feat(models): add Mistral AI as a native LLM provider

[xavierpestel-ai]

Summary

Adds Mistral AI as a first-class LLM provider across the Go controller, Go ADK runtime, Python ADK runtime, and UI. Mistral speaks the OpenAI-compatible wire protocol (POST /chat/completions with a Bearer token at https://api.mistral.ai/v1), so both runtimes reuse the OpenAI SDK client internally — no new HTTP transport, no new SDK dependency.

Changes

CRD (v1alpha2)

• New Mistral provider enum value + MistralConfig struct (baseUrl, temperature, topP, maxTokens, timeout)
• CEL rule enforcing mistral is nil for non-Mistral providers
• Default endpoint https://api.mistral.ai/v1
• Regenerated zz_generated.deepcopy.go and CRD manifests in both go/api/config/crd/bases/ and helm/kagent-crds/templates/

Controller translator

• New Mistral case in translateModel() — injects MISTRAL_API_KEY from Secret into the agent pod

Go ADK runtime

• New adk.Mistral type + models.MistralModel (wraps OpenAIModel internally)
• Wired into CreateLLM() in go/adk/pkg/agent/agent.go

Python ADK runtime

• New KAgentMistralLlm class subclassing BaseOpenAI (hardcodes Mistral base URL + MISTRAL_API_KEY env fallback)
• Registered in the ModelUnion Pydantic discriminator

Env vars, telemetry, catalog

• MISTRAL_API_KEY, MISTRAL_API_BASE
• OpenTelemetry GenAI semconv: gen_ai.provider.name=mistral_ai
• 10 Mistral models in the HTTP model catalog: mistral-large-latest, mistral-medium-latest, mistral-small-latest, magistral-medium-latest, magistral-small-latest, codestral-latest, ministral-8b-latest, ministral-3b-latest, pixtral-large-latest, open-mistral-nemo (all with function calling)

UI

• Mistral added to BackendModelProviderType, modelProviders, PROVIDERS_INFO
• New Mistral SVG icon (ui/src/components/icons/Mistral.tsx)
• Icon wired into both ModelProviderCombobox and ProviderCombobox icon maps

Helm

• providers.mistral entry in values.yaml. Default provider unchanged.

Docs

• New Mistral section in .claude/skills/kagent/references/providers.md (Helm keys, env vars, CLI examples, ModelConfig YAML)

Example ModelConfig

apiVersion: kagent.dev/v1alpha2
kind: ModelConfig
metadata:
  name: mistral-large
  namespace: kagent
spec:
  provider: Mistral
  model: mistral-large-latest
  apiKeySecret: kagent-mistral
  apiKeySecretKey: MISTRAL_API_KEY
  mistral:
    temperature: "0.3"
    maxTokens: 4096
    # baseUrl: https://api.mistral.ai/v1   # optional

Set MISTRAL_API_BASE to point at a self-hosted or regional Mistral endpoint.

Testing

All test suites green (evidence):

Suite	Result
go/api unit tests	121 passed — includes 4 new CEL admission cases for Mistral
go/adk unit tests	539 passed — includes new TestNewMistralModelWithLogger (7 cases) + TestMistralModel_NamePropagatesModelNotProvider
go/core unit tests (non-e2e)	1504 passed — includes new TestGoldenAdkTranslator/mistral_agent golden test
TestCreateLLMConfig_Mistral (agent_test.go)	✅ passes
Python test_mistral.py	10 passed
Python test_mistral.py + test_openai.py + test_anthropic.py	47 passed
E2E TestE2EInvokeWithMistralAgent	Compiles + skips cleanly when MISTRAL_API_KEY unset (verified). Modeled on foundry_test.go, uses a mock OpenAI-compatible server — never contacts the real Mistral API
make -C go generate + make -C go manifests + make controller-manifests	✅ regenerated CRDs
UI npx tsc --noEmit	✅ no new TypeScript errors (fixed 2 real "Property 'Mistral' is missing" errors caught in the combobox icon maps by wiring the new icon)
go vet ./...	✅ clean across the workspace
Local secret scan (gitleaks)	✅ clean, both staged + branch diff

Pre-existing issues observed (not fixed here)

• golangci-lint fails to load bin/kube-api-linter.so due to a Go runtime version drift in the vendored plugin. Unrelated to this PR (go vet is clean).
• 9 Python test_tls_e2e.py failures exist on main (verified via git stash + rerun). Unrelated to Mistral.

Not in this PR

• Embeddings: The Python EmbeddingConfig type does not discriminate on provider the same way Model does. Mistral embeddings (mistral-embed) is a straightforward follow-up but is not scoped here.
• Mistral-specific streaming e2e: Streaming works because the inner OpenAI client handles it; a real streaming e2e against Mistral can be added when a stable mock harness lands.

Type: feat (Conventional Commits).

[EItanya]

First off, thanks so much for the PR :)

If the wire protocol is the same as OpenAI I'm wondering if we really need a whole new API block for this? Typically we allow usage of chat completions compatible APIs via BaseURL.

I definitely see value in adding a mistral block to the UI and website, so it's clearer there that Mistral is supported, but I'm not sure it's worth adding the block to the API.

What do you think?

[xavierpestel-ai]

First off, thanks so much for the PR :)

If the wire protocol is the same as OpenAI I'm wondering if we really need a whole new API block for this? Typically we allow usage of chat completions compatible APIs via BaseURL.

I definitely see value in adding a mistral block to the UI and website, so it's clearer there that Mistral is supported, but I'm not sure it's worth adding the block to the API.

What do you think?

Hi. Yes it's just a proposition. Yes Mistral is openai compatible so it does the job like that (I already did a video to show Mistral Inference with kagent).

We can close the PR no worries.

[EItanya]

First off, thanks so much for the PR :)
If the wire protocol is the same as OpenAI I'm wondering if we really need a whole new API block for this? Typically we allow usage of chat completions compatible APIs via BaseURL.
I definitely see value in adding a mistral block to the UI and website, so it's clearer there that Mistral is supported, but I'm not sure it's worth adding the block to the API.
What do you think?

Hi. Yes it's just a proposition. Yes Mistral is openai compatible so it does the job like that (I already did a video to show Mistral Inference with kagent).

We can close the PR no worries.

So actually, after some more thought we realized that it's valuable to have more options in the API, so we'd love to have your contribution.

[xavierpestel-ai]

Rebased onto current main (648959e). Force-pushed as afdff3e.

What moved

Two upstream refactors landed between the original PR and today, so the rebase is not a small merge — most of the touched files on main are now in different locations.

Original PR target	New landing on main
v1alpha2/modelconfig_types.go + CEL tests + modelproviderconfig_types.go	v1alpha3/… (v1alpha3 is now the storage version; per AGENTS.md, do not add work to legacy API versions)
go/core/internal/httpserver/handlers/models.go (catalog)	go/core/internal/service/model/discovery.go
go/core/internal/controller/translator/agent/adk_api_translator.go	go/core/v2/translator/kagent/model.go (the legacy translator was removed in #2565)
ui/src/lib/providers.ts + ui/src/components/icons/Mistral.tsx + comboboxes	ui/src/components/model-form/providerInfo.ts + ProviderIcon.tsx (whole UI was rewritten on Vite in #2569; the icon is now an inline SVG that matches the sibling providers' pattern in ProviderIcon.tsx rather than a separate file)
.claude/skills/kagent/references/providers.md	helm/README.md (the references file was deleted upstream; there is no equivalent provider-docs page on main)

What was dropped, with reason

• go/core/internal/a2a/trace.go resolveProviderName addition — the whole helper was deleted in chore: remove legacy ACP and controller runtime #2565. The tracing middleware now takes provider as a constructor arg and is not wired to anything on main yet, so there is no place to inject the Mistral attribute.
• Legacy translator testdata (inputs/mistral_agent.yaml + outputs/mistral_agent.json) — the v2 translator has no golden-test harness today.
• go/core/test/e2e/mistral_test.go — the e2e suite now uses v2 AgentInstance with no v1alpha2 helpers (setupMistralModelConfig no longer exists). A native v2 e2e is follow-up work.

Also added on top of the original PR

• discovery_test.go + grpcserver/model_test.go: hard-coded provider counts and provider-name lists bumped from 10 → 11 to include Mistral.
• v2/translator/kagent/model.go: new case v1alpha3.ModelProviderMistral branch modeled on Anthropic (simplest OpenAI-compatible provider), injecting MISTRAL_API_KEY from Secret and passing BaseURL/Temperature/TopP/MaxTokens/Timeout through.
• helm/README.md: install example for --set providers.default=mistral.
• helm/kagent/values.yaml: providers.mistral block.
• Both .gitleaks.toml and all generated CRDs regenerated via make -C go generate + make -C go manifests + make controller-manifests.

Verification

• go vet ./... clean.
• go test ./... 1617 passed, 1 failed (TestFetchSourceReusesExistingMaterialization — verified pre-existing on clean main, macOS /private/var/… path issue unrelated to this PR), 16 skipped.
• pytest .../test_mistral.py 10 passed.
• Broader Python model tests: 183 passed, 9 failed (all in test_tls_e2e.py, pre-existing on main — matches the original PR's disclosure).
• npx tsc --noEmit in ui/ clean.

Not in this PR

Same as before — embeddings and a native streaming e2e are still separate follow-ups.

[EItanya]

Can we backout this file? I'm not against adding support for this, but I've never heard of this specific project and I'd rather discuss that separately so we can get this in

[github-actions]

This pull request has been marked as stale because of no activity in the last 15 days. It will be closed in the next 5 days unless it is tagged "no stalebot" or other activity occurs.

[github-actions]

This pull request has been closed after 15 days of inactivity and an additional 5 days after being marked as stale.