GLMFormatter

Formatter for Zhipu AI (Z.ai) GLM models (glm-5.2, glm-5.1, glm-5, glm-4.7, glm-4.6, etc.).

Adapted to the latest Zhipu open platform Chat Completions API (https://open.bigmodel.cn/api/paas/v4/chat/completions):

• At least one user message is required (error 1214 otherwise)

• Does NOT support the name parameter in messages

• tool_choice only supports "auto"; forced choices are degraded, while none removes tools to preserve the no-tool-call contract

• Does NOT support the strict parameter in tool definitions

• Does NOT support frequency_penalty / presence_penalty / thinking_budget; they are stripped from the request

• Only supports max_tokens; max_completion_tokens is mapped to max_tokens when the latter is not set

• temperature range is [0.0, 1.0] (OpenAI allows up to 2.0) and top_p range is [0.01, 1.0]; out-of-range values are clamped

Thinking mode: GLM-4.5 and later models accept a thinking object (e.g. {"type": "enabled"}). GLM-4.7 and GLM-5 series enable thinking by default. It can be controlled through GenerateOptions.additionalBodyParam("thinking", Map.of("type", "disabled")), which is passed through to the request body as-is. GLM-5.2 additionally supports the standard reasoning_effort option (GenerateOptions.reasoningEffort, default max) and streaming tool-call arguments via GenerateOptions.additionalBodyParam("tool_stream", true).

Usage:

`OpenAIChatModel.builder()
.formatter(new GLMFormatter())
.modelName("glm-5.2")
.baseUrl("https://open.bigmodel.cn/api/paas/v4")
.apiKey(apiKey)
.build();
`