@revenexx/studio-integrations 0.7.0 → 1.1.0

package/CHANGELOG.md

@@ -1,5 +1,638 @@
 # @revenexx/studio-integrations
 
+## 1.1.0
+
+### Minor Changes
+
+- c8f121b: The Schedules page shows the firing the platform actually holds, and the run it
+  really performed.
+
+  Both times in that list were made up. There was no telemetry to read, so the page
+  computed them from the schedule's own identity — arithmetic over the workflow id
+  and the trigger id — which told the truth about the _shape_ of the answer and
+  nothing about the value. A schedule that fires nightly at 03:00 could read "in
+  18m", and one that had never run could report a last run.
+
+  They are now the platform's: the next firing it holds, with the offset of the
+  schedule's own zone available as the cell's title, and the last run it really
+  performed. The header's live count and the timezones the filter offers are its
+  counts too, taken over the tenant rather than over the rows on screen — so
+  narrowing the list no longer shrinks the menu that narrows it.
+
+  Two promises changed with them.
+
+  A paused schedule used to state neither time, which looked like one rule and was
+  really an artefact of both being invented from the same flag. It has no next
+  firing, because the platform holds none for it — and it keeps the run it performed
+  before it was paused. Forgetting that is how a paused schedule reads as one that
+  never ran.
+
+  "A distance further out than a day is stated as the date and time instead" was a
+  promise with nothing to prove it: the fabricated times stayed inside twelve hours,
+  so no fixture could produce the day-scale form. A weekly schedule reaches it, and
+  it now has a test.
+
+  What the page still cannot say is _why_ a schedule will not fire. The platform
+  distinguishes six reasons beyond "somebody paused it" — the workflow is off, its
+  bundle never built, the expression cannot be read, the zone is not a real one, the
+  rhythm can never come round — and the page shows the absence of a next run and the
+  live/paused badge, which covers the first and leaves the others looking like it.
+
+  Review then found four more things a fully mocked test suite could not see, and
+  they are user-visible.
+
+  A schedule carrying no zone of its own showed a blank timezone and could be found
+  under none of the zones the filter offers — the platform counts such a row under
+  its own default and narrows to it there, so the studio puts that default back
+  rather than showing nothing.
+
+  A run the platform has created and no worker has started now says so, instead of
+  reading as a schedule that has never run: when a run _started_ is recorded only
+  once something picks it up.
+
+  A read the platform refuses says so. Until now the page fell through to the words
+  a tenant with no schedules gets, which on this page is the opposite state — an
+  operator told there are no schedules stops looking, and a tenant whose list merely
+  could not be reached is the one where something may be silently not firing.
+
+  And a date further out than a day is drawn in the schedule's own timezone, so it
+  no longer contradicts the zone stated beside it — or the exact moment on its own
+  hover — by an hour or two.
+
+- 82884b7: The Runs page, its Missed runs tab and the Dashboard read the platform's own
+  lists and figures instead of assembling them in the browser.
+
+  Runs are a per-workflow resource, so the tenant-wide views used to collect them:
+  a request per workflow, paged to the end of each one's history — up to a hundred
+  pages of a hundred rows apiece — and then ordered, filtered and counted here. A
+  tenant's ordering and totals were therefore only ever as complete as that
+  collecting, and nothing said how far back the page reached. The same went for the
+  missed firings and for the four figures on the landing page, four of which could
+  not start until one shared list of workflows had arrived.
+
+  The platform serves all of it now. One request for the runs, ordered by it; one
+  for the counts, taken over the tenant rather than over what happened to load; one
+  for the latest run of each workflow, which is what the Workflows list and the
+  dashboard's two panels actually want. The dashboard no longer loads a run list at
+  all.
+
+  Three things visibly change rather than merely getting faster:
+
+  The trigger column and its filter tell runs apart. Every row claimed to have been
+  started by hand, because the payload carried nothing about what started it, so
+  that menu had one populated option. The platform reports it, so the column is
+  true and narrowing by it works.
+
+  The Workflows list stops being able to say "never run" about a workflow that
+  simply ran a while ago. That cell used to pick the newest run out of the tenant's
+  whole history held in memory, which only worked while all of it was loaded.
+
+  The counts are the tenant's. The Missed runs badge, the runs in flight and the
+  failed figure are counted by the platform under the same narrowing as the list
+  beside them, so a figure and its rows can no longer disagree.
+
+  What is still bounded is depth: the Runs list reaches the most recent hundred
+  runs, and narrowing reaches what is on the page. Paging past that needs the shared
+  studio table to be able to report a narrowing back to the page, which it cannot
+  yet — that is the remaining half of PO-231.
+
+- 59133ea: The canvas says which steps a loop repeats.
+
+  An iterator draws its two outputs — Item and Done — like any other pair of ports,
+  so nothing said that everything hanging off Item runs once per element while
+  everything off Done runs once, afterwards. Somebody who had not drawn the workflow
+  could not see where the repeating part began or how far it reached.
+
+  Every step a loop repeats now carries a mark saying so, and naming the loop; a
+  step inside two loops names both, outermost first. The loop itself carries the
+  number of steps it repeats — including zero, which is what a loop nothing is wired
+  to yet looks like. Selecting the loop draws its whole body as part of the
+  selection, for as long as the selection lasts — and selecting any step of the body
+  does the same from the other end, drawing the loop and everything that repeats
+  alongside it. The marks are on all three canvases
+  that draw a workflow — the editor, the run replay and the revision preview — and
+  they follow a rewiring at once, before anything is saved.
+
+  What is marked is what the engine will actually repeat, which is not always the
+  body as it was drawn: the repeating part has no end anywhere in the workflow, only
+  a reach, so a step wired to the last step of the body runs once per element even
+  when it was meant to run once afterwards. A loop's Output tab now counts the steps
+  it repeats, names them, and says which of them are wired after the loop as well —
+  reported rather than refused, because a body that rejoins the main flow can be
+  exactly what its author wanted.
+
+  The connections themselves are unchanged: they carry importance, never meaning,
+  and colour on them stays reserved for a run that actually happened.
+
+### Patch Changes
+
+- 5385146: A missed run's state is said in the studio's own words.
+
+  The badge on a missed run showed the state as the platform spells it — `PENDING`,
+  `REPLAYED`, `DISCARDED` — while the badge on a run row one tab away said
+  "succeeded" or "running" in the operator's language. A German interface therefore
+  read English in this one list, and any state the platform starts recording
+  tomorrow would have appeared on screen untranslated with nothing to notice it.
+
+  Both lists that carry the entry — the tenant-wide Missed runs tab and the panel
+  inside a workflow's editor — now say the state through the message catalogue: a
+  firing still waiting, one replayed, one discarded, in English and in German. A
+  state the studio has no word for says exactly that and keeps the platform's own
+  value alongside it, so an operator can quote what arrived without the studio
+  passing it off as a word somebody chose.
+
+  The colours of those badges are unchanged, and the two surfaces still tint them
+  differently — that is a separate finding, not this one.
+
+- 630523f: A missed run's state is drawn the same way on both lists.
+
+  The same entry was coloured differently depending on where it was opened. Beside
+  the tenant's runs it read the cockpit-wide status palette, where a waiting firing
+  is tinted like a run that is about to start and the other two states are absent
+  altogether — so _replayed_ and _discarded_ were both grey. Inside a workflow the
+  panel carried a palette of its own: amber for waiting, green for replayed, grey
+  for discarded.
+
+  Both lists now draw a state from one place: waiting asks for attention, replayed
+  reads as finished well, discarded and anything the studio has no word for read as
+  over. The colours are the shared palette's, so this is one more vocabulary mapped
+  into the studio's tones rather than a second set of colours — and the word and the
+  colour for a state now come from a single map, which is what stops the next state
+  being added to one and forgotten in the other.
+
+  Visible change: on the tenant-wide _Missed runs_ tab a waiting firing turns from
+  blue to amber and a replayed one from grey to green. The panel inside a workflow
+  keeps the colours it had, give or take the shared palette's border and text tints.
+
+- 6e1a42e: A run's outcome is drawn the same way wherever it is shown.
+
+  A run terminated by an operator read _failed_ in a list and _failed_ at its own
+  record — in two colours. The word went through the studio's vocabulary, where the
+  platform's `terminated` becomes _failed_, and the colour did not: the palette was
+  handed the platform's word, which it has no branch for, so the badge fell back to
+  the grey a cancelled run wears. Clicking the row took the operator from a red
+  badge to a grey one for the same run.
+
+  The colour is now decided from the same word as the label, everywhere a run's or a
+  step's state is drawn: the runs list, the dashboard's latest runs, a run's record,
+  its timeline and the replay. Nothing changes on the surfaces that were already
+  right — a row's status is mapped before anything draws it — and a run whose state
+  the platform reports as `terminated` or `timed_out` now looks like the failure the
+  badge says it is.
+
+  A cancelled run was affected in the same way, in the _word_ rather than the
+  colour: the platform spells a cancellation the American way, so a run cancelled
+  from its row read _cancelled_ on the row and _canceled_ at its own record — and in
+  German the record said it in English. It now reads the studio's word wherever it
+  is shown.
+
+  One of those surfaces was worse in German than in English: the timeline's step
+  lightbox derived its colour from the word it had already translated, so a failed
+  step was drawn neutral for a German reader and red for an English one. The colour
+  is now taken from the status itself.
+
+  `statusClass` is unchanged and still exports the studio-wide status palette;
+  `runStatusClass` is the run-vocabulary door to it.
+
+- e69e407: A run is called a Run in German too.
+
+  The German interface used two words for one object. Three quarters of the
+  catalogue said _Run_ — the sidebar item, the editor's Runs tab, the Run button, the
+  breadcrumb, every audit entry — while 29 values said _Ausführung_, including the
+  tab sitting directly beside the one that said _Runs_, and two sentences that used
+  both words in the same breath ("Noch keine Ausführungen — im Editor auf Run
+  klicken").
+
+  Those 29 now say _Run_: the editor's _Verpasste Runs_ tab, the dashboard's _Letzte
+  Runs_ panel and its missed-run count, the schedules table's _Letzter Run_ and
+  _Nächster Run_, and the toasts and hints around starting, retrying and replaying
+  one.
+
+  _Ausführung_ stays where it names the act rather than the object —
+  _Ausführungsmodus_, _Ausführungsverhalten_, _Ausführungsverlauf_,
+  _Hintergrundausführung_, and the settings section _Ausführung_. Their English says
+  _Execution_ too, which is the test for the next one: if the English says _run_, the
+  German says _Run_. The rule is now written down in `docs/interface-language.md`
+  rather than living in one catalogue entry's description.
+
+  No English copy changed, and no key was added or removed.
+
+## 1.0.0
+
+### Major Changes
+
+- 91888a5: The workflow editor has one Save, and every dialog over the canvas is left the same
+  way.
+
+  "Save" used to mean three different things in this editor, and the one that read as
+  most final saved the least:
+
+  - A step's own dialog said **Save** and wrote nothing to the platform — it committed
+    into the workflow's unsaved state.
+  - The settings dialog said **Save changes** under its appearance block and really did
+    write, shipping the whole unsaved canvas along with it. It could also undo itself: a
+    workflow carries its name, tags and icon inside its own definition, so the next
+    header Save shipped the canvas draft it had been holding and carried the old values
+    straight back over what that button had just written.
+  - The variables dialog offered no footer at all and kept every keystroke, so escape
+    and the ✕ _kept_ the edits there while the same gestures threw a whole form away in
+    the dialog beside it.
+
+  Now the editor holds **one draft** — the name, the description, the dispatch mode and
+  the canvas together — and:
+
+  - **The header's Save is the only control that writes to the platform.** It persists
+    all four in one call, so there is one thing that can fail and one place that says
+    so.
+  - **Every dialog closes on Cancel or Done.** Done puts what was changed on the
+    workflow as an unsaved change; Cancel, escape and the ✕ abandon it. The same
+    gesture now means the same thing in all three.
+  - **The header names the workflow as entered** — a renamed workflow reads as renamed
+    before it is saved, with the unsaved marker beside it. The dispatch badge next to it
+    deliberately does _not_ follow the draft: Run fires against the workflow the platform
+    holds, so a badge showing a drafted mode would announce a dispatch that is not the
+    one about to happen.
+
+  The dispatch mode could honestly have gone on saving by itself, being the one field
+  that is not part of the workflow's own definition and so not revertible by a later
+  save. It was pulled into the draft anyway, so that nothing inside a dialog talks to
+  the platform.
+
+  `specs/workflow-save.md` AC-7 to AC-10 carry the promise.
+
+  **Why this is a major.** Two things a host has to react to. The `Save` accessible name
+  is gone from three dialogs — anything addressing it by name reaches nothing, and
+  accessible names are part of this module's surface. And the workflow settings dialog's
+  props and events changed: `modeDraft`, `savingMode` and `savingMeta` are gone along
+  with `change-mode` and `save-meta`, replaced by a required `current` and a single
+  `apply`.
+
+### Minor Changes
+
+- 6869846: A node's output ports say what flows through them, and so do the fields inside
+  them.
+
+  Node packages describe every output port they declare, and every field inside
+  one — what the branch is for, when it fires, what a value in the payload means.
+  None of it had ever appeared on screen: the Output tab in a node's configuration
+  drew the port's name and the field's type and discarded the sentence beside them.
+  It is now read in the reader's language, under the port's name and next to the
+  field's, without anything having to run first.
+
+  The two layers are drawn independently, because that is how they are published: of
+  the 84 node types registered today, 8 describe an output port and 22 describe a
+  field — 20 of those while the port holding the field says nothing at all. So a
+  port with nothing to say still shows what its fields have to say.
+
+  The same sentence is also on the tooltip of the port's handle on the canvas,
+  which is where somebody drawing a connection is actually looking. That one is a
+  convenience rather than a promise: a tooltip is not reachable by keyboard, so the
+  Output tab stays the place the text can be read.
+
+  **What a node expects to be given is shown too.** The panel of available data now
+  opens with what _this_ node takes: the kind of value, whether an upstream
+  connection is required, and the publisher's sentence about it where one exists.
+  That answers the opposite question to the list underneath it — the list says what
+  there is to feed from, this says what to feed in — and it had been published all
+  along: the studio was asking the catalogue for a node's input under a key the
+  service does not use, so every input description, data type and requirement
+  arrived and was dropped unread.
+
+  One rule was retired rather than repaired while doing it. A node card used to hide
+  its input dot when the manifest said an upstream connection was not required. It
+  read the same unused key, so it never once fired — and both halves of it were
+  wrong anyway: an optional input is still something you connect to, and a node that
+  consumes nothing may still be wired to fix the order it runs in. Every node card
+  offers somewhere to connect to, which is what they all did in practice.
+
+- 04bd108: The dashboard's shortcuts into Import & Export, the SFTP server, the event stream
+  and inbound webhooks are offered as a way in only where the host serves them. Those
+  four surfaces belong to the host, not to this module, so the studio now asks the
+  router which of the paths this host actually has: in the cockpit all four keep
+  working exactly as before, and in a host that has none of them — or two of them —
+  the remaining cards stay where they were, keep their name and their description, and
+  answer a click by naming the surface and saying that it is to be found in the
+  Revenexx Cockpit, instead of leading into a route nobody serves. They look exactly as
+  they did — the arrow included.
+
+  **For hosts that test the studio:** in that state the card is a `button`, not a
+  `link`. A suite addressing one of the four with `getByRole('link', { name: 'Import
+& Export' })` has to ask for the role its own host serves — which is why this is a
+  minor rather than a patch.
+
+- 3c8b669: A step in a template's flow drawing is called the same thing wherever the drawing
+  is shown.
+
+  The small drawing of what a template would build is on four surfaces: the card in
+  the gallery, the cards on the dashboard and in the New workflow dialog, and large
+  at the top of a template's own page. It named its steps from the node catalogue —
+  but only where the page around it had asked for the catalogue for some other
+  reason. The gallery and a template's own page had; the dashboard and the create
+  dialog had not, so the same step read "HTTP Request" in one place and "Http
+  Request" in the next, and a reader moving from the dashboard's card to the page it
+  opens saw one step under two names.
+
+  The drawing now asks for the names it needs itself, on every surface it appears
+  on. The request is shared, so a panel of four cards still costs one — and a page
+  that already holds the catalogue costs nothing. Two things come along with the
+  names on the two surfaces that were missing them: a step is drawn with its own
+  symbol and its vendor's logo instead of a generic mark, and with the tint of its
+  category behind it, so the four drawings now look alike as well as read alike.
+
+  The hero drawing on a template's own page also gained an accessible name ("The
+  flow this template builds"), which its two neighbouring panels already had. It was
+  the largest thing on the page and the only part of it a screen reader could not be
+  pointed at. A minor rather than a patch for that name and for the marks: both are
+  additive, and an accessible name is something a host or a test can address.
+
+- 8b8916f: A workflow's editor header says whether that workflow is active, and the play
+  symbol goes back to meaning "run it now".
+
+  The header used to carry one button that flipped between Activate and Pause, so
+  the state was never named — a reader had to invert the offered verb to work out
+  that "Activate" meant the workflow was paused, while the workflows list said
+  Paused outright. That button also wore the play triangle, the same glyph the
+  Run button in the canvas's control pane wears, so one symbol stood both for
+  "armed and waiting for its triggers" and for "execute now". A workflow that is
+  switched on is not running; it is waiting.
+
+  The header now names the state in the list's own words — Active or Paused — with
+  a switch beside it that flips it, and no glyph. The words come from the same
+  message as the list badge, so the two surfaces cannot drift apart in one language
+  and not in the other.
+
+  The activation gate is unchanged. A workflow with incomplete node settings still
+  cannot be switched on: the switch cannot be operated, and the reason — how many
+  node settings to fix first — is on the control both as a tooltip and as an
+  accessible description, because a disabled control shows no tooltip in Chrome.
+  The `shift+a` shortcut is still refused with the same message, and pausing an
+  active workflow still needs no validation.
+
+- 17e938f: Unsaved work in the workflow editor is no longer thrown away without being asked
+  about.
+
+  Everything done on a workflow's canvas lives in the browser until Save is pressed,
+  and until now every way out of the editor discarded it silently. Clicking the
+  Workflows breadcrumb lost an afternoon's arrangement without a word. So did
+  reloading the page.
+
+  Worse, so did a move that was not leaving at all: a workflow's runs, missed
+  firings and earlier versions sit on tabs beside the canvas, and going to one of
+  them and back rebuilt the canvas from the _saved_ workflow. The work was still
+  there — the header still said "Unsaved changes", and Save would still have written
+  it — but nobody could see it any more, so Save wrote a workflow the operator had
+  been given no chance to look at.
+
+  Now:
+
+  - **Moving between a workflow's own tabs keeps the work.** Coming back to the
+    canvas shows it as it was left.
+  - **Leaving the workflow asks first**, and staying is what the question falls back
+    to: escape or a click beside it keeps the work and the workflow open. Discarding
+    is a button somebody has to press.
+  - **A reload or a closed browser tab raises the browser's own warning**, which is
+    the only wording available on that exit.
+  - **Putting an earlier version back says that the unsaved changes go with it.**
+    That restore has to discard them — an earlier version is what was asked for —
+    but it happens from a tab where the canvas is not on screen, so the question now
+    names what is about to be lost instead of taking it silently.
+
+  Moves the studio makes itself are not questioned: opening one of the workflow's
+  runs and closing it again, and landing in the copy that duplicating it just
+  created.
+
+  The promise and its criteria are `specs/workflow-save.md`. Hosts that render the
+  workflow editor or the version-history panel themselves gain two optional props —
+  the draft the page is holding, and whether there is one.
+
+- 2c5d6e1: The node palette no longer offers a node version the platform cannot execute. Registering a node version and building its code are two separate steps, and a version stuck after the first one used to be offered like any other — it drew its settings form, saved into a workflow, and then failed the moment a step was tested, with an error advising a retry that could never help.
+
+  The palette now offers only versions that can run, and a node whose newest version is unbuilt is offered at its newest runnable one instead; a node with no runnable version at all is not offered. A workflow already pinned to such a version keeps drawing that node — the step is really there — and its output panel now says that this version has no built code rather than offering a test that would be refused.
+
+  Needs an Integrations API that reports the build state per node version. Against an older API nothing changes: every version counts as runnable, exactly as before.
+
+- 8d55660: Opening a node no longer lands the cursor in its name.
+
+  A node's configuration used to open with the whole name selected, so it read as an
+  edit already in progress and the first key pressed replaced the name of a step the
+  user had only opened to look at. The name is left alone now, and the keyboard goes
+  to the dialog itself — Escape still closes it, Tab still stays inside it, and it
+  announces which node was opened rather than announcing the name field.
+
+  Renaming is unchanged where it was asked for: Rename in a card's right-click menu,
+  and the shortcut that menu names, still open the configuration with the cursor in
+  the name and the old name selected, so typing replaces it. That intent is now spent
+  on the one node it was asked for — saving a rename and opening another node used to
+  carry it along.
+
+  A minor rather than a patch because the dialog gained an accessible name it did not
+  have — `Configure <node>` — which a host or a test can now address it by. Nothing
+  was renamed or removed.
+
+### Patch Changes
+
+- 9accf9e: The tenant's schedules no longer cost a request per workflow. The workflow list
+  already carries each workflow's triggers, so the Dashboard, the Schedules page and
+  the navigation rail read them from that one response instead of asking
+  `/workflows/{id}/triggers` once per workflow — a third of what opening the dashboard
+  cost on a cold visit, gone, with no change to what any of those surfaces show.
+
+  The per-workflow route is untouched for what it is actually for: a single workflow's
+  triggers in the editor, and creating, changing or removing one.
+
+- 706281a: A variable's description is saved in the language it is being read in, and replacing
+  one no longer glues another language to the typing.
+
+  Two fields also carry a promise they were relied on for and never had — the name a
+  trigger goes by, and a variable's description both hand back exactly what was typed,
+  the same promise a node's name already had. That is the specification half; what
+  follows is what changed in the product.
+
+  On a studio set to German, editing a description that already carried a German text
+  wrote the typing into the English one: the field went on showing the German text, and
+  reopening the dialog showed the old wording with nothing to say that the edit had gone
+  nowhere. Descriptions that arrive from a template carry several languages, so this was
+  reachable by anyone working in a language other than English.
+
+  Two consequences of the same rule, both fixed here. A studio set to a **regional**
+  language — `de-CH` — reads a `de` description through the region fall-back, and used to
+  write the edit into a new `de-CH` key: the Swiss reader saw the new text, everyone on
+  plain German went on seeing the old one. And **clearing** a description now takes only
+  the language being edited; it used to drop the description in every language at once,
+  so emptying a German text destroyed the English one the field had never shown.
+
+  A description is therefore written under the language it was read from, and a language
+  the map does not have yet is written under its base subtag — never the host's raw tag.
+  **Clearing** a description removes the text that was on screen, which is not always the
+  same key: where the studio is borrowing another language's text — every description
+  carries English, so this is the common shape — emptying the field removes that one and
+  the description goes with it. Languages the field never displayed are left alone.
+
+  **Replacing** a description was the gesture that suffered for it. The field was
+  re-derived from the saved workflow on every keystroke, so clearing a description that
+  carried a second language refilled the field from that language at once — and the next
+  character was appended to a sentence the author never typed, then saved under _their_
+  language. Ctrl+A, Backspace, "Neu" on a German studio saved `How often it retries.Neu`
+  as the German text. The field now holds what is being typed into it while the dialog is
+  open, so clearing leaves it empty and what follows is only what was typed.
+
+  One read-side improvement comes with it: a host locale written with an underscore
+  (`de_CH`) now resolves to its language like any other tag, where it used to fall
+  through to English.
+  That also keeps a host locale the API rejects (`de-ch`, `en_US`, `zh-Hans`) out of a
+  key it validates, where it would have failed the whole workflow save rather than just
+  the description.
+
+- f6a6fdb: The list of available data inside a node's setting opens again, and a setting
+  waiting on another one names it the way the form does.
+
+  A setting that takes a mapped value offers the workflow's data from the field
+  itself — earlier steps' outputs, the trigger payload, variables — and that list
+  silently stopped opening whenever an earlier step declared what its output
+  contains, which every published node does. Mapping a value then only worked
+  through the panel beside the form.
+
+  A setting whose options the platform has to be asked for tells the operator which
+  setting to fill in first, and it was naming it by its config key: `Set company_id
+first` under a field labelled _Company_. It now names the label the form actually
+  shows, falling back to the key only where no setting matches it.
+
+  The list of available data is also addressable by name — _Available data for this
+  setting_ — so a screen reader announces which of the two offers it is reading. The
+  panel beside the form describes itself as _Available data_, and heard rather than
+  seen the two were the same words.
+
+- 6be63aa: A sticky note's headings and lists are announced as headings and lists.
+
+  A note on the canvas rendered bold, italic, code and links as the real thing, but
+  a heading only as heavier text and a list item only as a bullet character beside
+  its words. A sighted reader saw the structure; anyone using a screen reader got an
+  unbroken run of text with no heading to jump to and no list to count. Notes are
+  where an author explains a workflow to whoever comes next, so this is exactly the
+  content that ought to be navigable.
+
+  A note's `#`, `##` and `###` are now real heading elements, starting at level two
+  so the page keeps the one top-level heading it promises, and a run of bulleted or
+  numbered lines is one list with its items inside it. The bullet is drawn as before
+  but no longer read out.
+
+  The note looks exactly as it did — this changed the markup, not the design.
+
+  It also closes a hole found while reviewing that rewrite, in code it did not
+  introduce: a note's link address was not escaped against the quote that ends an
+  HTML attribute, so a crafted address could put an event handler on the anchor the
+  note drew. Notes live in a workflow blob a tenant shares, so it reached whoever
+  opened the workflow. **If your users author notes, this is the reason to take this
+  release.**
+
+- 2658447: Removing a group frame removes the frame, and deleting a connection redraws the
+  ones that are left.
+
+  **A frame folded shut no longer takes its members' connections with it.** Pressing
+  Delete on a folded frame did what it was meant to — the frame went, the nodes it
+  held stayed — and also removed, from the saved workflow, every connection that ran
+  between a node inside the frame and a node outside it. Nothing warned. While the
+  frame was folded those connections were drawn as the card's own, so the loss could
+  not be seen until the frame was gone and the members were drawn again with a line
+  missing, or until the workflow ran. Dissolving a frame through its own toolbar was
+  never affected; the keystroke was.
+
+  **A connection that is deleted no longer leaves its siblings bending around it.**
+  Where several connections leave one node, the canvas spreads them apart so their
+  horizontal runs do not land on the same line. Deleting one of them left the
+  survivors with the spread they had been given while it was still there, so a single
+  remaining connection curved as if it were avoiding a neighbour that no longer
+  existed — until something unrelated redrew the canvas or the page was reloaded.
+
+  Both came from the same place. One deletion gesture arrives from the canvas as two
+  separate reports — the connections first, the node second — and each half was acting
+  on its own: one deleted a connection by an id that, for a folded frame, belonged to
+  a connection the user had not touched, while the other never re-ranked what was
+  left. The two are now applied together, so deleting a connection reaches the host as
+  one change rather than two.
+
+- ddbabc6: Deleting a selection of nodes in the workflow editor costs one canvas rebuild
+  instead of one per node.
+
+  The editor applied a Delete one id at a time, and each of those rebuilt the whole
+  canvas — nodes, edges, every group's member count, the handle bounds — and made a
+  structural clone of the whole workflow to report the change. A ten-node selection
+  therefore cost ten of each where one does. The batch path this should have taken
+  already existed and was already what the context menu used, so the two routes
+  into the same gesture were paying different prices.
+
+  Nothing about the gesture changes: the same nodes go, a group frame in the
+  selection still dissolves and leaves its members where they are, and the workflow
+  ends up in the same state. What is new is that the route has a test — pressing
+  the key on more than one node was not covered anywhere before.
+
+- aac0de4: A magnifier, a spinner, a back arrow, a bolt and a warning triangle keep their
+  shape from one screen to the next.
+
+  Five shapes were drawn from both of the studio's icon libraries at once: whoever
+  learned the magnifier on the templates page met a different magnifier in the node
+  palette, and the same held for the spinner on a node card, the back arrow in the
+  palette, the palette's trigger bolt and the amber warning badge. All five are now
+  drawn from Solar, at every place that draws them.
+
+  Two duplications turned up beside them and went with them. The magnifier had a
+  third drawing, in the node inspector's input search. And the trigger's glyph was
+  drawn on eight surfaces that agreed in none of them: one glyph on a trigger card
+  and another in the header that opens from it, a globe for the webhook on the
+  templates page against a link everywhere else, and a different drawing again in the
+  palette's trigger tiles, the create dialog's trigger select and the template flow
+  preview. A trigger is now the same picture from the moment it is picked to the card
+  it becomes.
+
+  One of those eight was telling you something untrue rather than something
+  inconsistent. In the run list, the row for a run is its trigger's glyph, its `#id`
+  and its status badge — so that glyph is the only thing on the row that says what
+  started the run, and an event-triggered run drew the **manual play button**: "you
+  started this yourself", about a run nobody started. Every run now shows the trigger
+  it actually had.
+
+  Nothing about the module's dependencies changes: every file that drew a Lucide
+  glyph before still draws one, so `lucide-vue-next` remains as obligatory a peer
+  dependency for a host as it was. What this buys is recognition, not a smaller
+  install. `docs/icons.md` records what the icon language now is, the eight
+  shapes still drawn twice — four of which a rule about glyph _names_ cannot see, and
+  that is the finding underneath the sweep — and the one thing a small Solar glyph
+  needs on a coloured background.
+
+- 276bf2f: A connection test that passed says what it reached
+
+  The service has always answered a successful test with a sentence, and the client
+  threw it away in three separate places: the composable returned a bare `{ ok: true }`
+  and discarded the response, `credentialState()` carried `last_test_message` only in
+  its `failed` branch, and the detail dialog typed the live result as `{ ok, error? }`
+  and rendered a fixed confirmation whenever `ok`. Each was written as "it worked, or
+  else here is a reason", so a message could only ever be a complaint.
+
+  Now a credential type that reached something can name it, and **all three surfaces
+  that offer the test show what it said** — a credential's detail view, the dialog it is
+  added and edited in, and the one a node's credential field opens. Where a type reports
+  nothing the wording is unchanged.
+
+  The same field carries the reason a test was **refused**, and all three dropped that
+  too: the service sends `message` whichever way the test went, and never sends the
+  `error` these surfaces were reading — that one is synthesised locally when the call
+  does not arrive at all. So a refusal showed "Connection failed." with the reason
+  sitting unread beside it. It is shown now.
+
+  And the detail view says it **once**. It used to print the sentence twice after a
+  test — the live result and the state block both had it — and past 497 characters the
+  two disagreed, because the platform stores a shortened copy of what it answered.
+
+  This is worth more than a nicer sentence where a credential's target is **derived
+  rather than configured**. A DeepL key decides by itself which of two API hosts every
+  translation goes to — the plans are different addresses and a key is refused on the
+  wrong one — so since the node stopped asking for an endpoint, the credential is the
+  only thing that knows, and its test result is the only place it can say so.
+
+  The promise is `specs/credential-test.md`, new: one spec rather than the same
+  criterion written into three, which is what the granularity rule asks for once a
+  promise reaches more than one surface.
+
 ## 0.7.0
 
 ### Minor Changes

package/dist/module.json

@@ -1,7 +1,7 @@
 {
   "name": "@revenexx/studio-integrations",
   "configKey": "studioIntegrations",
-  "version": "0.7.0",
+  "version": "1.1.0",
   "builder": {
     "@nuxt/module-builder": "1.0.2",
     "unbuild": "unknown"

package/dist/module.mjs

@@ -1,4 +1,4 @@
-import { existsSync, readdirSync, statSync } from 'node:fs';
+import { readdirSync, statSync, existsSync } from 'node:fs';
 import { join } from 'node:path';
 import { defineNuxtModule, createResolver, addPlugin, hasNuxtModule, installModule, addComponentsDir, addImportsDir, extendPages } from '@nuxt/kit';
 
@@ -22,6 +22,7 @@ function scanRoutes(dir, baseUrl, baseName) {
   }
   return routes;
 }
+
 const module$1 = defineNuxtModule({
   meta: {
     name: "@revenexx/studio-integrations",