[octo-patch]

Reason: new MiniMax text agents still selected M3, and the current catalog lacked MiniMax-M3.1-Flash and its adaptive-thinking effort contract.

Description

Add Flash before M3 in the direct MiniMax and verified OpenRouter catalogs, and use distinct Flash IDs for new ordinary text selections. Existing minimax / minimax/minimax aliases keep their M3 meaning. Explicit M2.7 and M2-her requests remain resolvable but do not appear in ordinary model lists. Other providers and the multimodal picker retain their existing choices.

The direct Anthropic-compatible adapter sends adaptive thinking and output_config.effort (default max) for Flash; unsupported none / minimal requests fail before HTTP. M3 keeps its thinking toggle without effort-depth fields. MINIMAX_API_KEY / MINIMAX_BASE_URL support direct accounts with the existing MINIMAX_PLAN_* settings as fallbacks. Arbitrary compatible-provider configurations remain intact; custom endpoint support does not imply regional or subscription entitlement.

OpenRouter uses the verified minimax/minimax-m3.1-flash:free route, zero gateway pricing, and the minimax provider pin. M3 retains its gateway-specific limits and minimax/fp8 pin. Direct M3's existing metadata and billing scope are not expanded in this patch.

Sources:

• https://platform.minimax.io/docs/guides/models-intro
• https://platform.minimax.io/docs/api-reference/text-anthropic-api
• https://platform.minimax.io/docs/pricing/overview
• https://openrouter.ai/api/v1/models/minimax/minimax-m3.1-flash:free/endpoints
• https://openrouter.ai/api/v1/models/minimax/minimax-m3/endpoints

Type of Change

• Bugfix
• New Feature
• Improvement
• Documentation Update

Validation

• 91 passed, 1 deselected in the focused Python 3.13.14 suite: real picker → provider → LangChain → SDK HTTP serialization, default/effort handling, saved direct/gateway/composite IDs, custom compatible endpoints, tools, images, reasoning history, catalog, picker, cost and provider-pin regressions. HTTP sends are intercepted offline; no paid inference.
• SDK pins from uv.lock: anthropic 1.6.0, langchain-anthropic 1.7.2, langchain-core 1.6.3, langchain-openrouter 0.2.8, openrouter 0.11.46, openai 3.14.1, langchain-openai 1.6.2. pytest 9.1.1 / pytest-asyncio 1.4.0.
• Repository-pinned Ruff 0.16.7 passes across all 841 Python files (format and lint). Python syntax, rollout-only whitespace, exact non-MiniMax catalog preservation, and unchanged multimodal-picker checks pass.
• Focused validation used a package namespace shell to bypass the unrelated eager engine/blockchain initializer; actual config/catalog/provider/picker source, repository conftest, and SDK implementations were loaded. The one deselected test imports the full agent model. This is not a full-suite result.
• Scoped basedpyright 1.40.1 remains incomplete: the trimmed environment lacks botocore.session, aws_secretsmanager_caching, langchain_deepseek, langchain_google_genai, and langchain_ollama. Full dependency/build setup is paused by the shared disk-capacity gate. Upstream Lint CI requires maintainer approval (action_required).

Checklist

• I have read the contributing guidelines.
• I have added tests to cover my changes.
• All new and existing tests passed.

Related Issue

No linked issue. This updates the existing text-model PR in place after merging current upstream.

[octo-patch]

Updated the MiniMax model registry to keep MiniMax M3 pricing on the catalog target values and removed the unsupported tier pricing path. I ran uv run --with psycopg-binary pytest tests/models/test_minimax_provider.py tests/models/test_llm_cost.py and uv run ruff check intentkit/models/llm.py tests/models/test_minimax_provider.py tests/models/test_llm_cost.py.