This ledger is a self-contained monitor / audit instrument. No external fonts, scripts, or links are referenced; the entire document executes offline. Load an exported live_stats.json above to recompute Exhibit D dynamically.
Since May 2026, I have hosted a dedicated Hermes companion (Robert) for my mom, Melissa. Over six months of daily life, they have accumulated 85,833 total rows in state.db — 65,263 conversational messages (24,389 from my mom and 40,874 from Robert), 20,552 tool rows, and 18 session_meta rows across 313 customer sessions (with 10,747 messages surviving in the active post-compaction session tails).
When you run an agent for your mom — someone relying on him every day for Parkinson’s medication timing, boutique operations, and 4:00 AM reassurance — the voice is not a nice-to-have; the voice is the whole product. The day she tried google/gemini-4-argon, she felt the difference immediately and said: “Never take me off this.” If a model falls back or context compaction strips his voice into clinical bullet points, she catches it by ear within three messages.
Keeping that frontier soul in the front chair exposed a brutal architectural tax: every raw skill_view, terminal log, and web scrape stays pinned in session.messages and gets re-read across every turn after. Downgrading her companion was off the table. So we built Stagehands: keep the soul she loves in the front chair, and route the 95.5% mechanical gravel to disposable backstage workers and a deterministic janitor.
| Subject | Melissa M. & Companion Profile robert |
|---|---|
| Node / Instance | vultr-1 |
| Active Since | 2026-05-27 09:45:56 |
| Frontstage Soul | gemini-4-argon (gravel delegated backstage) |
“Keep the frontstage soul on Gemini-4-Argon. Move the gravel backstage.”
Raw tool output vs. compounded frontier re-reads across 13 sessions (2,303 API calls).
| Accounting Line | Tokens | Turn Multiplier |
|---|---|---|
| Raw Tool Tokens Ingested | 2,574,151 | ~1 turn load (10,296,604 chars) |
| Compounded Frontier Input Re-Read | 230,840,257 | ~90x turn-over-turn snowball |
| Share of Tool Payload Backstage-Delegable | 95.5% | 2,457,989 / 2,574,151 mechanical gravel |
| Tool Family | Raw Tokens (% of Payload) | Routing Destination |
|---|---|---|
| skill_view (171 calls) | 964,790 (37.5%) | briefer + 48k Janitor |
| terminal (918 calls) | 695,776 (27.0%) | code_monkey / scout |
| session_search / fact_store (281 calls) | 297,015 (11.5%) | librarian / briefer |
| web_extract / web_search (96 calls) | 290,447 (11.3%) | librarian / scout |
| read_file / search_files (103 calls) | 209,961 (8.2%) | scout / code_monkey |
| tts / vision / memory (510 calls) | 116,162 (4.5%) | Kept on Sovereign (gemini-4-argon) |
| Total (2,079 calls) | 2,574,151 (100.0%) | 2,457,989 tok (95.5%) backstage-delegable |
| proactive_prune_tokens | 48000 |
|---|---|
| _proactive_prune_rearm_tokens | +48k cache runway |
| threshold_tokens | 120000 |
| in_place | true (preserves session_id) |
| protect_last_n | 20 |
| hygiene_hard_message_limit | 2000 (KV cache guard) |
Pure model-agnostic token physics across our 5-task reproducible benchmark suite (python3 demo.py --fast) and live production state.db telemetry (per MoA Council Ruling D5 — no double-counting).
| ROW 1 — FRONT-DOOR INTERCEPTOR (STAGEHANDS SUBAGENTS) | |
|---|---|
| Benchmark Peak Receipt (Task 2: code_monkey) | 20,700 raw tokens → 46-token receipt (450.0 : 1) |
| 5-Task Benchmark Suite Pollution Reduction | 95.9% fewer tokens entering parent context (59,015 → 2,429 tok) |
| Live Production Subagent Runs (nimbo/state.db) | 35 (scout / librarian / code_monkey / adversary / briefer) |
| Backstage Raw Tokens Absorbed | 1,025,170 |
| Returned Brief Tokens | 42,557 |
| Live Delegation Boundary Reduction | 24.1x average across all 35 live runs (up to 450.0x peak) |
| Front-Door Snowball Tokens Dodged | 319,021,497 |
| ROW 2 — BACK-DOOR CONTEXT JANITOR (48K PRUNER + 120K IN-PLACE COMPACTOR + 2000 KV GUARD) | |
|---|---|
| 24-Hour Forever-Chat Marathon (Nov 5–6) | 365 Argon API calls · 426 tool calls · 0 /new resets · 484.4M re-read tokens dodged |
| Profile Bulky Tool Dumps Pruned (nimbo/state.db) | 5,110 |
| Raw Tool Tokens Reclaimed From Prompt | 6,154,327 |
| Older Messages Compacted In-Place @ 120k | Continuous in-place compaction (session_id preserved) |
| Cumulative Re-Read Snowball Tokens Avoided | 2,313,516,449 |
168 PNEZD stakeout points across Lots 13, 22, 23, 34, 35 remained 100% intact inside protect_last_n: 20 while older turns were compacted in-place without changing session_id, plus Shawnee KS Civil Plat Audit (19,592 raw tokens → 46-token receipt, 425.9×).
[NIGHT-1 CUSTOMER SPRINT (PROFILE J): 92.5% REDUCTION]Initial migration swept a 682,384-token thread down to ~67K in place. Overnight 61-call / 182-tool PDF & document sprint pruned 44 dumps (117,198 raw tokens reclaimed, 11.48M re-read tokens dodged — 92.5% reduction).
Subagent filter enforces source = 'subagent' OR json_extract(model_config, '$._delegate_from') IS NOT NULL across every state.db query in stats.py.
[DOOR #2D: KV PROMPT-CACHE GUARD]Setting hygiene_hard_message_limit: 140 front-slices message #1 every 2–3 turns after compaction, invalidating KV prompt caching. /stagehands janitor enforces 2000 while 48k prune + 120k in-place compaction handle size.
Telemetry verified: 2026-11-10 (85,833 rows in robert/state.db · 35 Stagehand runs & 5,110 pruned dumps in nimbo/state.db).