Dreamina Provider Knowledge

Consume the current commission: outcome, materials, canonical targets, operation, fixed settings and explicit grants. Loading this skill neither changes role nor authorizes submission.

Actual generation requests establish the target operation without a second permission handshake. Short/series includes necessary asset images and local references, never paid video submission. A later manual generate-video request is persisted as the real initial grant. Diagnosis/config-only/retrieval does not imply generation; check/auto consume recorded grants for validated typed inputs. Scope, fixed settings, overwrite, snapshots, pending/inflight and retry limits still apply.

Read ${CLAUDE_PLUGIN_ROOT}/skills/_meta/rules/user-decision-relay.md for intake and decisions. Read-only capability discovery may precede intake; creating asset designs or prompts requires relevant known needs or explicit role/scope/constraints delegation, never silence. Missing needs return for clarification, not invented chat previews. Use selected models/parameters without budget, credit/balance, affordability or cheapest-option prerequisites or savings-driven downgrades. User-supplied limits remain binding; report actual account/provider failures. Pause only affected work; submission/overwrite/retry/inflight protections remain intact.

If the commission supplies config_path, validate that exact path with review-evidence.mjs config-path "{config_path}" rather than falling back to the subprocess environment. Conflicting explicit config selections need clarification. Forward its canonical result as SVD_CONFIG on each config-dependent command and to delegated tasks. Retrieval-only needs neither config nor this validation.

• capabilities.md: live discovery, configuration and scoped resolution.
• image.md: image execution, dependencies and pending recovery.
• video.md: Dreamina authoring materials, reference syntax, persisted video submission and retrieval contracts.
• shot-inputs.md: draft {shots,references} materials and final {shots,references,prompt} manifests with Creator-authored semantic prompt, full-group MP4, task clock, fidelity/completeness/integration review, capture and typed wrapper. Final --json passes prompt unchanged; drafts are not ready. Sources do not upload; GIF is unsupported. Retrieve by recorded ID/provider. This is tool knowledge, not a production pipeline.

For Dreamina video authoring, run node "${CLAUDE_PLUGIN_ROOT}/scripts/storyboard-materials-dreamina.mjs" STORYBOARD TASK_ID EP (three positional arguments, no mode flags; module export resolveDreaminaMaterials). It accepts draft/final manifests and returns JSON {task_id,shots,timeline,duration,references,assetCards,sources,inputPath,materials} without prompt. Dreamina's materials is {style,shotBlocks:[{shot,block}],referenceSlots}; identity slots are {media,path,slot,name,markdown}, local slots {media,path,slot,use}. Dreamina owns {图片N} / {视频N}, separate 1-based media counters, source binding and leading-cue rebasing; exact behavior is in video.md.

Creator uses sources, this pack and actual refs to write semantic manifest.prompt before independent review. Generic storyboard-to-prompt.sh/.mjs --json STORYBOARD TASK_ID EP returns the exact manifest string without generation or rewriting. Shared assembly data is token-free and internal; each future provider owns its material tool, without a universal adapter, registry or manifest-schema change.

Read only the relevant guidance. Retrieval does not require fresh generation capability discovery. New paid or destructive operations require current capabilities, actual authorization and current scoped production evidence.