8.4 KiB
Architecture — twenty-partners
Vertical-slice modules, one folder per domain. Adapted from twentyhq/twenty-eng
(the largest Twenty SDK app), which migrated flat → modular as it grew. Repo-wide
conventions (kebab-case files, named exports, no any, types over interface,
short // comments) live in the root CLAUDE.md and still apply — this file only
adds the partners-specific structure.
Status: migration complete — every domain lives under
src/modules/. The only top-levelsrc/entries left areconstants/universal-identifiers.ts,scripts/,roles/,skills/,workflows/,__tests__/, and the two root config files. Add new code to a domain module, never a flat top-levelsrc/<primitive>/folder.
Layout
src/
application-config.ts # app + application variables — unchanged
default-role.ts # unchanged
constants/universal-identifiers.ts # ⚠️ MUST NOT MOVE (see Invariants)
scripts/ # ⚠️ MUST NOT MOVE — seed/purge/rls/slugify
roles/ # app-wide RLS — stays at root (spans every object)
skills/ # bundled Claude skills — unchanged
workflows/ # bundled workflows — unchanged
__tests__/helpers/ # shared test setup (client, fixtures, cleanup)
modules/
shared/ # cross-cutting: secret-guard (http/), find-or-create, paginate,
# utils, front-components, cross-domain nav folders
<domain>/ # partner · application · opportunity · … (large domains add a <feature>/ tier)
<name>.logic-function.ts # thin SDK entrypoints — sit at the domain/feature root, NOT in a subfolder
objects/ fields/ view-fields/ views/ navigation-menu-items/ page-layouts/ # declarations
constants/ # value lists + declaration UUID maps (NOT types)
services/ # business logic (testable)
graphql/{queries,mutations}/ # the only place raw queries/mutations live
connector/ # OUTBOUND third-party API per folder (<name>.connector.ts · config · types)
mappers/ utils/ types/ front-components/
A domain is a self-contained slice: its declarations (objects, fields, views, …)
and its logic (logic-functions → services → graphql/connector) live together.
Adding a feature means adding a folder, not scattering files across ten flat piles.
Feature tier for large domains: when a domain holds several distinct features, insert a
<feature>/ level — modules/<domain>/<feature>/<primitive>/ (the twenty-eng standard,
e.g. modules/code-build/build-task/services/). It keeps graphql/ and services/ from
becoming flat piles. Partner uses partner/{self-service,marketplace,application-intake, directory}/…; opportunity uses opportunity/{intake,matching}/…. A small single-feature
domain (application) may keep <domain>/<primitive>/ directly.
The dependency rule (the one that matters — never import upward)
logic-functions → services → { graphql, connector, mappers } → shared · utils · types
- Entrypoints (
*.logic-function.ts, at the domain/feature root) — SDK entrypoint ONLY. Parse input, call one service, return. No business logic, no raw GraphQL, no external HTTP. >~40 lines means it's doing a service's job — extract it. services/— all business logic. Testable, no SDK/transport coupling. Must not import alogic-function.graphql/— the ONLY place raw queries/mutations live (named typed operations, not inlineclient.query(...)).connector/— the ONLY place outbound third-party APIs are called; one folder per API (<name>.connector.ts/config.ts/types.ts). Inbound webhooks are NOT connectors — a webhook the app receives is an ordinarylogic-functions/entrypoint guarded by the shared secret-guard util. (Today the one real connector is Discord; TFT/client-brief/ partner-application are inbound.)- Never import logic sideways between domains — share via
modules/shared/. (A relation field naturally referencing its target object's ID constant is a schema reference, not a logic dependency — that cross-module import is allowed.)
Naming
<name>.<primitive>.ts — e.g. submit-application.logic-function.ts,
resolve-candidacy.service.ts, partner.object.ts. Files are kebab-case (no PascalCase,
even for React components: profile-picture-upload.tsx).
Tests co-locate with their subject and split by kind: <name>.test.ts = unit (mocked
client, fast, no infra); <name>.integration-test.ts = real workspace (needs a live
server + global setup; excluded by tsconfig). The split is a selectable vitest project
(--project unit), not a file-count mandate — one vitest.config.ts with two projects is
preferred. Don't rename a mocked unit test to .integration-test.ts. Shared test setup
lives in src/__tests__/helpers/.
Invariants (moving code must not break these)
- The set of
universalIdentifierUUIDs must stay byte-identical across a move — the SDK tracks primitives by identifier, not path, so a puregit mvis safe but a changed/dropped UUID re-registers or orphans an object. - Do not move
src/constants/universal-identifiers.ts— the env-toolkit rewrites it by path per bundle. - Do not move
src/scripts/*—package.jsoninvokes them by path. - Keep importers of
src/constants/universal-identifiers.tspointing at the root file — don't relocate its UUID exports into a module. - Never split a relation across domains with a dangling ID import. Paired relation
fields export/import each other's field-ID constant; keep both sides of a relation in one
domain, or hoist the shared ID into
shared/. - One domain per commit when migrating (
git mvonly, imports fixed in the same commit).
Application review playbook
The Matching Admin Workspace playbook is a standalone page in
src/modules/partner/application-intake/ (front-component + page layout +
nav item). It must stay in lockstep with real behaviour. When you change any
of the following, update the playbook in the same change:
- Website form fields that land on Partner
- The Partner Applications view (filter, columns, sort)
validationStagevalues used as the review decision- Intake side effects the reviewer needs to know (Discord on create;
welcome email on
VALIDATED, linked fromWELCOME_EMAIL_WORKFLOW_URL) - Marketplace ranking bonuses in
packages/twenty-website/src/partners-marketplace/completeness-score.ts(case study, cover, introduction, service, picture, calendar, rate, category, ghost rule)
Facts on the page come from the same constants the view and intake use. Do not leave stale steps. Do not keep a second copy on twenty.com or in a Note.
Opportunity process playbooks
How to process (Matching Admin) and How to apply (Partner Workspace) are
standalone pages under src/modules/opportunity/how-to-process/. They must
stay in lockstep with real operator behaviour. Update the matching page and
its copy test in the same change when you change any of:
- Intake (marketplace brief already Listed + Discord; Twenty Internal
unlisted;
/twenty-lead-briefwrites Design Doc URL) - Listing (
Listedflip → Discord; daily digest; Open Briefs membership) - Intro (
Intro Sent Atturns Listed off; Invited applications) - Win (set Partner → Won / Declined; Backup stays; unassign reopens Won to Applied; Deals stages New → Screening → Meeting → Proposal → Customer)
- Apply (Open Briefs listed-only; pitch minimum; one pitch; state meaning)
- Bundled skills (
twenty-lead-brief,twenty-partner-shortlist,twenty-partner-intro): local run, expected output, GitHub URL ofSKILL.md - Match via MCP: the operator can ask the agent to write CRM changes
Pages stay operator language. Skill names, slash triggers, and GitHub URLs of those bundled skills are allowed on How to process. Open Briefs and My Applications links are required on How to apply. Do not open Apply from the standalone page — Apply is pinned on the brief record.
Where does new code go?
External API call → connector/. Raw query/mutation → graphql/. Any logic →
services/. New SDK trigger/function → a thin *.logic-function.ts at the domain/feature
root that calls a service. Shared across domains → modules/shared/.