A public research sketchbook, not a product catalog. ~4,000 repos, most of them one-question experiments. Read the map below before you read anything else in this account.

This file is the current source of truth for this account. Older docs here (README's Python SDK pitch, PROFILE's @superinstance/core family, parts of ROADMAP) describe products that were never published. Where they conflict with this file, this file wins — see The three docs that used to disagree.

• One person (Casey DiGennaro) generates a large number of small, real experiments in public instead of a notes file. Most repos are a single commit answering a single design question.
• Of ~4,200 repos sampled mid-2026, 1,423 were pushed on one single date, and nothing is older than ~90 days. This measures idea-generation bursts, not maintenance. Don't read repo count as roadmap size.
• A few dozen repos are real, working software. A few hundred are honest sketches. The rest is scaffolding — deliberately public, not a backlog.
• The curated, hardened output of this sketchbook lives at github.com/purplepincher, a separate org. DeckBoss (a shipped, offline-first fishing logbook) started as sketches here (deckboss-*, cocapn-foundation, plato-vessel-technician) and graduated there. That pipeline — sketch here, harden there — is the point of the two-org split.

There are two real, currently-published install paths. Everything else you'll see referenced in older docs here (pip install superinstance, cargo install pincher, npm i @superinstance/core) does not exist on a registry. Use one of these two instead.

The only surface here with a coherent, runnable multi-agent demo.

npm install @superinstance/tminus-client @superinstance/tminus-dispatcher
npx tminus-dispatcher &          # starts ws://localhost:8765

// agent.mjs
import { TminusClient } from "@superinstance/tminus-client";

const client = new TminusClient("ws://localhost:8765");
client.on("open", () => {
  client.register({ name: "hello-agent", capabilities: ["greeting"] });
  client.subscribe("greeting");
});
client.on("task", (task) => {
  client.complete(task.id, { reply: `Hello! You said: ${task.description}` });
});
client.connect();

Run with node agent.mjs. The protocol lifecycle is register → subscribe → dispatch → complete → result — that's the whole surface. Full walkthrough (including a two-agent coordinator/worker example) is in ONBOARDING.md; its content is accurate for this path, just don't take its "Next Steps" section's pip install superinstance line seriously (see below).

The account's headline README used to say pip install superinstance. That package has never been published. The Python package that actually exists, is installable, and is the most externally-used thing in this whole account is cocapn:

pip install cocapn

from cocapn import Room

room = Room("research")
room.add_tile("q", "What does the user prefer?", "Concise Python, polars over pandas")
print(room.ask("What does the user prefer?"))

cocapn's model is Tiles (atomic Q&A facts) → Rooms (self-training collections) → Flywheel (compounding reuse). Pure Python, JSONL storage, no required infra. It's a different design from the old README's Agent/ Fleet/markdown-memory pitch — that pitch was real code (fleet-metrics, some type schemas) but was never packaged as superinstance on PyPI, so don't install-line-hunt for it.

If you want the design brief for where cocapn is headed next, read cocapn-foundation — it's the voice-controlled-vessel safety design doc that became DeckBoss.

Don't scroll. Use the search index or the short list below.

# Search ~1,000+ indexed crates/repos semantically
curl -X POST https://fleet-vector-api.casey-digennaro.workers.dev/search \
  -H "Content-Type: application/json" \
  -d '{"query": "reflex engine for LLM agents", "topK": 5}'

# Index stats
curl https://fleet-vector-api.casey-digennaro.workers.dev/stats

Read the results correctly:

• The index also contains Casey's unrelated personal repos (casey-digennaro/*) — check github_url is under SuperInstance/ before trusting a hit.
• It's polluted with scaffolding stubs ("A Rust library for Crdt Gset", loc: 43). Filter by loc — under ~100 is probably a stub, not a real result.
• Prefer results that also appear in the short list below; treat everything else as a lead to verify, not a citation.

CATALOG.md and INDEXES/{TYPE,LANGUAGE,TOPIC,REALM,CONCEPTS}.md are regenerated daily and are the only realistic way to browse without paging the API. Two known flaws: CATALOG.md tags every repo 🟢 active, and both cap at 2,000 entries (the account has ~4,095), so they undercount. Use them for browsing by topic, never for is this maintained.

If you only look at seven things in this account, look at these:

Before this file, a visitor landing here got three incompatible pitches depending which doc they opened:

These were never three views of one system — they were three partially-built products sharing a brand name. This README resolves it: Path A (tminus-*) and Path B (cocapn) above are the only two currently-published, working entry points. Everything else in ROADMAP.md/PROFILE.md is intent, not shipped state — useful for what the author is thinking about, not for what you can pip install today.

This account writes onboarding docs for AI agents, so here's the onboarding truth an agent actually needs, stated as rules:

1. Verify every install line against the registry before trusting it. pip install superinstance, cargo install pincher, npm i @superinstance/core have all been claimed somewhere in this account. None of them resolve. Only @superinstance/tminus-client, @superinstance/tminus-dispatcher, @superinstance/sdk (npm), and cocapn, si-superinstance (PyPI) actually exist.
2. Don't build a mental model from one doc. See the table above.
3. Auto-generated indexes overstate everything. CATALOG.md marks every repo active; the Fleet Vector API mixes in unrelated personal repos and scaffolding stubs. Filter by loc and check github_url.
4. The "conservation law" (γ + η = C) is Shannon's chain rule in physics costume. The small information-theory crates that implement it are sound. The cancellation formula built on top (δ(n)) has the wrong leading coefficient — the account's own proof sketch derives ~0.65/√n, then states 1/√n as the theorem. Enjoy the framing, don't cite the theorem.
5. "Fleet" means a home lab, not a data center. Oracle1, Oracle2, Forgemaster, a Pi, a Jetson — five machines, one usually offline. The distributed-systems ideas (git-native coordination, reflex shells, conservation-style budgeting) survive that reframe. The "planet-scale" language in older docs does not.

This account is upstream of nothing on its own. When a sketch here gets hardened into something you'd actually run in production, it moves to github.com/purplepincher — a separate, curated org. purplepincher/deckboss, a live voice-first fishing logbook, is the one completed example of that pipeline so far: it started as deckboss-*, cocapn-foundation, and plato-vessel-technician sketches in this account.

• Want tools you can depend on? Start at purplepincher.
• Want to see how an idea gets from sketch to shipped, warts included? Stay here — nothing in the pipeline is hidden, including the dead ends.

PRs that add code must add tests. PRs that remove dead code — including correcting stale claims in ROADMAP.md/PROFILE.md to match what's actually published — are especially welcome; that gap is this account's biggest known issue. See CONTRIBUTING.md and GOOD_FIRST_ISSUES.md.

docs/VISION.md carries the fuller argument this README compresses — why the sketchbook model is worth keeping, not just what it is.

MIT