Skip to main content

GitHub Copilot best practices

Last updated 2 min read

GitHub Copilot emits two different response payload shapes, and which one you get determines how much token detail Coolhand can record. The newer shape carries `copilotUsage.tokenDetails`; the older one omits several fields, so some cost figures are estimates rather than reported values.

Response payload shape determines data quality

Coolhand extracts fields from two different Copilot response payload shapes:

Field Older payload Newer payload (copilotUsage.tokenDetails)
Model name From response_body.model From response_body.model
Input token breakdown Top-level inputTokens (net of cache) tokenDetails entry with tokenType: "input"
Cached Input Tokens Top-level cacheReadTokens tokenDetails entry with tokenType: "cache_read"
Cache Creation Tokens Top-level cacheWriteTokens tokenDetails entry with tokenType: "cache_write"
Reasoning Tokens response_body.reasoningTokens response_body.reasoningTokens
Thinking Response response_body.reasoningText response_body.reasoningText

Coolhand prefers the copilotUsage.tokenDetails array when present and falls back to top-level fields automatically. The branching is based on what the Copilot server returns in its response payload, not on a client SDK version.

Note: Token fields for cache reads/writes are inconsistently populated across Copilot environments — if they are blank, this is a known upstream behavior, not a Coolhand issue.

Temperature is always blank

Copilot manages generation parameters server-side. You cannot set or observe temperature — it will always be blank in Coolhand.

Structured outputs

Depends on the underlying model Copilot routes to. Not configurable directly — Copilot controls routing and model parameters.

Caching

Managed entirely by Copilot server-side — there is no caller API for enabling or configuring caching. Coolhand captures what Copilot reports:

  • Cached Input Tokens — from cache_read in tokenDetails, or cacheReadTokens as fallback
  • Cache Creation Tokens — from cache_write in tokenDetails, or cacheWriteTokens as fallback

TTL and invalidation are opaque — Copilot does not expose cache lifecycle details to callers.

Batch processing

No batch API. GitHub Copilot is designed for interactive, real-time use. There is no mechanism for submitting offline batch jobs.

Related articles