[nasser1941]

Why

The user wanted to pick a model straight from the /models list, and not have to type /model <name> afterwards. They also asked for 512K and 1M context options for qwen3.6-35b.

What changes

/models and /model (with no name)

• Both show the models table with a # column and ask "Switch to which model? (number or name, Enter to keep …)".
• The list holds the catalog first, then the other installed Ollama models that can chat and call tools.
• /model 2 or /models 2 switches straight away. Names still work.
• A catalog model that isn't installed says how to get it (lcode setup <key>).
• Choosing the current model says "Already using …".
• For OpenAI-compatible servers, the served models are numbered too.

Embedding models are left out. Ollama lists tools for qwen3-embedding:0.6b (it's in its template) but not completion, so lcode requires both.

/context says why the list stops where it does: "Not offered: 512K and 1M. lcode-qwen3.6-35b was trained on up to 256K tokens, and Ollama caps the window there whatever the memory…".

• I checked the user's Ollama 0.35.1. It contains requested context size too large for model and caps num_ctx at n_ctx_train, and the GGUF has no RoPE-scaling (YaRN) keys.
• So 512K would be capped back to 256K behind lcode's back. Its estimated 33 GB would fit that machine's 35 GB budget, but the model can't use it.
• 512K and 1M stay available for models trained on them, such as nemotron-3.5-lightning.

lcode models in a shell keeps its table without numbers.

Checks

• Unit tests: tests/test_model_picker.py (7 tests), plus a numbered served list in tests/test_backends.py. The full suite and the strict docs build pass.
• Real terminal (pty) against the user's real Ollama model list, through a stand-in server that answers loads without loading anything (their model stayed loaded):
  • /models then 6 switched to qwen3.5:9b;
  • /model then Enter kept it;
  • /model 1 switched back at 256K;
  • /context showed the note.
• Bug found in that run and fixed: the embedding model was listed.

🤖 Generated with Claude Code