Last TODO closed: the pre-TAPPaaS setup ran five years uninterrupted and never
had a name.
Guardrail 1 now says 2-3 artefacts rather than three — the Issue is always
there, the ADR and the plan only when the change warrants them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The graph replaces the prose slide that stood in for it — annotated with the
vacation that explains the Sept/Oct 2025 flat spot, and regenerated so August
runs to the 25th (126). Keeps the one line worth saving from the slide it
replaces: the services stayed up through the flat bits.
Also: "every commit" rather than "every line" on the never-push guardrail;
Home Assistant as the suggested first service; evaluation tier calls for two
disks.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The ADR-007 rewrite was the move from a flat pile of sequenced scripts to the
manager/controller paradigm — not the dependsOn change. That was an earlier
and separate rewrite, ADR-003: a plain install.sh per module replaced by
dependsOn/provides contracts the platform resolves into an order.
Also: a closing line on the hardware slide that this is one of five known
instances, and the September 2025 gap now answers itself.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Guardrail 1 is now the process that runs before the AI is allowed to type:
an Issue with a defined problem and proposed fix, an ADR reviewed by at least
two people for anything non-obvious, then a design and implementation plan
that says how it will be tested. The eight agent roles move here — they are
what "then the AI takes over" means.
Guardrail 5 replaces "I stayed the reviewer" with the test regime: quick test
on every change, deep test via test-module.sh --deep, test plan authored in
the ADR, and a full regression across every module after implementation.
Test figures reconciled to one scope across the deck (both repos): 195 suites,
34,490 lines, 27% of source.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Drop the commit-graph slide and the two cost slides. Replace all three with
one argument: the platform is expensive to build, but that cost is fixed and
per-site cost is near zero — so it only makes sense shared across many home
labs and SMB sites.
"What I got wrong" becomes "Did I get it all right?" — no, two large rewrites
(ADR-007 numbering to dependsOn, ADR-014 zone tiers to checked state) plus a
predecessor system run for years before this one.
36 slides; the agenda line and the quiet-months slide no longer refer to the
graph that left.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Split the single architecture slide into two: the iron (three nodes, switch,
AP, modem) and the software (foundation band with the AI and productivity
stacks above it). Both read off the running cluster via tappaas-cicd, not
from docs — CPU, RAM, pool topology, link speeds and switch ports are live
values.
The VM layer-cake is inline SVG rather than mermaid: mermaid ignores
`direction LR` inside a subgraph that has a cross-boundary edge, so the bands
came out as a 629x1258 column.
Module slide now counts all three sources — 8 foundation + 12 apps +
21 Community = 41 — and slide 3 is recomputed for today.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Talk announced at pretalx.varum.dk/sommerhack-2026/talk/WD8EVP/ —
27 August, Taler Teltet, 60 minutes.
Figures come from the repo rather than the abstract: commit counts and the
per-month graph from git log, source counts from src/STATISTICS.md as
regenerated 2026-08-23. Twelve TODO markers remain for the personal
material (costs, the autumn 2025 dip, war stories).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
