The spec is canonical.

Every piece of Najim's architecture, published openly. Code is bound to these documents — not the other way around. If code disagrees with the spec, the spec wins.

html

The Najim Brain — architecture + deploy runbook

Canonical spec. Architecture diagram, bootstrap procedure, per-tenant provisioning, integration points, verification, rollback. Spec v1.0.0.

563 lines · ~41 KB

json

Versions · pinned

Every container tag, gbrain SHA, port, env var, model name. Single source of truth — drift is a git diff.

Implementation checklist

5-phase tracker — pre-flight, bootstrap, engineering scaffolding, first-tenant provisioning, cutover & hand-off. Done when grep returns nothing.

bootstrap.sh — one-time VPS install

Executable §5 of the spec. Idempotent, fail-fast, [ok]/[skip]/[fatal] markers map 1:1 to numbered sub-steps.

258 lines

provision-tenant.sh — replicable per-tenant deploy

Executable §6. Four args (tenant_id, slug, name, owner_email), 7 verified steps, ≈ 10 min per tenant.

193 lines

gbrain.py — Python integration module

Locked function signatures for prompt-builder. get_context(), append_fact(), format_context_for_prompt(). Implementation fills bodies; cannot invent surface.

155 lines

sql

021_clients_gbrain.sql — schema migration

Real migration file. Adds clients.gbrain_token, gbrain_source_slug, gbrain_provisioned_at + two indexes. IF NOT EXISTS throughout.

38 lines

test_najim_brain.py — verification suite

§10 as pytest. Health, stats, retrieval relevance, p95 latency, Dream Cycle. Green tests are the only legitimate "done" signal.

154 lines

Hardwiring rules. If the doc, versions.json, and code disagree — doc wins, then JSON, then code. Never edit bootstrap.sh or provision-tenant.sh without updating §5 or §6 of the spec in the same commit. Green pytest is the only legitimate “done.”