24 KiB
Mosaic Stack — Developer Guide
Table of Contents
- Architecture Overview
- Local Development Setup
- Building and Testing
- Adding New Agent Tools
- Adding New MCP Tools
- Database Schema and Migrations
- Claude Code Skill Bridge
- Pi Persistent Goal Extension
- API Endpoint Reference
- Local Fleet Canary
Architecture Overview
Mosaic Stack is a TypeScript monorepo managed with pnpm workspaces and Turborepo.
mosaic-mono-v1/
├── apps/
│ ├── gateway/ # NestJS + Fastify API server
│ └── web/ # Next.js 16 + React 19 web dashboard
├── packages/
│ ├── agent/ # Agent session types (shared)
│ ├── auth/ # BetterAuth configuration
│ ├── brain/ # Structured data layer (projects, tasks, missions)
│ ├── cli/ # mosaic CLI and TUI (Ink)
│ ├── coord/ # Mission coordination engine
│ ├── db/ # Drizzle ORM schema, migrations, client
│ ├── design-tokens/ # Shared design system tokens
│ ├── log/ # Agent log ingestion and tiering
│ ├── memory/ # Preference and insight storage
│ ├── mosaic/ # Install wizard and bootstrap utilities
│ ├── prdy/ # PRD wizard CLI
│ ├── quality-rails/ # Code quality scaffolder CLI
│ ├── queue/ # Valkey-backed task queue
│ └── types/ # Shared TypeScript types
├── docker/ # Dockerfile(s) for containerized deployment
├── infra/ # Infrastructure configuration (for example, OTEL collector)
├── docker-compose.yml # Local services (Postgres, Valkey, OTEL, Jaeger)
└── CLAUDE.md # Project conventions for AI coding agents
Key Technology Choices
| Concern | Technology |
|---|---|
| API framework | NestJS with Fastify adapter |
| Web framework | Next.js 16 (App Router), React 19 |
| ORM | Drizzle ORM |
| Database | PostgreSQL 17 + pgvector extension |
| Auth | BetterAuth |
| Agent harness | Pi SDK (@mariozechner/pi-coding-agent) |
| Queue | Valkey 8 (Redis-compatible) |
| Build | pnpm workspaces + Turborepo |
| CI | Woodpecker CI |
| Observability | OpenTelemetry → Jaeger |
| Module resolution | NodeNext (ESM everywhere) |
Module System
All packages use "type": "module" and NodeNext resolution. Import paths must
include the .js extension even when the source file is .ts.
NestJS @Inject() decorators must be used explicitly because tsx/esbuild
does not support emitDecoratorMetadata.
Local Development Setup
Prerequisites
- Node.js 20+
- pnpm 9+
- Docker and Docker Compose
1. Clone and Install Dependencies
git clone <repo-url> mosaic-mono-v1
cd mosaic-mono-v1
pnpm install
2. Use the local PGlite tier
The supported local tier is in-process PGlite and requires no PostgreSQL service. Leave
DATABASE_URL unset for this route. Its default local configuration uses PGlite and performs no
external database probe.
If a local queue service is useful, start only that non-PostgreSQL service:
docker compose up -d valkey
Do not use the current Compose PostgreSQL service: it mounts legacy infra/pg-init SQL and is
not qualified for KBN-101. Start OTEL Collector or Jaeger individually only when needed and
without starting PostgreSQL.
3. Gateway/Web local process (held)
Do not start the current Gateway or web process as a local PGlite route. Gateway first loads the
daemon configuration and then project environment files without a tier guard; a pre-existing
DATABASE_URL can select PostgreSQL, where current startup still reaches runtime DDL/migrations.
Creating a root .env that omits DATABASE_URL does not make this safe, so neither a local
credential file nor a web environment file is a current developer procedure.
PGlite remains the supported in-process data-layer implementation, and the optional Valkey command
above remains safe because it does not start PostgreSQL. A safe Gateway/Web local procedure is held
until KBN-101-02 rejects a daemon, inherited, root, or app-local PostgreSQL DSN and any non-local
tier before connection or DDL; KBN-101-05 then supplies the production renderer/Vault process-exec
or LoadCredential boundary.
Held future procedure
PostgreSQL local and federated deployment are held until KBN-101-00 (external bootstrap), KBN-101-03 (runner), and KBN-101-05 (renderer-backed deployment) land. The following is the held, non-operative future activation order with no current command authority:
external bootstrap → TLS/roles → mosaic-db-migrator --run →
mosaic-db-migrator --verify → Gateway/Compose readiness.
Neither current Compose nor this development guide authorizes PostgreSQL initialization SQL, manual DDL, or a pre-runner start.
5. Gateway/Web start (held)
No Gateway/Web start command is currently authorized for the local PGlite route. Do not use root
pnpm dev as a workaround: it additionally starts configured integrations and cannot establish the
required local-tier/DSN isolation. Resume this section only after KBN-101-02 provides its
fail-closed local-startup evidence.
Building and Testing
TypeScript Typecheck
pnpm typecheck
Runs tsc --noEmit across all packages in dependency order via Turborepo.
Lint
pnpm lint
Runs ESLint across all packages. Config is in eslint.config.mjs at the root.
Format Check
pnpm format:check
Runs Prettier in check mode. To auto-fix:
pnpm format
Tests
pnpm test
Runs Vitest across all packages. The workspace config is at
vitest.workspace.ts.
Build
pnpm build
Builds all packages and apps in dependency order.
Pre-Push Gates (MANDATORY)
All three must pass before any push:
pnpm format:check && pnpm typecheck && pnpm lint
A pre-push hook enforces this mechanically.
Adding New Agent Tools
Agent tools are Pi SDK ToolDefinition objects registered in
apps/gateway/src/agent/agent.service.ts.
1. Create a Tool Factory File
Add a new file in apps/gateway/src/agent/tools/:
// apps/gateway/src/agent/tools/my-tools.ts
import { Type } from '@sinclair/typebox';
import type { ToolDefinition } from '@mariozechner/pi-coding-agent';
export function createMyTools(): ToolDefinition[] {
const myTool: ToolDefinition = {
name: 'my_tool_name',
label: 'Human Readable Label',
description: 'What this tool does.',
parameters: Type.Object({
input: Type.String({ description: 'The input parameter' }),
}),
async execute(_toolCallId, params) {
const { input } = params as { input: string };
const result = `Processed: ${input}`;
return {
content: [{ type: 'text' as const, text: result }],
details: undefined,
};
},
};
return [myTool];
}
2. Register the Tools in AgentService
In apps/gateway/src/agent/agent.service.ts, import and call your factory
alongside the existing tool registrations:
import { createMyTools } from './tools/my-tools.js';
// Inside the session creation logic where tools are assembled:
const tools: ToolDefinition[] = [
...createBrainTools(this.brain),
...createCoordTools(this.coordService),
...createMemoryTools(this.memory, this.embeddingService),
...createFileTools(sandboxDir),
...createGitTools(sandboxDir),
...createShellTools(sandboxDir),
...createWebTools(),
...createMyTools(), // Add this line
...mcpTools,
...skillTools,
];
3. Export from the Tools Index
Add an export to apps/gateway/src/agent/tools/index.ts:
export { createMyTools } from './my-tools.js';
4. Typecheck and Test
pnpm typecheck
pnpm test
Adding New MCP Tools
Mosaic connects to external MCP servers via McpClientService. To expose tools
from a new MCP server:
1. Run an MCP Server
Implement a standard MCP server that exposes tools via the streamable HTTP
transport or SSE transport. The server must accept connections at a /mcp
endpoint.
2. Gateway MCP configuration (held)
Do not configure MCP endpoint credentials, write them to a local environment file, or restart the
Gateway from this guide. Gateway/Web startup is held until KBN-101-02 supplies fail-closed
local-tier/DSN isolation and KBN-101-05 supplies the renderer/Vault process-exec or
LoadCredential secret-consumer interface. The future authenticated MCP route requires verified
HTTPS and certificate validation; plaintext bearer-token examples are forbidden.
Tool Naming
Bridged MCP tool names are taken directly from the MCP server's tool manifest.
Ensure names do not conflict with built-in tools (check
apps/gateway/src/agent/tools/).
Database Schema and Migrations
The schema lives in a single file:
packages/db/src/schema.ts
Schema Overview
| Table | Purpose |
|---|---|
users |
User accounts (BetterAuth-compatible) |
sessions |
Auth sessions |
accounts |
OAuth accounts |
verifications |
Email verification tokens |
projects |
Project records |
missions |
Mission records (linked to projects) |
tasks |
Task records (linked to projects and/or missions) |
conversations |
Chat conversation metadata |
messages |
Individual chat messages |
preferences |
Per-user key-value preference store |
insights |
Vector-embedded memory insights |
agent_logs |
Agent interaction logs (hot/warm/cold tiers) |
skills |
Installed agent skills |
summarization_jobs |
Log summarization job tracking |
The insights table uses a vector(1536) column (pgvector) for semantic search.
PostgreSQL schema work (held)
Do not prepare or run a PostgreSQL target from this branch. The sole runner, bootstrap, and
renderer are future KBN-101 artifacts, not current commands. When KBN-101-00/-03/-05 land, the
owned activation documentation will require external bootstrap → TLS/roles → runner --run →
runner --verify → Gateway/Compose readiness.
Generating migration artifacts
pnpm --filter @mosaicstack/db db:generate is an offline artifact-generation command. It does
not authorize connecting to or initializing PostgreSQL. A future reviewed PostgreSQL procedure
will determine when its output is applied.
Drizzle Config
Config is at packages/db/drizzle.config.ts. The schema file path and output directory are
defined there.
Adding a New Table
- Add the table definition to
packages/db/src/schema.ts. - Export it from
packages/db/src/index.ts. - Generate the offline artifact with
pnpm --filter @mosaicstack/db db:generate. - Do not apply it to PostgreSQL until the future KBN-101 activation artifacts and their owned procedure are available. Direct schema push is not a production-like workflow.
Claude Code Skill Bridge
The framework's canonical skill root is ~/.config/mosaic/skills/; Claude Code
requires registrations under ~/.claude/skills/. The implementation in
packages/mosaic/src/commands/skill.ts owns only direct-child symlinks whose
resolved target remains inside the canonical root.
Security invariants:
- Validate the user-supplied name before filesystem access against
[A-Za-z0-9][A-Za-z0-9._-]*. Separators, control characters, whitespace,.., absolute paths, and leading-are invalid; filesystem-derived invalid names are escaped before terminal output. - Never replace a real file, directory, foreign symlink, or live misdirected symlink in the Claude skill directory.
- Repair a dangling link only when its lexical target is inside the canonical Mosaic skills root.
- Unregister only a symlink pointing inside that root.
- Enumerate canonical directories at runtime; never hardcode framework skill names.
finalizeStage reconciles after wizard/framework synchronization, and
runFrameworkReseed reconciles after the sync-only mosaic update path. A
foreign conflict is reported but does not prevent unrelated canonical skills
from registering. Filesystem tests use injected temporary roots in
skill.spec.ts, finalize-skills.spec.ts, and update-checker.reseed.spec.ts.
M1 intentionally manages Claude Code only. Pi's Mosaic launcher can discover the canonical root directly. Codex still relies on the existing full skill-sync linker and needs separate parity analysis before this lifecycle API is extended.
Pi Persistent Goal Extension
The source of the Mosaic-owned Pi goal controller is:
packages/mosaic/framework/runtime/pi/goal-extension.ts
The framework manifest classifies runtime/** as framework-owned. Both the bash installer and the
TypeScript file adapter therefore deploy the same reviewed source to:
$MOSAIC_HOME/runtime/pi/goal-extension.ts
# default: ~/.config/mosaic/runtime/pi/goal-extension.ts
Do not copy or link this extension into ~/.pi/agent/extensions/. The launcher function
discoverPiExtensionArgs() emits the core mosaic-extension.ts first and the optional
goal-extension.ts second, preserving compatibility with an older installed framework that does
not have the goal file yet.
Lifecycle design
| Pi API | Goal-controller responsibility |
|---|---|
registerCommand('goal') |
Set, inspect, pause, resume, or cancel one branch-specific goal |
registerTool(...) |
Record a terminating structured progress report with evidence |
context |
Inject the active goal contract before every provider request |
turn_end |
Record every turn, reject mixed final reports, and enforce the turn bound |
agent_settled |
Start one deduplicated continuation only after Pi has no retry/compact/queue work |
session_compact |
Record the compact check, reset provisional verification, and defer idle work |
session_start/session_tree |
Rebuild state from custom entries on the active branch |
session_shutdown |
Invalidate deferred callbacks and clear UI state |
State is appended as mosaic-goal-state custom entries, which do not enter model context. The
context hook creates a fresh hidden mosaic-goal-context message for each request instead of
trusting compaction summaries. The mosaic_goal_report result uses terminate: true; when it is the
sole final tool call, Pi avoids an unnecessary model response before the controller decides whether
to verify, continue, or stop.
Before state is appended or displayed, the controller applies bounded credential-pattern redaction
to the goal statement, report summary/evidence/next step, and stop reason. Fingerprints are computed
over redacted report content. Pi session entries are append-only, so a credential-bearing legacy
entry cannot honestly be erased by the extension: restoration fails closed, emits a warning, and
requires removal of the affected session before setting a new goal. This is defense-in-depth rather
than a secret-storage contract, and it does not rewrite Pi's separate model-message/tool-call
history. Goal prompts tell the agent not to submit credentials or raw sensitive output, and tests use
canaries to prove known forms do not reach new custom entries, status text, context, or tool details
while ordinary typed fields such as token: string remain intact.
Completion remains evidence-gated but semantic: two consecutive achieved reports are required,
and the second run is explicitly a verification pass. This avoids an extra judge-model request after
every turn. Deterministic validator commands are intentionally not accepted as /goal input in this
slice, so never describe this mechanism as proof of arbitrary natural-language completion.
Tests and local smoke workflow
pnpm --filter @mosaicstack/mosaic exec vitest run \
src/runtime/pi-goal-extension.spec.ts \
src/commands/launch.spec.ts \
src/config/file-adapter.test.ts
bash packages/mosaic/framework/tools/quality/scripts/test-install-migration.sh
For an additive local smoke test without reseeding unrelated live framework files:
install -D -m 0644 \
packages/mosaic/framework/runtime/pi/goal-extension.ts \
~/.config/mosaic/runtime/pi/goal-extension.ts
pi --extension ~/.config/mosaic/runtime/pi/goal-extension.ts
Use /goal help, /goal set ..., and /goal status in that test session. A released framework
sync installs the file, and a released Mosaic CLI loads it automatically through mosaic pi.
API Endpoint Reference
All endpoints are served by the gateway at http://localhost:14242 by default.
Authentication
Authentication uses BetterAuth session cookies. The auth handler is mounted at
/api/auth/* via a Fastify low-level hook in
apps/gateway/src/auth/auth.controller.ts.
| Endpoint | Method | Description |
|---|---|---|
/api/auth/sign-in/email |
POST | Sign in with email/password |
/api/auth/sign-up/email |
POST | Register a new account |
/api/auth/sign-out |
POST | Sign out (clears session cookie) |
/api/auth/get-session |
GET | Returns the current session |
Chat
WebSocket namespace /chat (Socket.IO). Authentication via session cookie.
Events sent by the client:
| Event | Payload | Description |
|---|---|---|
message |
{ content, conversationId?, provider?, modelId? } |
Send a message |
Events emitted by the server:
| Event | Payload | Description |
|---|---|---|
token |
{ token, conversationId } |
Streaming token |
end |
{ conversationId } |
Stream complete |
error |
{ message } |
Error during streaming |
HTTP endpoints (apps/gateway/src/chat/chat.controller.ts):
| Endpoint | Method | Auth | Description |
|---|---|---|---|
/api/chat/conversations |
GET | User | List conversations |
/api/chat/conversations/:id/messages |
GET | User | Get messages for a conversation |
Admin
All admin endpoints require role = admin.
| Endpoint | Method | Description |
|---|---|---|
GET /api/admin/users |
GET | List all users |
GET /api/admin/users/:id |
GET | Get a single user |
POST /api/admin/users |
POST | Create a user |
PATCH /api/admin/users/:id/role |
PATCH | Update user role |
POST /api/admin/users/:id/ban |
POST | Ban a user |
POST /api/admin/users/:id/unban |
POST | Unban a user |
DELETE /api/admin/users/:id |
DELETE | Delete a user |
GET /api/admin/health |
GET | System health status |
Agent / Providers
| Endpoint | Method | Auth | Description |
|---|---|---|---|
GET /api/agent/providers |
GET | User | List all providers and their models |
GET /api/agent/providers/models |
GET | User | List available models |
POST /api/agent/providers/:id/test |
POST | User | Test provider connectivity |
Projects / Brain
| Endpoint | Method | Auth | Description |
|---|---|---|---|
GET /api/brain/projects |
GET | User | List projects |
POST /api/brain/projects |
POST | User | Create a project |
GET /api/brain/projects/:id |
GET | User | Get a project |
PATCH /api/brain/projects/:id |
PATCH | User | Update a project |
DELETE /api/brain/projects/:id |
DELETE | User | Delete a project |
GET /api/brain/tasks |
GET | User | List tasks |
POST /api/brain/tasks |
POST | User | Create a task |
GET /api/brain/tasks/:id |
GET | User | Get a task |
PATCH /api/brain/tasks/:id |
PATCH | User | Update a task |
DELETE /api/brain/tasks/:id |
DELETE | User | Delete a task |
Memory / Preferences
| Endpoint | Method | Auth | Description |
|---|---|---|---|
GET /api/memory/preferences |
GET | User | Get user preferences |
PUT /api/memory/preferences |
PUT | User | Upsert a preference |
MCP Server (Gateway-side)
| Endpoint | Method | Auth | Description |
|---|---|---|---|
POST /mcp |
POST | User (session cookie or Authorization header) | MCP streamable HTTP transport |
GET /mcp |
GET | User | MCP SSE stream reconnect |
Skills
| Endpoint | Method | Auth | Description |
|---|---|---|---|
GET /api/skills |
GET | User | List installed skills |
POST /api/skills |
POST | Admin | Install a skill |
PATCH /api/skills/:id |
PATCH | Admin | Update a skill |
DELETE /api/skills/:id |
DELETE | Admin | Remove a skill |
Coord (Mission Coordination)
| Endpoint | Method | Auth | Description |
|---|---|---|---|
GET /api/coord/missions |
GET | User | List missions |
POST /api/coord/missions |
POST | User | Create a mission |
GET /api/coord/missions/:id |
GET | User | Get a mission |
PATCH /api/coord/missions/:id |
PATCH | User | Update a mission |
Observability
OpenTelemetry traces are exported to the OTEL collector (OTEL_EXPORTER_OTLP_ENDPOINT).
View traces in Jaeger at http://localhost:16686.
Tracing is initialized before NestJS bootstrap in
apps/gateway/src/tracing.ts. The import order in apps/gateway/src/main.ts
is intentional: import './tracing.js' must come before any NestJS imports.