fix: make reasoning effort options model-aware

[zwj110610]

Summary

This PR makes reasoning_effort options model-aware so DeerFlow does not send unsupported reasoning effort values to providers such as DeepSeek.

Previously, the frontend could send reasoning_effort: "minimal" for Flash or stale local settings, but some providers only accept values such as low, medium, high, max, and xhigh. That could cause requests to fail with a provider-side 400 error.

Changes

• Add optional reasoning_efforts to model configuration.
• Expose reasoning_efforts through the Gateway /api/models response and embedded Python client.
• Filter unsupported reasoning_effort values in the backend model factory as a defensive guard.
• Update the frontend to render only the reasoning effort values supported by the selected model.
• Normalize stale local reasoning effort values, such as minimal, to a supported value for the current model/mode.
• Document the new config/API field and update config.example.yaml.
• Add backend and frontend regression tests.

Testing

• make doctor
• cd backend && uv run --group dev pytest tests/test_model_factory.py -k reasoning_effort
• cd backend && uv run --group dev pytest tests/test_client.py -k "list_models or get_model"
• cd backend && uv run --group dev ruff check app/gateway/routers/models.py packages/harness/deerflow/client.py packages/harness/deerflow/config/model_config.py packages/harness/deerflow/models/factory.py tests/test_client.py tests/test_model_factory.py
• cd frontend && pnpm vitest run tests/unit/core/models/reasoning-effort.test.ts
• cd frontend && pnpm check
• pre-commit hooks passed during commit

[Copilot]

Pull request overview

Makes reasoning_effort model-aware to prevent sending unsupported values (e.g., "minimal") to providers like DeepSeek. Adds an optional reasoning_efforts allowlist on model config that flows through the gateway/client APIs and is enforced both in the backend factory and in the frontend UI.

Changes:

• Add optional reasoning_efforts field to ModelConfig, surface via gateway /api/models and embedded client, and filter unsupported values in the model factory.
• Centralize a new ReasoningEffort type plus getModelReasoningEfforts/getDefaultReasoningEffort/normalizeReasoningEffort helpers, and refactor InputBox to render and normalize per-model allowed values (adds max/xhigh i18n labels).
• Update docs/config example and add backend + frontend regression tests.

Reviewed changes

Copilot reviewed 19 out of 19 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
backend/packages/harness/deerflow/config/model_config.py	Adds optional reasoning_efforts list field.
backend/packages/harness/deerflow/models/factory.py	Strips runtime/config reasoning_effort values not in allowlist; excludes new field from kwargs.
backend/packages/harness/deerflow/client.py	Adds helper and includes reasoning_efforts in list_models/get_model.
backend/app/gateway/routers/models.py	Exposes reasoning_efforts in ModelResponse and example payloads.
backend/tests/test_model_factory.py	Adds tests for allowlist filtering and preservation.
backend/tests/test_client.py	Asserts reasoning_efforts propagation through client/gateway.
backend/docs/CONFIGURATION.md, backend/docs/API.md	Documents the new field.
config.example.yaml	Example commented config for reasoning_efforts.
frontend/src/core/threads/types.ts	New shared ReasoningEffort type including max/xhigh.
frontend/src/core/models/types.ts	Adds reasoning_efforts to Model.
frontend/src/core/models/reasoning-effort.ts	New helpers for available/default/normalized efforts.
frontend/src/core/models/index.ts	Re-exports the new helpers.
frontend/src/core/settings/local.ts	Uses shared ReasoningEffort type.
frontend/src/components/workspace/input-box.tsx	Renders model-aware efforts, normalizes stale values on model/mode/init/early-submit.
frontend/src/core/i18n/locales/{types,en-US,zh-CN}.ts	Adds max/xhigh labels and descriptions.
frontend/tests/unit/core/models/reasoning-effort.test.ts	Unit tests for helper behavior.