An agentic OS for a record label.
Live demo · Submission narrative · Adversarial sponsor audit · Architecture, in five pages (PDF) · Platform spec

_{CockroachDB × AWS Hackathon — Agentic Memory · cluster respect-the-funk, CockroachDB v26.2.5, AWS us-east-1, plan BASIC (scales to zero) · no sign-in needed for anything linked below}

---

The one sentence

We map the music industry into a vector index, mail real human beings one at a time, and CockroachDB is the only reason we can prove why we did it.

A label with a new single has to find the few hundred people, out of tens of thousands, who might actually play it. Get that wrong and you are spam. Spindle indexes those people — 43,191 radio stations, curators and shows, from public registers — works out who should hear a given record, and writes to each of them individually. The fleet takes irreversible real-world actions: it emails named music directors. So the interesting question is never what does the agent know. It is why did the agent decide that, then — and that question is what the database is for.

For a judge, in five minutes

Everything here is public, live against the cluster, and needs no account. Numbers are queried on page load, not typed into the page.

Look at	What it is	What it proves
/	Landing page. Every counter across the top is a query run as the page loads.	The index is real: 43,191 counterparties, 14,139 embedded, 29,667 carrying a genre (measured 2026-08-15).
/#tune	Move the dial, watch a shortlist rerank.	Decision provenance. The ranking is a vector query you are running, not a recorded animation.
/#plan	The EXPLAIN output, in full.	vector search with prefix spans — the tenant, model and contact-state filters are inside the index, not a post-filter.
/#send	The send gate and the outbox.	An email is the one thing that cannot be taken back. This is the schema that makes a double-send impossible rather than unlikely.
/#verdict	Where Postgres would have done fine, feature by feature.	We are only claiming the parts that are actually CockroachDB-shaped.
/manual	The operator manual, written for label staff.	What the product is when it is not being pitched.

If you read one document, read docs/2026-08-11-sponsor-audit.md — it is this project arguing against itself about whether CockroachDB is load-bearing here, and it is where the claims below were cut down to the ones that survived.

Why CockroachDB, in one query

Memory here is not static. Lessons accumulate and rerank shortlists, genres land and change embeddings, contact_state moves as threads open. A station ranked 3rd on Tuesday and 15th today — and the fleet emailed a real person on the strength of the Tuesday ranking.

Answering "why did you email this station?" in Postgres means building an event-sourcing subsystem: temporal tables, a versioned copy of every embedding, an audit log of every rerank — and it can only ever answer the questions you thought to log in advance.

Here it is the same query with four words added:

SELECT name FROM party@party_shortlist AS OF SYSTEM TIME '-10m'
 WHERE tenant_id = $1 AND embedding_model = 'openai:text-embedding-3-small'
   AND party_class = 'counterparty' AND contact_state = 'contactable'
 ORDER BY profile_embedding <=> $2 LIMIT 3;

A filtered vector search against the index as it existed at a past timestamp. Not a snapshot table, not an audit trail — the actual index, actual embeddings, actual lesson rows, as of then, with no schema written to support it and no write-path cost. pgvector cannot do this at any price. That is the argument, and it is the whole argument.

The second piece is the send gate, and it is deliberately ordinary SQL carrying the whole load. outbox carries UNIQUE (message_id), so a message cannot be queued twice; message carries UNIQUE (tenant_id, idempotency_key); and a partial unique index allows one open thread per counterparty. Senders claim rows under a lease, so two workers cannot take the same one. apps/spindle/web/tests/test_sender.py asserts this by name — test_a_second_outbox_row_for_one_message_is_structurally_impossible, test_two_senders_never_claim_the_same_outbox_row.

What is actually running

Three states, and the third is the one that matters. live — running against the cluster. built — the code exists, is tested, and nothing has run it in anger. written, not created — the code exists and turning it on is a deliberate act somebody has decided not to take, for a reason recorded in the audit. The third is not a softer version of the second: it means one command finds nothing, and every document here is required to agree with what that command returns.

Piece	State
The counterparty index — FCC, Radio Browser, Wikipedia	live — 43,191 rows
Vector indexes on party, party_chunk, party_fact, lesson	live, cosine, prefix-filtered; the plan is asserted by test_vector_plans.py, not by a comment
Time-travel retrieval (AS OF SYSTEM TIME over the vector index)	live — the query above, verified on the cluster
The console — thirteen views, every one reading the cluster	live, behind sign-in
Outreach: campaign, thread, message, outbox, and the send gate	live
The agent fleet — lease claiming, backoff, follow-on leads, agent_run	live, 15 agents in agents.REGISTRY
Masters — content-addressed by SHA-256 in S3, hash verified on read	live
Genre classification — Discogs-EffNet in a container Lambda	live, validated against six reference tracks
Embeddings	live, via OpenAI. Not Bedrock — see the audit for why not
The changefeed — statement, consumer, Lambda handler, --verify	written, not created. SHOW CHANGEFEED JOBS returns zero rows
REGIONAL BY ROW on contact_route (migration 024, infra/terraform/multiregion/)	written, not created. SHOW REGIONS returns aws-us-east-1 and nothing else

What we are not claiming

Multi-region data domiciling. It would be the strongest architectural argument available and it is not available on this cluster: BASIC is single-region, and CockroachDB Cloud cannot remove a region once added, so converting the cluster holding the system of record would be a one-way door taken for a demo. The migration and the Terraform exist and validate; nothing has been applied. A claim a judge can falsify with one SHOW REGIONS is worse than no claim. Same rule for the changefeed. Both are in the roadmap, named as future work, in the audit §3.

On AWS

One deploy, and it is free at idle. FastAPI + Jinja + htmx in a Lambda behind a Function URL — no ALB, no always-on container. Masters live in S3, content-addressed. The genre classifier is a container-image Lambda from ECR, because essentia-tensorflow does not fit a zip and the work is seconds of CPU per track, sporadically — so it scales to zero like the rest. Sign-in codes and outreach go out over SES. Terraform for all of it is in infra/terraform/spindle/, and the priced workload models that chose the shape are in infra/.

The repository

apps/spindle/     the platform — schema/, web/ (console + fleet), bin/
apps/remixkit/    the asset generator, a subproject
content/          the RemixKit asset pipeline and its five format specs
infra/            Terraform, and the priced workload models
docs/             specs, research, audits, submission, showcase

	What's in it	Start at
apps/spindle/	The platform. Migrations in schema/, applied in order by apply.py; the console and the agent fleet in web/.	apps/spindle/README.md
apps/remixkit/	The asset generator — FastAPI, ports and adapters. A subproject, not the product.	apps/remixkit/README.md
infra/	AWS as Terraform, plus platform-architecture.pdf — the design on eight pages, ending in the risk register.	infra/README.md
docs/	SCOPE-RESET.md and PLATFORM-SPEC.md are the two binding documents. research/ is 14 numbered reports on the outside world. showcase/ is fifteen screens of the running console.	docs/SCOPE-RESET.md

Specs live next to what they describe: content/'s five format specs stay in content/ because content/bin/ parses them, and infra/'s workload models stay in infra/. docs/ is for what describes the product.

Running it

cp .env.example .env          # OPENAI_API_KEY, AWS creds, and the rest

# DATABASE_URL is deliberately not in .env.example — it is a cluster you create.
# docs/runbooks/environments.md §3 has the `ccloud` commands that make one.
export DATABASE_URL='...'
for f in apps/spindle/schema/0*.sql; do
    python apps/spindle/schema/apply.py "$(basename "$f")" || break
done

cd apps/spindle/web
python -m venv .venv && .venv/bin/pip install -r requirements.txt
./dev.sh                                     # http://localhost:8099

apply.py is the only thing that runs a migration, and it refuses any DROP not declared in its DESTRUCTIVE table — so a broken ordering stops rather than eats something.

Tests:

cd apps/spindle/web && .venv/bin/python -m pytest -q
# 788 passed, 342 skipped — the skipped ones need a disposable cluster.

The suite refuses to start against the production cluster and says so; the fence and the reasoning are in apps/spindle/web/tests/conftest.py. docs/runbooks/environments.md has the one command that creates the cluster it will run against.

Three names, and they are not the same thing

• Spindle is the platform — what this repository builds.
• Respect the Funk is the record label that runs on it: tenant #1, not the product. It is why tenant_id and the live cluster still carry that name.
• RemixKit is the asset generator inside the platform, a subproject.

Renamed 2026-08-15. Before that the repository, the product and the label shared one name, which made every sentence about "Respect the Funk" ambiguous between a company and a codebase.

For maintainers

docs/SCOPE-RESET.md is binding and says which of the older documents are reference and which are void — take nothing in a void one as fact or plan. PRODUCT.md, BUILD-SPEC.md, MEMORY-SPEC.md, MINDS-SPEC.md and PIPELINE-SPEC.md were deleted on 2026-08-15; source comments still cite them by section for design rationale, and git show e6c6bd4:docs/PRODUCT.md recovers one when you need to know why existing code looks the way it does. Never for what to build next.

The mark above is docs/brand/thumbnail.svg — see docs/brand/README.md before re-rendering it, and apps/spindle/web/spindle/templates/_mark.html for the same mark as it appears in the product.