Add community XML tools and reorganize parsers (#85)

* Add community XML tools and reorganize parsers

* Update middleware references from `xmlToolMiddleware` to `morphXmlToolMiddleware` across documentation and examples for consistency and clarity.

* Update packages/parser/src/community/index.ts

Co-authored-by: Copilot <[email protected]>

* Fix typo in middleware reference from `sijawaraConsiseXmlToolMiddleware` to `sijawaraConciseXmlToolMiddleware` in example and community index files for consistency.

* fix: update morphXmlProtocol to enhance HTML tag handling

- Disabled self-closing tag behavior for HTML elements in the morphXmlProtocol, ensuring they are treated as regular tags.
- This change improves the parsing accuracy of XML tool calls by accommodating standard HTML tag structures.

* chore: attribute co-author

Co-authored-by: sijawara <[email protected]>

---------

Co-authored-by: Copilot <[email protected]>
Co-authored-by: sijawara <[email protected]>

Author: 3 people

.changeset/twenty-turkeys-open.md

@@ -0,0 +1,5 @@
+---
+"@ai-sdk-tool/parser": patch
+---
+
+Add community XML tools and reorganize parsers

AGENTS.md

@@ -30,7 +30,7 @@ This project provides middleware for AI SDK to enable tool calling with models t
 **@ai-sdk-tool/parser** (`packages/parser/`):
 
 - Main middleware package for tool call parsing
-- Exports pre-configured middlewares: `gemmaToolMiddleware`, `hermesToolMiddleware`, `xmlToolMiddleware`
+- Exports pre-configured middlewares: `gemmaToolMiddleware`, `hermesToolMiddleware`, `morphXmlToolMiddleware`
 - Core factory function: `createToolMiddleware()` for custom protocols
 - Protocol system with pluggable parsers for different model formats (JSON-mix, XML, etc.)
 

README.md

@@ -15,7 +15,7 @@ Note: Requires AI SDK v5. For AI SDK v4, use `@ai-sdk-tool/[email protected]`.
 
 ## Packages
 
-- `packages/parser` — core tool‑call parsing middleware and prebuilt middlewares (`gemmaToolMiddleware`, `hermesToolMiddleware`, `xmlToolMiddleware`).
+- `packages/parser` — core tool‑call parsing middleware and prebuilt middlewares (`gemmaToolMiddleware`, `hermesToolMiddleware`, `morphXmlToolMiddleware`).
   - Quickstarts: [packages/parser/README.md](packages/parser/README.md)
   - Official docs reference: [Custom tool call parser](https://ai-sdk.dev/docs/ai-sdk-core/middleware#custom-tool-call-parser)
 - `packages/eval` — evaluation utilities (BFCL, JSON generation).  
@@ -24,7 +24,7 @@ Note: Requires AI SDK v5. For AI SDK v4, use `@ai-sdk-tool/[email protected]`.
 ### Choose a middleware (at a glance)
 
 - **gemmaToolMiddleware**: JSON tool calls inside markdown fences. Best for Gemma-like models.
-- **xmlToolMiddleware**: Plain XML tool calls. Good fit for GLM/GLM-like models.
+- **morphXmlToolMiddleware**: Plain XML tool calls. Good fit for GLM/GLM-like models.
 - **hermesToolMiddleware**: JSON payload wrapped in `<tool_call>` XML tags. Hermes/Llama-style prompts.
 
 ## Install (per package)

docs/concepts/middleware.md

@@ -83,7 +83,7 @@ See concrete implementations in `packages/parser/src/protocols/json-mix-protocol
 - Preconfigured:
   - `gemmaToolMiddleware` — JSON-mix with fenced code blocks for Gemma-style models.
   - `hermesToolMiddleware` — JSON-mix with XML-tag wrapping of tool calls.
-  - `xmlToolMiddleware` — Morph-XML protocol for XML-native formats.
+  - `morphXmlToolMiddleware` — Morph-XML protocol for XML-native formats.
 - Protocol factories: `jsonMixProtocol`, `morphXmlProtocol`.
 
 Defined in `packages/parser/src/index.ts`.

docs/concepts/protocols.md

@@ -101,7 +101,7 @@ Preconfigured middlewares exported from `packages/parser/src/index.ts`:
 
 - `gemmaToolMiddleware` — Uses `jsonMixProtocol` with markdown code fences labeled `tool_call` and a prompt tailored for Gemma.
 - `hermesToolMiddleware` — `jsonMixProtocol` with `<tool_call>` tags and a Hermes-style prompt.
-- `xmlToolMiddleware` — `morphXmlProtocol` with an XML-oriented prompt.
+- `morphXmlToolMiddleware` — `morphXmlProtocol` with an XML-oriented prompt.
 
 ## Implementing a Custom Protocol
 

docs/examples.md

@@ -38,6 +38,6 @@ cd examples/eval-core && pnpm dlx tsx src/json-generation.ts
 
 ### [dev] Notes
 
-- **Middleware switch**: parser examples show how to swap between `xmlToolMiddleware` and `gemmaToolMiddleware` via commented lines. Choose a provider/model your account can access.
+- **Middleware switch**: parser examples show how to swap between `morphXmlToolMiddleware` and `gemmaToolMiddleware` via commented lines. Choose a provider/model your account can access.
 - **Providers**: OpenRouter base URL is `https://openrouter.ai/api/v1`. Friendli serverless base URL is `https://api.friendli.ai/serverless/v1`.
 - **Streaming vs non‑streaming**: `00-stream-tool-call.ts` uses streaming (`streamText`), while `00-tool-call.ts` is non‑streaming (`generateText`).

docs/getting-started.md

@@ -119,7 +119,7 @@ main().catch(console.error);
 ## Choose a middleware (protocol)
 
 - **gemmaToolMiddleware**: JSON tool calls inside markdown fences. Best for Gemma-like models.
-- **xmlToolMiddleware**: XML tool calls. Works well with GLM-style models.
+- **morphXmlToolMiddleware**: XML tool calls. Works well with GLM-style models.
 - **hermesToolMiddleware**: JSON payload wrapped in XML tags. For Hermes/Llama-style prompts.
 
 Swap by importing a different middleware; your tool definitions stay the same.

docs/guides/tool-calling.md

@@ -6,7 +6,7 @@ Use the middleware to enable tool calls on models without native `tools` support
 
 - `gemmaToolMiddleware` — JSON-mix in markdown fences (```tool_call)
 - `hermesToolMiddleware` — JSON-mix with XML wrappers (<tool_call>)
-- `xmlToolMiddleware` — Morph-XML protocol (native XML elements per tool name)
+- `morphXmlToolMiddleware` — Morph-XML protocol (native XML elements per tool name)
 
 ## Generate mode
 
@@ -112,7 +112,7 @@ Set env variables:
 - `hermesToolMiddleware` (JSON-mix with `<tool_call>`):
   - System prompt describes `<tools>` block and requires returning JSON inside `<tool_call> ... </tool_call>` tags.
   - Tool responses are formatted with `<tool_response>` tags.
-- `xmlToolMiddleware` (Morph-XML):
+- `morphXmlToolMiddleware` (Morph-XML):
   - Tool call must be an XML element named after the tool (e.g., `<get_weather>...</get_weather>`).
   - Arguments are parsed by RXML (Robust XML) via `RXML.parse` and then coerced via JSON Schema. On parse/coercion issues, the protocol falls back to emitting the original text and reports via `options.onError`.
   - To improve coercion accuracy, pass `originalToolSchemas` under `providerOptions.toolCallMiddleware`.

docs/index.md

@@ -6,7 +6,7 @@ Monorepo documentation for AI SDK tool-calling middleware and evaluation utiliti
 - Prebuilt Middlewares (choose one):
   - **gemmaToolMiddleware**: JSON tool calls in markdown fences — Gemma-like models
   - **hermesToolMiddleware**: JSON payload wrapped in XML tags — Hermes/Llama-style
-  - **xmlToolMiddleware**: XML elements per tool — GLM/XML-friendly models
+  - **morphXmlToolMiddleware**: XML elements per tool — GLM/XML-friendly models
 - Guides
   - [dev] Tool Calling: [guides/tool-calling.md](guides/tool-calling.md)
   - [dev] Handlers Overview: [guides/handlers.md](guides/handlers.md)

examples/eval-core/src/bfcl-simple.ts

@@ -7,7 +7,14 @@ import {
   evaluate,
   type ReporterType,
 } from "@ai-sdk-tool/eval";
-import { createToolMiddleware, morphXmlProtocol } from "@ai-sdk-tool/parser";
+import {
+  hermesToolMiddleware,
+  morphXmlToolMiddleware,
+} from "@ai-sdk-tool/parser";
+import {
+  sijawaraConciseXmlToolMiddleware,
+  sijawaraDetailedXmlToolMiddleware,
+} from "@ai-sdk-tool/parser/community";
 import { wrapLanguageModel } from "ai";
 
 const friendli = createOpenAICompatible({
@@ -16,34 +23,10 @@ const friendli = createOpenAICompatible({
   baseURL: "https://api.friendli.ai/serverless/v1",
 });
 
-const morphExpToolMiddleware = createToolMiddleware({
-  protocol: morphXmlProtocol,
-  toolSystemPromptTemplate(tools: string) {
-    return `You are a function calling AI model.
-
-Available functions are listed inside <tools></tools>.
-<tools>${tools}</tools>
-
-# Rules
-- Use exactly one XML element whose tag name is the function name.
-- Put each parameter as a child element.
-- Values must follow the schema exactly (numbers, arrays, objects, enums → copy as-is).
-- Do not add or remove functions or parameters.
-- Each required parameter must appear once.
-- Output nothing before or after the function call.
-
-# Example
-<get_weather>
-  <location>New York</location>
-  <unit>celsius</unit>
-</get_weather>`;
-  },
-});
-
 // const xmlGemma27b = wrapLanguageModel({
 //   model: friendli("google/gemma-3-27b-it"),
 //   // model: openrouter("z-ai/glm-4.5-air"),
-//   middleware: xmlToolMiddleware,
+//   middleware: morphXmlToolMiddleware,
 // });
 
 // const jsonGemma27b = wrapLanguageModel({
@@ -52,23 +35,42 @@ Available functions are listed inside <tools></tools>.
 //   middleware: gemmaToolMiddleware,
 // });
 
-// const morphExpGemma27b = wrapLanguageModel({
+// const morphXmlGemma27b = wrapLanguageModel({
 //   model: friendli("google/gemma-3-27b-it"),
-//   middleware: morphExpToolMiddleware,
+//   middleware: morphXmlToolMiddleware,
 // });
 
-// const compareDifferentMiddlewares = { xml: xmlGemma27b, morphExp: morphExpGemma27b, json: jsonGemma27b };
+// const compareDifferentMiddlewares = { xml: xmlGemma27b, morphXml: morphXmlGemma27b, json: jsonGemma27b };
+
+const testTargetModel = friendli(
+  "meta-llama/Llama-4-Maverick-17B-128E-Instruct"
+);
+const hermes = wrapLanguageModel({
+  model: testTargetModel,
+  middleware: hermesToolMiddleware,
+});
+
+const morphXml = wrapLanguageModel({
+  model: testTargetModel,
+  middleware: morphXmlToolMiddleware,
+});
 
-const morphExp = wrapLanguageModel({
-  model: friendli("LGAI-EXAONE/EXAONE-4.0.1-32B"),
-  middleware: morphExpToolMiddleware,
+const sijawaraDetailed = wrapLanguageModel({
+  model: testTargetModel,
+  middleware: sijawaraDetailedXmlToolMiddleware,
 });
 
-const original = friendli("LGAI-EXAONE/EXAONE-4.0.1-32B");
+const sijawaraConcise = wrapLanguageModel({
+  model: testTargetModel,
+  middleware: sijawaraConciseXmlToolMiddleware,
+});
 
 const compareWithNativeToolCalling = {
-  morphExp: morphExp,
-  original: original,
+  hermes: hermes,
+  morphXml: morphXml,
+  sijawaraDetailed: sijawaraDetailed,
+  sijawaraConcise: sijawaraConcise,
+  original: testTargetModel,
 };
 
 async function main() {