Skip to content

Workflow Node Config — Design Spec

Milestone: workflow-node-config Design-led spec (Testing Constitution, #4619 wave 1). The design lead owns this intent; tests enforce it; code makes the tests pass. One line per behavior; each maps to a pinning test cited by name in the test’s docstring. Status: APPROVED — creative director ratified; shipped.

Tags: [OK] behaves this way today · [BROKEN] regression, code contradicts the line · [GAP] intended behavior never built. Untagged = design intent whose current state is not yet pinned either way.

Changelog 2026-09-16 — wave-1 + safe-wave fixes landed. The behaviors below flipped from [BROKEN]/[GAP] to [OK], each now pinned by a named test: F1 effective/ghost prompt visible before registry (shows-effective-prompt, visible-before-registry — NodePromptEditorTests); F2 transcribe prompt shows for auto/llm/provider-present (transcribe.prompt.llm-only — TranscribeNodeConfigTests); F4 default is a ghost, opening is read-only for the prompt (default-is-ghost-not-text — NodePromptEditorTests / WorkflowReadOnlySavePolicyTests); F6 usesLLM is a stable tool fact and the preview shows for LLM tools (uses-llm-is-tool-fact, preview.shown-for-llm-tools — NodeProviderModelSelectorVisionModeTests); F7 node model list == Settings list, same labels (same-list-as-settings, same-labels-as-settings — NodeModelListParityTests); F13 Compare-Models Apply updates the picker (compare-apply-updates-picker — NodeSubtitleAndApplyTests.testCompareApplyUpdatesPickerSelection); F9b canvas alias/Apple subtitle no longer blank (subtitle-reflects-provider — NodeSubtitleAndApplyTests). Commits: 719094241 (F1/F2/F4/F6), 9388a2679 (F7 shared one-step picker), 5559a5908 (F13/F9b). The Findings entries below describe the pre-fix state and are kept as historical code evidence. Still open (NEEDS-EYES): F3 auto-mode chip label, F5 folder/collection prompt (server decision), F8 vision-requirement-from-server, F9a canvas icon/colour registry drift, F11 stale removed-provider warning, F16 kraken model field, and the E-section config-summary list row. (The F4 open-is-read-only residuals — transcribe legacy-language + search seed-on-open — were fixed 2026-09-16.)

Intent (the design)

The node popover is the ONE place a step is configured. Opening it shows every field the tool actually honours, in the state the run will use — the prompt the model will receive, the provider/model that will answer, and the tool’s own knobs. Nothing is hidden that the run reads; nothing is shown that the run ignores. The provider/model chooser is the same instrument everywhere (popover, AI Settings, workflow bar): same list, same labels, same rules. Looking at a node never changes it; editing it always survives close → reopen → save → reload.

Surfaces: NodePopover (++Comparison), NodeProviderModelSelector → shared ModelPicker, PromptPreviewPanel, the per-tool NodeConfigs/*, and the canvas representations (WorkflowNodeView subtitle, WorkflowNodeCard, WorkflowNodeRow).

Behaviors (each → one pinning test)

A. Fields present + editable, per node type

Common chrome (every tool) - nodeconfig.fields.label.editable — [PARTIAL] (implemented, unpinned; #4800) the Label field shows label ?? tool and edits write node.label. - nodeconfig.fields.enabled.toggle — [PARTIAL] (implemented, unpinned; #4800) the header switch toggles node.enabled; a disabled node renders at 50% on the canvas. - nodeconfig.fields.tool-name.advanced — [PARTIAL] (implemented, unpinned; #4800) the Advanced disclosure names the tool id; input mappings appear there only behind the advanced-views flag. - nodeconfig.fields.schema-driven.fallback — [PARTIAL] (implemented, unpinned; #4800) a tool without a custom config view renders every configSchema field (string/enum/int/number/array/bool) via DynamicConfigView; unknown types say so in red rather than vanishing. - nodeconfig.fields.only-honoured-keys — every field shown maps to a key the server reads for that tool (no dead knobs); every key the server honours for a user-facing option is shown. (Audit line — see Findings F5/F16.)

Transcribe (transcribe) - nodeconfig.fields.transcribe.language — [OK] Language/Locale picker; legacy codes (es, es_mx) normalise to canonical (es-ES, es-MX); auto allowed. (Pinned: TranscribeNodeConfigTests.) - nodeconfig.fields.transcribe.image-size.llm-only — Max Image Size appears whenever the run will go through an LLM, hidden when Apple Vision / Kraken will read the page. - nodeconfig.fields.transcribe.prompt.llm-only — [OK] the Prompt editor appears whenever the run will go through an LLM — including vision_mode = "auto" (the server default for new nodes) and $vision_* alias selections; TranscribeNodeConfig.showsLLMFields now treats llm/auto and any chosen provider/alias as LLM paths, while apple/kraken hide the LLM fields. (F2; pinned: TranscribeNodeConfigTests.) - nodeconfig.fields.transcribe.kraken-model — [GAP] (#4707) when the engine offers kraken mode, the kraken_model choice the server schema declares is offered; today the custom view omits it.

Describe (describe) - nodeconfig.fields.describe.detail-level — [PARTIAL] (implemented, unpinned; #4800) segmented Brief/Detailed/Comprehensive → detail_level. - nodeconfig.fields.describe.focus — [PARTIAL] (implemented, unpinned; #4800) optional Focus text → focus; clearing removes the key. - nodeconfig.fields.describe.thinking-mode — [PARTIAL] (implemented, unpinned; #4800) Thinking Mode picker; off removes thinking_mode. - nodeconfig.fields.describe.prompt — [PARTIAL] (implemented, unpinned; #4800) Prompt editor always shown (LLM-only tool); now routed through the shared NodePromptEditor, so the effective/ghost content rules in section B (F1/F4) apply.

Summarize File / Folder / Collection - nodeconfig.fields.summarize.style — [PARTIAL] (implemented, unpinned; #4800) segmented style → style (file/folder: brief/detailed/bullets; collection: executive/detailed/narrative). - nodeconfig.fields.summarize.max-length — [PARTIAL] (implemented, unpinned; #4800) slider → max_length with per-tool ranges (50–1000 / 100–2000 / 200–3000). - nodeconfig.fields.summarize.thinking-mode — [PARTIAL] (implemented, unpinned; #4800) Thinking Mode picker on all three. - nodeconfig.fields.summarize-folder.include-themes — [PARTIAL] (implemented, unpinned; #4800) toggle → include_themes. - nodeconfig.fields.summarize-collection.include-statistics — [PARTIAL] (implemented, unpinned; #4800) toggle → include_statistics. - nodeconfig.fields.summarize-file.prompt — [OK] the prompt editor shows the tool’s default (ghosted) so the user can see what they are overriding: SummarizeFileNodeConfig now composes the shared NodePromptEditor (backendPrompt + toolInfo?.defaultPrompt). Fixed in 719094241; pinned by NodePromptEditorTests. (F5) - nodeconfig.fields.summarize-folder+collection.prompt — [GAP] (#4708) these two have no prompt surface at all, and the server builds their prompt from an f-string with no override and no prompt_builder, so even the Prompt Preview shows an empty string. Design decision needed: make the prompt editable end-to-end (server prompt_builder + inputs.get("prompt")), or state in the popover that this tool’s prompt is fixed. Never an empty preview. (F5)

Extract Entities (extract_entities) - nodeconfig.fields.entities.targets — [PARTIAL] (implemented, unpinned; #4800) built-in targets plus library-registered custom types as checkboxes → entity_types (sorted array). - nodeconfig.fields.entities.add-custom — [PARTIAL] (implemented, unpinned; #4800) adding a target registers it in the library, checks it, clears the field; failure is shown inline, not swallowed. - nodeconfig.fields.entities.remove-custom — [PARTIAL] (implemented, unpinned; #4800) the chip disappears only after the backend confirms; built-ins have no remove button. - nodeconfig.fields.entities.include-context — [PARTIAL] (implemented, unpinned; #4800) toggle → include_context. - nodeconfig.fields.entities.prompt — [PARTIAL] (implemented, unpinned; #4800) the server honours a prompt override for this tool, and the popover now offers an editor: ExtractEntitiesNodeConfig composes the shared NodePromptEditor (backendPrompt + toolInfo?.defaultPrompt), matching Describe/SummarizeFile. Fixed in a265cf7f3; pinned by NodeSubtitleAndApplyTests.testEntitiesNodeComposesPromptEditor. (F5)

Sources: Files / Collection / Search - nodeconfig.fields.files.selection-default — [PARTIAL] (implemented, unpinned; #4800) with no pinned files the node states it uses the library selection at run time, with a “Pin specific files…” affordance. - nodeconfig.fields.files.pin-picker — [PARTIAL] (implemented, unpinned; #4800) the picker stages a selection (Cancel discards, Add commits); rows can be removed; drops of doc:<id> are accepted only for files. - nodeconfig.fields.collection.folder-picker — [PARTIAL] (implemented, unpinned; #4800) indented folder tree, Inbox first; a folder that disappears clears the selection rather than pointing at a ghost. - nodeconfig.fields.search.saved-search — [PARTIAL] (implemented, unpinned; #4800) saved-search picker → search_id (+ query copied for display); empty state offers Reload.

Zoom (zoom) - nodeconfig.fields.zoom.live-preview — [OK] tile-grid preview mirrors the server’s strip math and updates as rows/overlap/scale change; region mode explains itself instead of drawing strips. (Pinned: ZoomTileGridTests.)

B. Prompt text — surfaced and editable

  • nodeconfig.prompt.shows-effective-prompt — [OK] the editor’s initial text is the prompt the run will actually send for the CURRENT config, not the frozen registry default: NodePopover now fetches the server-assembled default (config minus the node’s own override, via WorkflowNode.defaultPromptConfigDict) on open and 300 ms after any prompt edit. (F1; pinned: NodePromptEditorTests.)
  • nodeconfig.prompt.visible-before-registry — [OK] if the tool registry has not loaded when the popover opens, the prompt appears once it does; the ghost falls back to the registry default live rather than initialising once in onAppear. (F1; pinned: NodePromptEditorTests.)
  • nodeconfig.prompt.default-is-ghost-not-text — [OK] when the node has no prompt override, the default is rendered as a placeholder (ghost) and config.prompt stays ABSENT. The new NodePromptEditor + pure NodePromptOverride reducer replace the hand-rolled Transcribe/Describe/Summarize-file editors, so only a user edit writes config.prompt; merely opening the popover no longer pins a stale default (no autosave) and the prompt tracks later Language/Detail edits. (F4; pinned: NodePromptEditorTests / WorkflowReadOnlySavePolicyTests.)
  • nodeconfig.prompt.edit-writes-override — [OK] typing writes config.prompt; clearing to empty removes the key (“Reset to default”). Pinned: NodePromptEditorTests.
  • nodeconfig.prompt.reset-affordance — one explicit “Reset to default” action when an override exists (present in DynamicConfigView, absent in the hand-rolled editors).
  • nodeconfig.prompt.one-editor-component — all prompt editing goes through one component (the DynamicConfigView.promptEditor pattern); Transcribe/Describe/Summarize-file stop hand-rolling divergent editors.
  • nodeconfig.prompt.preview.assembled — [OK] the Prompt Preview disclosure fetches the assembled prompt (config included, arrays and nested values converted recursively), re-fetches 300 ms after any config change while open, and offers Copy. Pinned: NodePromptEditorTests.
  • nodeconfig.prompt.preview.shown-for-llm-tools — [OK] the preview shows for every LLM tool regardless of provider choice; node.usesLLM is now a stable tool fact (no longer flipped to false on “Default”), so a Default-provider summarize node keeps its preview. (F6; pinned: NodeProviderModelSelectorVisionModeTests.)
  • nodeconfig.prompt.preview.never-empty — [BROKEN] (#4709) a tool with no prompt states “This tool has no prompt” instead of rendering an empty box (folder/collection summaries today). (F5)

C. Provider / model selection — cross-surface invariant with Settings

  • nodeconfig.model.one-picker — [PARTIAL] (implemented, unpinned; #4800) popover, AI Settings and workflow bar all render ModelPicker; per-surface differences are flags, not forks.
  • nodeconfig.model.same-list-as-settings — [OK] the popover lists exactly the models Settings lists for a provider: the user-configured models (listProviderModels(providerId:)), loaded through the one shared ModelPicker.ModelChoice.configured mapping Settings also uses, via the shared one-step picker. (F7; pinned: NodeModelListParityTests.)
  • nodeconfig.model.same-labels-as-settings — [OK] a model is labelled the same everywhere: the configured fullName, not the raw id. (F7; pinned: NodeModelListParityTests.)
  • nodeconfig.model.default-entry — [OK] “Default” (tag "") clears providerName/modelName and the model picker hides; the runtime resolves the tier default. Pinned: NodeProviderModelSelectorVisionModeTests.
  • nodeconfig.model.aliases — [OK] $small/$large always; $vision_small/medium/large only for vision tools; choosing one persists the alias as providerName, clears modelName, hides the model picker. (Pinned: NodeProviderModelSelectorVisionModeTests.)
  • nodeconfig.model.apple-vision-entry — [OK] tools that support on-device OCR list “Apple Vision (On-Device)”; choosing it sets vision_mode = "apple" and clears provider/model; the catalog’s Apple Intelligence row is hidden to avoid a duplicate Apple choice. Pinned: NodeProviderModelSelectorVisionModeTests.
  • nodeconfig.model.vision-only-filter — [PARTIAL] (implemented, unpinned; #4800) a vision tool lists only vision-capable providers, shows vision aliases, and says “No vision-capable providers available” when none.
  • nodeconfig.model.vision-requirement-from-server — [GAP] (#4710) which tools need vision / support Apple Vision comes from the served tool definition (server already knows supports_apple_vision, requires_generative_model, category); today the popover keeps two hard-coded sets (visionTools, appleVisionTools) that drift from the registry — the same client-hardcodes-a-server-fact pattern the node-editor connection-contract issue (since fixed) forbade for port conversions. (F8)
  • nodeconfig.model.uses-llm-is-tool-fact — [OK] node.usesLLM describes the TOOL and never changes with the provider choice; NodeProviderModelSelector.apply (pure static) no longer sets usesLLM=false on “Default”, so the provider section, Compare Models button and Prompt Preview survive reopen. (F6; pinned: NodeProviderModelSelectorVisionModeTests.)
  • nodeconfig.model.picker-is-one-step-not-drill-down — [PARTIAL] (#4899) the node’s model picker is a single shared chip/popover, not a two-step drill-down (choose provider, then choose model in a second screen). Built: 9388a2679 rewrote NodeProviderModelSelector.swift (232 insertions, 16 deletions) to render the shared picker directly. PARTIAL because nothing tests it: the commit touches zero test files, and NodeModelListParityTests/ NodeProviderModelSelectorVisionModeTests — the tests that DO exist for this file — cover a SIBLING surface (the popover’s model-list mapping) and are explicitly unchanged by this commit per its own message; a test that only exercises those would keep passing even if this rewrite fully reverted to the old drill-down. The test shape that would catch a regression: drive the picker through the real SharedModelListBuilder/SharedModelRow and assert the rendered choice set, PLUS an explicit companion asserting the production call site still routes through the shared component rather than a re-inlined drill-down — the same pure-function-plus-call-site-routing pairing AISettingsSelectionTests uses for its own “productionCallSiteRoutesOnlyThroughThePureFunction” check.
  • nodeconfig.model.auto-mode-representation — [BROKEN] (#4711) a transcribe node whose vision_mode is "auto" (server default for new nodes) shows “Default” in the picker AND its LLM-only fields (prompt, image size), since auto resolves to LLM unless the resolved provider is Apple. Today it shows Default with the LLM fields hidden. (F2/F3)
  • nodeconfig.model.provider-switch-picks-first-model — [PARTIAL] (implemented, unpinned; #4800) selecting an LLM provider auto-selects its first model so the node is never provider-without-model.
  • nodeconfig.model.stale-provider-visible — [BROKEN] (#4712) if the node’s saved provider is no longer among the loaded providers (disabled/removed), the picker shows the stale value with a warning; today it silently shows “Default” while providerName still carries the stale id the run will use. (F11)
  • nodeconfig.model.compare-apply-updates-picker — [PARTIAL] (#4895, retagged 2026-09-19 by the spec-led test audit: its sole pinning test reads NodePopover+Comparison.swift as text and asserts two substrings appear — it never drives the comparison sheet or a real apply action) applying a Compare Models result updates the visible provider/model selection immediately; NodeComparisonSheet’s onApply now moves both selectedProviderId/selectedModelId as well as the node fields, so the chip no longer stays on the old model until reopen. (F13; pinned: NodeSubtitleAndApplyTests testCompareApplyUpdatesPickerSelection.)
  • nodeconfig.model.reload — [PARTIAL] (implemented, unpinned; #4800) the header refresh re-loads providers; loading shows a spinner; an empty list says “No providers configured”.

D. Config round-trip (edit → persist → reopen)

  • nodeconfig.roundtrip.open-is-read-only — [PARTIAL] (#4895, retagged 2026-09-19 by the spec-led test audit: both its pinning tests read node-config Swift files as text and count a guard-string’s occurrences — neither opens a node, mutates a real binding, and asserts no autosave call actually fired) opening and closing a popover without touching anything leaves node byte-identical (no autosave fires). The prompt half of F4 — Transcribe/Describe no longer write the default prompt on open (via NodePromptEditor, no autosave); AND the config-seeding nodes (Transcribe’s legacy-language normalise, Search’s search_id/query) raise an isLoadingConfig guard during loadInitialState, cleared a main-hop later, so the on-open seed never trips the pickers’ config-writing onChange (2026-09-16). Pinned by WorkflowReadOnlySavePolicyTests (prompt + transcribe + search). Normalisation of legacy values happens only as an explicit user action, never as a side effect of looking.
  • nodeconfig.roundtrip.every-field — each field in section A reads its saved value on open and shows it (language, image size, prompt override, style, max length, thinking mode, toggles, entity types, file ids, collection id, search id, provider, model). (One parameterised test per node type.)
  • nodeconfig.roundtrip.legacy-values — [OK] legacy shapes still load: doc:-prefixed file ids are stripped; provider_name inside config is read when the typed field is empty; legacy language codes normalise. Pinned: NodeProviderModelSelectorVisionModeTests, TranscribeNodeConfigTests.
  • nodeconfig.roundtrip.autosave — [PARTIAL] (implemented, unpinned; #4800) any real edit autosaves after a 300 ms debounce; autosave never fires for a locked system preset.
  • nodeconfig.roundtrip.duplicate-copies-config — [PARTIAL] (implemented, unpinned; #4800) Duplicate copies config, provider, model, usesLLM, output schema and mappings, and opens the copy for editing.
  • nodeconfig.roundtrip.remove-means-absent — [OK] “off”/empty choices REMOVE the key (thinking_mode, prompt, focus) rather than writing an empty value, so the server default applies. Pinned: NodePromptEditorTests.
  • nodeconfig.roundtrip.alias-survives — [PARTIAL] (implemented, unpinned; #4800) a node saved with an alias provider reopens with the alias selected and no model picker.

E. Canvas representations agree with the config

  • nodeconfig.canvas.subtitle-reflects-provider — [OK] the node subtitle states what will answer: “Apple Vision”, the model name, or the alias rendered readably (“Vision · Large”). The rule is extracted to a pure static (WorkflowNodeView.nodeSubtitle) that falls back to the alias when modelName is nil, so alias nodes are no longer blank. (F9b; pinned: NodeSubtitleAndApplyTests.)
  • nodeconfig.canvas.icon-color-from-registry — [BROKEN-partial] (#4749) icon/colour come from the served registry on every canvas representation; WorkflowNodeCard/WorkflowNodeRow still carry their own hard-coded maps (the canvas node already reads the registry).
  • nodeconfig.canvas.list-row-shows-config-summary — [GAP] (#4713) the list row shows the same provider/model summary as the canvas subtitle instead of the (x, y) position badge.

First wave to pin (proposed — awaiting the design lead)

  1. nodeconfig.model.uses-llm-is-tool-fact + nodeconfig.prompt.preview.shown-for-llm-tools (F6) — the trap that hides the provider section for good.
  2. nodeconfig.prompt.default-is-ghost-not-text + nodeconfig.roundtrip.open-is-read-only (F4) — looking must not write.
  3. nodeconfig.prompt.shows-effective-prompt + visible-before-registry (F1) — the live “no prompt shown” complaint.
  4. nodeconfig.fields.transcribe.prompt.llm-only + nodeconfig.model.auto-mode-representation (F2/F3).
  5. nodeconfig.model.same-list-as-settings (F7) — cross-surface invariant.

Everything else is the larger design, pinned in later waves.

Open questions

  1. Recorded request, not decided here (from the source-model spec work, branch spec/page-model, specs/source/models-chains-and-projects.md — not in this tree): that workflow steps declare a job with a TYPED input and output, that a chain is checked before it runs, and that ONE general segment-cut-and-hand-off step replace what the request describes as three or four duplicate paths — see workflows.md’s own Open Question 7 for the full request and the verified kraken_model/kraken_recognition_model parameter-name drift; not restated here since that spec owns the cross-cutting record.

Findings (code evidence)

Paths are relative to fichero/fichero/ unless noted; server paths under fichero-server/src/fichero_server/.

  • F1 — backend prompt never fetched. Views/Workflow/Nodes/NodePopover.swift:32-33 declare backendPrompt/isLoadingPrompt; they are passed at :117 and :123 but no code assigns them (grep backendPrompt = → 0 hits). WorkflowService.getToolPrompt (Services/WorkflowService.swift:39) exists and is used only by PromptPreviewPanel. Result: TranscribeNodeConfig.currentDefaultPrompt (NodeConfigs/TranscribeNodeConfig.swift:80-87) and DescribeNodeConfig (:15-22) only ever see the static toolInfo.defaultPrompt; the onChange(of: backendPrompt) repopulation (Transcribe :188-196, Describe :104-112) is dead code. If toolInfo is nil at onAppear, the editor stays empty.
  • F2 — LLM-only gate uses the literal "llm". TranscribeNodeConfig.swift:89-97 (isLLMMode) and the if isLLMMode gates at :126/:155. Server default for new transcribe nodes is vision_mode: "auto" (workflows/tools/transcribe.py:273-274, applied through Models/WorkflowTypes.swift:235), and auto resolves to llm unless the provider is Apple (workflows/tools/vision_base.py:3502-3504). Alias selection removes vision_mode (Nodes/NodeProviderModelSelector.swift:83). Both states hide Prompt and Max Image Size.
  • F3 — popover has no representation for auto. NodePopover.swift:154-168 selects Apple Vision only when vision_mode == "apple" or the key is absent; "auto" falls through to “Default”.
  • F4 — opening writes the default prompt. TranscribeNodeConfig.swift:210-215 sets promptText = defaultPrompt in onAppear; .onChange(of: promptText) at :168-177 writes config["prompt"]. Same in DescribeNodeConfig.swift:126-131 + :85-94. Autosave fires 300 ms later (Views/Workflow/Editor/WorkflowEditor.swift:333-338). Server treats any prompt as a custom prompt: transcribe.py:339 bypasses the language-aware builder and :373 changes page-content promotion. Language rewrite-on-open: TranscribeNodeConfig.swift:200-203 + :113-118. Contrast the correct ghost pattern in Nodes/DynamicConfigView+FieldRendering.swift:119-161.
  • F5 — prompt missing or fixed per tool. SummarizeFileNodeConfig.swift:54-76 empty editor, no default. SummarizeFolderNodeConfig.swift, SummarizeCollectionNodeConfig.swift, ExtractEntitiesNodeConfig.swift have no prompt field. Server: entities.py:260 honours prompt; summarize.py:145 honours it for summarize_file only; summarize.py:287 and :452 are f-strings with no override, and only summarize_file registers default_prompt/prompt_builder (summarize.py:123-124). ToolResponse.default_prompt defaults to "" (api/routes/workflow/workflows.py:109), so the preview renders an empty string.
  • F6 — provider choice rewrites a tool fact. NodeProviderModelSelector.swift:62-69 sets node.usesLLM = false on “Default”. NodePopover.swift:313-324 gates the whole provider section on node.usesLLM for non-vision tools; PromptPreviewPanel.swift:22 gates the preview on it; NodePopover.swift:265 gates Compare Models. Registry sets uses_llm=True for all these tools (summarize.py:117/250/406, entities.py:190, describe.py:107).
  • F7 — model list source differs from Settings. Popover: Nodes/NodePopover+Comparison.swift:15-37 → listProviders() + listAvailableModels(providerType:) (catalog), labels = raw id (:23-25), available: true hard-coded, keyed by provider.id. Settings: Views/Settings/AI/AISettingsView+Helpers.swift:191-193 and :231-233 → listProviderModels(providerId:) (user-configured only; rationale at :217-221), labels = fullName (:61), keyed by providerType (:55).
  • F8 — client-side vision/Apple sets. NodePopover.swift:292-297 (visionTools) and :305 (appleVisionTools). Server knows supports_apple_vision (transcribe.py:44, describe.py:35) and requires_generative_model but ToolResponse (workflows.py:95-127) serves neither a vision requirement nor Apple-Vision support.
  • F9 — canvas duplicates. Nodes/WorkflowNodeCard.swift:48-74 and Nodes/WorkflowNodeRow.swift:96-122 hard-code icon/colour maps; Nodes/WorkflowNodeView.swift:162-172 subtitle shows nothing for alias nodes.
  • F11 — stale provider silently shown as Default. NodePopover+Comparison.swift:43-51 only sets the selection when the saved id is an alias or is in the loaded list.
  • F13 — Compare apply bypasses the picker state. NodePopover+Comparison.swift:71-75 writes node.providerName/modelName without touching selectedProviderId/selectedModelId.
  • F16 — schema/UI drift. Server transcribe config_schema declares language and kraken_model (transcribe.py:53, :84) but no prompt/max_image_dimension; the custom view shows prompt + image size and omits kraken_model. The only-honoured-keys line is the audit for this class.

Existing coverage (kept, all pass today): fichero/Tests/Unit/general/Views/Workflow/TranscribeNodeConfigTests.swift (locale table), NodeProviderModelSelectorVisionModeTests.swift (alias recognition, legacy provider read, vision_mode cleared on Default/alias — a source-text test), ZoomTileGridTests.swift. Nothing pins prompt visibility, prompt round-trip, usesLLM stability, or the Settings/popover model-list invariant.