@kolbo/mcp 1.88.1 → 1.88.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kolbo/mcp",
-  "version": "1.88.1",
+  "version": "1.88.3",
   "description": "Kolbo AI MCP Server - Generate images, videos, music, speech, and sound effects from Claude Code",
   "main": "src/index.js",
   "bin": {

package/skill/GENERATED.md

@@ -1,6 +1,6 @@
 # AUTO-GENERATED — do not edit
 
-This tree is mirrored from kolbo-code@06fbd86, the single source of truth.
+This tree is mirrored from kolbo-code@0ac4fba, the single source of truth.
 Canonical source: packages/opencode/skills/kolbo/
 Distribution: .github/workflows/sync-skill-to-plugin.yml
 

package/skill/SKILL.md

@@ -1,5 +1,5 @@
 ---
-version: 0.9.14
+version: 0.9.15
 name: kolbo
 description: |
   Generate, edit, analyze, and direct creative media through Kolbo AI: images,

package/skill/VERSION

@@ -1 +1 @@
-0.9.14
+0.9.15

package/skill/references/workflows/personal-fonts.md

@@ -34,6 +34,31 @@ The backend renders internal specimens. Do NOT render specimens, attach them as
 
 Font upload/preparation has no separate credit charge; normal image generation remains billable under the existing approval rules. No automatic paid regeneration to improve typography.
 
+## Making the typeface actually come through
+
+Measured 2026-09-10 by rerunning one customer ad (Hebrew, custom family, three reference
+images) across models and settings. The backend renders a specimen and the model IMITATES
+it — nothing installs the font — so these are the levers that decide how close it lands.
+
+- **Model choice is the biggest one.** GPT Image 2 reproduced the uploaded letterforms
+  clearly better than GPT Image 2.5 Sunburst / Flare, which drift toward a default bold
+  Hebrew. Recommend GPT Image 2 whenever the typeface matters.
+- **Quality does not compensate.** 2K + `high` on GPT Image 2 beat both 2.5 rows at
+  `max`. Do not sell a higher tier as a fix for typography.
+- **Weight words in the prompt beat the specimen.** "bold", "medium weight", "very large
+  bold headline" read as typeface instructions and usually win — an ad that said bold five
+  times came back in a generic sans. Coach the user to describe size, placement, colour and
+  glow, and to choose the weight by selecting the uploaded STYLE (Bold / Medium / Light)
+  instead of writing it. Keep their exact-copy line ("EXACTLY letter for letter").
+- **Busy layouts drift; calm ones do not.** The same font on a simple prompt reproduced
+  almost exactly, and on a split-screen ad with three competing references it was ignored.
+  Fewer competing reference images and fewer text blocks buy real fidelity.
+- **Emoji never block a generation** and are drawn from the platform emoji set; they are
+  excluded from the specimen by design. No font carries them.
+- **When it must be exact, say so.** For client-final work where the typeface cannot drift,
+  generate the layout with the text areas empty and set the type over it. Never promise
+  faithful reproduction — the model is imitating a picture of the letters.
+
 ## SDK / REST
 
 The account-authenticated server-side SDK exports `createFontClient`: `list`, `get`, `upload(Blob, filename)`, `status`, `rename`, `delete`, `createUploadTicket`, and `grantToApp`. Keep account API keys on the server. The dedicated REST root is `/api/v1/fonts`; multipart upload is POST to that root. App end-user credentials do not grant access to an owner's personal library; use explicit app font grants. Image SDK calls use the same optional `font_ids`.

package/src/apps/widgets/generation.js

@@ -1330,6 +1330,64 @@ function openPromptRow(placeholder, onSend) {
    The host mounts this iframe as soon as the tool is CALLED; the result can
    take many seconds (model resolution, file upload, submit). Show a live
    shell immediately instead of a blank card. */
+// Which tool args carry a reference the browser can actually load. The kind
+// here is only a FALLBACK: refKind() still lets the file extension win, exactly
+// like the result path, so a .mp4 handed to the files arg renders as video.
+var PRE_IMAGE_KEYS = ['source_images', 'reference_images', 'image_url', 'mask_image_url',
+  'additional_images', 'first_frame', 'last_frame', 'seed_reference_image_url',
+  'elements', 'files', 'keyframes', 'source'];
+var PRE_VIDEO_KEYS = ['source_video', 'reference_videos'];
+var PRE_AUDIO_KEYS = ['audio', 'audio_url', 'reference_audio_urls', 'seed_reference_audio_urls'];
+
+// The card mounts the moment the tool is CALLED, so the only thing it knows is
+// the raw tool input - and for an edit/elements call the submit that follows is
+// the LONGEST wait on the card (local files are re-hosted first). Map the input
+// onto the same shape renderChips already reads for a server payload so the
+// references, DNA count and settings are on screen immediately instead of after
+// a minute of blank skeleton.
+// Deliberately NOT shown here: model, voice, DNA and moodboard NAMES. Those are
+// resolved server-side and arrive with the result (which overwrites all of
+// this) - rendering the raw identifier the caller passed would put an id on the
+// card, which is never allowed.
+function preRefSc(toolName, a) {
+  var img = [], vid = [], aud = [];
+  var take = function (v, bucket) {
+    if (typeof v === 'string') {
+      // http(s) only - an absolute local path is not loadable from the iframe.
+      if (/^https?:/i.test(v) && bucket.indexOf(v) < 0) bucket.push(v);
+      return;
+    }
+    if (Array.isArray(v)) {
+      v.forEach(function (item) {
+        take(typeof item === 'string' ? item : (item && (item.image_url || item.url)), bucket);
+      });
+    }
+  };
+  PRE_IMAGE_KEYS.forEach(function (k) { take(a[k], img); });
+  PRE_VIDEO_KEYS.forEach(function (k) { take(a[k], vid); });
+  PRE_AUDIO_KEYS.forEach(function (k) { take(a[k], aud); });
+  return {
+    tool: toolName,
+    kind: kindFromTool(toolName, null),
+    count: a.num_images || (Array.isArray(a.prompts) ? a.prompts.length : 1),
+    reference_images: img,
+    reference_videos: vid,
+    reference_audio: aud,
+    settings: {
+      duration: a.duration,
+      resolution: a.resolution,
+      aspect_ratio: a.aspect_ratio,
+      quality: a.quality,
+      mode: a.mode,
+      cinematic: a.cinematic,
+      visual_dna_ids: a.visual_dna_ids,
+      moodboard_ids: a.moodboard_ids,
+      moodboard_id: a.moodboard_id,
+      preset_id: a.preset_id
+    }
+  };
+}
+
 function bootPre(toolName, args) {
   if (toolName) originTool = toolName;
   if (args) originArgs = args;
@@ -1347,6 +1405,7 @@ function bootPre(toolName, args) {
     setPrompt(promptHTML(raw), raw);
   }
   setPhaseChip('Preparing', true);
+  renderChips(preRefSc(toolName, args || {}));
   if (!el('stage').innerHTML) {
     el('stage').innerHTML = '<div class="k-gen-grid n1"><div class="k-skel video" style="min-height:100px;max-height:140px"></div></div>';
   }