[diw-zw]

Motivation

Previously, users had no way to discover which model IDs and run modes were available before running kubectl rbg llm svc run. When a user-defined model config overlapped with a built-in one, the entire built-in model was silently replaced (model-level merge), causing loss of non-overridden modes. This PR addresses both issues by:

1. Introducing a discovery command so users can inspect supported models, engines, and resource presets before deployment.
2. Switching to mode-level merge so user modes override only same-named built-in modes while preserving others.

Modifications

New Command: llm svc model-configs

• Lists all available model configurations in a tabular format.
• Default view shows MODEL ID, MODE, DESCRIPTION.
• -o wide shows additional columns: ENGINE, SOURCE.

Documentation

• Updated cmd/cli/README.md and doc/cli/kubectl-rbg-llm-svc.md with:
  • llm svc model-configs command description, flags, and examples.
  • Cross-reference explaining that model-configs helps discover deployable models before llm svc run.

Tests

• Updated model_config_test.go assertions to reflect the new mode-level merge behavior:
  • Overridden modes carry the user file source.
  • Non-overridden built-in modes are preserved.
  • Last-file-wins semantics for duplicate user modes.
• go test ./cmd/cli/... passes.

Files Changed

• cmd/cli/cmd/llm/svc/model_configs.go (new)
• cmd/cli/cmd/llm/svc/run/model_config.go
• cmd/cli/cmd/llm/svc/run/model_config_test.go
• cmd/cli/cmd/llm/svc/svc.go
• cmd/cli/README.md
• doc/cli/kubectl-rbg-llm-svc.md

[gemini-code-assist]

Code Review

This pull request introduces the kubectl rbg llm svc model-configs command to list available model configurations and refactors the model loading logic to support mode-level merging between built-in and user-defined configurations. Feedback identifies a bug in the mode merging logic where duplicates within the same file are not correctly handled, suggests validating the --output flag, recommends sorting modes for deterministic output, and points out a shallow copy issue with the Resources map in the configuration loader.

[Syspretor]

/lgtm