Model picker: hard to tell models apart + easy to land on (and silently run) a weak one

[Svector-anu]

Context

Built a full project with ZERO v0.1.0 (macOS, x64) driving edits via OpenRouter — overall a genuinely great experience, and the code output quality was solid. The one consistently rough spot was model selection, so here's concrete feedback.

What's rough

1. Picker descriptions are truncated and near-identical. In /model, several entries show the same cut-off blurb (e.g. "Flagship Claude model for deep reasoning, coding, and long-horizon ag…"), so I couldn't tell claude-sonnet-4.5 / 4.6 / sonnet-5 apart. The only reliable way to pick a specific model was typing the exact slug into search.

2. Navigating the picker seems to auto-commit the model. While arrowing through, I ended up saved onto ai21/jamba-large-1.7 then aion-labs/aion-1.0 without meaning to — easy to land on a weak model by accident.

3. A weak model runs with no quality signal. On those models, replies leaked raw control tokens into the output (e.g. <|eom|>) and nothing flagged it — I only caught it from the garbage text.

Suggestions

• Show the model slug (and ideally price / context length) inline in the picker, not just a truncated blurb.
• Commit the model on Enter only, not on hover/navigation.
• Consider a hint/filter for models that leak control tokens.

Env

ZERO 0.1.0 · macOS 15 (x64) · provider: openrouter

[anandh8x]

Just checked this on the latest main and the picker issues seem to be fixed. I also couldn’t reproduce the model leakage.