Bug

When a custom provider is created through the UI/CLI, the model name is saved as-is (preserving the original case, e.g. GLM-5.1). However, when goose resolves the model's context_limit at runtime, it performs a strict case-sensitive comparison against GOOSE_MODEL from config.yaml, which is stored in lowercase. This causes the lookup to fail, silently falling back to the default 128_000 context limit.

Root cause

Two locations that disagree on normalization:

1. create_custom_provider / update_custom_provider (declarative_providers.rs) — saves the model name without normalization:

   // Line ~217 (create) and ~266 (update)
   .map(|name| ModelInfo::new(name, 128000))

2. normalize_model_config (provider_registry.rs) — performs a strict equality check:

   if let Some(info) = self.metadata.known_models
       .iter()
       .find(|m| m.name == model.model_name && m.context_limit > 0)
   {
       model.context_limit = Some(info.context_limit);
   }

   model.model_name comes from ModelConfig::new() → GOOSE_MODEL in config.yaml, which is lowercased. But m.name in known_models retains the original case from the JSON config.

Impact

Custom provider models with uppercase letters in their name (e.g. GLM-5.1, Qwen3-Coder-480B) will always use the default 128k context limit, ignoring the context_limit value set in their JSON config file.

Suggested fix

Add case-insensitive comparison in normalize_model_config:

.find(|m| m.name.to_lowercase() == model.model_name.to_lowercase() && m.context_limit > 0)

Or alternatively, normalize the model name to lowercase in create_custom_provider / update_custom_provider when writing the JSON config.

Workaround

Manually edit the custom provider JSON file (~/.config/goose/custom_providers/<provider>.json) to use the exact same case as GOOSE_MODEL in config.yaml (typically lowercase).

Environment

• goose version: 1.31.1
• OS: Debian 12