[octo-patch]

Reason: Make MiniMax M3.1 Flash the first and default ordinary text model while preserving explicit M3, M2.7, saved/custom/legacy selections, and existing multimodal inputs.

Type of Change

New Feature

Description

• Adds the verified direct MiniMax M3.1 Flash catalog entry with 1,000,000 context, 524,288 maximum output tokens, adaptive-only thinking, and published standard pricing.
• Makes MiniMax-M3.1-Flash the provider default and first ordinary catalog choice; keeps MiniMax M3 as the second ordinary choice and hides legacy M2.7 from the ordinary catalog.
• Sends Flash reasoning_effort: max through the OpenAI-compatible boundary and output_config.effort: max through the Anthropic-compatible boundary.
• Applies Flash request metadata only to the matching catalog model ID, preserving explicit M3 multimodal payloads and custom or legacy model IDs; explicit legacy M2.7 remains creatable even though it is hidden from ordinary choices.

Testing

• python3 -m py_compile python/valuecell/adapters/models/factory.py python/valuecell/adapters/models/tests/test_minimax_provider.py
• YAML catalog assertions for default/order/context/output/thinking/effort/pricing and legacy M2.7 filtering
• Ruff format check and import-order check for touched Python files
• git diff --check
• Focused pytest: isolated Python 3.12 environment and 42 light SDK/test wheels installed, but collection stopped because the router import chain requires pinned SQLAlchemy; the bounded SQLAlchemy-only install timed out before completion. No paid inference or credentialed provider call was made.
• Full Python/frontend suites: not run because the rollout environment is below the 110 GiB heavy-setup threshold and no existing project environment is available.

Checklist

• I have read the Contributing Guidelines.
• My changes follow the project coding style for touched code.
• I have added relevant mocked SDK/HTTP boundary tests.

[octo-patch]

Corrected the MiniMax-M3 per-token pricing in the provider config to match the published rates (input $0.6/M, output $2.4/M, cache read $0.12/M) and updated the corresponding test assertion. Ran the MiniMax provider tests (6 passed) and ruff on the changed Python files.

[octo-patch]

Updated the MiniMax model availability probe to support both regional OpenAI-compatible and Anthropic-compatible endpoints, with matching request/response handling, and pushed commit fbd2c77. I ran the focused MiniMax tests (10 passed), the full Python suite (411 passed), Ruff, and the frontend lint, typecheck, and build.

[octo-patch]

Thanks. I aligned the MiniMax provider config to the target pricing fields and pushed a new commit. I ran YAML assertions for the MiniMax provider config and git diff --check; focused pytest could not collect in the local system Python because agno is not installed, and a full uv sync could not complete because dependency extraction exhausted the local disk.

[octo-patch]

Updated the MiniMax provider metadata to include all supported Flash input modalities and current M3 pricing. I also adjusted the Anthropic boundary test setup so the provider's regional base URL remains under test. I ran the 12 MiniMax provider tests, Ruff, and git diff --check.