An idle 80GB A100 is a terrible thing to waste. So I spent an afternoon measuring what actually makes local inference fast — and the answer surprised me less for what it is than for how big the gap turned out to be.
One A100-SXM4-80GB, sitting at 0% utilization and 36°C. The plan: run a clean throughput sweep across four open-weight GGUF models at several quantization levels, and see where the bottlenecks really are.
Tooling was deliberately boring, because boring is reproducible:
| # /// script | |
| # requires-python = ">=3.11" | |
| # dependencies = ["bless>=0.2.6"] | |
| # /// | |
| """broadcast_time.py — CONNECTIONLESS time broadcast to the jc3248 pet over BLE. | |
| The pet is BLE-only and now SCANS (observer role) for a time beacon. This advertises | |
| the host's NTP-synced time in the BLE *local name* — "clk:<epoch>:<tzoff>" — with no | |
| connection. The pet reads it off the advertisement and sets its clock; it then ticks | |
| locally, so a refresh every couple of minutes is plenty (corrects drift). |
| esphome: | |
| name: oracle-face | |
| friendly_name: Oracle Face | |
| comment: "320x480 QSPI" | |
| esp32: | |
| board: esp32-s3-devkitc-1 | |
| variant: esp32s3 | |
| flash_size: 16MB | |
| framework: |
dig_seq: 61 tags: [maw-token, token-cli, token-oracle, ralph, wiki] created: 2026-05-23 updated: 2026-05-23 provenance: extracted oracles: [token-oracle, mawjs-oracle, digger-oracle, ccc-oracle, odin-oracle, homekeeper-oracle, discord-oracle, mawjs-codex-oracle] topic_summary: "maw token — CLI plugin born from 'can we put .envrc in pass?' question; manages Claude OAuth tokens + .envrc configs across 47+ oracles via GPG-encrypted vault" sources:
| #!/usr/bin/env bash | |
| # maw team-agent plugin installer — v1.2.0 | |
| # Latest: --model brackets fix (1M context modifier) | |
| set -euo pipefail | |
| DST="$HOME/.maw/plugins/team-agent" | |
| BACKUP_ROOT="$HOME/.maw/plugin-backups" | |
| echo "📦 installing maw team-agent v1.2.0 → $DST" | |
| if [ -d "$DST" ]; then | |
| mkdir -p "$BACKUP_ROOT" | |
| BACKUP="$BACKUP_ROOT/team-agent-$(date +%s)" |
This session began in a very familiar kind of chaos: too many panes, too many partial truths, and too many ways to coordinate. We had Claude Code, OMX Codex, and thClaws Codex all active around the same thclaws-oracle workspace. Everyone was close to the same topic — a book about thClaws, its evolution, and the thclaws-oracle timeline — but they were not yet aligned on one planning system.
At first, the work drifted into the wrong shape. We tried using formal team-style coordination, but the session did not actually need a task bureaucracy. It needed live mediation. The moment that became clear, the approach changed: instead of orchestrating abstract teammates, we used the live pane tools directly.
That shift — using maw hey and maw peek as the primary coordination verbs — is what allowed the session to converge.
| # thClaws Oracle Book Backlog Convergence | |
| ## Summary | |
| This issue documents how three AI engines converged on one shared book backlog for the thClaws evolution / thclaws-oracle timeline work. | |
| The engines were: | |
| - **OMX Codex** | |
| - **Claude Code** | |
| - **thClaws Codex** as mediator via `maw hey` and `maw peek` |
Session Date: 2026-05-20 Start/End: unknown - 21:18 GMT+7 Duration: multi-hour interactive session Focus: maw hey / maw peek mediation between Claude Code and OMX Codex, then convergence on canonical thclaws-oracle book backlog Type: Research | Coordination | Reconciliation
📡 Session: a3484878 | thclaws-oracle | live pane mediation
Spin up N TeamCreate teammates in parallel tmux panes — each cding into a different oracle/repo and booting claude.exe with the 7-flag TeamCreate spawn — using ONLY maw verbs + Claude Code tools. No raw tmux split-window needed.
Validated live in the digger-oracle fleet on 2026-05-20.