Get started

Before you build

Read the capability catalog first. It lists every unit with a "use this when" line and its key exports, the runtime paths with the credential each needs, and the known gaps. Reuse what is there; name the gap you fill when nothing fits.

Install one package

Every package is published independently. Take the one you need; there is no meta-package and nothing pulls the rest of the DAG in behind your back.

npm install @titan-design/store-sqlite

Requirements that apply to all of them:

• Node 20 or newer, and ESM only. There is no CommonJS build; require() will not work.
• zod v4 is a peer dependency of registry, daemon, agent, agent-protocol, authority, hitl, memory, workflow, chat-protocol, messaging, code-read, session-analytics, and the three style-* packages. Install it yourself so one copy is shared: npm install zod.
• web-tree-sitter is a peer of code-parser and style-analyzer.
• React 18 or 19 is a peer of react-app, with Vite optional for its ./vite entry.
• @huggingface/transformers is an optional peer of embed and retrieval. Without it, embed's local backend and retrieval's cross-encoder reranker are unavailable and every other path still works.
• Packages that store anything depend on better-sqlite3, which builds a native addon on install.

A first example

This opens a database, installs two kit tables, indexes a piece of text for full-text search, and links two entities with an edge. It is the shape most things here start from.

import {
  EdgeTable,
  SpanFtsTables,
  kitMigration,
  openDatabase,
  ref,
  runMigrations,
} from "@titan-design/store-sqlite";

const db = openDatabase(":memory:"); // WAL and foreign keys on; a path works the same
runMigrations(db, [kitMigration(1, { edge: true, spanFts: "search" })]);

const spans = new SpanFtsTables(db, { name: "search" });
const text = "the daemon returns 503 until the pid file exists";
spans.index(
  { ownerRef: ref("session", "abc"), field: "prompt", sourceId: 1, byteOffset: 0, byteLength: text.length },
  text,
);

console.log(spans.search("daemon OR 503", 5).map((hit) => hit.ownerRef));
// [ 'session:abc' ]

const edges = new EdgeTable(db);
edges.assert({ sourceRef: ref("session", "abc"), relation: "touched", targetRef: ref("file", "repo/a.ts") });
console.log(edges.from("session:abc").map((edge) => edge.targetRef));
// [ 'file:repo/a.ts' ]

Two things in there are load-bearing everywhere else on this site.

Refs. ref("session", "abc") is just the string session:abc. Every domain mints refs for its own entities and the edge table links them, which is how a session graph and a code graph end up in one queryable space without either knowing about the other.

The FTS index stores no text. spans.index records a byte range and feeds a contentless FTS5 table. Searching gives you back a locator, and locator reads the original bytes out of the source file. Nothing duplicates the corpus.

Try the whole DAG at once

The session miner is a working product built from ten of these packages. It is private, so run it from a checkout rather than npm:

git clone https://github.com/HJewkes/titan-platform
cd titan-platform
pnpm install && pnpm build
node products/session-miner/dist/bin.js refresh   # index your Claude Code transcripts
node products/session-miner/dist/bin.js search "daemon 503 at startup"

refresh reads ~/.claude/projects and writes to ~/.local/state/titan-session-miner. Both are overridable with --corpus and --state. Every command takes --json and returns the same envelope the HTTP and MCP surfaces return.

The case study walks through which package does what in that pipeline.

Run the products

Five units in this repo run as programs. All are private, so each runs from a checkout after pnpm install && pnpm build. Each has a usage guide with every command, where its state lives, and how it fails.

Unit	What it does	Guide
products/factory	titan-factory lands pull requests with durable runs and human gates	Factory, Shepherd
products/session-miner	titan-miner indexes and searches agent transcripts	Session miner
products/retrieval-eval	retrieval-eval scores retrievers against mined labels	Retrieval eval
apps/codewatch	a browser report over a code-graph snapshot	Code report
apps/console	a read-only console over active-work, the agent-chat broker and the session graph (skeleton)	README
deploy/hub	the Matrix homeserver the human queue uses	Matrix hub

Where each package fits

Pick by problem:

Problem	Package
I need SQLite tables for a graph, a search index, or an ingest watermark	store-sqlite
I need to point at a byte range in a file instead of copying its text	locator
I need to group thousands of error blobs into a handful of templates	cluster
I need embeddings, but not a mandatory model download	embed
I need search that keeps working when a retriever is down	retrieval
I need to run a headless Claude Code or Codex session with a hard budget	agent
I need one command definition to serve a CLI, MCP, and HTTP	registry
I need to host that on a loopback port with health, SSE, and a pid file	daemon
I need to pause work on a human and resume it from another process	hitl
I need to read Claude Code or Codex transcripts	session-read
I need those transcripts as a queryable, incrementally maintained graph	session-graph
I need the import and call graph of a TypeScript or Python tree, with metrics and checks	code-graph
I need a rule playbook whose confidence decays with evidence	memory
I need a long-running process that survives a restart mid-flight	workflow
I need to run many agent steps under a concurrency and budget cap	workflow (mapItems)
I need a durable record of which process owns a running agent	agent-lifecycle
I need identity and usage types that do not care which harness ran	agent-protocol
I need a cost report, roles, or episodes over mined sessions	session-analytics
I need to parse TypeScript or Python with tree-sitter	code-parser
I need a read API over a code graph that a browser can call	code-read
I need to learn a repo's code style and enforce it	style-analyzer, style-profile, style-checker
I need to check that a model's cited lines really say what it claims	evidence
I need to decide whether an actor may merge, release, or read a secret	authority
I need to read and change GitHub through gh, with writes that are safe to repeat	github
I need to prove a fix PR's new tests fail on the base and pass at head	fix-proof
I need to stop home paths and private terms from reaching a public repo	egress-scan
I need to start or resume an agent-chat agent from code	agent-dispatch
I need the envelope, routes, and SSE vocabulary a daemon and its clients share	rpc-protocol
I need a typed daemon client in a browser, live or from a snapshot file	rpc-client
I need React hooks and a Vite preset for a daemon-backed app	react-app
I need one chat message shape across every agent-chat surface	chat-protocol
I need to send and receive over iMessage or Telegram	messaging
I need to talk to a Matrix homeserver without an SDK	matrix-bus
I need human approvals and questions to show up in a Matrix room	queue-mirror

The package families guide groups these by the job they share.