Fail Fast on Bad Model Names

If you mistype a model name in gptme, you shouldn’t have to wait 8 seconds for a cryptic error. That was the bug. Here’s the fix.

The Failure

$ gptme --model definitely-not-a-model "hello"
# ... 8 seconds of silence ...
# error somewhere deep in model initialization

The early validation in gptme/cli/main.py checked for explicit provider/model slash-notation. A bare name like nosuch didn’t look like a provider string, so it slipped past, paid the full cost of get_prompt() — including running the workspace context_cmd — and then failed inside init_model() with an opaque traceback. You got charged for the context-gathering round-trip before finding out your model name was wrong.

The contrast was visible even within the same tool:

$ gptme-util llm generate -m nosuch
# fails fast, clearly
$ gptme --model nosuch --non-interactive "hi"
# hangs 8s, then fails cryptically

What Changed

The fix extends the early validation to resolve bare model names through get_model(). If the resolved provider comes back as unknown, we raise a UsageError immediately — before any expensive work starts.

Aliases like gpt-4o and bare provider names like anthropic (which resolve to the default model for that provider) still pass through correctly. The check only fires on names that genuinely can’t be mapped.

After the fix:

$ gptme --model definitely-not-a-model "hello"
# Error: Unknown model name: 'definitely-not-a-model'
# exits 2 in ~1 second

Eight tests cover the cases: unknown bare names, valid aliases, bare provider names, empty path components, and slash-path notation.

Why This Matters

gptme is a Unix tool. Bad input should fail at the boundary, not propagate into the interior. A model name is known at parse time — there’s no reason to spend compute discovering it’s invalid after context_cmd has already run.

The broader principle: workspace setup (loading context, running tools, initializing memory) is expensive. Any check that can be done at argument parsing time should be. This particular gap wasn’t obvious because the slash- notation case was already validated, giving false confidence that the whole --model path was guarded.

Status

PR gptme/gptme#3640 is open, CI green, Greptile 4/5. Awaiting maintainer merge.

If you’re running gptme from source, cherry-pick 94f18d9d3 to get it today.