We are starting to put too many services in common folder. Version and
step building should be separated into different services and go to the
builder folder. Today builder folder only manage schema.
We should:
- keep services responsible for only one action
- keep modules based on the actual action these provide rather than
having common module
This PR:
- creates a service for workflow version builder
- moves version and step builders to workflow builder folder rather than
commun
- creates separated folders for schema, version and steps
No logic has been added. Only modules created and functions moved.
- create form action
- add `pendingEvent` to executor output
- when receiving pendingEvent, exit workflow execution
- let the workflow in running status for now before taking a decision
Actions will now:
- receive the complete input
- get the step they want to execute by themself
- check that the type is the right one
- resolve variables
These all share a common executor interface.
It will allow for actions with a special execution process (forms, loop,
router) to have all required informations.
Main workflow executor should:
- find the right executor to call for current step
- store the output and context from step execution
- call next step index
Should be possible to end workflow, not matter what the current status
is.
On failure before the workflow was started, this error prevent the
workflow to be marked as failed with the right error message.
We need the version trigger and steps to be stored in the output. We
should not rely on the version itself because some run are made on draft
versions. Which means versions could be edited afterwards.
Close
https://discord.com/channels/1130383047699738754/1334441759484149793
Using refetch queries was not working for certain use cases.
To find manual active workflows in cmd+k, we use a query with specific
filters that was complicated to refetch.
Finally I will update the cache manually. It was not properly updated
before because the json value of the version trigger was stringified
without spaces. So the entity was not found in apollo cache.
Refresh of `objectmetadataitems` was not happening fast enough. Page was
breaking when enabling the feature flag. Instead of not storing worklow
objects in state, we will use the feature flag to block on read. This
way we avoid race conditions
<img width="1511" alt="Capture d’écran 2025-02-04 à 14 11 56"
src="https://github.com/user-attachments/assets/912cc59a-f422-48ab-84b7-7fdd7bbc35c1"
/>
Workflow version statuses are too complex to handle in cache. You always
need to:
- modify entity in cache
- modify queries in cache
- do the same for all other version statuses that should also be updated
This complex behavior leads to bugs because of a cache wrongly updated.
Let's simply do a refetch versions on status update.
We create draft without emitting creation event. So workflow statuses
are not properly updated
This service needs a refacto. Will do it once we get out from fast
follows rush
- Sync metadata to create workflow entities, since those are not behind
a flag anymore
- Seed workflow views
- Seed workspace favorite for workflow
- Put all steps in upgrade command
From now on workflow entities and views will be seed for every new
workspace. What will prevent user to see those is the feature flag used
in frontend. It will prevent workflow objects to be stored in the recoil
state.
Without feature flag, workflows will:
- remain invisible in metadata
- not be accessible through views or show page
- remain invisible on side menu
We were using metadata client by legacy. Architecture is not great for
Core engine: workflows are available both in data and metadata client.
It makes more sense to use the data client since workflows are part of
standard objects
After hitting use as draft, we redirect the user to the workflow page.
If the user already accessed that page, the workflow and its current
version will be stored in cache. So we also need to update that cache.
I tried to update it manually but it was more complex than expected.
Steps and trigger are not fully defined objects.
I ended with a simple refetch query that I wanted to avoid but that is
at least fully working with minimum logic.
- catch error on action execution. We will log the error and return it
in the step
- catch error on workflow run
- remove the catch in the action. All actions should simply throw and
let the executor do the job
<img width="1512" alt="Capture d’écran 2025-01-14 à 17 35 53"
src="https://github.com/user-attachments/assets/dcf79567-a309-45f1-a640-c50b7ac4769b"
/>
When a workflow was activate, the previous activated version was still
available in cmd+k
This is because we were only updating entities in cache after
activation.
We also need to update the queries stored in cache, because this is
where the cmd+k look to know which workflow is active.
https://github.com/user-attachments/assets/750cc24a-7583-4641-856c-0f9892d26331
When switching between the same step detail component, even if the props
is updated, component is not re-created.
This is an issue since form fields often keep an internal state. Which
means that even if form data are updated, the displayed value may still
be the value of the previous step detail.
Adding a key to for re-creation.
We want to avoid infinite loops using workflows. Adding a throttler with
a limit of 10 executions / sec by default for each workflow.
We were not emitting events on workflow actions so loops could not
happen. Since throttler is there we can now and these.
Adding an error message so the user knows when it happens.
<img width="1284" alt="Capture d’écran 2024-12-27 à 17 05 20"
src="https://github.com/user-attachments/assets/dafa837b-5b4c-48be-8207-c90f5c71a236"
/>
Delete all workflow sub objects when workflow is deleted. Other sub
objects cannot be deleted otherwise.
We do not listen to deletion events so I am not adding them. Those post
hooks should be deleted Q1 once we properly handle cascade for soft
deletion
Actions using tiptap trigger updates when called.
Flow is following:
- component using tiptap receive default value
- we separate the default value because we need to add specific
attributes to the editor when inserting breaks or variables
- editor performs an update for each value
It means that initialize our editor performs several updates. We need to
avoid persisting until the init finished.
- composite field need to be formatted before being saved
- repository.create() does not do it. So we simply lose the composite
fields on the way
- save() does it directly and doing create() before does not change
anything
Having a global record crud action adds complex logic.
We decided to split those actions. I only kept a common folder / module
in backend.
⚠️ this may break existing workflows if these were using previous
actions!
Output schema is now separated in two sections:
- object, that gather all informations on the selectable object
- fields, that display object fields in a record context, or simply the
available fields from the previous steps
The dropdown variable has now a new mode:
- if objectNameSingularToSelect is defined, it goes into an object mode.
Only objects of the right type will be shown
- if not set, it will use the already existing mode, to select a field
When an object is selected, it actually set the id of the object
https://github.com/user-attachments/assets/1c95f8fd-10f0-4c1c-aeb7-c7d847e89536
- Rename all parts using the name "relation" to "record" when component
is only selecting record
- Remove the use of scope states in folder
- Rename entities to records
This PR prepares the use of the record picker in workflows
Steps to enable workflows:
- enable feature flags
- run metadata sync
- run this command
If the feature flag is not true, the command will fail. Which will be
useful to prevent seeding a wrong workspace.
Including a little fix for send email aciton error
- add `inputSchema` column in serverless function. This is an array of
parameters, with their name and type
- on serverless function id update, get the `inputSchema` + store empty
settings in step
- from step settings, build the form
TODO in next PR:
- use field type to decide what kind of form should be printed
- have a strategy to handle object as input
https://github.com/user-attachments/assets/ed96f919-24b5-4baf-a051-31f76f45e575
Tiptap uses non breaking spaces between nodes (like variables). Those
html characters are not properly handles in emails. Replacing by regular
spaces during parsing.
I tried to fix it in settings but looks like this is only for preserving
those nbsp and not for removal (see
https://github.com/ueberdosis/tiptap/pull/254)
Expected behavior:
- workflows can be added and deleted. Only name field is editable
- versions and runs cannot be added nor deleted. No fields are editable
Added two new utils for those needs:
- `isReadOnlyObject` the similar logic between remote objects, versions
and runs
- `isFieldReadonlyFromObjectMetadataName` to easily block field edition
from object context
When object is not part of the workspace favorite list, we want to show
it in the "opened section" while its record page is accessed.
This PR:
- adds a new component `NavigationDrawerOpenedSection`
- makes workflow versions and runs not system object + creates a
prefilled view index for these
- do not create workspace favorites for these so these do not appear in
the workspace section
<img width="1129" alt="Capture d’écran 2024-09-26 à 11 45 25"
src="https://github.com/user-attachments/assets/c84d773c-0bef-4dce-b66a-55d7d00b0fb6">
Thomas Trompette