Feature: UI dynamic provider model discovery by nujragan93 · Pull Request #1274 · kagent-dev/kagent

nujragan93 · 2026-02-09T15:10:12Z

Summary

Introduces a new Provider Custom Resource Definition (CRD) that enables automatic model discovery from LLM providers. Users can configure provider endpoints (e.g., AI gateways, self-hosted LLMs) and kagent will dynamically discover available models.

Changes

Backend (Go)

New CRD: Provider type (go/api/v1alpha2/provider_types.go) with fields for provider type, endpoint, and secret reference
Controller: ProviderController watches Provider CRs and triggers model discovery
Model Discovery: provider/discoverer.go - discovers models from OpenAI, Anthropic, Ollama, Azure, Gemini, and Bedrock APIs
Provider Manager: provider/manager.go - caches discovered models and manages provider state
API Endpoints:
- GET /api/providers/configured - list configured providers
- GET /api/providers/configured/{name}/models - get models for a provider

Frontend (React/Next.js)

ProviderCombobox: New dropdown grouping "Stock Providers" and "Configured Providers"
ModelCombobox: Displays discovered models with function-calling indicator
Dynamic Model Fetching: Fetches models from configured provider endpoints
Auto-selection: Defaults to stock OpenAI provider on page load

Infrastructure

CRD manifest: kagent.dev_providers.yaml
Helm chart templates for CRD and RBAC
Sample Provider manifests for OpenAI and Anthropic

Testing

Unit tests for model discoverer (17 tests)
Unit tests for provider manager
Provider selection behavior tests

Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

nujragan93 · 2026-02-09T15:17:28Z

EItanya

These changes are a super exciting first step for this feature! Really excited to see it all coming together. I have some structural changes as well as some nits. I think there will probably be multiple rounds of reviews, but this is a good start.

go/api/v1alpha2/provider_types.go

go/internal/controller/provider/discoverer.go

go/internal/controller/reconciler/reconciler.go

go/internal/controller/provider/manager.go

Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

…/nujragan93/kagent into feature/provider-model-discovery Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

Signed-off-by: Nagarjun Krishnan <159933832+nujragan93@users.noreply.github.com>

Copilot

Pull request overview

This PR introduces dynamic model discovery for LLM providers through a new Provider Custom Resource Definition (CRD). It enables users to configure custom provider endpoints (e.g., AI gateways, self-hosted LLMs) and automatically discover available models from those providers.

Changes:

Added Provider CRD with controller for automatic model discovery from OpenAI, Anthropic, Ollama, Azure, Gemini, and Bedrock APIs
Implemented backend API endpoints to list configured providers and their models
Created new React components (ProviderCombobox, ModelCombobox) to support provider and model selection in the UI
Added auto-selection of stock OpenAI provider on page load and fixed provider selection persistence bug

Reviewed changes

Copilot reviewed 30 out of 31 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
go/api/v1alpha2/provider_types.go	Defines Provider CRD structure with spec and status fields
go/internal/controller/provider_controller.go	Controller that watches Provider CRs and triggers reconciliation
go/internal/controller/reconciler/reconciler.go	Reconciliation logic for model discovery and status updates
go/internal/controller/provider/discoverer.go	Model discovery implementation for various LLM providers
go/internal/httpserver/handlers/providers.go	API handlers for listing providers and fetching models
go/config/crd/bases/kagent.dev_providers.yaml	Base CRD manifest with validation rules
helm/kagent-crds/templates/kagent.dev_providers.yaml	Helm template for Provider CRD
ui/src/types/index.ts	TypeScript type definitions for providers and models
ui/src/components/ProviderCombobox.tsx	New component for provider selection with grouping
ui/src/components/ModelCombobox.tsx	New component for model selection
ui/src/components/models/new/BasicInfoSection.tsx	Updated to use separate provider and model selection
ui/src/app/models/new/page.tsx	Main page logic with auto-fetch and provider management
ui/src/app/actions/providers.ts	Server actions for fetching configured providers and models

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-02-11T21:23:34Z

helm/kagent-crds/templates/kagent.dev_providers.yaml

+    - jsonPath: .spec.endpoint
+      name: Endpoint
+      type: string
+    - jsonPath: .status.discoveredModels[?(@)]


The JSONPath for the "Models" additionalPrinterColumn is inconsistent with the base CRD. In the base CRD (go/config/crd/bases/kagent.dev_providers.yaml), this uses .status.modelCount, but here it uses .status.discoveredModels[?(@)].

The JSONPath .status.discoveredModels[?(@)] is invalid and will not work correctly. It should be .status.modelCount to match the base CRD definition and to display the count of models as an integer.

Suggested change

- jsonPath: .status.discoveredModels[?(@)]

- jsonPath: .status.modelCount

Copilot · 2026-02-11T21:23:35Z

go/config/crd/bases/kagent.dev_providers.yaml

+            - endpoint
+            - secretRef
+            - type
+            type: object
+            x-kubernetes-validations:
+            - message: endpoint must be a valid URL starting with http:// or https://
+              rule: self.endpoint.startsWith('http://') || self.endpoint.startsWith('https://')


The CRD defines endpoint as a required field (line 96-97), but in the Go API definition (provider_types.go line 77), the Endpoint field is marked as omitempty and has validation rules that explicitly allow it to be empty (line 65: !has(self.endpoint) || self.endpoint == ''). Additionally, the GetEndpoint() method provides default values when the endpoint is not specified.

This creates an inconsistency where the CRD will reject Provider resources without an endpoint, even though the code is designed to support optional endpoints with defaults. The endpoint field should not be in the required list in the CRD.

Suggested change

- endpoint

- secretRef

- type

type: object

x-kubernetes-validations:

- message: endpoint must be a valid URL starting with http:// or https://

rule: self.endpoint.startsWith('http://') || self.endpoint.startsWith('https://')

- secretRef

- type

type: object

x-kubernetes-validations:

- message: endpoint must be a valid URL starting with http:// or https://

rule: "!has(self.endpoint) || self.endpoint == '' || self.endpoint.startsWith('http://') || self.endpoint.startsWith('https://')"

Copilot · 2026-02-11T21:23:35Z

helm/kagent-crds/templates/kagent.dev_providers.yaml

+            - endpoint
+            - secretRef
+            - type
+            type: object
+            x-kubernetes-validations:
+            - message: endpoint must be a valid URL starting with http:// or https://
+              rule: self.endpoint.startsWith('http://') || self.endpoint.startsWith('https://')
+            - message: secretRef.name and secretRef.key are required
+              rule: has(self.secretRef) && has(self.secretRef.name) && size(self.secretRef.name)
+                > 0 && has(self.secretRef.key) && size(self.secretRef.key) > 0


Same issue as in go/config/crd/bases/kagent.dev_providers.yaml: the endpoint and secretRef fields are marked as required, but the Go API definition treats them as optional with defaults. This will prevent users from creating Provider resources without explicitly specifying an endpoint, even though the code supports defaults via the GetEndpoint() method.

Additionally, secretRef should only be required for providers that need authentication (not Ollama), as indicated by the XValidation rule in the Go type definition.

Suggested change

- endpoint

- secretRef

- type

type: object

x-kubernetes-validations:

- message: endpoint must be a valid URL starting with http:// or https://

rule: self.endpoint.startsWith('http://') || self.endpoint.startsWith('https://')

- message: secretRef.name and secretRef.key are required

rule: has(self.secretRef) && has(self.secretRef.name) && size(self.secretRef.name)

> 0 && has(self.secretRef.key) && size(self.secretRef.key) > 0

- type

type: object

x-kubernetes-validations:

- message: endpoint must be a valid URL starting with http:// or https://

rule: !has(self.endpoint) || self.endpoint.startsWith('http://') || self.endpoint.startsWith('https://')

- message: secretRef.name and secretRef.key are required for providers that need authentication (not Ollama)

rule: self.type == 'Ollama' || (has(self.secretRef) && has(self.secretRef.name) && size(self.secretRef.name) > 0 && has(self.secretRef.key) && size(self.secretRef.key) > 0)

Copilot · 2026-02-11T21:23:36Z

go/internal/controller/provider_controller.go

+func providerReferencesSecret(provider *v1alpha2.Provider, secretObj types.NamespacedName) bool {
+	// Secrets must be in the same namespace as the provider
+	return provider.Namespace == secretObj.Namespace &&
+		provider.Spec.SecretRef.Name == secretObj.Name
+}


The function providerReferencesSecret directly accesses provider.Spec.SecretRef.Name without checking if SecretRef is nil. Since SecretRef is an optional pointer field (e.g., for Ollama providers that don't require authentication), this will cause a nil pointer dereference panic when processing providers without a SecretRef.

Add a nil check before accessing SecretRef fields.

Copilot · 2026-02-11T21:23:36Z

ui/src/types/index.ts

 export interface TypedLocalReference {
  kind?: string;
  apiGroup?: string;
  name: string;
-  namespace?: string;
 }


The namespace field was removed from the TypedLocalReference interface, but it is still being accessed in ui/src/app/actions/agents.ts (line 72) and ui/src/lib/toolUtils.ts (line 157). This will cause TypeScript compilation errors or runtime issues.

The code attempts to access agent.namespace and tool.agent.namespace, which no longer exist on the type definition. This breaking change needs to be addressed by either:

Keeping the namespace field in TypedLocalReference, or

Updating all code that references it to use an alternative approach

Copilot · 2026-02-11T21:23:36Z

ui/src/app/models/new/page.tsx

+
+    fetchProviderModels();
+    return () => { isMounted = false; };
+  }, [selectedProvider, isEditMode, providerModelsData]);


The auto-fetch models effect (lines 279-332) includes providerModelsData in its dependency array (line 332). This creates a problematic cycle:

When a provider is selected, it checks if models exist in providerModelsData

If not, it fetches models and updates providerModelsData (lines 308-311, 323)

The update to providerModelsData triggers the effect to run again

This could lead to unnecessary re-renders and potential infinite loops if the hasModels check (line 288) doesn't work as expected. The dependency array should only include selectedProvider and isEditMode, not providerModelsData.

Suggested change

}, [selectedProvider, isEditMode, providerModelsData]);

}, [selectedProvider, isEditMode]);

Copilot · 2026-02-11T21:23:37Z

go/go.mod

 	github.com/jedib0t/go-pretty/v6 v6.7.8
 	github.com/kagent-dev/kmcp v0.2.6
 	github.com/kagent-dev/mockllm v0.0.3
+	github.com/mark3labs/mcp-go v0.33.0


The dependency github.com/mark3labs/mcp-go v0.33.0 was added to go.mod but appears to be unused in the codebase. This adds unnecessary bloat to the dependency tree. If this was added by mistake or for future use, it should be removed. If it's actually needed, there should be import statements using it.

Suggested change

github.com/mark3labs/mcp-go v0.33.0

Copilot · 2026-02-11T21:23:37Z

ui/src/app/models/new/__tests__/providerSelection.test.tsx

+ * Fix: Remove `providers` from dependency array so effect only runs on mount.
+ */
+
+import { describe, it, expect, jest } from '@jest/globals';


Unused import jest.

Suggested change

import { describe, it, expect, jest } from '@jest/globals';

import { describe, it, expect } from '@jest/globals';

Copilot · 2026-02-11T21:23:37Z

ui/src/app/models/new/__tests__/providerSelection.test.tsx

+    it('demonstrates the bug: user selection gets overwritten', () => {
+      const providers = [stockOpenAI, configuredOpenAI];
+
+      const { result, rerender } = renderHook(


Unused variable rerender.

Suggested change

const { result, rerender } = renderHook(

const { result } = renderHook(

EItanya

These changes are looking great! I would love to test it all locally but it looks like there are some build failures

EItanya · 2026-02-12T13:26:40Z

go/api/v1alpha2/provider_types.go

+
+// Provider is the Schema for the providers API.
+// It represents a model provider configuration with automatic model discovery.
+type Provider struct {


I think we should rename this to model provider so it's more clear what the purpose is, what do youthink?

Agree, provider seems too vague, I will change this to modelprovider similar to modelconfig

EItanya · 2026-02-12T21:45:38Z

go/api/v1alpha2/provider_types.go

+
+	// Key is the key within the secret that contains the API key or credential
+	// +required
+	Key string `json:"key"`


Another API nit, why is this required? If there's only one key in the secret we can use that. Just small usability things.

EItanya · 2026-02-12T21:49:17Z

go/internal/controller/reconciler/reconciler.go

+		return ctrl.Result{}, fmt.Errorf("failed to update provider status: %w", err)
+	}
+
+	// No periodic requeue - discovery only on-demand via HTTP API


llms behind a provider would not change very frequently, this is why I opted to have a force refresh

Signed-off-by: Nagarjun Krishnan <159933832+nujragan93@users.noreply.github.com> Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

Signed-off-by: Nagarjun Krishnan <159933832+nujragan93@users.noreply.github.com>

…/nujragan93/kagent into feature/provider-model-discovery Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

Nagarjun Krishnan and others added 2 commits February 9, 2026 08:06

Add model discovery feature for UI

cf5ca24

Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

Merge branch 'kagent-dev:main' into feature/provider-model-discovery

772d4a2

nujragan93 marked this pull request as ready for review February 9, 2026 15:14

nujragan93 requested review from EItanya, ilackarms, peterj and yuval-k as code owners February 9, 2026 15:14

EItanya requested changes Feb 9, 2026

View reviewed changes

Nagarjun Krishnan added 3 commits February 11, 2026 09:28

default envpoints and fix manager

2833040

Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

more fixes

68df742

Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

Merge branch 'feature/provider-model-discovery' of https://github.com…

bb38ad2

…/nujragan93/kagent into feature/provider-model-discovery Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

Copilot AI review requested due to automatic review settings February 11, 2026 21:18

Copilot started reviewing on behalf of nujragan93 February 11, 2026 21:19 View session

Merge branch 'main' into feature/provider-model-discovery

1de291b

Signed-off-by: Nagarjun Krishnan <159933832+nujragan93@users.noreply.github.com>

Copilot AI reviewed Feb 11, 2026

View reviewed changes

nujragan93 requested a review from EItanya February 11, 2026 22:22

EItanya reviewed Feb 13, 2026

View reviewed changes

nujragan93 and others added 3 commits February 13, 2026 10:41

Merge branch 'main' into feature/provider-model-discovery

d870da4

Signed-off-by: Nagarjun Krishnan <159933832+nujragan93@users.noreply.github.com> Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

Merge branch 'main' into feature/provider-model-discovery

1b302d7

Signed-off-by: Nagarjun Krishnan <159933832+nujragan93@users.noreply.github.com>

Merge branch 'feature/provider-model-discovery' of https://github.com…

3269e81

…/nujragan93/kagent into feature/provider-model-discovery Signed-off-by: Nagarjun Krishnan <nkrishnan@ancestry.com>

	- jsonPath: .status.discoveredModels[?(@)]
	- jsonPath: .status.modelCount

	}, [selectedProvider, isEditMode, providerModelsData]);
	}, [selectedProvider, isEditMode]);

	import { describe, it, expect, jest } from '@jest/globals';
	import { describe, it, expect } from '@jest/globals';

	const { result, rerender } = renderHook(
	const { result } = renderHook(

Conversation

nujragan93 commented Feb 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

nujragan93 commented Feb 9, 2026

Uh oh!

EItanya left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

EItanya left a comment

Choose a reason for hiding this comment

Uh oh!

EItanya Feb 12, 2026

Choose a reason for hiding this comment

Uh oh!

nujragan93 Feb 13, 2026

Choose a reason for hiding this comment

Uh oh!

EItanya Feb 12, 2026

Choose a reason for hiding this comment

Uh oh!

EItanya Feb 12, 2026

Choose a reason for hiding this comment

Uh oh!

nujragan93 Feb 13, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

nujragan93 commented Feb 9, 2026 •

edited

Loading