@acedatacloud/skills 2026.815.5 → 2026.815.7

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@acedatacloud/skills",
-  "version": "2026.815.5",
+  "version": "2026.815.7",
   "description": "Agent Skills for AceDataCloud AI services — music, image, video generation, LLM chat, web search. Compatible with Claude Code, GitHub Copilot, Gemini CLI, OpenAI Codex, and 30+ AI coding agents.",
   "keywords": [
     "agent-skills",

package/skills/capability-spotlight-video/SKILL.md

@@ -35,7 +35,7 @@ Read `{{run_count}}`, `{{date_iso}}`, and `{{last_output}}`. Call `maestro_list_
 
 Extract the last 12 runs' fields:
 
-`FAMILY_ID TOPIC_ID DEMO_ID STYLE_ID LAYOUT_ID PALETTE_ID VOICE_ID ASSET_HASHES`
+`FAMILY_ID TOPIC_ID DEMO_ID STYLE_ID LAYOUT_ID PALETTE_ID VOICE_ID ASSET_HASHES EXECUTED_APIS EVIDENCE_URLS`
 
 Select from the topic registry using these rules:
 
@@ -65,11 +65,12 @@ Do not accept a generic beautiful image/clip when the registry calls for typogra
 
 ### Async hero evidence
 
-Before creating generated hero media, query that capability's recent tasks and match the selected `TOPIC_ID`, `DEMO_ID`, model, and normalized request fingerprint:
+Before creating generated hero media, the first capability-specific tool call MUST be that provider's `list_tasks`/batch-list tool with a 24-hour `created_at_min` window (or the closest supported recent-task filter). Do not call generate/create before this lookup. Compare each candidate's stored request fields—model, prompt, size/resolution, quality, count/duration, input/reference URLs, and action—against the normalized request fingerprint for the selected `TOPIC_ID` and `DEMO_ID`.
 
-- reuse a matching completed task's accepted URL/MP4;
-- resume a matching pending task instead of creating a duplicate;
-- create a new task only when no match exists or the prior match is terminal-failed.
+- reuse a matching completed task's accepted URL/MP4 and do not call generate/create;
+- resume a matching pending task by ID instead of creating a duplicate;
+- create a new task only when the list call proves no match exists or the prior match is terminal-failed;
+- if the provider has no list tool, use a `SOURCE_TASK` ID from the current task's `last_output`; do not pretend artifacts from other scheduled tasks are visible.
 
 After submission, follow the MCP's returned poll interval and poll until terminal. Do not poll once and give up. For activation tests, reserve up to two minutes (for example eight 15-second polls) for hero evidence. If it is still pending after that lease, record one `ADC-SPOTLIGHT-SOURCE:v1` draft artifact containing the provider task ID, IDs/fingerprint, and no completion claim; end without Maestro. The next run must resume that task. A completed source-prep run does not count as a published Spotlight episode or consume its topic/demo diversity slot.
 
@@ -113,7 +114,7 @@ Immediately after Maestro returns an accepted task, call `publish_artifact` **be
 
 Record the accepted Maestro task as a draft. The artifact **summary itself** must begin verbatim with the complete marker below; putting IDs only in tags does not satisfy history evidence:
 
-`ADC-SPOTLIGHT:v1 | FAMILY_ID=<id> | TOPIC_ID=<id> | DEMO_ID=<id> | STYLE_ID=<id> | LAYOUT_ID=<id> | PALETTE_ID=<id> | VOICE_ID=<id> | ASSET_HASHES=<hashes> | MAESTRO_TASK=<uuid> | SUBMISSION=accepted`
+`ADC-SPOTLIGHT:v1 | FAMILY_ID=<id> | TOPIC_ID=<id> | DEMO_ID=<id> | STYLE_ID=<id> | LAYOUT_ID=<id> | PALETTE_ID=<id> | VOICE_ID=<id> | ASSET_HASHES=<hashes> | EXECUTED_APIS=<public paths> | EVIDENCE_URLS=<accepted URLs> | MAESTRO_TASK=<uuid> | SUBMISSION=accepted`
 
 Follow the marker with live docs/API references and service→asset mapping. The outer Producer must then end; it must not poll Maestro. Artifact recording is part of submission, not a post-production step.
 

package/skills/capability-spotlight-video/references/topic-registry.md

@@ -12,9 +12,9 @@ Every topic block must keep this schema: `TOPIC`, `FAMILY`, `SERVICE`, `APIS`, `
 - PROMISE: legible text and faithful edit/composite workflows, not generic text-to-image beauty shots.
 - DEMOS:
   - `gpt2-editorial-typography`: a representative editorial poster, packaging system, or wayfinding composition with a short exact headline; inspect every character and spacing. Reject ordinary architecture/still life.
-  - `gpt2-faithful-composite`: pass the canonical A symbol plus a real product/screenshot/QR source to the edits endpoint; prove the source remains faithful while the surrounding campaign changes.
+  - `gpt2-faithful-composite`: pass a real product/screenshot/QR source to `/openai/images/edits`; both the source URL and the accepted edits output URL are mandatory. This demo ID is forbidden when only `/openai/images/generations` executed—use `gpt2-editorial-typography` instead. Prove the source remains faithful while the surrounding campaign changes.
   - `gpt2-consistent-campaign`: edit one accepted hero into 2–3 placements while preserving product, typography, and brand geometry.
-- EVIDENCE: exact source(s), accepted full-resolution result, character/crop checklist, generation/edit request, authored API panel.
+- EVIDENCE: exact source(s), accepted full-resolution result, character/crop checklist, executed API path(s), accepted evidence URL(s), and authored API panel. `gpt2-faithful-composite` specifically requires `/openai/images/edits` in `EXECUTED_APIS` and its accepted output in `EVIDENCE_URLS`.
 - CREATIVE: styles `editorial,vibrant,swiss`; layouts `kinetic-type,split-proof,comparison-field`; palettes must follow the chosen recipe; voices `energetic-male,bright-female,clean-female`.
 - LIMITS: maximum 3 image calls; typography failure gets one targeted replacement; never redraw the canonical logo.