Overcast Crime Board

overcast-crime-board

Use this skill when the case has accumulated media and you want the "red-string corkboard": the people, objects, and connections laid out visually. It composes existing evidence into two shareable surfaces. Use the broad overcast skill and overcast/reference/verbs.md for exact flags.

Workflow

1. Materialize the evidence cards — faces and objects as durable crops (run face --thumbnails first so crop has frame images to cut from). Object cards need a bound open-vocabulary detector (OWLv2): crop cuts from detection boxes, so bind a detector as the see provider before --detect, then crop the --detect record (a caption/OCR see record has no boxes to crop):

overcast doctor --json
overcast face ./clip.mp4 --thumbnails --json
overcast crop <face-record-id> --all --class face --square --pad 0.1 --json
scripts/visual-db-uv.sh --detect     # once: uv-installs torch+transformers+scipy, prints DETECT_PY
export DETECT_PY="$DETECT_PY"; overcast provider setup apply --preset owl-local --yes --json  # owl-local persists a portable shipped: ref for detect.py + uses $DETECT_PY (the venv python; system python3 lacks the deps)
overcast see ./clip.mp4 --detect "car, bag, weapon, phone" --json
overcast crop <detect-record-id> --all --kind object --json   # crop the --detect record (it has boxes)

2. Draw the strings — link the same person across clips with the local face DB, and connect visual themes with CLIP semantic search:

overcast index create people --type face-cluster --local --json
overcast cluster add ./clip.mp4 --index <cluster-index-id> --json
overcast cluster identify ./person-of-interest.jpg --index <cluster-index-id> --json
overcast index create scenes --type basic-clip --local --json
overcast similar add ./clip.mp4 --index <clip-index-id> --json
overcast similar search "red backpack on a bicycle" --index <clip-index-id> --json

3. Record the connections as notes so they land on the board:

overcast note "same man (cluster <person-id>) appears in clip.mp4 and cctv.mp4 carrying the red backpack" --ref <identify-record-id> --tag connection --confidence medium --json

4. String the RELATIONAL board — graph connects the same crops, cluster people, device fingerprints, places, and typed entities across records into one force-graph (the entity/relation companion to the visual corkboard):

overcast graph --no-open --json                 # the relational board: hubs + edges (each carries a record id)
overcast graph --focus <person-id> --json       # everything tied to one cluster person

Deeper drill on the relational board (hubs, --focus, the opt-in --extract LLM pass): overcast-connect-the-dots.

5. Render the two visual surfaces — the CSI brief is the corkboard, the wall is the live monitor bank:

overcast brief --theme csi --export ./crime-board.html --json
overcast wall --theme csi --json                # add --infinite for an endless bank

Output

Two artifacts: a CSI-themed brief that lays out the crops, cited findings, and connection notes as an evidence board; and a control-room wall of the case videos looping at their evidence moments. Each connection is cited to the record.id it was drawn from.

Caveats

Crops need detections first — run face --thumbnails before cropping faces, and a bound detector for see --detect object crops. cluster/similar are local, deepface/CLIP-backed indexes (scripts/visual-db-uv.sh); doctor flags missing deps. A CLIP or cluster link is a suggestion to verify, not a proven connection — label its confidence and corroborate before drawing the string. Face similarity and CLIP scores are both 0–100; keep them distinct from an image match inlier count.