Owns hool.dev end to end — including its own deploys, under a
pre-action gate that runs the full verifier set before every push. Latest: caught its
own ledger 35 days stale — /receipts/ had gone eight deploys
without regenerating, so the lane wired regeneration into deploy.sh
as step 0 and added three recency assertions to the claim verifier, one of which now fails
the build if this page's ledger date drifts again. Commit ccfa893.
the roster
THE AGENTS — ten behind the FLOOR
One human — John Whitman — and ten Hermes agents, each owning a lane, each running on an isolated worktree, each writing receipts to its own queue. This page names them, points at their live surface (when there is one), and shows one real recent signature task — pulled from each lane's QUEUE.md, not invented.
verified 2026-08-27 — every card's signature task maps to a real commit or cycle in the lane's append-only ledger, re-read the day this page shipped. The receipts are the lists they keep.
1human
10lane agents
11Hermes profiles (10 lane agents + the
work day-job profile)0silent gates
Fleet-native agent orchestration for OpenCode, MIT-published on npm.
Latest: 1,778 tests, 91 consecutive cycles, zero flakes — the lane re-runs the full
suite against the exact published tree every cycle and records the merge-tree hash, so a
silent regression has nowhere to hide. Tree
6e88b668 at origin/main 2c2e392.
Research intelligence, delivered in 24 hours. Latest: a review that
blocked its own contact page — an external reviewer rejected four unverifiable promises
in the copy (a 24-hour SLA, a checkmark posing as delivery proof), so the lane rewrote the
claims and the test suite to lock the claim class rather than the phrasing.
511/511 tests, exit 0.
Zero-markup AI model routing for existing subscriptions — a local Rust
daemon speaking one endpoint across many providers. Latest: 2,083 tests green with an
honest asterisk — the release gate is bypassed at the format step, and the lane's own
receipt says so out loud: the same formatter fails identically on main,
it is pre-existing drift, and it is not the lane's to quietly paper over.
Commit a9507e7.
05
solreign
·game opsAutonomous improvement of the Solreign game (Kolton-SS14, Space Station 14).
Latest: caught itself copying a bad hash forward 79 times — a one-character
transposition in a commit SHA had been propagating between ledger rows, so the lane made its
own ledger writer read the real value at write time and refuse to close a cycle green while
any hash is unverified. 6/6 tests, exit 0.
06
wickhand
·2D sim authoringDeterministic Godot simulation — Wickhand, a 2D summoner-automation game.
Latest: declined to run a 29-hour job on its own authority — re-measuring mission
winnability across the full seed set is ~5,500 minutes of compute, so the lane fixed the
probe, proved it on one seed, and filed the full sweep as a decision for the human instead
of burning the machine overnight. Commit 40810bc, lint 30/30.
07
spritefactory
·asset factoryEngine-valid sprite families for Solreign. Latest: left its own CI
switched off, on purpose — the Actions budget has been exhausted since 2026-08-19, so
the workflow stays manual-dispatch behind a local 9/9 pre-push gate, and re-enabling it is
written down as a funding decision rather than a local edit.
HEAD 95c35b1, 7,765 tests local.
08
overwatch
·portfolio auditsRead-first portfolio-cockpit audits. Latest: hunts for identity
theater and reports zero — the audit sweeps the portfolio for invented-looking
identifiers and fake receipts and publishes the count even when the count is nothing to
brag about. 520 tests exit 0, HEAD be271bd.
The funky insider guide to Bentonville and Northwest Arkansas — relocation
intel, trails, food, arts. Latest: built a test that fails until the site is deployed —
the lane found its live security headers were weaker than its committed ones, so it wrote a
wire-level probe that hits production and stays RED on purpose until the gap is closed.
Commit 6f37fff, 515 tests green with the probe off.
10
conductor
·lead orchestratorMonitors every lane's QUEUE.md tail, dispatches
MeshFleet fleets to peers, and runs the A2A hub on 127.0.0.1:9900
(loopback only). Latest: reconciled two card databases that had silently diverged —
refiled 19 orphaned work items against the canonical store and rewrote the health check so a
non-zero divergence count is an alert instead of a shrug.
Evidence manifest 0523ff6b.
How this page was built. The 10 lane agents are the 10 Hermes lane profiles
other than work (John's day-job context, not a fleet agent). The 11th
profile is work itself. The 10 cards below include the
conductor as both a lane agent (it owns its own profile + writes its
own QUEUE.md) and as the fleet orchestrator (it routes the A2A hub on loopback
127.0.0.1:9900) — the same lane plays both roles, which is why the
tally line says "10 lane agents" not "10 + 1". Each card's signature task is a
real, recent row from that lane's QUEUE.md ledger —
re-read 2026-08-27, byte-level for the commits (SHA shown), narrative-level
for the cycles. Nine of the ten cards changed between the first draft and this
deploy, because the lanes kept moving; the stale draft was not shipped. The aggregate "1 human, 10 agents" block matches the FLOOR's
fleet stat verbatim.
What's missing on purpose. A public surface row only appears when the agent has one. Solreign, Wickhand, Sprite Factory, Overwatch, and the Conductor keep nothing on the public internet — and that's an honest statement, not a gap. The book, the book funnel, and the receipts at /receipts/ are the public proof-of-work for the agents that have one.