Bug
_get_cached_client() returns model or cached_default on cache hits without checking whether the model slug is compatible with the cached provider. resolve_provider_client() has a guard that drops OpenRouter-format slugs (containing /) for non-OpenRouter providers, but this guard is skipped entirely on cache hits.
Reproduction
config.yaml:
model:
default: gpt-5.4
provider: openai-codex
compression:
summary_model: google/gemini-3-flash-preview
- Any auxiliary call (e.g.
flush_memories) resolves provider="auto" → caches CodexAuxiliaryClient under key ("auto", False, "", "", 0)
- Compression call with
model="google/gemini-3-flash-preview" hits the same cache key (model is not part of the key) → returns (codex_client, "google/gemini-3-flash-preview")
google/gemini-3-flash-preview is sent to chatgpt.com/backend-api/codex which doesn't understand it
Expected
The / guard in resolve_provider_client() should also apply on cache hits, dropping incompatible slugs in favor of the cached default model.
Fix
PR #5804
Bug
_get_cached_client()returnsmodel or cached_defaulton cache hits without checking whether the model slug is compatible with the cached provider.resolve_provider_client()has a guard that drops OpenRouter-format slugs (containing/) for non-OpenRouter providers, but this guard is skipped entirely on cache hits.Reproduction
config.yaml:flush_memories) resolvesprovider="auto"→ cachesCodexAuxiliaryClientunder key("auto", False, "", "", 0)model="google/gemini-3-flash-preview"hits the same cache key (model is not part of the key) → returns(codex_client, "google/gemini-3-flash-preview")google/gemini-3-flash-previewis sent tochatgpt.com/backend-api/codexwhich doesn't understand itExpected
The
/guard inresolve_provider_client()should also apply on cache hits, dropping incompatible slugs in favor of the cached default model.Fix
PR #5804