📊 Claude Token Usage Report2026-06-16

[github-actions]

Overview

Period: 2026-06-15 10:30 UTC to 2026-06-16 04:47 UTC (~18.3 hours)
Runs analyzed: 26 (14 had token data — 54% coverage; 12 runs missing token data)
Total tokens: 876K across all workflows
Total Actions minutes: 427 min | Total turns: 40 | Total errors: 25
Estimated cost: not available — estimated_cost and token_usage_summary fields absent from this snapshot

🚨 Failure rate at new high — 25/26 runs failed (96.2%), up from 89.5% in yesterday's report. The one success was on the main branch (§27553866481).

⚠️ Token coverage dropped sharply — Only 54% of runs recorded token data (down from 95% in yesterday's report). All 12 runs with missing token data are single-turn failures, suggesting they crash before the api-proxy can record usage.

⚠️ Another burst cluster — 12 runs fired between 13:42–15:07 UTC on June 15 (~85 minutes), all failing, repeating the retry-storm pattern seen in previous reports.

---

Workflow Summary

Workflow	Runs	w/ Tokens	Total Tokens	Avg Tokens/Run	Avg Turns	Errors	Failures
Smoke Claude	26	14	876K	62.5K	1.5	25	25/26 🚨
Total	26	14	876K	—	—	25	25/26

---

🔍 Optimization Opportunities

1. Smoke Claude — 96.2% failure rate; chronic deterministic failure pattern persists

   • All 14 runs with token data consumed exactly ~62.5K tokens in exactly 2 turns (range: 62,346–62,674 — 0.5% variance). This is not flakiness; it is a hard-coded failure path
   • The 12 single-turn failures (no token data) represent a worse failure mode: the run crashes before even completing a turn — no useful work is done, but runner time is still consumed (~12 min × 12 = ~144 lost action-minutes)
   • The June 15 burst cluster (13:42–15:07 UTC) shows 12 back-to-back runs across multiple copilot branches, all failing within 85 minutes
   • Recommendation: Investigate and resolve the underlying failure before any further optimization; the 96% failure rate makes all other optimizations irrelevant

2. Token coverage drop — 54% of runs now have token data (was 95% yesterday)

   • The 12 runs without token data all have turns=1 and conclusion=failure, and span diverse branches (copilot/*, fix/*, feat/*), not just one bad branch
   • This suggests a new failure mode that terminates before the api-proxy emits usage data
   • Recommendation: Check api-proxy logs and agent container startup for single-turn crash signatures; compare agent container version across token-present vs token-absent runs

3. Estimated cost not instrumented — estimated_cost is null for all 26 runs (persistent across all historical reports)

   • Recommendation: Verify api-proxy sidecar cost instrumentation is emitting to run metadata

Per-Workflow Details

Smoke Claude

• Runs: 26 total (14 with token data, 12 without)
• Total tokens: 875,722 (avg ~62,551/run for data-bearing runs; ~33.7K averaged across all 26)
• Token range (14 runs): 62,346 – 62,674 (extremely low variance — fully deterministic)
• Total turns: 40 (avg 1.5/run — runs without token data are all 1-turn crashes)
• Total Action minutes: 427 (avg 16.4 min/run)
• Estimated cost: not available
• Model/cache breakdown: not available (token_usage_summary absent from all runs)
• Error rate: 25/26 runs reported errors
• Failure rate: 25/26 (96.2%)

Runs with token data (newest first):

Run	Created (UTC)	Branch	Tokens	Conclusion
§27555559405	2026-06-15 15:02	copilot/optimize-copilot-token-usage	62,493	❌ failure
§27555485028	2026-06-15 15:01	copilot/duplicate-code-token-limit-proxy-tests	62,674	❌ failure
§27555456601	2026-06-15 15:01	copilot/remove-unused-export-squidlogcheckresult	62,597	❌ failure
§27555408122	2026-06-15 15:00	copilot/remove-unused-export-aggregatelogs	62,618	❌ failure
§27552347965	2026-06-15 14:12	copilot/duplicate-code-token-limit-proxy-tests	62,595	❌ failure
§27553866481	2026-06-15 14:36	main	62,346	✅ success
§27551885971	2026-06-15 14:05	no token data	—	❌ failure
§27551523853	2026-06-15 13:59	copilot/fix-duplicate-auth-error-tests	62,606	❌ failure
§27551255384	2026-06-15 13:55	copilot/optimize-copilot-token-usage	62,555	❌ failure
§27551021234	2026-06-15 13:51	copilot/remove-unused-export-aggregatelogs	62,559	❌ failure
§27550971808	2026-06-15 13:50	copilot/remove-unused-export-copilotmodelvalidationresult	62,574	❌ failure
§27550843311	2026-06-15 13:48	copilot/fix-duplicate-code-schema-tests	62,554	❌ failure
§27550615099	2026-06-15 13:44	copilot/remove-unused-export-iscontainerstartupfailureerro	62,469	❌ failure
§27550498228	2026-06-15 13:42	copilot/remove-unused-export-squidlogcheckresult	62,515	❌ failure
§27540168877	2026-06-15 10:30	test-coverage/squid-acl-security-d3ff004e8d3a1289	62,567	❌ failure

Runs Without Token Data (12 runs)

All 12 are single-turn failures — these runs crash before the api-proxy records any token usage. Diverse branches suggest a systemic early-exit issue, not branch-specific:

Run	Created (UTC)	Branch	Conclusion
§27594833083	2026-06-16 04:47	copilot/allowed-and-disallowed-models-again	❌ failure
§27593367930	2026-06-16 04:05	fix/ghes-copilot-auth-prefix	❌ failure
§27588088857	2026-06-16 01:35	main	❌ failure
§27587614658	2026-06-16 01:22	fix/ghes-copilot-auth-prefix	❌ failure
§27584084227	2026-06-15 23:50	copilot/refactor-finalize-http-tracking	❌ failure
§27581415328	2026-06-15 22:46	fix/ci-gate-reviewer-login	❌ failure
§27579831037	2026-06-15 22:12	copilot/refactor-validate-domain-or-pattern	❌ failure
§27573923031	2026-06-15 20:21	copilot/refactor-split-api-proxy-options	❌ failure
§27565954996	2026-06-15 18:00	feat/ci-gate-label-command	❌ failure
§27562384992	2026-06-15 16:56	fix/acl-generator-test-assertions	❌ failure
§27555855276	2026-06-15 15:07	copilot/smoke-claude-optimization	❌ failure
§27551885971	2026-06-15 14:05	copilot/smoke-claude-optimization	❌ failure

Note: the copilot/smoke-claude-optimization branch appears twice among single-turn failures — somewhat ironic given the branch name.

---

Historical Trend

Date	Runs	w/ Token Data	Total Tokens	Failure Rate	Notes
2026-06-16 (today)	26	14 (54%)	876K	96.2% 🚨	Token coverage dropped; 12 single-turn crashes
2026-06-15	19	18 (95%)	1,125K	89.5% 🚨	Retry storm Jun 14 afternoon
2026-06-14	16	~16	1,001K	94% 🚨	—
2026-06-13	—	—	—	—	See §4862

Trend: Smoke Claude has failed at ≥89% for at least 4 consecutive days. This is no longer a transient issue — it is a chronic systemic failure. The drop in token coverage (95% → 54%) suggests the failure mode is also evolving, with runs now dying earlier before even completing a turn.

Previous Report

§5011 — 📊 Claude Token Usage Report 2026-06-15

---

References:

• Workflow run for this report
• Previous report #5011
• Previous previous report #4926

Generated by Daily Claude Token Usage Analyzer · ◷