Prompt caching not working: app does not send anthropic-beta header on /anthropic/v1/messages requests

まず、/anthropic/v1/messages エンドポイントに対する Copilot アプリのリクエスト構築と、supports_prompt_caching の読み取り方法を追跡します。対象となるカスタムプロバイダーのモデルへのリクエストに anthropic-beta: prompt-caching-2024-07-31 が含まれ、それ以外のリクエストは変更されないことを確認します。既存のリクエストまたはプロキシのログ設定を使用して、送信されるヘッダーを検証します。

Summary

The Copilot app correctly uses the /anthropic/v1/messages endpoint when a custom model provider is configured with an Anthropic-compatible base URL. However, it does not send the anthropic-beta: prompt-caching-2024-07-31 header, so prompt caching is never activated — even when cache_control: {type: ephemeral} markers are present in the request.

Evidence (from local hai-desktop proxy logs)

The local hai-desktop proxy forwards requests and adds ?beta=true to the upstream URL only when it sees the anthropic-beta header from the client. Comparing the two callers:

Claude Code sends anthropic-beta: prompt-caching-2024-07-31; the Copilot app does not.

Confirmed by live test (same proxy, same model)

With anthropic-beta header (Claude Code / manual curl):

• Request 1: cache_creation_input_tokens: 1202, cost 0.005942
• Request 2: cache_read_input_tokens: 1202, cost 0.000587 — 10× cheaper

Without anthropic-beta header (Copilot app):

• Request 1: cache_creation_input_tokens: 0
• Request 2: cache_read_input_tokens: 0 — no savings

Workaround

I added a local HTTP caching proxy on a separate port that sits in front of hai-desktop to avoid redundant calls. This is a symptom fix, not a solution.

Request

When the Copilot app is configured with a custom provider using an Anthropic-format endpoint, and a model reports supports_prompt_caching: true, the app should send:

anthropic-beta: prompt-caching-2024-07-31

on requests to /anthropic/v1/messages. This is the only missing piece — the endpoint, the model, and the cache_control markers in the request body are all already correct.