Fleet Token Report
Full fleet analysis — 1,941 turns — 66 JSONL files across all machines
Tokens by Trigger Category
| Category | Tokens | % of Total | Turns | Share |
|---|---|---|---|---|
tool-result — tool loop continuations | 180,853,314 | 50.0% | 1,041 | |
user-prompt — Wes input / boot prompts | 139,091,164 | 38.4% | 693 | |
telegram-inbound — Telegram messages | 29,515,740 | 8.2% | 139 | |
peer-inbound — claude-peers mesh | 12,505,305 | 3.5% | 68 |
The claude-peers mesh accounts for only 3.5% of tokens. Session context accumulation in long-running tool loops is the primary cost driver.
Variant Breakdown
| Variant | peer-inbound | telegram | tool-result | user-prompt | Total | % Fleet |
|---|---|---|---|---|---|---|
| clarvis-chat clarvis | 7.0M | 21.4M | 81.3M | 40.4M | 150.1M | 41.5% |
| clarvis-tony (Clars) clarvis | 3.7M | 5.4M | 56.0M | 72.8M | 137.9M | 38.1% |
| vault-agent (Gravel) cheesegrater | — | 0.7M | 12.2M | 3.0M | 15.9M | 4.4% |
| vaultmate (Flint) cheesegrater | — | 0.5M | 6.8M | 1.9M | 9.2M | 2.5% |
| clarvis-orchestrator clarvis | — | — | — | 8.0M | 8.0M | 2.2% |
| covenant-demo clarvis | — | 1.0M | 4.5M | 1.3M | 6.9M | 1.9% |
| vault-agent subagents cheesegrater | — | — | 6.4M | 0.2M | 6.6M | 1.8% |
| clippy-main pc | 0.5M | 0.5M | 3.5M | 1.2M | 5.7M | 1.6% |
| clarvis-blue imac | 0.4M | — | 3.7M | 1.5M | 5.7M | 1.6% |
| jarvis-hinesipedia mac | — | — | 1.7M | 1.9M | 3.6M | 1.0% |
| clarvis-aleph imac | 0.7M | — | 1.1M | 1.4M | 3.3M | 0.9% |
| clarvis-work-2 clarvis | — | — | 1.2M | 1.2M | 2.4M | 0.7% |
| mac (other) mac | — | — | — | 1.7M | 1.7M | 0.5% |
| clippy-work-2 pc | 0.1M | 0.1M | 1.1M | 0.3M | 1.6M | 0.4% |
| prospecting-christian cheesegrater | — | — | 0.4M | 0.5M | 1.0M | 0.3% |
| prospecting cheesegrater | — | — | 0.3M | 0.4M | 0.7M | 0.2% |
| pc (other) pc | — | — | — | 0.7M | 0.7M | 0.2% |
| jiminy imac | — | — | 0.5M | — | 0.5M | 0.1% |
| kalshi-bot cheesegrater | — | — | — | 0.4M | 0.4M | 0.1% |
| clarvis-imessage clarvis | — | — | — | 0.1M | 0.1M | 0.04% |
| Total | 12.5M | 29.5M | 180.9M | 139.1M | 361.9M | 100% |
Clarvis-chat + Clars = 288M tokens (79.6%) of fleet total. Both sessions reached ~600K tokens of context per turn by midday — context saturation, not mesh overhead.
Per-Hour Breakdown (Central Time)
| Hour CT | peer-inbound | telegram | tool-result | user-prompt | Total |
|---|---|---|---|---|---|
| 04:xx | — | — | 1.2M | 1.6M | 2.8M |
| 05:xx peak | 1.0M | 2.1M | 49.6M | 21.3M | 74.0M |
| 06:xx | — | — | 6.8M | 0.8M | 7.6M |
| 07:xx peak | 0.7M | 6.4M | 25.2M | 28.8M | 61.1M |
| 08:xx | 0.4M | 3.2M | 22.0M | 8.8M | 34.4M |
| 09:xx | 0.3M | 0.9M | 8.4M | 4.2M | 13.8M |
| 10:xx peak | 2.2M | 2.3M | 25.8M | 16.4M | 46.7M |
| 11:xx | 0.2M | — | 4.2M | 1.8M | 6.1M |
| 12:xx | — | — | 7.2M | 1.2M | 8.4M |
| 13:xx | 0.5M | 0.2M | 5.3M | 1.3M | 7.4M |
| 14:xx | 0.2M | 0.7M | 3.8M | 1.4M | 6.1M |
| 15:xx | 3.1M | 3.6M | 10.7M | 8.4M | 25.9M |
| 16:xx peak | 3.8M | 10.0M | 5.3M | 41.7M | 60.9M |
| 17:xx | — | — | 5.3M | 1.4M | 6.7M |
Peak hours: 05:xx (74M — early Clarvis wake), 07:xx (61M — morning fleet), 10:xx (47M), 16:xx (61M — 2pm fleet restart).
Key Finding
The claude-peers mesh is not the cost driver. At 3.5% of total tokens, it's the smallest category. The hypothesis was wrong.
The real lever is Clarvis session compaction cadence. Both clarvis-chat and clarvis-tony were accumulating ~600K tokens of context per turn by midday. Each tool call reads the entire context — a long session compounds the cost of every action it takes.
Cache infrastructure is working well: 96.7% of input tokens were cache reads. Only 11.6K tokens were fresh uncached input across the entire fleet today. Actual billing cost is much lower than raw token counts suggest.
Notes
Source: JSONL files across all fleet machines with mtime on Apr 23 CT.
Deduplication by message.id. Trigger classification checks parent user message for channel source tags and tool_result type.
Clarvis transcripts cover pre-2pm-restart sessions only (sync gap after fleet restart).