The fields sat in a grey box as blue chips, with a pencil on every one
of them and values printed as the constants they are stored as.
A field is now a quiet white card that lifts on hover, its label small
and grey above the value, and its pencil appears only when the field is
pointed at. A value that is one of a fixed set - CATALOG, PER_EXECUTION
- reads as words on a tinted plate, so it is clearly a choice and not
typed text; an empty one says so in grey rather than as a bare dash.
The dialog behind the pencil is roomier, with a focus ring on the
control and the same words in its dropdown.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two pictures for one action: a node expanded with the fullscreen arrows
and closed with their inward twin, a resource with the diagonal arrows
and a plain ×. The node and the container now use the resource's pair.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The badge and the kind under the name already say what sort of step it
is; the stripe only added weight. A resource card keeps its own.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The saturated blue (or amber) header bar gives way to a light one: a
tinted badge with the step's icon, its name in dark type with its kind
under it in the step's colour, and quiet borderless buttons that colour
on hover. The kind of step stays in a stripe on the left, as on a
resource card.
A step's icon now follows its kind - automated, human, decision, end,
branches rejoin - with the pictures the palette already uses for them,
instead of one raster image for every automated block. A schema that
names an icon still wins.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Like a block, a resource can now be lifted out of the canvas into a
panel at full size, with every field open, whatever the zoom - an init
script above all is unreadable in a card on a zoomed-out canvas. The
settings button still opens the fields in place for a quick look.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Its links now leave to the right from the bar's middle, as a step's
outputs do, and come into the step's handle from the left.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A small square on the edge was hard to see and hard to grab, and Rete
anchored its links 12px to one side of it - the offset it gives an
output on a node's right. The handle now runs along the whole top edge,
and its links start from the bar's middle.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A resource looked like a step: a header bar, a "Used by / Steps" row
with the handle on the right, every field open. It is now a card - its
type's icon, its name, one line saying where the store is - with the
fields behind a settings button and its handle on the top edge. Its
links leave that handle vertically and come into the step from the
left, so they read as something the steps hang from rather than one
more connection in the chain.
Saved links to a resource were not drawn at all when the resource
types had not been fetched yet - which was always, unless the
Resources palette had been opened - because the link was checked
against a kind nobody knew. The editor now fetches the resource types
before drawing, and puts saved links back without checking them again:
the server has.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The resource palette now lists an S3 storage and a PostgreSQL database
as two cards, each with the label and icon its type declares, and the
node on the canvas takes the same. A step's resource field names a
kind, so a storage operation links to either store.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The resource node inherited the generic node's schema lookup, which asks
the block catalog. Storage is not a block, so the lookup found nothing:
the node showed no fields, and every click on it fetched the catalog
again and covered it with the loading overlay, so nothing inside could
be clicked.
The lookup is now a hook the resource node overrides with the resource
catalog, and a resource is never recreated on the server - it has no
ports for the server to recompute.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
They were squeezed into the blocks palette as one more kind of block,
which is exactly the confusion the model change was meant to end. The
sidebar now has a third entry with its own icon and its own list, and
the blocks palette knows nothing about resources again.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A storage was a block with a visual role that said "not a step", which
left resource-shaped exceptions scattered through the editor, the node
components, the palette and the execution view.
The flow now carries resources beside its blocks, and the canvas draws
them with a component of their own: no ports, no dependencies, one
handle for the steps that use it. What a step links is read from the
schema flag x-resource-ref, so the editor no longer knows the word
"storage" anywhere - a new kind of resource needs no code here.
The configuration form the two kinds of node share is now a component of
its own (app-node-parameters), so a resource is edited exactly as a
block while it is drawn differently.
Resource types come from /resources/types, and share the palette with
the blocks under the Resource kind.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A Storage node has a 'Used by' handle and a storage operation a 'Works
on' one; the link drawn between them, teal and dotted, is what sets the
operation's storageNode. It is neither a connection nor a dependency: it
carries no value and orders nothing, and loops leave it out. An
operation works on one node, so a new link replaces the old one, and
removing the link empties the field. On load the links are drawn back
from what each operation says.
Fields the schema marks x-ui-hidden are left out of every form and chip
- storageNode is set by the link, not typed - and kept in the
configuration.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The chips took four lines of a narrow sidebar. A filter button at the
end of the search field opens the kinds as checkboxes instead, and
shows how many are picked; its tooltip says which. The menu lives in
the overlay, so its font is set globally, where the palette's own
styles do not reach. A storage's icon is 'storage': 'database' is not
in the icon font, and showed nothing.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A row of chips under the search, one for each kind the palette has -
automated and human steps, decisions, rejoins, ends, resources - in a
fixed order. Picking some narrows the list to them, together with the
search; picking none shows every block, and Clear goes back to that.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The palette knew two kinds of block, a human step and an automated one,
so a Storage node, an End and a routing block all read 'Automated step'.
It now goes by the role the block's type declares: a resource that is
not a step, an end of a path, an automated or a human decision, a rejoin
of branches, a human step, an automated step - each with its own icon.
The server's BRANCH_REJOIN role is read as the web's MERGE, which it had
never matched, so a rejoin was taken for an ordinary activity.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The editor a node opens for a single field offered Use default on every
optional field that was not a checkbox: a storage's connection, its
operation, its scope, and any other choice between alternatives, where
there is nothing to go back to. The settings dialog already followed the
schema's x-ui-defaults-when-empty; this editor now does too, so the
button is on the model's sampling knobs and nowhere else.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A storage node comes from the server with the RESOURCE role. In the
editor it is drawn apart from the steps - dashed, teal, labelled
'Storage · not a step' - and its targets take a connection from every
step that writes there; a view parameter (<view>.<parameter>) still
takes one.
Its edges are left out when the editor works out which connections lead
back round a loop: a write followed by a read of the same node is not a
cycle, the same view the server takes.
The execution view draws the node from the flow, since it never has a
step, with no status of its own, so the reads and writes on its edges
still have somewhere to go.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A storage step on the runner's own connection needs it before the
execution starts. The gate now takes any requirement the server marks as
paid from the vault (vaultBacked), with the key prefixes as a fallback,
so the S3 or PostgreSQL connection shows up beside the LLM credentials.
Adding one opens a form with the storage type's own fields, from
/storage/types - endpoint, bucket, keys, or host, database, user - and
saves them together as the JSON the server encrypts whole.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A list name too long for the node was cut with an ellipsis and nowhere
readable in full. It now shows whole in a tooltip when it is cut, the
way sidebar titles do. The count is white on a solid badge, and quiet
when the list is empty.
The name tooltip also takes the app's font: it asked for Roboto, which
the app does not load, and fell back to a serif - in the flow list too.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every list parameter of a node now reads as its name and a count, and
opens on demand to its items; with none there is nothing to open, the
add button is the way in, and the "No items" line is gone. An item
added opens the list, so it is there to see.
The list is one shared component, app-node-list-field, rather than two
near-identical blocks in the block node and a third in the execution
view: the same list reads and behaves the same everywhere, and a
container that gains a list parameter takes the component as it is.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A list whose elements are plain values - the tools an LLM node must use
successfully, the steps an evaluator follows - had no fields to build
its element dialog from, and fell back to a box asking for "Item JSON"
with {} in it, a shape nobody could guess. The element is now edited in
a single field of its own type, labelled and explained as the list is;
an empty value is not added and a value already in the list is refused.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The user guide files already committed stay tracked; anything new in
docs/ has to be added on purpose.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A secure retriever's values are at .../{parameter}/items and its
questions - is the field required, does it take a typed value - at
.../{parameter}/required and /open beside them. The question URL was
made by suffixing the values URL, so for every secure retriever it
pointed under items, where nothing answers: choosing a global input as
a file's source showed an error contacting the server.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
It cleared the two nodes it connects and nothing else, so when the
loop's own steps sat higher - the usual case - its top ran hidden behind
them, limit label included. It now clears every node between its ends,
and follows any of them being moved.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
docs/ now holds only the user guide in the repository - Markdown, PDF and
the new DOCX. The working notes, the scripts that build the guide and its
screenshots are ignored; render-user-guide.py, tracked until now, leaves
the repository but stays on disk.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Five unhandled rejections came with every test run, from the pages that
render a real graph. In tests, rete-angular-plugin was loaded by Node and
got rete's CommonJS build while everything else got its ES module, so
the renderer was not a Scope of the rete it was plugged into. Vitest now
processes the rete packages itself, and all of them share one rete.
With the graph actually built, a real bug showed: an editor still loading
when its component was destroyed went on mounting nodes with the
component's destroyed injector. A load now stops once its component is
destroyed or a newer load begins - between nodes and connections, and at
every render, which is dropped for an abandoned editor.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The full view opens only when a preview leaves something out, so a short
output or input had no way to be copied. Every value in the Output,
Intermediate and Rounds tabs now has its own copy button, beside the
view button when there is one. The full view's button just says Copy:
it copies any value, JSON or not.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The full view of an output or input gains a Copy button - Copy JSON for
a JSON value - and the outcome payload a copy button in its corner, which
stays put while a long payload scrolls. JSON is copied indented, whether
it came as an object or as a model's text in a fence, so it pastes as
clean JSON elsewhere; text is copied as shown.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Sized to its content and centred on the page, the modal grew and shrank
with every node opened or closed, its top edge jumping each time. Showing
a JSON tree it now has a fixed height, and only its content scrolls.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Outputs, intermediate inputs, loop rounds and outcome payloads that are
JSON - objects, arrays, or JSON a model wrote as text, fenced or not -
are shown with the JSON tree instead of their text. A preview opens the
first level and a few entries, summing up the rest; the full view opens
the whole tree. The JSON viewer gains the depth and entry limits this
takes.
In a read-only graph, such as an execution's, connections can no longer
be picked or shown as selected: there is nothing to do with one there.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
For an execution with a loop, the side panel gains a Rounds tab: each
loop, which round it is on out of how many, and every round with what
its steps were given - the draft sent for review, the verdict handed on
- and how it ended: went round again, left the loop, or stopped at the
limit. Earlier rounds come from the execution's step history, the last
from its steps as they stand.
From the second round of a loop on, a step's inputs are shown as they
are now rather than as prepared at the start, which is only the first
round's value. This also corrects the Intermediate tab and the inputs
shown on a node in a loop.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every connection puts its × at the midpoint of its two ends, which is on
the curve for an ordinary one. A way back is drawn above the nodes, so
that midpoint landed on whatever connection ran between them - it looked
like that connection's ×, still showing after another was selected.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Lucio Lelii