diff --git a/CLAUDE.md b/CLAUDE.md
index 7bd8b1afeef..f5b1acf5fb9 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -125,6 +125,7 @@ LiteLLM is a unified interface for 100+ LLM providers with two main components:
 - Use `RecordNotFoundError` (not bare `except Exception`) when catching "already deleted" in credential delete endpoints.
 
 ### ChatGPT / Copilot OAuth Credentials
+- **`STORE_MODEL_IN_DB=True` is required** for the proxy UI OAuth flow to survive a restart. The `proxy_config.get_credentials(prisma_client)` call that reloads `litellm.credential_list` from `LiteLLM_CredentialsTable` at startup lives inside the same `if store_model_in_db is True:` block as the model-deployment scheduler in `proxy_server.py`. Without the env var: the DB write succeeds, but nothing reads it back — credentials "disappear" on restart, and any model configured with `api_key: oauth:<name>` fails at request time because the name isn't in the in-memory cache. Also needed for the Add Model UI (`STORE_MODEL_IN_DB=True` → `LiteLLM_ProxyModelTable` → `add_deployment` job). The OAuth flow silently no-ops without it; document this prominently when rolling out.
 - Two providers use OAuth device-code flows with tokens stored in `LiteLLM_CredentialsTable`: `chatgpt` (ChatGPT / Codex backend) and `github_copilot`. Each has an `Authenticator` (filesystem-backed, used by CLI / direct SDK use) and a `DBAuthenticator` subclass (DB-backed, used by the proxy).
 - **Request-time dispatch**: the convention is `api_key: "oauth:<credential_name>"` in the model's `litellm_params`. When the transformation sees this prefix it resolves via `resolve_authenticator(...)` in `llms/chatgpt/db_authenticator.py` (or `llms/github_copilot/db_authenticator.py`), which swaps in a `DBAuthenticator` bound to that credential. Any other `api_key` falls through to the filesystem authenticator. Do not invent alternative markers — reuse the `oauth:` prefix so both providers stay symmetric.
 - **Copilot has two use-sites**: `_get_openai_compatible_provider_info` may rewrite `api_key` to the resolved Copilot key before `validate_environment` runs. The Copilot `resolve_authenticator` therefore takes both `api_key` and `litellm_params` and checks both; the ChatGPT version only needs `litellm_params`.