You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Period: 2026-04-04T10:20Z to 2026-04-05T10:20Z Runs analyzed: 30 total (1 had token data) Total tokens: 1,399K Estimated total cost: $2.02 (Copilot rates: $2.50/M input, $10/M output)
🆕 New workflow with token data: Pelis Agent Factory Advisor appears for the first time with token data — 11 requests, $2.02 per run.
ℹ️ Secret Digger (Copilot) absent: Only Claude and Codex variants of Secret Digger ran today. The Copilot variant did not run in this period (not a failure — it simply wasn't scheduled or triggered).
First request always cold (0 cache reads, 39,950 input tokens) — full system prompt cost each run
Cache warms progressively from request 2 onwards (reaching ~49% by req 6–11)
Recommendation: Ensure the system prompt prefix is stable across requests so Anthropic's prompt cache can pre-warm it; consider structuring the first request to maximize the cacheable prefix
Pelis Agent Factory Advisor — high I/O ratio (110:1 > 100:1 threshold)
759K input tokens vs only 13K output tokens across 11 requests
Input context grows from 40K → 83K over 11 turns (conversation history accumulation)
Recommendation: Review whether all prior conversation history needs to be included each turn; consider summarizing or truncating older turns
No cache_write tokens (Copilot provider limitation)
cache_write_tokens is 0 across all requests — the Copilot inference endpoint does not report this separately
Cost estimates may be slightly understated (cache write overhead not billed separately here)
Cache hit rate: 45.2% (cold start on req 1, warming progressively)
I/O ratio: 110:1
Avg latency: 20,490ms/request (median ~8,000ms; req 10 took 88,691ms for a 5K-token response)
Estimated cost: $2.02
Per-request breakdown:
Req
Input
Output
Cache Read
Cache Rate
Latency
Cost
1
39,950
400
0
0%
7,045ms
$0.10
2
59,297
520
33,976
36%
8,090ms
$0.15
3
61,343
505
46,636
43%
7,688ms
$0.16
4
69,222
513
53,989
44%
8,043ms
$0.18
5
71,086
497
61,605
46%
7,772ms
$0.18
6
68,647
362
66,345
49%
6,177ms
$0.18
7
73,335
443
67,410
48%
14,526ms
$0.19
8
76,012
1,794
70,372
48%
33,149ms
$0.21
9
78,685
2,227
73,051
48%
35,288ms
$0.22
10
78,289
4,959
75,657
49%
88,691ms
$0.25
11
83,324
348
78,288
48%
8,920ms
$0.21
Observation: Input context grows monotonically (39K → 83K), consistent with conversation history accumulation. Requests 8–10 show significantly higher output tokens (1,794 / 2,227 / 4,959) and latency, suggesting heavier model work in later turns. Request 10 took 88.7 seconds for a 4,959-token response.
Workflows Without Token Data
The following Copilot-engine workflows ran but produced no agent-artifacts with token-usage.jsonl:
Workflow
Runs
Conclusion
Reason
Secret Digger (Copilot)
0
—
Did not run this period (only Claude/Codex variants ran)
Smoke Services
1
❌ failure
Has artifacts (agent, detection, firewall-audit-logs) but no agent-artifacts — likely failed before AWF agent completed
Plan Command
2
⏭️ skipped
Skipped — condition not met
Agentic Maintenance
5
✅ success
No artifacts — does not use --enable-api-proxy
Firewall Issue Dispatcher
2
✅ success
Has agent artifact but not agent-artifacts — does not use --enable-api-proxy
Secret Digger (Copilot) absent today — daily cost down from $1.16 to $2.02 net, but driven by new Pelis workflow rather than efficiency gains
Pelis Agent Factory Advisor ($2.02/run) is more expensive per run than Secret Digger success runs ($0.18/run), reflecting its longer agentic session (11 vs 2 requests, 4 minutes of work)
Documentation Maintainer ($3.91/run) has not re-appeared since 2026-04-03 — most expensive workflow seen so far
The recurring retry-loop bug in Secret Digger (Copilot) cannot be assessed today (no Copilot variant runs)
Overview
Period: 2026-04-04T10:20Z to 2026-04-05T10:20Z
Runs analyzed: 30 total (1 had token data)
Total tokens: 1,399K
Estimated total cost: $2.02 (Copilot rates: $2.50/M input, $10/M output)
Workflow Summary
🔍 Optimization Opportunities
Pelis Agent Factory Advisor — low cache hit rate (45.2% < 50%)
Pelis Agent Factory Advisor — high I/O ratio (110:1 > 100:1 threshold)
No cache_write tokens (Copilot provider limitation)
cache_write_tokensis 0 across all requests — the Copilot inference endpoint does not report this separatelyPer-Workflow Details
Pelis Agent Factory Advisor
schedule, actor:dsyme)claude-sonnet-4.6via Copilot endpointPer-request breakdown:
Observation: Input context grows monotonically (39K → 83K), consistent with conversation history accumulation. Requests 8–10 show significantly higher output tokens (1,794 / 2,227 / 4,959) and latency, suggesting heavier model work in later turns. Request 10 took 88.7 seconds for a 4,959-token response.
Workflows Without Token Data
The following Copilot-engine workflows ran but produced no
agent-artifactswithtoken-usage.jsonl:agent,detection,firewall-audit-logs) but noagent-artifacts— likely failed before AWF agent completed--enable-api-proxyagentartifact but notagent-artifacts— does not use--enable-api-proxyExcluded from scope (Claude/Codex engine): Secret Digger (Claude) (6 runs), Secret Digger (Codex) (6 runs), Smoke Claude (1 run), Smoke Codex (1 run)
Historical Trend
Trend observations:
Previous Report
📊 Copilot Token Usage Report 2026-04-04
References: