feat(minimax): add MiniMax provider support

[Bhavyashah20]

Summary

Adds first-class MiniMax support to instructor via from_minimax() and from_provider("minimax/...").

• MINIMAX_TOOLS (default): uses the model's OpenAI-compatible tool-calling API for structured output
• MINIMAX_JSON: injects the Pydantic schema into the system prompt — MiniMax does not support response_format, so tool calling is the preferred mode but JSON mode is available as a fallback
• Think-tag stripping: MiniMax reasoning models prepend <think>…</think> blocks before their answer in JSON mode; these are stripped automatically before JSON parsing so callers never see them
• No new SDK dependency — uses the standard openai package pointed at https://api.minimax.io/v1

Usage

import os, openai, instructor
from pydantic import BaseModel

class User(BaseModel):
    name: str
    age: int

# one-liner
client = instructor.from_provider("minimax/MiniMax-M2.7")

# or explicit
raw = openai.OpenAI(api_key=os.environ["MINIMAX_API_KEY"], base_url="https://api.minimax.io/v1")
client = instructor.from_minimax(raw)

user = client.chat.completions.create(
    model="MiniMax-M2.7",
    messages=[{"role": "user", "content": "Extract: Jason is 25 years old"}],
    response_model=User,
)
# User(name='Jason', age=25)

Files changed

File	Change
instructor/v2/core/mode.py	MINIMAX_TOOLS, MINIMAX_JSON added to Mode enum, tool_modes(), json_modes(), and DEPRECATED_TO_CORE
instructor/v2/core/providers.py	Provider.MINIMAX added, wired in provider_from_mode() and get_provider()
instructor/v2/core/provider_specs.py	PROVIDER_SPECS entry for MiniMax — supported/unsupported modes, legacy mode aliases, factory function reference
instructor/v2/providers/minimax/handlers.py	MiniMaxToolsHandler (extends OpenAIToolsHandler) and MiniMaxMDJSONHandler (injects schema into system prompt, strips think-tags)
instructor/v2/providers/minimax/client.py	from_minimax() factory using _from_openai_compat(), defaults to Mode.TOOLS
instructor/v2/providers/minimax/__init__.py	Exports from_minimax, both handler classes
instructor/v2/auto_client.py	_build_minimax() builder added to _PROVIDER_BUILDERS dict
instructor/providers/minimax/client.py	Thin shim re-exporting from instructor.v2.providers.minimax.client
instructor/__init__.py	from_minimax added to _LAZY_IMPORTS and _add_optional_export
docs/integrations/minimax.md	Full integration docs with sync/async/nested examples
tests/llm/test_minimax/	Integration tests (auto-skipped without MINIMAX_API_KEY)
mkdocs.yml	Navigation entry under Integrations
CHANGELOG.md	Unreleased entry

Closes #2260

[Bhavyashah20]

Just wanted to flag the architecture note for reviewers: this follows the v2 provider pattern introduced in the recent refactor. MiniMax is registered via PROVIDER_SPECS in instructor/v2/core/provider_specs.py, the handlers live in instructor/v2/providers/minimax/, and instructor/providers/minimax/client.py is a thin shim for backward compatibility.

The one non-obvious design choice: MINIMAX_JSON mode injects the Pydantic schema directly into the system prompt rather than using response_format, because MiniMax's API doesn't support response_format on all models. Think-tag stripping happens in the handler before JSON parsing so callers never see raw <think>…</think> output.

Happy to adjust anything if this doesn't align with where the v2 architecture is heading.

[Bhavyashah20]

Hi, just following up on this PR — it's been open for a few weeks and wanted to check if there's anything needed from my side or if there are any concerns with the changes. Happy to revise if anything looks off. Thanks!