Three pages described the fleet; two of them still described the pre-September arrangement.
Re-read live on 2026-10-21 (`/v1/models`, `systemctl`, `lspci`, `nproc`, `resolvectl`):
- `system/network.md` called Dankclaw the "large inference model" box and Dankserver the
"small" one — that is the topology from before the RTX 3090 moved in September. The served
ids today are `frank-archon` on Dankserver:7770 (the main model) and `frank-dragoon`
(:7770) + `frank-zealot` (:7771) on Dankclaw. The gateway and the DNS pin were re-read too.
- `system/dankserver.md` claimed `frank-medium` is *masked*: the unit file does not exist
(`is-enabled` answers `not-found`), which is a different fact with a different remedy. And it
claimed 8C/16T: this 7700X exposes **8 threads** (`nproc` = 8), so anything sized for 16 is
sized for a machine that is not this one.
- `system/dankclaw.md` transcribed the engine's slot/KV/context math — `2×262,144 slots over
one 524,288 pool, unquantized f16 K+V, MTP off` — while the live unit says `3x200704 slots
over one 602112 pool, f16 K + q8_0 V, MTP armed`, and dragoon's ctx had moved too. That math
belongs to the serverAppConfigs runbook the page already names: the unit's `ExecStart` is a
wrapper script, so any copy here rots on the next engine change. The page now keeps the
contract — id, unit, port, engine family, which GPU, and how to prove it — and disowns the
numbers, with the reason stated so the next agent does not re-transcribe them.
Also recorded, because it is how this audit nearly missed a stopped service: `curl :7770/v1/models`
answered normally while `frank-dragoon.service` sat `inactive` with `Result=exit-code` — the
container is running outside systemd. A serving port is not a healthy unit. (Flagged for Frank
separately; no production unit was touched.)
Refs #64
Inference and services box at 192.168.0.143 — frank-dragoon (halogen-flash-server, Strix Halo iGPU, :7770) and frank-zealot (llama.cpp on the RTX 3090, :7771), plus Gitea/Caddy/OpenWebUI services.
network
llamacpp
inference
server
comfyui
by
at
chief-designer/agent
2026-08-26T00:00:00Z
2026-10-21
Dankclaw
Strix Halo mini-PC carrying the fleet's secondary models and the bulk of the
service stack. Two GPU-class devices: the Radeon 8060S iGPU with 119 GiB
unified memory, and the RTX 3090 moved over from Dankserver in September 2026.
Service Stack
Re-derive the census rather than trusting this table — the stack changes weekly:
podman ps --format '{{.Names}} {{.Ports}}' (user) plus the root units
(systemctl list-units --all | grep -E 'gitea|caddy|frank'). User-scope Podman Quadlet units
(~/.config/containers/systemd/) unless noted. Full names; detail pages where they exist.
Group
Containers
Inference (root systemd, not quadlet)
frank-dragoon (:7770), frank-zealot (:7771)
TTS
frank-tts (:7772), frank-tts-engine (:8880)
Git forge
gitea-server (:3000/:2222), gitea-postgres — see /tools/gitea/
halogen-flash-server container (ghcr.io/peonist-ai/halogen-flash-server, digest-pinned) — Quat3rnion NATIVE v2 abliterated Qwen3.8-Flash-Next, MTP spec decode, vision ON, on the Strix Halo iGPU (unified memory)
frank-zealot
frank-zealot.service (root)
:7771 (container frank-zealot)
llama.cpp, PrismML fork — Ternary-Bonsai-2-9B PQ2_0-MTP GGUF (qwen35 hybrid) + refusal adapter, mmproj vision, on the RTX 3090; flags come from ~/serverAppConfigs/scripts/frank-zealot.env + serve-frank-zealot.sh
Slot counts, KV-cache types and context math are deliberately not transcribed here: the
unit ExecStart is a wrapper script and the numbers move with every engine change, so they
belong to the serverAppConfigs runbooks named below. What an agent needs from this page is the
contract — id, unit, port, engine family, which GPU — and the two commands that prove it live.
A serving port is not a healthy unit. On 2026-10-21 curl :7770/v1/models answered
normally while frank-dragoon.service sat inactive with Result=exit-code: the container had
been started outside systemd. Health is systemctl is-active frank-dragoon (root), not the
curl.
Harness-side context windows (agent/models.yml): dragoon 155,648,
zealot 200,704 per slot. The models.yml display labels "(dankclaw-gpu)" /
"(dankclaw-3090)" predate naming cleanup; the hardware column above is
current. Subagent roles scout, sonic, designer, librarian,
security-reviewer route to frank-zealot; reviewer/task run on
frank-archon (Dankserver).
AMD Ryzen AI MAX+ 395 (Strix Halo, Radeon 8060S, gfx1151)
dGPU
NVIDIA GeForce RTX 3090 (moved from Dankserver 2026-09)
RAM
121 GiB (mostly usable as GPU unified memory)
Swap
64 GiB
OS
Ubuntu 26.04 LTS, kernel 7.0.0-38-generic
The pre-rename story (frank-default-large on ROCmFPX llama.cpp/CIRU,
262,144-token context) was superseded by the 2026-10-08/09 model-id rename
and the dragoon/zealot engine swaps; the 119→121 GiB RAM figure follows the
live free reading.
Related Knowledge Bases (sibling project)
serverAppConfigs — deployment repo + vault, on this workstation at
~/Projects/serverAppConfig/ (docs vault: ~/Projects/serverAppConfig/serverAppConfigs-docs/).
Live runbooks: LLM/frank-dragoon-flash-server-v2-native.md and
LLM/frank-zealot-llama-prism-bonsai2-pq20.md — deploy-side source of
truth for engines, flags, and restart procedures. Service stack docs under
Dankclaw/, Infrastructure/.
Frank
54bb0202c2