gollm/openai.go (4 lines):
	- line 157: // TODO: Implement listing OpenAI models using c.client
	- line 228: // TODO: Handle other content types if necessary?
	- line 248: // TODO: Check if error is retryable using cs.IsRetryableError
	- line 306: // TODO: Implement actual retry logic if needed


gollm/gemini.go (2 lines):
	- line 275: // TODO: remove this hack once gemma-3-27b-it supports system prompt
	- line 445: // TODO: Should we try to coalesce parts when we have a streaming response?


gollm/factory.go (2 lines):
	- line 112: // TODO (droot): Adjust the signature to allow underlying client to relay the backoff
	- line 272: // TODO: Retry logic


gollm/azopenai.go (2 lines):
	- line 175: // TODO: Implement this
	- line 180: // TODO: Implement streaming


kubectl-utils/pkg/kube/discovery.go (1 line):
	- line 29: // TODO: share cache with kubectl?


gollm/persist.go (1 line):
	- line 26: // TODO: Structured data?


k8s-bench/eval.go (1 line):
	- line 336: // TODO: Wait for idle between sending steps?


k8s-bench/main.go (1 line):
	- line 44: // TODO: support namespaces also


k8s-bench/pkg/model/results.go (1 line):
	- line 48: // TODO: Maybe different styles of invocation, or different temperatures etc?


gollm/ollama.go (1 line):
	- line 197: // TODO: Implement streaming


pkg/agent/conversation.go (1 line):
	- line 478: buffer = "" // TODO: any trailing text?


kubectl-utils/pkg/kel/expression.go (1 line):
	- line 29: // TODO: Can we / should we do better than AnyType?