         "npm": "@ai-sdk/openai-compatible",
+//         "options": {
+//           "baseURL": "https://<resource>.services.ai.azure.com/api/projects/<project>/openai/v1"
+//         },
+//         "models": { "gpt-5.5": { "name": "GPT-5.5", "reasoning": false } }
+//       }
+//     }
+//   }
+//
+// Foundry rejects some params that the openai-compatible SDK emits for gpt-5.x
+// deployments (`reasoningSummary`, `reasoningEffort`, `textVerbosity`) plus
+// `max_tokens` (must be `max_completion_tokens`, which the SDK doesn't rename).
+// The `chat.params` hook strips them when the provider id is `foundry`.
+export async function FoundryAuthPlugin(_input: PluginInput): Promise<Hooks> {
+  return {
+    auth: {
+      provider: "foundry",
+      methods: [
+        {
+          type: "api",
+          label: "API key",
+          prompts: [
+            {
+              type: "text",
+              key: "resourceName",
+              message: "Foundry resource hostname",
+              placeholder: "e.g. faraday-resource.services.ai.azure.com",
+              validate: (value: string) => (value.includes(".") ? undefined : "Must be a fully-qualified hostname"),
+            },
+            {
+              type: "text",
+              key: "projectName",
+              message: "Foundry project name",
+              placeholder: "e.g. faraday-agents",
+              validate: (value: string) => (value.trim() ? undefined : "Project name is required"),
+            },
+            {
+              type: "text",
+              key: "deployments",
+              message: "Deployment names (comma-separated, used in opencode.json)",
+              placeholder: "e.g. gpt-5.5,gpt-4o-mini",
+            },
+          ],
+        },
+      ],
+    },
+
+    "chat.params": async (incoming, output) => {
+      if (incoming.model.providerID !== "foundry") return
+      const opts = output.options as Record<string, any> | undefined
+      if (opts) {
+        delete opts.reasoningEffort
+        delete opts.reasoningSummary
+        delete opts.textVerbosity
+        delete opts.include
+        delete opts.promptCacheKey
+      }
+      // The openai-compatible SDK emits `max_tokens`; gpt-5.x Foundry deployments
+      // require `max_completion_tokens` instead. Dropping the cap lets the model
+      // use its default output budget rather than failing the request.
+      output.maxOutputTokens = undefined
+    },
+  }
+}

packages/opencode/src/plugin/index.ts

@@ -19,6 +19,7 @@ import { PoeAuthPlugin } from "opencode-poe-auth"
 import { CloudflareAIGatewayAuthPlugin, CloudflareWorkersAuthPlugin } from "./cloudflare"
 import { AzureAuthPlugin } from "./azure"
 import { DigitalOceanAuthPlugin } from "./digitalocean"
+import { FoundryAuthPlugin } from "./foundry"
 import { Effect, Layer, Context, Stream } from "effect"
 import { EffectBridge } from "@/effect/bridge"
 import { InstanceState } from "@/effect/instance-state"
@@ -66,6 +67,7 @@ const INTERNAL_PLUGINS: PluginInstance[] = [
   CloudflareAIGatewayAuthPlugin,
   AzureAuthPlugin,
   DigitalOceanAuthPlugin,
+  FoundryAuthPlugin,
 ]
 
 function isServerPlugin(value: unknown): value is PluginInstance {