@crazy-goat/nexos-provider 1.2.1 → 1.3.0

package/README.md

@@ -1,41 +1,24 @@
 # nexos-provider
 
-Custom [AI SDK](https://sdk.vercel.ai/) provider for using [nexos.ai](https://nexos.ai) Gemini models with [opencode](https://opencode.ai).
+Custom [AI SDK](https://sdk.vercel.ai/) provider for using [nexos.ai](https://nexos.ai) models with [opencode](https://opencode.ai).
 
-## Problem
+## What it does
 
-When accessing Gemini models through the nexos.ai API proxy, two issues prevent them from working with opencode (and likely other AI SDK-based tools):
+Fixes compatibility issues when using Gemini, Claude, and ChatGPT models through nexos.ai API in opencode:
 
-1. **Missing `data: [DONE]` in SSE streaming** — Gemini responses via nexos don't emit the standard `data: [DONE]` signal at the end of a streaming response. The AI SDK's `EventSourceParserStream` waits indefinitely for more data, causing opencode to hang forever.
-
-2. **`$ref` in tool schemas** — opencode sends JSON Schemas with `$ref` / `$defs` for tool parameters. Gemini (Vertex AI) rejects these with: `Schema.ref was set alongside unsupported fields`.
-
-## Solution
-
-This provider wraps `@ai-sdk/openai-compatible` and intercepts `fetch` to:
-
-- **Append `data: [DONE]\n\n`** to the end of streaming responses from Gemini models (via a `TransformStream` flush handler)
-- **Inline `$ref` references** in tool parameter schemas before sending them to the API
-
-No proxy, no extra processes — everything runs inline inside opencode.
+- **Gemini**: appends missing `data: [DONE]` SSE signal (prevents hanging), inlines `$ref` in tool schemas (rejected by Vertex AI), fixes `finish_reason` for tool calls (`stop`→`tool_calls`)
+- **Claude**: converts thinking params to snake_case (`budgetTokens`→`budget_tokens`), fixes `finish_reason` (`end_turn`→`stop`, prevents infinite retry loop), strips `thinking` object when disabled
+- **ChatGPT**: no fixes needed — `reasoningEffort` is handled natively by opencode
 
 ## Setup
 
-### 1. Clone this repo
-
-```bash
-git clone <this-repo> ~/nexos-provider
-cd ~/nexos-provider
-npm install
-```
-
-### 2. Set your API key
+### 1. Set your API key
 
 ```bash
 export NEXOS_API_KEY="your-nexos-api-key"
 ```
 
-### 3. Configure opencode
+### 2. Configure opencode
 
 Add the provider to your `~/.config/opencode/opencode.json`:
 
@@ -43,9 +26,9 @@ Add the provider to your `~/.config/opencode/opencode.json`:
 {
   "$schema": "https://opencode.ai/config.json",
   "provider": {
-    "nexos-gemini": {
-      "npm": "file:///absolute/path/to/nexos-provider/index.mjs",
-      "name": "Nexos Gemini",
+    "nexos-ai": {
+      "npm": "@crazy-goat/nexos-provider",
+      "name": "Nexos AI",
       "env": ["NEXOS_API_KEY"],
       "options": {
         "baseURL": "https://api.nexos.ai/v1/",
@@ -56,13 +39,25 @@ Add the provider to your `~/.config/opencode/opencode.json`:
           "name": "Gemini 2.5 Pro",
           "limit": { "context": 128000, "output": 64000 }
         },
-        "Gemini 3 Flash Preview": {
-          "name": "Gemini 3 Flash Preview",
-          "limit": { "context": 128000, "output": 64000 }
+        "Claude Sonnet 4.5": {
+          "name": "Claude Sonnet 4.5",
+          "limit": { "context": 200000, "output": 16000 },
+          "options": {
+            "thinking": { "type": "enabled", "budgetTokens": 1024 }
+          },
+          "variants": {
+            "thinking-high": { "thinking": { "type": "enabled", "budgetTokens": 10000 } },
+            "no-thinking": { "thinking": { "type": "disabled" } }
+          }
         },
-        "Gemini 3 Pro Preview": {
-          "name": "Gemini 3 Pro Preview",
-          "limit": { "context": 128000, "output": 64000 }
+        "GPT 5": {
+          "name": "GPT 5",
+          "limit": { "context": 400000, "output": 128000 },
+          "options": { "reasoningEffort": "medium" },
+          "variants": {
+            "high": { "reasoningEffort": "high" },
+            "no-reasoning": { "reasoningEffort": "none" }
+          }
         }
       }
     }
@@ -70,74 +65,46 @@ Add the provider to your `~/.config/opencode/opencode.json`:
 }
 ```
 
-> **Note:** The `npm` path must be an absolute `file://` URL pointing to `index.mjs`.
+> **Tip:** You can automatically generate the config with all available nexos.ai models using [opencode-nexos-models-config](https://github.com/crazy-goat/opencode-nexos-models-config).
+
+> **Warning:** Gemini 3 models (Flash Preview, Pro Preview) are currently unavailable — tool calling through nexos.ai does not work for these models.
 
-### 4. Use it
+### 3. Use it
 
+Simple prompt:
 ```bash
-opencode run "hello" -m "nexos-gemini/Gemini 2.5 Pro"
+opencode run "hello" -m "nexos-ai/Gemini 2.5 Pro"
 ```
 
-Or select the model interactively in opencode with `Ctrl+X M`.
-
-## Listing available models
-
-To see all models available on nexos.ai:
-
+With tool calling:
 ```bash
-npx @crazy-goat/nexos-provider
+opencode run "list files in current directory" -m "nexos-ai/Gemini 2.5 Pro"
 ```
 
-Or if you have the repo cloned:
-
+Claude with thinking:
 ```bash
-npm run list-models
+opencode run "what is 2+2?" -m "nexos-ai/Claude Sonnet 4.5" --variant thinking-high
 ```
 
-Requires `NEXOS_API_KEY` to be set.
-
-## GPT and Claude models
-
-GPT and Claude models work fine through nexos.ai without this provider — they correctly emit `data: [DONE]` and handle `$ref` schemas. Use the standard `@ai-sdk/openai-compatible` provider for those:
-
-```json
-{
-  "nexos-ai": {
-    "npm": "@ai-sdk/openai-compatible",
-    "name": "Nexos AI",
-    "env": ["NEXOS_API_KEY"],
-    "options": {
-      "baseURL": "https://api.nexos.ai/v1/",
-      "timeout": 300000
-    },
-    "models": {
-      "Claude Opus 4.6": {
-        "name": "Claude Opus 4.6",
-        "limit": { "context": 128000, "output": 64000 }
-      },
-      "GPT 5.2": {
-        "name": "GPT 5.2",
-        "limit": { "context": 128000, "output": 64000 }
-      }
-    }
-  }
-}
+GPT with reasoning effort:
+```bash
+opencode run "what is 2+2?" -m "nexos-ai/GPT 5" --variant high
 ```
 
+Or select the model interactively in opencode with `Ctrl+X M`.
+
 ## How it works
 
-The provider exports `createNexosAI` which creates a standard AI SDK provider with a custom `fetch` wrapper:
+The provider exports `createNexosAI` which creates a standard AI SDK provider with a custom `fetch` wrapper. Per-provider fixes are in separate modules:
 
 ```
-Request flow:
-  opencode → createNexosAI → fetch wrapper → nexos.ai API
-                                 │
-                                 ├─ Resolves $ref in tool schemas (for Gemini)
-                                 └─ Appends data: [DONE] to SSE stream (for Gemini)
+opencode → createNexosAI → fetch wrapper → nexos.ai API
+                               │
+                               ├─ fix-gemini.mjs: $ref inlining, finish_reason fix
+                               ├─ fix-claude.mjs: thinking params, end_turn→stop
+                               └─ fix-chatgpt.mjs: passthrough (no fixes needed)
 ```
 
-Only Gemini model requests are modified — all other models pass through unchanged.
-
 ## License
 
 MIT

package/index.mjs

@@ -1,69 +1,13 @@
 import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
-
-function resolveRefs(schema, defs) {
-  if (!schema || typeof schema !== "object") return schema;
-  if (Array.isArray(schema)) return schema.map((s) => resolveRefs(s, defs));
-
-  if (schema.$ref || schema.ref) {
-    const refName = (schema.$ref || schema.ref)
-      .replace(/^#\/\$defs\//, "")
-      .replace(/^#\/definitions\//, "");
-    const resolved = defs?.[refName];
-    if (resolved) {
-      const merged = { ...resolveRefs(resolved, defs) };
-      if (schema.description) merged.description = schema.description;
-      if (schema.default !== undefined) merged.default = schema.default;
-      return merged;
-    }
-  }
-
-  const result = {};
-  for (const [k, v] of Object.entries(schema)) {
-    if (k === "$defs" || k === "definitions" || k === "$ref" || k === "ref")
-      continue;
-    result[k] = resolveRefs(v, defs);
-  }
-  return result;
-}
-
-function fixToolSchemas(body) {
-  if (!body.tools?.length) return body;
-  return {
-    ...body,
-    tools: body.tools.map((tool) => {
-      if (tool.type !== "function" || !tool.function?.parameters) return tool;
-      const params = tool.function.parameters;
-      const defs = params.$defs || params.definitions || {};
-      return {
-        ...tool,
-        function: {
-          ...tool.function,
-          parameters: resolveRefs(params, defs),
-        },
-      };
-    }),
-  };
-}
-
-function fixFinishReason(text) {
-  return text.replace(/data: ({.*})\n/g, (match, jsonStr) => {
-    try {
-      const parsed = JSON.parse(jsonStr);
-      let changed = false;
-      if (parsed.choices) {
-        for (const choice of parsed.choices) {
-          if (choice.finish_reason === "stop" && choice.delta?.tool_calls?.length) {
-            choice.finish_reason = "tool_calls";
-            changed = true;
-          }
-        }
-      }
-      if (changed) {
-        return "data: " + JSON.stringify(parsed) + "\n";
-      }
-    } catch {}
-    return match;
-  });
+import { isGeminiModel, fixGeminiRequest, fixGeminiThinkingRequest, fixGeminiStream } from "./fix-gemini.mjs";
+import { fixClaudeRequest, fixClaudeStream } from "./fix-claude.mjs";
+import { fixChatGPTRequest, fixChatGPTStream } from "./fix-chatgpt.mjs";
+
+function fixStreamChunk(text) {
+  text = fixGeminiStream(text);
+  text = fixClaudeStream(text);
+  text = fixChatGPTStream(text);
+  return text;
 }
 
 function appendDoneToStream() {
@@ -75,7 +19,7 @@ function appendDoneToStream() {
       let text =
         typeof chunk === "string" ? chunk : new TextDecoder().decode(chunk);
       if (text.includes("[DONE]")) sawDone = true;
-      text = fixFinishReason(text);
+      text = fixStreamChunk(text);
       controller.enqueue(encoder.encode(text));
     },
     flush(controller) {
@@ -86,10 +30,6 @@ function appendDoneToStream() {
   });
 }
 
-function isGeminiModel(model) {
-  return typeof model === "string" && model.toLowerCase().includes("gemini");
-}
-
 function createNexosFetch(baseFetch) {
   const realFetch = baseFetch || globalThis.fetch;
 
@@ -102,17 +42,30 @@ function createNexosFetch(baseFetch) {
     }
 
     const gemini = isGeminiModel(requestBody.model);
+    let needsStreamFix = gemini;
 
     if (gemini) {
-      if (requestBody.tools) {
-        requestBody = fixToolSchemas(requestBody);
-      }
+      requestBody = fixGeminiRequest(requestBody);
+      const geminiThinking = fixGeminiThinkingRequest(requestBody);
+      requestBody = geminiThinking.body;
+      if (geminiThinking.hadThinking) needsStreamFix = true;
+    }
+
+    const claudeResult = fixClaudeRequest(requestBody);
+    requestBody = claudeResult.body;
+    if (claudeResult.hadThinking) needsStreamFix = true;
+
+    const beforeChatGPT = requestBody;
+    requestBody = fixChatGPTRequest(requestBody);
+    const chatgptChanged = requestBody !== beforeChatGPT;
+
+    if (gemini || claudeResult.hadThinking || chatgptChanged) {
       init = { ...init, body: JSON.stringify(requestBody) };
     }
 
     const response = await realFetch(url, init);
 
-    if (gemini && requestBody.stream) {
+    if (needsStreamFix && requestBody.stream) {
       const fixedBody = response.body.pipeThrough(appendDoneToStream());
       return new Response(fixedBody, {
         status: response.status,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@crazy-goat/nexos-provider",
-  "version": "1.2.1",
+  "version": "1.3.0",
   "description": "Custom AI SDK provider for nexos.ai Gemini models in opencode",
   "type": "module",
   "main": "index.mjs",