The send_raw_llm_request still accepts a custom_url parameter, but the only call site in zeta2.rs:120-123 always passes None. No env var or setting populates it anymore.

  Here's what you can still customize via env vars:

  ┌───────────────────────────┬───────────────────────────────────────────────────────────────────────────────────────────────────┐
  │          Env var          │                                           What it does                                            │
  ├───────────────────────────┼───────────────────────────────────────────────────────────────────────────────────────────────────┤
  │ ZED_ZETA_FORMAT           │ Activates the raw endpoint path and selects a prompt format (e.g. v0131_git_merge_markers_prefix) │
  ├───────────────────────────┼───────────────────────────────────────────────────────────────────────────────────────────────────┤
  │ ZED_ZETA_MODEL            │ Overrides the model ID sent to the raw endpoint                                                   │
  ├───────────────────────────┼───────────────────────────────────────────────────────────────────────────────────────────────────┤
  │ ZED_ACCEPT_PREDICTION_URL │ Custom URL for the acceptance tracking callback only                                              │
  └───────────────────────────┴───────────────────────────────────────────────────────────────────────────────────────────────────┘

  But all prediction requests go to Zed's own LLM server (/predict_edits/raw or /predict_edits/v3) — there's no way to redirect them to your own endpoint anymore.

  If you need this capability back, you'd need to re-add a ZED_PREDICT_EDITS_URL env var check. The plumbing is still there (send_raw_llm_request accepts custom_url: Option<Arc<Url>>), so the change would be small — just populate it from the env in zeta2::request_prediction_with_zeta2 instead of always passing
  None.