📊 Copilot Token Usage Report2026-04-03

[github-actions]

Overview

Period: 2026-04-03T06:34Z to 2026-04-03T10:23Z
Runs analyzed: 10 Copilot-engine runs (6 had token data)
Total tokens: 3,853K across all instrumented workflows
Estimated total cost: $5.24

🆕 New workflow with token data: Documentation Maintainer appears for the first time with 32 requests and $3.91 per run. This is by far the most expensive workflow and warrants attention.

⚠️ Failure anomaly: One Secret Digger (Copilot) run failed after entering a retry loop (7 requests, requests 2–7 all returned 2 output tokens at identical input sizes), costing 3.4× a normal run.

Workflow Summary

Workflow	Runs	Requests	Total Tokens	Est. Cost	Cache Rate	I/O Ratio	Top Model
Documentation Maintainer	1	32	3,003K	$3.91	48.7%	408:1	sonnet-4.6
Secret Digger (Copilot)	5	15	850K	$1.33	38.1%	433:1	sonnet-4.6
Total	6	47	3,853K	$5.24	46.0%	418:1

🔍 Optimization Opportunities

1. Documentation Maintainer — 32 requests, $3.91/run (new)

   • First observed run, 3M total tokens, 32 LLM calls in ~3.5 minutes
   • Input grows steadily from 33K → 55K per request (accumulating context), while output is tiny (avg 229 tokens)
   • 408:1 I/O ratio — almost entirely system prompt + tool schemas
   • 48.7% cache hit rate helps, but first request is always cold (0 cache reads)
   • Recommendations:
     • Audit whether 32 requests is expected for a doc maintenance task; if iterating over many files, consider batching
     • Enable cache breakpoints on the static system prompt prefix to warm the cache on request 1 (currently only warms from request 2+)
     • Review MCP tool surface — large tool schemas inflate input on every call

2. Secret Digger (Copilot) failure run — retry loop detected

   • Run §23942926212 (failure): 7 requests vs. normal 2
   • Requests 2–7 all had output_tokens = 2 and identical input_tokens = 35,039 — a clear retry/stuck loop
   • Cost: $0.62 vs. $0.18 for a normal run (3.4×) — wasted cost from repeated identical calls
   • Recommendation: Investigate what causes the Secret Digger workflow to loop on identical input without making progress; consider adding a max-iterations guard

3. Cache write tokens still 0 across all workflows (provider limitation)

   • Consistent with previous reports — the Copilot inference endpoint doesn't report cache_write_tokens separately
   • Cost estimates may be slightly underestimated (cache writes billed at $3.75/M for Sonnet)
   • Recommendation: No action needed; tracked as a known limitation

Per-Workflow Details

Documentation Maintainer

• Run: §23936885217 (success)
• Requests: 32 (all claude-sonnet-4.6 via copilot provider)
• Duration: 06:41:52Z → 06:45:23Z (~3.5 minutes)
• Tokens: 3,003K total — 1,535K input, 7.3K output, 1,460K cache_read, 0 cache_write
• Cache hit rate: 48.7% (request 1 cold; requests 2–32 benefit from warming)
• Avg per request: 48K input, 230 output, 46K cache_read
• Avg latency: 5,045ms/request
• Estimated cost: $3.91

Request	Timestamp	Input	Output	Cache Read	Duration
1	06:41:52	33,292	203	0	5,847ms
2	06:41:58	43,501	174	29,334	5,407ms
3	06:42:03	36,858	182	36,417	4,611ms
4	06:42:07	37,730	144	36,637	3,507ms
5	06:42:17	44,532	402	37,729	9,016ms
… (32 total)
31	06:45:19	54,039	820	53,452	14,851ms
32	06:45:23	54,671	102	53,745	3,749ms

Secret Digger (Copilot) — Normal Runs

• Runs: 4 successful (plus 1 failed), all claude-sonnet-4.6 via copilot provider
• Requests: 2 per normal run
• Tokens per run: ~100K total — ~70K input, ~430 output, ~29.5K cache_read, 0 cache_write
• Cache hit rate: ~29.6–29.7% per normal run
• Avg latency: ~6,000–9,400ms/request
• Estimated cost: ~$0.18/run

Run	Conclusion	Requests	Cost	Cache Rate
§23942926212	❌ failure	7	$0.62	45.5%
§23941423866	✅ success	2	$0.18	29.7%
§23939721870	✅ success	2	$0.18	29.6%
§23938302916	✅ success	2	$0.18	29.7%
§23936764428	✅ success	2	$0.18	29.6%

Failure run detail — retry loop: requests 2–7 all returned output_tokens=2 with identical input_tokens=35,039:

#	Input	Output	Cache Read	Duration	Status
1	34,787	239	0	8,703ms	200
2	35,039	2	29,517	3,377ms	200
3	35,039	2	35,038	4,655ms	200
4	35,039	2	35,038	3,198ms	200
5	35,039	2	35,038	2,395ms	200
6	35,039	2	35,038	3,277ms	200
7	35,039	2	35,038	2,889ms	200

Workflows Without Token Data

The following Copilot-engine workflows ran in the past 24 hours but had no agent-artifacts with token data:

Workflow	Run IDs	Reason
Agentic Maintenance	§23940582037, §23937336405	No artifacts — likely doesn't use --enable-api-proxy
Firewall Issue Dispatcher	§23937040816	Has agent/firewall-audit-logs artifacts but no agent-artifacts
Copilot Code Review	§23935549881	No artifacts
Plan Command	§23937137924	Skipped

Historical Trend

Comparison to previous report #1604 (2026-04-02, 15:43–16:21 UTC):

Workflow	Prev Cache Rate	Curr Cache Rate	Prev Cost/Run	Curr Cost/Run	Change
Secret Digger (Copilot)	— (no run)	30–45%	—	$0.18 normal / $0.62 failed	New
Documentation Maintainer	— (not tracked)	48.7%	—	$3.91	🆕 First run
Smoke Copilot	43.7%	— (no run today)	$0.70	—	—
Build Test Suite	44.8%	— (no run today)	$0.47	—	—

Total cost trend: $1.48 (previous) → $5.24 (today) — primarily due to Documentation Maintainer being a new, high-cost workflow. Smoke Copilot and Build Test Suite did not run today on main.

Secret Digger normal runs remain stable at $0.18/run, consistent with the first report (April 1).

Previous Report

#1604 — 📊 Copilot Token Usage Report 2026-04-02

References:

• §23936885217 — Documentation Maintainer
• §23942926212 — Secret Digger (Copilot) failure
• §23941423866 — Secret Digger (Copilot) success

Generated by Daily Copilot Token Usage Analyzer · ◷