Feat hzw 20260810

[lijiayang619]

模型配置修改

[Copilot]

🟡 Changes recommended

Unresolved critical and moderate findings affect runtime behavior, API contracts, tests, security, and deployment.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This pull request expands model configuration across the backend, SDK, frontend, persistence, catalog, and deployment layers.

Changes:

• Adds model catalogs, provider discovery, and capacity suggestions.
• Adds inference parameters and agent/knowledge-base overrides.
• Updates model UI, connectivity checks, migrations, tests, and deployment wiring.

Unresolved findings include catalog mismatches, provider and TLS issues, dropped overrides, exception/configuration problems, an invalid test, an incorrect frontend port, and an unpinned catalog dependency.

File summaries

File	Reviewed change
test/backend/test_model_consts.py	Tests model inference-parameter schemas and defaults.
test/backend/configs/test_model_catalog_loader.py	Tests catalog loading and defaults.
test/backend/configs/__init__.py	Initializes the backend test package.
sdk/nexent/core/models/openai_llm.py	Adds inference parameters to connectivity checks.
frontend/types/modelConfig.ts	Defines model, catalog, and inference types.
frontend/types/agentConfig.ts	Defines agent override types.
frontend/stores/agentStore.ts	Tracks agent draft model overrides.
frontend/stores/agentConfigStore.ts	Tracks editable agent configuration state.
frontend/services/modelService.ts	Handles model, catalog, batch, and inference API operations.
frontend/services/api.ts	Provides frontend API helpers.
frontend/services/agentConfigService.ts	Maps agent configuration API data.
frontend/server.js	Configures the production frontend server.
frontend/hooks/model/useModelCatalog.ts	Loads model catalog data.
frontend/hooks/model/useInferenceFieldSpecs.ts	Loads inference field specifications.
frontend/app/[locale]/newchat/assistant-ui/thread-list.tsx	Adds batch conversation selection and deletion.
frontend/app/[locale]/models/components/model/ModelOverrideModal.tsx	Provides model override editing.
frontend/app/[locale]/models/components/model/ModelListCard.tsx	Renders model list cards.
frontend/app/[locale]/models/components/model/ModelItemCard.tsx	Renders individual model cards.
frontend/app/[locale]/models/components/model/ModelEditDialogV2.tsx	Edits model configuration and inference defaults.
frontend/app/[locale]/models/components/model/ModelCapacityFields.tsx	Renders model capacity settings.
frontend/app/[locale]/models/components/model/ModelAdvancedSettings.tsx	Renders advanced inference settings.
frontend/app/[locale]/models/components/model/DefaultModelDialog.tsx	Configures default models.
frontend/app/[locale]/knowledges/KnowledgeBaseConfiguration.tsx	Adds knowledge-base embedding overrides.
frontend/app/[locale]/knowledges/components/document/DocumentList.tsx	Provides knowledge-base document controls.
frontend/app/[locale]/agents/components/agent-prompt.tsx	Adds per-model agent overrides.
deploy/sql/migrations/v2.6.0_0806_add_model_inference_params.sql	Persists model inference parameters.
deploy/images/dockerfiles/main/Dockerfile	Bundles the runtime model catalog.
backend/services/providers/openai_provider.py	Discovers models from OpenAI-compatible providers.
backend/services/model_provider_service.py	Coordinates provider discovery.
backend/services/model_management_service.py	Builds provider model lists.
backend/services/model_health_service.py	Handles model health and connectivity checks.
backend/services/model_gateway_service.py	Maps inference parameters into gateway requests.
backend/services/model_capacity_suggestion_service.py	Provides catalog and LiteLLM capacity suggestions.
backend/management/services/agent/service.py	Persists agent override updates.
backend/management/services/agent/management.py	Manages agent configuration operations.
backend/database/model_management_db.py	Filters and persists model parameters.
backend/database/db_models.py	Defines model and agent database fields.
backend/database/agent_db.py	Persists agent model overrides.
backend/consts/model.py	Defines model catalog and inference schemas.
backend/consts/const.py	Defines shared configuration constants.
backend/configs/model_catalog.json	Provides preset provider and model metadata.
backend/configs/model_catalog_loader.py	Loads and normalizes the model catalog.
backend/configs/__init__.py	Initializes the backend configuration package.
backend/apps/model_managment_app.py	Exposes catalog and management endpoints.
backend/agents/create_agent_info.py	Applies runtime agent model overrides.

Review details

Suppressed comments (13)

backend/services/model_gateway_service.py:95

• This helper drops every fixed inference field and returns only extra_params.__custom__. enable_thinking is explicitly stored in extra_params and the model/agent UI sends it there, so gateway adapters silently ignore that setting (including connectivity probes) even though the model record persists it. Merge the supported fixed fields into the adapter request as well, with provider-specific translation where required.

    custom = extra_params.get("__custom__")
    return custom if isinstance(custom, dict) and custom else None

backend/services/providers/openai_provider.py:50

• When the request supplies model_type, the caller skips name inference, but this provider intentionally leaves model_type off every row. Typed discovery therefore returns rows without their requested type; ModelAddDialogV2 falls back to llm, so embedding/rerank/voice requests can be presented and saved as LLMs. Preserve the requested type in the annotation when it is supplied.

    try:
        ip = ipaddress.ip_address(host)

deploy/images/dockerfiles/main/Dockerfile:78

• Fetching this runtime dependency from @main makes otherwise identical image builds consume different catalog contents over time. Worse, a transient build-network failure silently writes {}, producing an image with no LiteLLM capacity fallback. Pin a release/commit or vendor a known artifact, and make an unavailable catalog visible rather than silently changing behavior.

    https://cdn.jsdelivr.net/gh/BerriAI/litellm@main/model_prices_and_context_window.json \
    || echo '{}' > /opt/nexent/litellm_models.json

deploy/sql/migrations/v2.6.0_0806_add_model_inference_params.sql:42

• The frontend now exposes and edits per-knowledge-base embedding overrides, but this migration adds storage only to ag_tenant_agent_t; createKnowledgeBase sends no override field and the knowledge record has no corresponding column. Values changed in the new KB UI are therefore discarded when the KB is created. Add a KB persistence/API path or remove the UI until it is supported.

ALTER TABLE nexent.ag_tenant_agent_t
ADD COLUMN IF NOT EXISTS model_params_override JSONB DEFAULT NULL;

frontend/app/[locale]/agents/components/agent-prompt.tsx:149

• This effect only depends on configuringModelId, so it can initialize while useInferenceFieldSpecs is still empty and never rerun when the specs request resolves. Opening the override modal during that race leaves the form without the model defaults/spec fields; include the loaded model/specs and the relevant override entry in the initialization dependencies.

  }, [configuringModelId]);

frontend/app/[locale]/models/components/model/ModelAddDialogV2.tsx:924

• On the custom add path, buildInferenceParamsPayload returns voice fields as snake_case (model_factory, model_appid, access_token), but addCustomModel serializes only the camelCase modelFactory/modelAppid/accessToken properties. Unlike the batch path, these fields are never mapped, so STT/TTS models are created with the default OpenAI-compatible factory and lose required credentials. Add the camelCase voice fields before submitting.

          ...inferencePayload,

frontend/app/[locale]/models/components/model/ModelAddDialogV2.tsx:920

• The custom-submit path never copies the successful customConnectivity result into modelParams, unlike the batch path. Consequently a model that just passed the connectivity probe is created without connect_status and is stored as not_detected despite the new API field intended to persist the verified status.

        const modelParams: any = {
          name: customForm.name,
          type: resolvedModelType,
          url: customForm.url,
          apiKey: customForm.apiKey.trim() === "" ? "sk-no-api-key" : customForm.apiKey,

frontend/app/[locale]/models/components/model/ModelAdvancedSettings.tsx:38

• This documentation says temperature, top_p, and extra_params are not included, but the same component classifies temperature/top_p as dedicated keys and buildInferenceParamsPayload sends them. The contradiction makes the storage contract misleading and can lead future changes to remove working fields; update the comment to describe dedicated columns plus extra_params JSONB accurately.

//  - temperature / top_p / extra_params and other v2.6.0 additions are NOT
//    included — the new dialog matches the original ModelAddDialog's parameter
//    set without adding new inference parameters.

frontend/services/modelService.ts:387

• The method now advertises type as optional and says each model carries its own type, but /model/batch_create still validates BatchCreateModelsRequest.type as required and the backend batch service overwrites every model's model_type with the top-level value. Omitting type therefore yields a 422, while per-model types are not honored; update the backend contract/service or keep this field required.
  frontend/services/modelService.ts:1301
• The manage batch endpoint has the same contract mismatch: this signature allows type to be omitted, but ManageBatchCreateModelsRequest.type is required and batch_create_models_for_tenant indexes the key unconditionally. A caller following this type definition receives a validation error before the request reaches the service; either update the backend for per-row types or keep type required.
  frontend/services/modelService.ts:115
• This helper is used by the update paths as well as create. Clearing an existing field in ModelEditDialogV2 removes it from buildInferenceParamsPayload, so temperature, top_p, or an emptied extra_params map becomes undefined here and is omitted from the PATCH-like request; the backend then leaves the old database value intact. Update requests need an explicit clear representation (for example null) distinct from an omitted field.
  frontend/services/modelService.ts:1406
• This manage-list method also advertises an optional type, but /model/manage/provider/list still uses ManageProviderModelListRequest.model_type as a required field. A caller following this signature without type receives a 422; keep it required or update that request model and the service to support all-type discovery.
  frontend/types/modelConfig.ts:240
• The backend catalog profile emits chunk_batch (see ModelCatalogProfile and model_dump()), but this interface declares chunking_batch_size and getFullCatalog() performs no mapping. The documented 1:1 API type is therefore wrong and typed consumers reading the declared field will get undefined.

• Files reviewed: 47/50 changed files
• Comments generated: 11
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[codecov]

Codecov Report

❌ Patch coverage is 57.69231% with 308 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
backend/configs/model_catalog_loader.py	64.09%	53 Missing and 26 partials ⚠️
backend/apps/model_managment_app.py	26.08%	68 Missing ⚠️
...kend/services/model_capacity_suggestion_service.py	45.88%	42 Missing and 4 partials ⚠️
backend/agents/create_agent_info.py	10.25%	32 Missing and 3 partials ⚠️
backend/consts/model.py	76.56%	28 Missing and 2 partials ⚠️
backend/services/providers/openai_provider.py	79.22%	9 Missing and 7 partials ⚠️
backend/services/model_management_service.py	50.00%	11 Missing and 3 partials ⚠️
backend/database/model_management_db.py	27.27%	6 Missing and 2 partials ⚠️
backend/services/model_gateway_service.py	78.57%	2 Missing and 1 partial ⚠️
backend/services/model_health_service.py	62.50%	1 Missing and 2 partials ⚠️
... and 3 more

📢 Thoughts on this report? Let us know!