[BUG] Ollama ModelConfig `options` (e.g. `num_ctx`) are defined but not propagated to runtime

[egallis31]

📋 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 have tried to clear cache/cookies or used incognito mode (if ui-related)
• I can consistently reproduce this issue

🎯 Affected Service(s)

Controller Service

🚦 Impact/Severity

Minor inconvenience

🐛 Bug Description

The OllamaConfig.Options field in the ModelConfig CRD allows users to specify Ollama-specific parameters such as num_ctx, temperature, etc. However, these options are currently not forwarded to the runtime or Ollama API calls.

Only the host (via OLLAMA_API_BASE) and defaultHeaders are passed through the ADK translator. As a result, any values defined under ollama.options are silently ignored at runtime, despite being part of the public API schema.

This creates a mismatch between the CRD definition and actual behavior.

🔄 Steps To Reproduce

Define an Ollama ModelConfig Resource with the options spec:

# ModelConfig for granite4:3b (for general tasks, not agents)
apiVersion: kagent.dev/v1alpha2
kind: ModelConfig
metadata:
  name: granite4-3b-ollama
  namespace: kagent
spec:
  model: "granite4:3b"
  provider: Ollama
  apiKeySecret: kagent-secrets
  apiKeySecretKey: litellm-api-key
  ollama:
    host: "http://ollama:11434"
    options:
      num_ctx: "128000" # bug with passing the options to runtime agent runner

🤔 Expected Behavior

When a user defines ollama.options in a ModelConfig, those values should be passed to the Ollama provider client (LiteLLM / HTTP request) so they affect model execution. This should result in the Ollama request being executed with the specified context size and sampling parameters.

📱 Actual Behavior

Actual Behavior

• The options map is accepted by the CRD and appears in test data
• The ADK translator does not forward options
• The Python runtime (LiteLLM) never receives num_ctx or other Ollama parameters
• Only keys passed through:
  • OLLAMA_API_BASE
  • defaultHeaders

This makes ollama.options effectively a no-op.

💻 Environment

• OS: TalosOS
• Kuberenets: 1.34.1
• kagent Helm Version: 0.7.8

🔧 CLI Bug Report

No response

🔍 Additional Context

DeepWiki

Relevant code paths:

• CRD defines OllamaConfig.Options
go/api/v1alpha2/modelconfig_types.go
• Test YAML includes num_ctx
go/internal/controller/translator/agent/testdata/inputs/ollama_agent.yaml
• Translator only sets OLLAMA_API_BASE and headers
  go/internal/controller/translator/agent/adk_api_translator.go
• Python runtime initializes LiteLLM with headers only
  python/packages/kagent-adk/src/kagent/adk/types.py

📋 Logs

Ollama:

time=2026-01-19T17:20:45.615Z level=DEBUG source=sched.go:290 msg="runner with non-zero duration has gone idle, adding timer" runner.name=registry.ollama.ai/library/granite4:3b runner.inference="[{ID:GPU-aafce598-4e49-4405-5ed6-7d8c9af80b95 Library:CUDA}]" runner.size="7.5 GiB" runner.vram="7.5 GiB" runner.parallel=1 runner.pid=1199 runner.model=/root/.ollama/models/blobs/sha256-6c02683809a8dc4eb05c78d44bc63bcd707703b078998fa58829c858ab337bb0 runner.num_ctx=40000 duration=5m0s


Manually set context via Ollama env variable to reach 40000

📷 Screenshots

No response

🙋 Are you willing to contribute?

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

[EItanya]

closed by #1245