Overview

Period: 2026-04-02T15:43Z to 2026-04-02T16:21Z (push-triggered runs on fix/openai-cache-token-tracking)
Runs analyzed: 4 Copilot-engine runs (3 had token data)
Total tokens: 786,861 across all instrumented workflows (449,616 billable)
Estimated total cost: $1.48

✅ Major improvement vs. previous report: Cache hit rates jumped from 0% → 38–45% across all instrumented workflows. The fix/openai-cache-token-tracking branch appears to have fixed cache token reporting — caching is now correctly tracked and billing reflects actual cache discounts. Per-request billable tokens dropped ~1–2% compared to the previous report's first runs (where no cache was active yet).

Workflow Summary

🔍 Optimization Opportunities

1. All workflows — first request in each run has 0 cache reads

   • Requests 1 in all three workflows show 0 cache_read_tokens — the cache is cold on each fresh run
   • Subsequent requests within the same run benefit from the warm cache (38–45% hit rate)
   • Recommendation: If runs share a consistent system prompt, consider pre-warming the cache or using persistent cache keys across runs. Enabling cache_write_tokens tracking at the Copilot endpoint would also improve cost attribution.

2. No cache_write_tokens reported in any workflow

   • All workflows show cache_write_tokens = 0, meaning we can only see cache reads but not writes
   • The Copilot provider may not return write counts separately from input tokens
   • Recommendation: Investigate whether the api-proxy can track cache writes via response headers or usage metadata for the Copilot inference endpoint. This would make cost attribution more accurate.

3. High I/O ratios persist (233–305:1)

   • All workflows show very high input-to-output ratios, driven by large system prompts / tool schemas (31–43K input tokens/request vs 29–494 output tokens)
   • This is partially mitigated by caching, but the underlying context size is large
   • Recommendation: Review MCP tool surface in Smoke workflows — reducing tool schemas would lower input tokens for all requests including the cold-cache first request

Historical Trend

Comparison to previous report #1591 (2026-04-02, 01:06 UTC window):

The cache hit rate improvement from 0% to ~40–45% is the direct result of the fix/openai-cache-token-tracking fix correctly parsing prompt_tokens_details.cached_tokens from OpenAI/Copilot API responses. Per-request billable tokens are essentially unchanged (within 1–2%), confirming the cache was not newly enabled but was already active and now being correctly measured.

Previous Report

#1591 — 📊 Copilot Token Usage Report 2026-04-02

References:

• §23908898227 — Smoke Copilot
• §23908898271 — Build Test Suite
• §23908898275 — Smoke Chroot

Generated by Daily Copilot Token Usage Analyzer · ◷