[Feat]: Support schema-constrained structured output

[vvlrff]

Is your feature request related to a problem? Please describe.

Summary

Today the chat completions API only supports response_format of {"type": "text"} and {"type": "json_object"}. JSON mode guarantees syntactically valid JSON, but it does not let the caller specify a schema the model must conform to. Every other major provider now supports schema-constrained structured output:

• OpenAI — response_format: {"type": "json_schema", "json_schema": {...}, "strict": true}
• Google Gemini — response_schema / responseJsonSchema
• xAI (Grok) — response_format: {"type": "json_schema", ...}

This means GLM models cannot be used as a drop-in replacement when an application relies on guaranteed-shape output. I'm integrating Z.AI as a provider in AG2, which exposes a single structured-output interface across providers; without schema support, Z.AI is the odd one out and we fall back to prompt-engineered json_object + client-side validation, which is unreliable (the model can return valid JSON that doesn't match the requested schema).

What I'm asking for

1. API: Support response_format = {"type": "json_schema", "json_schema": {...}} on chat completions, where json_schema carries a standard JSON Schema and the decoder constrains output to conform to it (ideally a strict flag for guaranteed conformance, as OpenAI does).
2. SDK (z-ai-sdk-python): Accept and forward this shape with proper typing. Currently response_format is typed as object | None and passed through untyped, so even once the API supports it there's no typed surface or documentation for it. (No opinion on whether a Pydantic-model convenience helper is added — the core need is just being able to pass a JSON Schema and have it enforced.)

Desired usage

from zai import ZaiClient

client = ZaiClient(api_key="your-api-key")

schema = {
    "type": "object",
    "properties": {
        "sentiment":  {"type": "string", "enum": ["positive", "neutral", "negative"]},
        "confidence": {"type": "number"},
        "keywords":   {"type": "array", "items": {"type": "string"}},
    },
    "required": ["sentiment", "confidence", "keywords"],
    "additionalProperties": False,
}

response = client.chat.completions.create(
    model="glm-4.6",
    messages=[
        {"role": "user", "content": "Analyze the sentiment of: 'I love this product!'"},
    ],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "sentiment_analysis",
            "schema": schema,
            "strict": True,
        },
    },
)

# Guaranteed to parse AND match `schema`:
import json
data = json.loads(response.choices[0].message.content)

Current behavior

# Only these are supported today:
response_format={"type": "text"}
response_format={"type": "json_object"}   # valid JSON, but shape is not enforced

Passing a json_schema object is silently ignored / not honored by the API, and the SDK's response_format: object | None typing gives no indication of what's accepted.

Why it matters

Schema-constrained output is now the standard way applications get reliable, machine-parseable results from LLMs (tool/agent frameworks, extraction pipelines, multi-provider routers). Supporting it would make GLM models usable as a drop-in replacement in these stacks instead of requiring provider-specific fallbacks.

Thanks for considering this!

Describe the solution you'd like

schema = {
    "type": "object",
    "properties": {
        "sentiment":  {"type": "string", "enum": ["positive", "neutral", "negative"]},
        "confidence": {"type": "number"},
        "keywords":   {"type": "array", "items": {"type": "string"}},
    },
    "required": ["sentiment", "confidence", "keywords"],
    "additionalProperties": False,
}

response = client.chat.completions.create(
    model="glm-5.2",
    messages=[{"role": "user", "content": "Analyze the sentiment of: 'I love this product!'"}],
    response_format={
        "type": "json_schema",
        "json_schema": {"name": "sentiment_analysis", "schema": schema, "strict": True},
    },
)
# response.choices[0].message.content is guaranteed to parse AND match `schema`

Describe alternatives you've considered

No response

Additional context

No response

Code of Conduct

• I agree to follow this project's Code of Conduct