[BUG] Agent pod CrashLoops with pydantic ValidationError when RemoteMCPServer ref omits toolNames (http_tools[*].tools=null) #1797

📋 Prerequisites

• I have searched the existing issues to avoid creating a duplicate
• By submitting this issue, you agree to follow our Code of Conduct
• I am using the latest version of the software
• I can consistently reproduce this issue

🎯 Affected Service(s)

Multiple services / System-wide issue

🚦 Impact/Severity

Blocker

🐛 Bug Description

When an Agent (kagent.dev/v1alpha2) lists a tool of type: McpServer
referencing a RemoteMCPServer without an explicit toolNames: filter,
the controller materializes the agent's config.json Secret with
http_tools[*].tools = null (and sse_tools[*].tools = null for the SSE
protocol). The Python ADK runtime image
(cr.kagent.dev/kagent-dev/kagent/app:0.9.1) then crashes on startup with a
pydantic ValidationError, leaving the agent pod in CrashLoopBackOff.

Whether toolNames is intended to be required or optional, the current
behavior is a bug because omitting it silently produces a config that the
runtime cannot load, with no controller-side validation or admission error.

The mismatch is between two sides of the same wire format:

• Controller (Go) — HttpMcpServerConfig.Tools is []string with the
  JSON tag json:"tools" (no omitempty). When mcpServerTool.ToolNames
  is nil in
  go/core/internal/controller/translator/agent/adk_api_translator.go,
  encoding/json marshals the nil slice as JSON null.
• Runtime (Python) —
  HttpMcpServerConfig.tools
  is typed list[str] = Field(default_factory=list), which rejects None
  with a list_type validator error.

The CRD schema for McpServerTool.toolNames does not appear to mark the
field as required, so users (and tooling) have no signal that omitting it
is unsafe.

🔄 Steps To Reproduce

1. Stand up kagent v0.9.1 against any working RemoteMCPServer
   (e.g. cr.kagent.dev/kagent-dev/kagent/tools).

2. Apply the following Agent (no toolNames on the tool ref):

   apiVersion: kagent.dev/v1alpha2
   kind: Agent
   metadata:
     name: repro
     namespace: kagent
   spec:
     type: Declarative
     description: repro
     declarative:
       modelConfig: <existing>
       systemMessage: "test"
       tools:
         - type: McpServer
           mcpServer:
             apiGroup: kagent.dev
             kind: RemoteMCPServer
             name: <existing-rms>
             # NOTE: no toolNames

3. Observe the agent pod enter CrashLoopBackOff.

4. kubectl logs shows the pydantic validation error below.

5. Inspect the rendered config:
   kubectl get secret repro-config -o jsonpath='{.data.config\.json}' | base64 -d | jq '.http_tools[0].tools'
   prints null.

Adding any non-empty toolNames: [<one>] to the tool ref makes the pod
start successfully — that's a viable workaround.

🤔 Expected Behavior

The agent should not CrashLoop on a config the controller itself produced.
Possible resolutions (deferring to maintainers on the intended semantics):

1. If toolNames is meant to be required, validate it in the CRD
   schema or in a controller-side admission check so users get a clear
   error at apply-time rather than a pydantic crash at pod-start.
2. If toolNames is meant to be optional (with omission meaning "all
   discovered tools"), the controller should emit an explicit list (or
   the runtime should tolerate None) rather than handing the runtime a
   null it can't parse.

📱 Actual Behavior

Agent pod CrashLoops on startup with:

ValidationError: 1 validation error for AgentConfig
http_tools.0.tools
  Input should be a valid list [type=list_type, input_value=None,
  input_type=NoneType]

The rejected blob is:

{
  "http_tools": [
    {
      "params": {"url": "...", "headers": {}, "terminate_on_close": true},
      "tools": null
    }
  ]
}

💻 Environment

• Kubernetes: Talos Linux, k8s 1.34
• kagent: v0.9.1 (cr.kagent.dev/kagent-dev/kagent/app:0.9.1 runtime,
  controller image cr.kagent.dev/kagent-dev/kagent/controller:0.9.1)
• Helm chart: kagent 0.9.1
• Topology: agents reference a single RemoteMCPServer named
  agentgateway-mcp that fronts several upstream MCP servers via
  agentgateway. The RMS reports ~50
  discoveredTools and is Healthy.

🔍 Additional Context

• Same field exists on sse_tools[*].tools and the same nil-slice → JSON
  null defect applies there.
• Workaround: pin explicit toolNames per agent.
• I'll open a PR shortly that addresses both sides defensively (controller
  expands discovered tools when toolNames is empty; runtime coerces
  None → [] so older controllers don't crash newer runtimes). Happy
  to rework toward whichever direction maintainers prefer if the intended
  semantics are different.

📋 Logs

$ kubectl logs -n kagent repro-<hash>
Traceback (most recent call last):
  ...
  File ".../kagent/adk/types.py", line ..., in to_agent
    config = AgentConfig.model_validate(...)
pydantic_core._pydantic_core.ValidationError: 1 validation error for AgentConfig
http_tools.0.tools
  Input should be a valid list [type=list_type, input_value=None, input_type=NoneType]
    For further information visit https://errors.pydantic.dev/2.x/v/list_type

🙋 Are you willing to contribute?

• I am willing to submit a PR to fix this issue