univa-agent Codex UniVA Media Ops
Use this skill when Codex operates UniVA for video, image, audio, speech, understanding, editing, or compound media tasks.
Video Generation
Prefer scripts/codex_video_runner.py for Codex-driven video generation. The runner loads relevant skills, records skill context, creates research and storyboard artifacts, and calls existing univa/mcp_tools functions in process. Run it first in review mode so it writes research_brief.json, shot_plan.json, storyboard_validation.json, and storyboard_review.json, then stop for approval. Generate only by rerunning with --approved-plan <run_dir>/shot_plan.json.
For approved generation with audio enabled, follow UniVA's audio priority: keep native audio returned by the video API first, fall back to the registered dedicated audio API second, and use local FFmpeg synthetic audio only when the generated video has no audio stream and the dedicated audio path is unavailable or fails. Single-shot runs may pass the approved auto_audio options directly to the video generation tool. Multi-shot runs keep intermediate clip full mixes disabled and apply cohesive audio at merge time.
Video Understanding and Editing
Prefer scripts/codex_video_task_runner.py --task understand|edit. Understanding writes an analysis plan, calls vision2text_gen, and produces reports. Editing writes edit_proposal.json and edit_proposal_review.json, then stops for approval. Execute edits only with an approved plan.
For durable long-video indexing and time-coded lookup, also read
skills/core/media-index.md. Use index_video_media, enrich only with trusted
timed analysis through update_video_index_segments, then search. Do not claim
visual or semantic retrieval from an empty text index.
ASR and Localization
For prompts containing ASR, transcription, subtitle translation, bilingual
captions, dubbing, or localization, read skills/core/localization.md and
skills/pipelines/localization/executive-producer.md.
- ASR-only work is read-only: call
univa.mcp_tools.localization.transcribe_mediadirectly with the configured UniVA Python environment; no Web or backend server is required. - Save real transcript/caption output under
results/localization/. Do not fabricate text whenfaster-whisperor a model download is unavailable. - Translation, sidecar captions, voiceover, and rendering are separate user intents. Do not add them to an ASR-only request.
- Voiceover, muxing, and rendered subtitles mutate/create media and require the universal plan/review gate before execution.
Durable Runtime and Provenance
Read skills/meta/artifact-provenance.md for version, hash, receipt, cost, and
approval binding. Use the backend plus scripts/univa_agent_client.py only
when durable sessions, auth, pause/resume, or automatic Pipeline Artifact
records are required. Keep the current single-use continuation token and do
not retry resume blindly.
Other Media
For image, audio, speech, and non-video mutation, apply the same media review gate with media_plan.json, media_plan_validation.json, media_plan_review.json, or equivalent checkpoint artifacts before calling MCP tools.
Rules
Read skills/INDEX.md, skills/meta/media-review-gate.md, the matching core contract, and task-relevant creative/theme/pipeline/user skills. Do not bypass approval, do not shorten approved prompts, and do not invent alternate runtimes.