2026-07-31 — The Seven Brains

Seven brain models stood in a row on a workbench under a single amber lamp, each one built of a different material — one of clockwork gears, one crystalline, one woven wire, the rest in varying weaves of bone and stone — no two alike.

The council is not five voices in one robe. It is seven brains that disagree on purpose. This note is the ship record for the 2026-07-31 alignment: one map, three layers, no reboot gap.

The map

Seat	Alias	Upstream (Mini live)
Yoda	`council-max-thinking`	Fable 5 · effort max · Max bridge `:3456`
Windu	`council-spacial`	Gemini 3.1 Pro · `:6543`
Qui-Gon	`council-code`	Devstral · cathedral `:3301`
Ki-Adi-Mundi	`council-finance`	Grok 4.5 · OAuth `:4200`
Cilghal	`council-heretic`	ablated 27B champion · `:6669`
Jocasta	`council-brain`	Opus 5 · effort medium · `:3456`
Mon Mothma	`council-brain`	Opus 5 · effort medium · `:3456`

Jocasta and Mothma share the Opus-medium brain on purpose (archives and ops are the same subscription tier). Everyone else is a different family. CLI brainstorm counts families, not robes — three Claude seats still count as one designed family with a redundant badge.

What was wrong

CLI (sanctum brainstorm / sanctum council) still had Mundi on Yoda’s Fable seat and Cilghal on plain cathedral Qwen. Brainstorm only summoned five seats — no Jocasta, no Mothma.
OpenClaw had five agents; Signal could not reach Jocasta or Mothma.
proxyd had Devstral as council-devstral while council-code pointed at Qwen — so “code” did not mean Devstral.
The biggest reboot trap: openclaw.json is gitignored. The SoT is openclaw.json.enc. A plaintext edit survives only until the next gateway start where the .enc is newer — then SOPS overwrites the roster from the encrypted source. Without re-encrypt, a Mini reboot quietly undoes the agents.

What shipped

Layer	Change	Landed
`sanctum-cli`	7-seat brainstorm + council models; tests updated	`7df48ef` on `origin/main`
Mini proxyd	`council-code` → Devstral; `council-heretic` alias	`f28376c` on Mini `sanctum-config`
OpenClaw	7 agents; `council-tiered` model ids; re-encrypt	`a6b8b85` on `openclaw-config` (`.enc` + sample)
Mini CLI venv	non-editable install of `7df48ef` into `cli-venv`	on-box

Operational context windows (64 GB dual-cathedral)

These are door caps, not the models’ theoretical 256k. They keep OpenClaw turns inside client patience (~120 s) and keep both seats resident.

Seat path	Process	`max_prompt_tokens`	Cache slots	OpenClaw `contextWindow` ad
Cilghal / local Qwen / heretic	27B `:1337` + `:6669`	16384	4 (binary default)	16384
Qui-Gon (`council-code`)	Devstral `:3301`	32768	4 (binary default*)	32768

*Deployed sanctum-mlx does not yet expose --prompt-cache-slots (flag crash-loops the LaunchAgent). Target remains 2 slots for Devstral once the binary ships the flag; until then dual residency relies on Metal memory limits (--metal-memory-limit-mb 22528).

Why not higher: 27B prefills ~97 tok/s (16k ≈ 170 s prefill alone). Devstral is faster but pays ~2.5 GiB KV per 16k slot. Live OpenClaw tool turns often hit 10–30k on Qui-Gon; ads must match the door so agents stop packing past 413.

Durable pins:

27B: ~/.sanctum/launchagents/com.sanctum.mlx.plist → --max-prompt-tokens 16384
Devstral: ~/.sanctum/launchagents/com.sanctum.mlx-devstral-rust.plist → --max-prompt-tokens 32768 (tracked SoT, Library → symlink)
OpenClaw: openclaw.json.enc (re-encrypt after any ad change)

Cloud OpenRouter Qwen seats keep their large windows — only local cathedral paths are clamped.

Reboot permanence checklist

proxyd loads ~/.sanctum/sanctum-proxy/config.yaml from disk (tracked).
OpenClaw gateway decrypts openclaw.json.enc → openclaw.json at start (scripts/sops-start.sh). After re-encrypt, the seven-agent map and the 16k/32k context ads are the encrypted SoT — not a floating plaintext edit.
sanctum on Mini imports from cli-venv site-packages (not a /tmp worktree). Pulling a new CLI release still requires an explicit pip install (or the house upgrade path) — that is intentional, not a silent reboot rewrite.
Backends that must already be up: Max bridge :3456, Gemini :6543, Devstral :3301, Grok OAuth :4200, heretic :6669, cathedral :1337. LaunchAgents for those services are independent of this roster ship.

Operator notes

Live CLI fan-out: sanctum brainstorm "…" (seven panels) and sanctum council "…" (same seven + Yoda synthesis).
After any future OpenClaw edit:
cd ~/.openclaw && bash scripts/update-config.sh && git add openclaw.json.enc openclaw.json.sample && git commit && git push.
Devstral cold-prefill after a :3301 restart can take minutes; proxyd ttfb_timeout_secs: 300 on council-code is the intentional budget.
Context overflow on Qui-Gon/Cilghal: /reset or shrink tool dumps — do not raise the operational cap on a dual-resident 64 GB Mini.

EETISMAD score for this ship

Gate	Status
E2e Tested	Proxyd smoke: max-thinking, spacial, heretic, brain green; finance returns; code needs long TTFB after cold start. CLI unit suite green on the roster commit. OpenClaw ads assert 16k/32k after decrypt.
In Sanctum-docs	This field note (incl. 64 GB context-cap table).
Merged	`sanctum-cli` + `openclaw-config` + Mini proxyd + Devstral LaunchAgent pushed.
Deployed	Mini proxyd config live; OpenClaw re-encrypted with 16k/32k ads; cathedral argv 16k/32k; Devstral slots=2 in LaunchAgent SoT.

The map is the point, not the hardware. Seven seats thinking with one set of weights would agree with each other by construction, and a council that cannot disagree is a single opinion with extra latency. Each seat now reasons on a brain chosen for its job — including Cilghal on the ablated health port, which is a seat that had to stop refusing before it could be useful. That is what makes the disagreement worth listening to.