plugin-custom-llm 1.1.1 → 1.2.0

package/README.md

@@ -7,7 +7,9 @@ NocoBase plugin for integrating external LLM providers that support OpenAI-compa
 - **OpenAI-compatible**: Works with any LLM provider exposing `/chat/completions` endpoint
 - **Auto content detection**: Handles both string and array content blocks (`[{type: 'text', text: '...'}]`)
 - **Response mapping**: Transform non-standard API responses to OpenAI format via JSON config (supports streaming SSE and JSON)
-- **Reasoning content**: Display thinking/reasoning from DeepSeek-compatible providers
+- **Reasoning content**: Display thinking/reasoning from DeepSeek-compatible providers (multi-path detection)
+- **Stream keepalive**: Prevent proxy/gateway timeouts during long model thinking phases
+- **Tool calling support**: Gemini-compatible tool schema fixing (Zod + JSON Schema)
 - **Configurable**: JSON config editors for request and response customization
 - **Locale support**: English, Vietnamese, Chinese
 
@@ -23,6 +25,11 @@ Upload `plugin-custom-llm-x.x.x.tgz` via NocoBase Plugin Manager UI, then enable
 |---|---|
 | **Base URL** | LLM endpoint URL, e.g. `https://your-llm-server.com/v1` |
 | **API Key** | Authentication key |
+| **Disable Streaming** | Disable streaming for models that return empty stream values |
+| **Stream Keep Alive** | Enable keepalive to prevent timeouts during long thinking phases |
+| **Keep Alive Interval** | Interval in ms between keepalive signals (default: 5000) |
+| **Keep Alive Content** | Visual indicator text during keepalive (default: `...`) |
+| **Timeout** | Custom timeout in ms for slow-responding models |
 | **Request config (JSON)** | Optional. Extra request configuration |
 | **Response config (JSON)** | Optional. Response parsing and mapping configuration |
 
@@ -72,6 +79,26 @@ Upload `plugin-custom-llm-x.x.x.tgz` via NocoBase Plugin Manager UI, then enable
 
 Standard OpenAI-compatible parameters: temperature, max tokens, top P, frequency/presence penalty, response format, timeout, max retries.
 
+## Changelog
+
+### v1.2.0
+
+- **Fix**: Keepalive no longer interferes with tool call sequences (prevents tool call corruption)
+- **Fix**: Gemini-compatible tool schema fixing — handles Zod schemas via dual-phase approach (pre/post conversion)
+- **Fix**: Keepalive content no longer contaminates saved messages in DB
+- **Fix**: Response metadata extraction with long ID sanitization (>128 chars truncated)
+- **Fix**: Multi-path reasoning content detection (`additional_kwargs` + `kwargs.additional_kwargs`)
+- **Fix**: Improved error recovery in keepalive consumer (immediate error propagation)
+
+### v1.1.1
+
+- Stream keepalive proxy for long thinking phases
+- Response mapping for non-standard LLM APIs
+
+### v1.0.0
+
+- Initial release with OpenAI-compatible LLM provider support
+
 ## License
 
 Apache-2.0

package/dist/externalVersion.js

@@ -8,13 +8,13 @@
  */
 
 module.exports = {
-  "@nocobase/client": "2.0.15",
-  "@nocobase/plugin-ai": "2.0.15",
+  "@nocobase/client": "2.0.20",
+  "@nocobase/plugin-ai": "2.0.20",
   "react-i18next": "11.18.6",
-  "@nocobase/server": "2.0.15",
-  "@nocobase/flow-engine": "2.0.15",
-  "@nocobase/database": "2.0.15",
+  "@nocobase/server": "2.0.20",
+  "@nocobase/flow-engine": "2.0.20",
+  "@nocobase/database": "2.0.20",
   "react": "18.2.0",
-  "@nocobase/utils": "2.0.15",
+  "@nocobase/utils": "2.0.20",
   "antd": "5.24.2"
 };

package/dist/server/llm-providers/custom-llm.js

@@ -236,9 +236,16 @@ function wrapWithStreamKeepAlive(model, options) {
     let streamDone = false;
     let streamError = null;
     let notifyReady = null;
+    let hasToolCallChunks = false;
+    let hasErrored = false;
     const consumer = (async () => {
+      var _a, _b;
       try {
         for await (const chunk of baseIterator) {
+          const msg = chunk == null ? void 0 : chunk.message;
+          if (((_a = msg == null ? void 0 : msg.tool_call_chunks) == null ? void 0 : _a.length) || ((_b = msg == null ? void 0 : msg.tool_calls) == null ? void 0 : _b.length)) {
+            hasToolCallChunks = true;
+          }
           buffer.push(chunk);
           if (notifyReady) {
             notifyReady();
@@ -247,6 +254,11 @@ function wrapWithStreamKeepAlive(model, options) {
         }
       } catch (err) {
         streamError = err;
+        hasErrored = true;
+        if (notifyReady) {
+          notifyReady();
+          notifyReady = null;
+        }
       } finally {
         streamDone = true;
         if (notifyReady) {
@@ -260,6 +272,7 @@ function wrapWithStreamKeepAlive(model, options) {
         while (buffer.length > 0) {
           yield buffer.shift();
         }
+        hasToolCallChunks = false;
         if (streamDone) break;
         const waitForChunk = new Promise((resolve) => {
           notifyReady = resolve;
@@ -273,9 +286,14 @@ function wrapWithStreamKeepAlive(model, options) {
         ]);
         if (timer) clearTimeout(timer);
         if (result === "timeout" && !streamDone && buffer.length === 0) {
+          if (streamError || hasErrored) break;
+          if (hasToolCallChunks) continue;
           const keepAliveChunk = new ChatGenerationChunk({
-            message: new AIMessageChunk({ content: KEEPALIVE_PREFIX + keepAliveContent }),
-            text: KEEPALIVE_PREFIX + keepAliveContent
+            message: new AIMessageChunk({
+              content: KEEPALIVE_PREFIX,
+              additional_kwargs: { __keepalive: true }
+            }),
+            text: KEEPALIVE_PREFIX
           });
           yield keepAliveChunk;
         }
@@ -296,37 +314,51 @@ function fixEmptyToolProperties(model) {
   var _a;
   const originalBind = (_a = model.bindTools) == null ? void 0 : _a.bind(model);
   if (!originalBind) return model;
+  const PLACEHOLDER_PROP = {
+    _placeholder: { type: "string", description: "No parameters required" }
+  };
+  function fixPropertiesInSchema(schema) {
+    if (!schema || typeof schema !== "object") return;
+    if (schema.properties && typeof schema.properties === "object" && Object.keys(schema.properties).length === 0) {
+      schema.properties = { ...PLACEHOLDER_PROP };
+    }
+    for (const key of ["anyOf", "oneOf", "allOf"]) {
+      if (Array.isArray(schema[key])) {
+        schema[key].forEach((sub) => fixPropertiesInSchema(sub));
+      }
+    }
+  }
   model.bindTools = function(tools, kwargs) {
+    var _a2;
     const fixedTools = tools.map((tool) => {
-      var _a2, _b;
+      var _a3, _b;
       if (!tool || typeof tool !== "object") return tool;
+      if (typeof ((_a3 = tool.schema) == null ? void 0 : _a3.safeParse) === "function") {
+        return tool;
+      }
       const schema = tool.schema;
-      if (schema && typeof schema === "object") {
-        const props = schema.properties || (schema == null ? void 0 : schema.shape);
+      if (schema && typeof schema === "object" && !schema.safeParse) {
+        const props = schema.properties;
         if (props && typeof props === "object" && Object.keys(props).length === 0) {
           return {
             ...tool,
             schema: {
               ...schema,
-              properties: {
-                _placeholder: { type: "string", description: "No parameters required" }
-              }
+              properties: { ...PLACEHOLDER_PROP }
             }
           };
         }
       }
-      if ((_b = (_a2 = tool.function) == null ? void 0 : _a2.parameters) == null ? void 0 : _b.properties) {
-        const params = tool.function.parameters;
-        if (typeof params.properties === "object" && Object.keys(params.properties).length === 0) {
+      const funcParams = (_b = tool.function) == null ? void 0 : _b.parameters;
+      if (funcParams == null ? void 0 : funcParams.properties) {
+        if (typeof funcParams.properties === "object" && Object.keys(funcParams.properties).length === 0) {
           return {
             ...tool,
             function: {
               ...tool.function,
               parameters: {
-                ...params,
-                properties: {
-                  _placeholder: { type: "string", description: "No parameters required" }
-                }
+                ...funcParams,
+                properties: { ...PLACEHOLDER_PROP }
               }
             }
           };
@@ -334,7 +366,22 @@ function fixEmptyToolProperties(model) {
       }
       return tool;
     });
-    return originalBind(fixedTools, kwargs);
+    const result = originalBind(fixedTools, kwargs);
+    try {
+      const config = (result == null ? void 0 : result.kwargs) ?? (result == null ? void 0 : result.defaultOptions);
+      if ((config == null ? void 0 : config.tools) && Array.isArray(config.tools)) {
+        for (const tool of config.tools) {
+          if ((_a2 = tool == null ? void 0 : tool.function) == null ? void 0 : _a2.parameters) {
+            fixPropertiesInSchema(tool.function.parameters);
+          }
+          if (tool == null ? void 0 : tool.parameters) {
+            fixPropertiesInSchema(tool.parameters);
+          }
+        }
+      }
+    } catch {
+    }
+    return result;
   };
   return model;
 }
@@ -404,11 +451,12 @@ class CustomLLMProvider extends import_plugin_ai.LLMProvider {
     const resConfig = this.responseConfig;
     const text = extractTextContent(chunk, resConfig.contentPath);
     if (isKeepAlive(text)) {
-      return null;
+      return KEEPALIVE_PREFIX;
     }
     return stripToolCallTags(text);
   }
   parseResponseMessage(message) {
+    var _a, _b;
     const { content: rawContent, messageId, metadata, role, toolCalls, attachments, workContext } = message;
     const content = {
       ...rawContent ?? {},
@@ -429,6 +477,10 @@ class CustomLLMProvider extends import_plugin_ai.LLMProvider {
       content.content = content.content.replace(new RegExp(escapedPrefix + ".*?(?=" + escapedPrefix + "|$)", "g"), "");
       content.content = stripToolCallTags(content.content);
     }
+    if (((_b = (_a = content.metadata) == null ? void 0 : _a.additional_kwargs) == null ? void 0 : _b.__keepalive) !== void 0) {
+      const { __keepalive, ...cleanKwargs } = content.metadata.additional_kwargs;
+      content.metadata = { ...content.metadata, additional_kwargs: cleanKwargs };
+    }
     return {
       key: messageId,
       content,
@@ -436,15 +488,43 @@ class CustomLLMProvider extends import_plugin_ai.LLMProvider {
     };
   }
   parseReasoningContent(chunk) {
-    var _a;
+    var _a, _b, _c;
     const resConfig = this.responseConfig;
     const reasoningKey = resConfig.reasoningKey || "reasoning_content";
-    const reasoning = (_a = chunk == null ? void 0 : chunk.additional_kwargs) == null ? void 0 : _a[reasoningKey];
+    const reasoning = ((_a = chunk == null ? void 0 : chunk.additional_kwargs) == null ? void 0 : _a[reasoningKey]) ?? ((_c = (_b = chunk == null ? void 0 : chunk.kwargs) == null ? void 0 : _b.additional_kwargs) == null ? void 0 : _c[reasoningKey]);
     if (reasoning && typeof reasoning === "string") {
       return { status: "streaming", content: reasoning };
     }
     return null;
   }
+  /**
+   * Extract response metadata from LLM output for post-save enrichment.
+   * Sanitizes overly long message IDs from Gemini or other providers.
+   */
+  parseResponseMetadata(output) {
+    var _a, _b;
+    try {
+      const generation = (_b = (_a = output == null ? void 0 : output.generations) == null ? void 0 : _a[0]) == null ? void 0 : _b[0];
+      if (!generation) return [null, null];
+      const message = generation.message;
+      let id = message == null ? void 0 : message.id;
+      if (!id) return [null, null];
+      if (typeof id === "string" && id.length > 128) {
+        id = id.substring(0, 128);
+      }
+      const metadata = {};
+      if (message == null ? void 0 : message.response_metadata) {
+        metadata.finish_reason = message.response_metadata.finish_reason;
+        metadata.system_fingerprint = message.response_metadata.system_fingerprint;
+      }
+      if (message == null ? void 0 : message.usage_metadata) {
+        metadata.usage_metadata = message.usage_metadata;
+      }
+      return Object.keys(metadata).length > 0 ? [id, metadata] : [null, null];
+    } catch {
+      return [null, null];
+    }
+  }
   parseResponseError(err) {
     return (err == null ? void 0 : err.message) ?? "Unexpected LLM service error";
   }

package/package.json

@@ -3,7 +3,7 @@
   "displayName": "AI LLM: Custom (OpenAI Compatible)",
   "displayName.zh-CN": "AI LLM：自定义（OpenAI 兼容）",
   "description": "OpenAI-compatible LLM provider with auto response format detection for external LLM services.",
-  "version": "1.1.1",
+  "version": "1.2.0",
   "main": "dist/server/index.js",
   "nocobase": {
     "supportedVersions": [