- name
- nocobase-dsl-reconciler
- description
- NocoBase 2 only; never use in a NocoBase 3 project. **Opt-in DSL path** for NocoBase app building. Use ONLY when the user explicitly asks for YAML / DSL / committed-to-git / `cli push` / spec files — e.g. "use the DSL reconciler", "I want YAML I can commit", "build this as a workspaces/ project". For any other UI authoring request (new page, new block, tweak an existing screen), default to `nocobase-ui-builder` instead — this reconciler is still in active development and has rough edges that the live-UI path avoids.
When the user opts in: produces/changes files under `workspaces/<project>/`, supports new pages, menus, modules, whole systems, collections, tables, sub-tables, popups, dashboards, approval workflows, recordActions, and deploys them via `cli push`.
# NocoBase Application Builder (DSL path)
## Before you use this skill
This is the **opt-in DSL path**. Default is `nocobase-ui-builder`.
Stay on this skill only when the user explicitly wants YAML files they
can commit + `cli push`. If you arrived here from a generic "build me a
NocoBase app" request without the user naming DSL/YAML/git, switch to
`nocobase-ui-builder` instead — it's the default entry point.
## Golden rule
`templates/crm/` is a **read-only reference library**. When you're
unsure how a specific field, block, or popup is shaped, open the
closest CRM example, read it, then write your own adapted version in
your workspace. Per-scenario pointers live in the workflow sections below.
**Never** copy CRM files wholesale into your workspace. Do not
`cp -r templates/crm/...` to get started, do not duplicate
`collections/nb_crm_*.yaml`, do not base your `routes.yaml` on CRM's.
Bulk-copying drags unrelated leads/opportunities/orders state and
workflows into your project — you then spend the whole session fighting
hundreds of irrelevant validator errors instead of building your module.
The pre-deploy spec validator catches most structural mistakes with a
clear error message. **Trust the validator**: when it errors, fix what
it says rather than guessing — don't grep through `src/deploy/*.ts`.
## When NOT to use this skill
Hand off to the matching skill when the user's request is orthogonal:
| User asks for… | Skill |
|---|---|
| One-off live-UI tweak on an already-running page (move / reorder / reconfigure a single block, field, action) — **no DSL commit wanted** | `nocobase-ui-builder` |
| ACL / role permissions / route permissions | `nocobase-acl-manage` |
| Workflow create / update / revision / execution | `nocobase-workflow-manage` (this skill only wires the trigger button; authoring the graph goes there) |
| Collection / field / relation authoring outside a DSL project | `nocobase-data-modeling` |
| Plugin development (`.tsx` components, server code) | `nocobase-plugin-development` |
| Install / enable plugin | `nocobase-plugin-manage` |
| Environment setup / app install / upgrade | `nocobase-env-manage` |
Any change that should live as a committed YAML file under
`workspaces/<project>/` — stays here.
## Environment
```bash
cd <skill-dir>/src
export NB_USER=admin@nocobase.com NB_PASSWORD=admin123 NB_URL=http://localhost:14000
```
## Quick start — new build
**Default path for "build me a NocoBase app"**: copy the starter and
modify it. Do not hand-write the skeleton; do not study CRM first.
```bash
cd <skill-dir>/src
cp -r ../templates/starter ../workspaces/<name>
# First push: --copy bypasses validation rules that only matter once
# popups exist (m2o popup binding, clickToOpen file presence). The
# starter ships with its own popups so the first push is actually fully
# valid — --copy is for your *extensions* before popups are wired.
npx tsx cli/cli.ts push <name> --force
```
The starter is a complete minimal CRUD — 1 collection (Projects),
1 Dashboard page (4 KPI tiles + 2 charts), 1 list page with
filterForm / table / addNew popup / detail popup / 2 updateRecord
recordActions. Push as-is → visible in NB → then edit.
### Customizing the starter (the agile loop)
Iterate one concern at a time, push between each:
1. **Rename identifiers** to match the user's domain.
- `collections/nb_starter_projects.yaml` → your collection name
(match the `nb_<module>_<entity>` convention)
- `routes.yaml`: change "Starter" / "Projects" titles
- `pages/starter/` directory name if you want (match `routes.yaml`
key)
- Find-replace `nb_starter_projects` across all files
2. **Adjust the field list** in your collection YAML to match the
user's entity. Update `pages/<...>/layout.yaml` `fields:` and the
popup templates' `field_layout:` accordingly.
3. **Add a 2nd entity** only when the first one works end-to-end.
Create a new `collections/*.yaml` + `pages/<module>/<entity>/` dir,
copy `pages/starter/projects/layout.yaml` as a starting point.
4. **Extend incrementally**: add a tab, a chart, a workflow trigger.
Push after every change. See "Advanced patterns" below for which
CRM file matches which pattern.
**Never write the whole module in one shot.** For customer-facing
builds — land the skeleton, show the user, gather feedback, iterate.
The starter push takes minutes; a hand-built module takes hours.
### Fast-track: when `--copy` helps
Pass `--copy` when the workspace has **no popup files yet** (early
stage — validator would fire errors about "m2o field X has no popup
binding" that the user will fix in the next push). The reconciler
auto-bypasses spec errors in this state. Once any `popups/*.yaml`
exists, drop `--copy` and let validation run.
## Incremental edits — existing workspace
- **Add** a block / field / action / popup → write the DSL → push.
- **Remove** from DSL → push. The reconciler destroys the matching
live model on the NB side and cleans `state.yaml`. Manual NB-UI
authored elements (not tracked in `state.yaml`) are left alone.
- **Rename** → not supported automatically. Delete + re-add.
Targeted pushes:
- `--group <key>` scopes to one menu subtree
- `--page <key>` scopes to one page
- `--incremental` skips pages whose DSL hasn't changed since last push
**For pure live-UI tweaks without a DSL commit, hand off to
`nocobase-ui-builder` instead** (see the routing table above).
## Advanced workflow — when the starter isn't enough
Triggers for going beyond the agile starter loop:
- More than ~3 collections with cross-relations (m2m, tree structures)
- Dedicated workflow / approval / permission logic that the user wants
designed up-front
- Multi-tab pages, sub-tables, or cross-module navigation
- Dashboard with bespoke KPIs mapped to the user's domain language
Progression: Round 0 design → Round 1 scaffold → Round 2 fill →
Round 2' seed → Round 3 JS. Each round is a deployable state.
### Round 0: System architecture — confirm with user
Write a `DESIGN.md` (markdown, not YAML) covering:
1. **Collections** — every table, its fields, and its relations.
See `nocobase-data-modeling` skill for field-interface reference.
2. **Page list** — every page, one-line purpose each, grouped by menu.
3. **Navigation wiring** — which m2o fields open which popup
templates; which pages link to each other.
Wait for user confirmation before writing YAML. A single design pass
saves 3× redesigns.
Skip Round 0 if the user's ask fits the starter shape (single entity,
basic CRUD).
Example:
```markdown
## Collections
- nb_lib_books (title, author, isbn, category, status, loans: o2m → nb_lib_loans)
- nb_lib_members (name, email, phone, join_date, loans: o2m → nb_lib_loans)
- nb_lib_loans (loan_no, book: m2o, member: m2o, borrowed_at, due_date, returned_at, status)
## Pages (under menu "Library")
- Books list — browse + search books, add new
- Members list — browse members, their loan history
- Loans list — active/overdue loans, return action
- Dashboard — KPIs + charts
## Navigation
- books.table.title → books detail popup (shared template)
- loans.table.book → books detail popup (shared via defaults.yaml)
- loans.table.member → members detail popup (shared via defaults.yaml)
```
### Round 0.5: Sub-agent CWD (only when spawning)
If launching a sub-agent (kimi TUI, Claude Code subprocess), its
CWD becomes the default write target. **Set it before launch:**
```bash
mkdir -p <user-workdir>
cd <user-workdir>
kimi --yolo # or claude, codex
```
Skip the `cd` and the agent writes to the parent project root.
### Round 1: Scaffold — still start from the starter
Even in the advanced path, don't hand-write `routes.yaml` +
`collections/*.yaml` from scratch. Copy the starter, then grow:
| Step | What to do |
|---|---|
| Base | `cp -r templates/starter workspaces/<name>` — push once |
| Add collection | Write `collections/<next_coll>.yaml` (format matches starter's). Match `nb_<module>_<entity>` convention. |
| Add page | `mkdir pages/<module>/<page>/` + `layout.yaml`; mirror starter's projects layout. |
| Update routes | Add entry under the existing group in `routes.yaml`. |
| Deploy | `cli push <name> --force` after each collection/page addition |
CRM references (consult only when stuck on structure):
- `templates/crm/collections/nb_crm_leads.yaml` — collection format
- `templates/crm/routes.yaml` — multi-group routes (shape only)
### Round 2: Fill content — blocks, popups, templates
For each page beyond the starter basics:
| Building | CRM reference |
|---|---|
| Main list table + filter | `templates/crm/pages/main/leads/layout.yaml` |
| Multi-tab page | `templates/crm/pages/main/customers/` (`page.yaml` + `tab_*/layout.yaml`) |
| Create-form with inline sub-table for o2m children | `templates/crm/templates/block/form_add_new_opportunities_quotations_quotations.yaml` — `items` is an o2m field listed in `fields:` and rendered as an inline editable sub-table. Also `templates/crm/pages/main/products/` for master/child UX. |
| Detail-popup template | `templates/crm/templates/popup/activity_view.yaml` |
| m2o auto-popup bindings | `templates/crm/defaults.yaml` |
| Parent-detail + child-list popup | `templates/crm/pages/main/customers/tab_customers/popups/` |
| addNew + field click-popup pattern | `templates/crm/pages/main/leads/popups/` |
Rules for copying from CRM:
- Copy 10–30 lines, adapt names. **Never copy whole files.**
- Don't copy `uid:` / `targetUid:` / `route_id:` — deployer assigns fresh.
- For every m2o field displayed in a table, either set
`clickToOpen: templates/popup/popup_detail_<target>.yaml` OR add
`popups.<target>: ...` in `defaults.yaml`. Validator errors otherwise.
#### Per-row actions (`recordActions`)
By default a table's row-action column is empty — NB won't render
any action buttons unless `recordActions:` lists them. "Just edit
and delete" is a common but weak default: for most list tables the
user actually wants a **one-click state change** (Mark Done,
Approve, Archive) sitting next to edit.
Decision order:
1. **Is there a boolean / enum status field?** → add `updateRecord`
with `linkageRules` to show the matching button only when the
record is in the right state. One button per state transition
(Mark Done hidden when already done; Reopen hidden when not done).
2. **Does the record need a second detail/form view different from
the default edit popup?** → `popup` with `templateRef`.
3. **Tree/hierarchy collection?** → add `addChild`.
4. **Need to navigate elsewhere with this record's id/filter?** →
`link` with `url: /admin/...?filter={{ctx.record.id}}`.
5. **Want to let the user clone a complex record?** → `duplicate`.
6. **Need to start a workflow manually?** → `workflowTrigger`.
**Prefer `updateRecord + linkageRules` over custom JS buttons** for
state changes. linkageRules covers 80% of row-level UX (conditional
show/hide, field-based gating, role checks via `ctx.user`) without
touching JS. JS actions are only needed for multi-step logic
(query, then update, then navigate) and are not currently
deployable by this reconciler.
Full per-row action palette (declared in `recordActions:`):
| DSL `type` | Purpose | CRM reference |
|---|---|---|
| `edit` | Edit popup (default shape) | many |
| `view` | Read-only detail popup | `templates/crm/templates/popup/opportunity_view.yaml` |
| `delete` | Single-row delete with confirm | many |
| `updateRecord` | Assign fields + optional linkageRules — the state-change workhorse | `templates/crm/pages/main/overview/layout.yaml` (Done / Undone pair) |
| `popup` | Open a custom popup with `templateRef` (form/detail different from edit) | `templates/crm/pages/main/leads/popups/table.name.yaml` |
| `link` | Navigate to another admin page, carrying record context in URL | `templates/crm/pages/lookup/layout.yaml` |
| `addChild` | Tree collection: add a child node under this row | `templates/crm/pages/main/products/tab_categories/layout.yaml` |
| `duplicate` | Clone the record into a new form | — |
| `workflowTrigger` | Manually trigger a workflow on this record | (toolbar in `templates/crm/pages/main/customers/tab_customers/layout.yaml` — same shape works per-row) |
| `historyExpand` / `historyCollapse` | Record-history plugin inline expand | — |
| `ai` | AI employee button (tasks_file + employee) | `templates/crm/pages/main/leads/layout.yaml` (toolbar — same shape per-row) |
Toolbar actions (declared in block-level `actions:`) use the same
type names plus `filter`, `refresh`, `addNew`, `bulkDelete`,
`export`, `import`. Not every type makes sense in both contexts —
`updateRecord` as a toolbar action would apply to *no* specific row,
so put it in `recordActions`.
**Goal of Round 2**: all pages have working CRUD — add / edit / view
popups wired correctly, row-action columns reflect real per-record
operations (not just edit+delete). Validator clean, NB UI shows no
"Collection may have been deleted" banners.
### Round 2': Test data (parallel with Round 2)
Can run concurrently with page-filling. Insert data via API:
```bash
TOKEN=$(curl -sS -X POST $NB_URL/api/auth:signIn \
-H 'Content-Type: application/json' -H 'X-Authenticator: basic' \
-d '{"account":"'$NB_USER'","password":"'$NB_PASSWORD'"}' \
| python3 -c 'import json,sys;print(json.load(sys.stdin)["data"]["token"])')
# Always GET existing record IDs first — they're snowflake integers
# (e.g. 359571523764224), NEVER 1/2/3.
curl -sS -X POST $NB_URL/api/<collection>:create -H "Authorization: Bearer $TOKEN" \
-H 'Content-Type: application/json' -d '{...fields..., "owner":{"id": <real-user-id>}}'
```
Parent tables first; fill every FK on children (leaving it null
orphans the row). Then:
```bash
npx tsx cli/cli.ts verify-data <name> # FK & completeness check
```
Why parallel: Round 3 JS (charts, KPIs) needs data to render
anything. Start the seed once you have collections (end of Round 1);
by the time Round 2 finishes pages, you have records to test against.
### Round 3: JS — where CRM uses it, you probably need it
Now that CRUD + data work, audit where JavaScript adds value. **Walk the
CRM template and ask "does the CRM have JS here?" for each spot in your
project.** Three typical JS opportunities:
| Spot | CRM has JS? | Your project likely needs JS if... |
|---|---|---|
| **Field renderer / column** (e.g. color-coded status tag, days-until-due badge) | ✅ in most list tables | Any field whose display depends on a derived value (date math, status-to-color, multi-field composite) |
| **Block** (KPI card, custom widget inside a form) | ✅ overview, analytics, per-form tips | A summary widget, inline chart, or "helper panel" that reads from multiple collections |
View on GitHub