extensible-content-mcp

Read extensible-content://skills/publish-workflow before an empty pipeline or template→publish tasks (selected group: one template per size → that group's Merge Data → export → queue_qc_render → social copy → approve → schedule). If After videos is Leave in review, do not queue_auto_schedule unless the user asks. Read extensible-content://skills/clip-pipeline before campaign → clip ideas → generate_videos_from_clips. Read extensible-content://skills/template-authoring before any editor_* call, and extensible-content://skills/tool-reference when you are unsure a capability exists — the editor surface is large (clips, tracks, text, effects, colour grading, masks, crop, audio, speed, motion tracking, captions, markers, undo/redo) and guessing a tool name is slower than looking it up. Read extensible-content://skills/template-components before template_component_list / compose / insert — pick kind slugs and motion recipes; do not invent coordinates or keyframes. Read extensible-content://skills/scene3d before editor_add_scene3d or editor_add_object3d: 3D is the one place where the canvas is normalized but the world is metres, and where a model's scale multiplies its authored size rather than setting one. Prefer editor_add_object3d op:'take' for a walk with corners, and op:'speech' to lip-sync an avatar to existing audio. Read extensible-content://skills/choose-a-voice before narration or cloning (record-link voice-clone refuses under 12s). Read extensible-content://skills/ai-content before pipeline_generate_ai_content, editor_ai_scene, editor_animate_scene, or avatar_* generate. Pipeline tab → storyboard hops; an open canvas → editor_ai_scene; editor_animate_scene only for flags the wrappers omit. Read extensible-content://skills/media-understand before media_understand (Qwen vision: describe / objects / OCR / ask / compare — not queue_qc_render or InsightFace). After each tool call, follow result.nextStep for the immediate next action (workflow tools only — not every result). Before creating a canvas (editor_begin, template_author), settle its size: aspect ratio, orientation and resolution are three overlapping ways to say it, the canvas is fixed at creation, and resizing later rescales every layer. video_format_resolve turns whatever the user said into exact pixels and names anything still ambiguous. When you don't know the size, ask WHERE the video is going rather than asking someone to think in aspect ratios — a platform answers both the shape and the resolution (YouTube Shorts and TikTok are 9:16, YouTube is 16:9, all at 4k by default) and the tool returns the platforms as suggested answers. A result whose format block carries askUser means you guessed, so ask rather than letting the default stand. When the bytes are somewhere you cannot fetch — a file on the user's machine, a camera card, or a YouTube/Vimeo/private-Drive link — mint account_upload_link and hand them the URL rather than asking them to paste a link you can't download; account_record_link does the same for microphone/screen/webcam capture. Handoff tokens are reusable until expires_at (30 min). Tools marked longRunning return { status: 'running', jobId } — poll mcp_job_status until done: true, then read result.nextStep. Many editor_set_* / editor_state calls are sync and do not return a jobId. A pipeline can have many template groups; only the selected group generates. After editor_save_template, assign that canvas ratio on the selected group with pipeline_set_generate_section (group templatesByRatio — not reframe.targetRatios), then generate_run, which renders every row at every assigned size — do NOT jump to template_render_video (that is an escape hatch for one-off exports that skip the table). Stay a merge row when only cells change; pipeline_fork_row when a row needs edits beyond merge fields. What differs between videos is merge fields: editor_set_project op:bindMergeField with `key` (not mergeKey) binds a text or media clip to a {{KEY}}; use the SAME key on the 9:16 and 16:9 templates so one table column fills both. The authored content is the fallback for a blank cell. An open footage slot is a video clip bound to e.g. CLIP_SRC with editor_set_project op:setTiming length:source. There is no planner and no slides — intro, footage and CTA are ordinary clips on that one template. Author captions and the presenter on the template (editor_add_captions / editor_add_voiceover / editor_add_presenter); they are not pipeline toggles. Call worker_status before editor_generate_mesh / editor_generate_product_mesh / editor_animate_scene: a feature whose weight download has not finished is not enabled. First-boot pull is miners → NudeNet 320n → Qwen 3.8 27B → Qwen3-VL 8B → realesrgan → ambient-beds → hunyuan3d → trellis → mia → latentsync → qwen-image → qwen-image-edit → Wan I2V 14B → Wan T2V 14B → HunyuanVideo 1.5. Mesh, rig, LatentSync talking, and Qwen-Image come up before the Wan 14B pairs; HunyuanVideo 1.5 is last. closeUp+script still runs a short idle I2V then still-hold then LatentSync — it is not LatentSync-only. worker_status lists Clore boxes with usable/activity; worker_use occupies one (omit instanceId for the recommended idle Clore). Then editor_begin with preferContainer: true and that instanceId. worker_prefetch asks the box to pull or retry a bundle (prefetch id is the bundle, e.g. realesrgan, not comfy_upscale). worker_recycle drops leftover Chrome (including sticky activeJobs) so CPU+GPU mining can resume; leftover Chrome recycle is allowed while a waiting Comfy job occupies an idle box. Refuses livestream live, gpuMode=llm, and in-flight Comfy. Unpin Qwen with worker_llm enabled=false before recycle.