Report 1h vs 5m Anthropic cache-creation token split in telemetry

[bhavyaus]

Adds per-request 1h vs 5m cache-creation token split to telemetry, enabling exact COGS attribution for the chat.anthropic.promptCaching.extendedTtl A/B experiment without inferring rates from arm assignment.

The new fields live on a nested `anthropic_cache_creation?` object on `APIUsage.prompt_tokens_details`, namespaced to make the provider-specificity explicit at the type level. Other providers leave it undefined; telemetry uses optional chaining so missing values drop cleanly from the row.

prompt_tokens_details?: {
    cached_tokens: number;
    cache_creation_input_tokens?: number;
    anthropic_cache_creation?: {
        ephemeral_1h_input_tokens?: number;   // 2x rate
        ephemeral_5m_input_tokens?: number;   // 1.25x rate
    };
};

[Copilot]

Pull request overview

This PR extends Copilot Chat’s Anthropic Messages API usage accounting to capture (and emit in existing response.success telemetry) a per-request split of prompt cache-creation (write) tokens by TTL (1h vs 5m), enabling more accurate cost attribution for prompt caching experiments.

Changes:

• Add anthropic_cache_creation to APIUsage.prompt_tokens_details to represent Anthropic’s usage.cache_creation TTL breakdown.
• Parse and preserve the 1h/5m cache-creation token breakdown in both non-streaming and streaming Anthropic Messages API response handling.
• Emit two new response.success telemetry measurements and add unit tests covering the new parsing/streaming semantics.

Show a summary per file

File	Description
extensions/copilot/src/platform/networking/common/openai.ts	Extends APIUsage typing to include Anthropic-specific cache-creation TTL breakdown.
extensions/copilot/src/platform/endpoint/node/messagesApi.ts	Parses/accumulates 1h vs 5m cache-creation token counts for streaming + non-streaming responses and surfaces them in prompt_tokens_details.
extensions/copilot/src/extension/prompt/node/chatMLFetcherTelemetry.ts	Adds two new response.success telemetry measurements sourced from anthropic_cache_creation.
extensions/copilot/src/platform/endpoint/test/node/messagesApi.spec.ts	Adds unit tests for non-streaming + streaming preservation/override behavior of the TTL breakdown.

Copilot's findings

• Files reviewed: 4/4 changed files
• Comments generated: 1