Omni Deep Research
The host agent runs Omni Deep Research; MCP tools supply media evidence and rendering. Read each stage's linked reference before executing it. See parallel orchestration for host-specific execution.
Inputs and order
Inputs: video path/URL, question/focus, optional source URL and output directory.
Use omnidr_download_video, then omnidr_media_info; resume with inspect_omnidr_report.
Read perception and parallel orchestration, probe/plan, then launch source readers. Batch task creation and prepare the workspace while readers run; external searches wait for the source checkpoint.
On resume, preserve successes and index manifests/notes/ledger by path, status and gaps. Integrate findings and evidence references; work files alone do not meet research floors. Reserve time for reopening, fresh reviews after edits and delivery checks.
prepare_omnidr_report creates a draft. Keep its source at media/source.mp4,
external originals in _assets/{videos,images,docs}/ and working evidence in _work/.
Unless requested otherwise, Chinese speech → Chinese report; otherwise English.
Apply this to visible text and meta.lang; retain quotations in the source language.
1. Read the source completely
omnidr_plan_video makes consecutive jobs ≤300 seconds without orphan subsecond tails.
With native agents, run one video-reader per job in parallel, passing exact intervals.
Use the supplied Claude reader or give a native agent the
video-reader role. Do not serialize independent jobs.
Readers reconcile omnidr_read_video's concurrent 32-frame captions and ASR.
Without native agents, use one whole-source call with its internal pool.
Reuse successful passes; never repeat completed ASR. Failed clips remain gaps.
Cover the full source in semantic segments, preserving emphasis, OCR, numbers, units, scene changes and uncertainty. Read closer only where needed. Save synthesis, manifests and full transcripts; return concise text/paths, never Base64 or page dumps.
Save the source checkpoint before broad acquisition. Read report.md
and the schema; write grounded notes, source-time boundaries and available representative
images to js/data.js. Link reading manifests in the ledger and mark failed intervals
incomplete. Run the coverage pass now and before final review;
conversation alone is not a checkpoint, and a checkpoint is not delivery approval.
2. Select text and visual research anchors
Select anchors from reader visual candidates and the transcript. Retain six categories: background, additional support, similar-product/event comparison, community viewpoints,