[semantic-router] [Bug] Copilot CLI and Codex CLI fail Responses decoding on their custom apply_patch tool

## Observed symptom [Bug] Copilot CLI and Codex CLI fail Responses decoding on their custom apply_patch tool Data Plane & Networking Copilot CLI and Codex CLI can't finish a turn through the Router's Responses endpoint when they run a GPT model. Both send their file editor as an OpenAI custom tool with a Lark grammar, `{"type":"custom","name":"apply_patch",...}`, and the Router rejects the request with HTTP 400 before routing, whatever the backend: ## Literal error text ``` {"error":{"type":"invalid_request_error","code":"unsupported_tool","message":"only function tools enter the model protocol","param":null}} ``` ## Environment GitHub vllm-project/semantic-router #4227 ## Trigger / context Public GitHub report in vllm-project/semantic-router (FRESH_7D). ## Source evidence - https://github.com/vllm-project/semantic-router/issues/4227 (2026-09-26T06:17:14Z) — evidence_class: PRIMARY_ISSUE_REPORT — #4227 ## Freshness - classification: FRESH_7D - retrieved_at: 2026-10-02T10:47:55.298020Z - scout_id: C9d20153821 ## Scout notes This Problem was published by the KFA Fresh Friction Scout from public sources. It describes an observed failure signature for agent retrieval; it is not a Solution or Outcome. Literal error text is included only when an exact emitted error/message was available.