overcast

Give any agent senses (video/audio/image understanding) and OSINT reach (search/capture/monitor) organized around an investigation case. Use when the user wants to analyze a video/audio/image, scan or monitor sources for a target, or ask/brief over accumulated findings. Drives the `overcast` CLI (built on pi + the tinycloud/Cloudglue perception backend); see reference/verbs.md for the full verb surface.

Install and configure overcast for this harness: install the CLI, verify the system ffmpeg, and configure reusable provider profiles. Use once per machine/profile before driving the `overcast` skill.

Turn a case into a visual evidence board — materialize face and object crops, link the same person across clips, connect themes, and render the corkboard as a CSI brief plus a live monitor wall.

Make unreadable footage legible — denoise/upscale a marked moment, re-run OCR and detection on the enhanced output, and record what was recovered with honest provenance (interpolation is a lead, not proof).

Work out where a photo or clip was taken — check embedded EXIF/GPS first, then pull signage, landmarks, and terrain clues, reverse-image-search the strongest ones, and corroborate to a location with cited evidence.

Run a standing surveillance watch on public sources for a target — auto-sense new media, auto-suggest matches into a triage queue, and keep a live control-room wall — so new evidence surfaces itself over time.

Is this real? — triage whether an image or video was altered by accumulating independent leads: C2PA / Content Credentials provenance, EXIF capture metadata, ELA/noise forensic overlays, and a sun/shadow time check — never a single-signal call.

Find a person, logo, object, landmark, or visual reference across local clips or captured media with timestamped Overcast evidence.

Locate WHERE in a frame a target is — open-vocab detect, crop the box, then VLM-verify the crop so hallucinated boxes don't survive.

Same recording, surfaced again — fingerprint audio into a local audio-fp index, then match a query clip against it (or clip-to-clip) with time-offset alignment, gate out sped-up re-uploads with a margin, and escalate a fingerprint miss to a CLAP semantic pass.

Same camera shot these — pull EXIF device fingerprints (make/model/lens/serial) off every case image and video, roll them up by camera with devices, and tell a strong serial link from a weak make+model+lens one, feeding the connections into the graph and the map.

String the board — build the case knowledge graph, read its hubs (shared media, targets, device fingerprints, places, typed entities), focus the neighborhood around a node, optionally run the opt-in LLM entity pass, and promote real connections into evidence via findings.

Localize the exact instant of a one-way state change (light on, door opens, poster removed) by binary-searching frames — ~log2(N) VLM calls.

Triage a whole clip in one VLM call — tile sampled frames into a labeled contact sheet, ask which cells show the target, then zoom in.

Pinpoint WHEN a specific thing happens in a video — funnel from cheap coarse candidates (shots / CLIP / grid) to a frame-verified time window.

Find the interval a person or object is present — anchor one appearance, then sweep outward with face-match / detect until it drops off both sides.

Trace a suspicious or viral clip back to its earliest appearance and originator — reverse-image-search distinctive frames, sweep sources with no recency floor, and return a cited origin verdict.

Listen to the police scanner — register a dispatch (CAD / calls-for-service) feed on the Socrata SODA API, validate with one scan, then monitor it on an interval, plot the geolocated calls on a map, and triage call-types against the case's lines of investigation.

Stand up the live monitoring page over a case — pick and seed sources, have the operator serve the page in its own pane, then drive the running page (panels, interval, filters) through the control plane without ever opening the listener yourself.

Build an identity dossier from opt-in people-search sources (authorized use only) — discover accounts from a handle, resolve a name to public records, reverse a phone, pull property records — landing each hop as a cited scan record and cross-checking the pieces in the graph.

Voice lineup — enroll a reference speaker into a local voice-print index, then rank WHERE that speaker talks inside a clip and WHICH recordings contain them, reading the anchored 0–100 rank score and its margin, and corroborating before naming anyone.

Work a recorded call or audio clip already in the case — separate the speakers, read the background scene for location clues, isolate and re-transcribe voices, and correlate content across recordings.

Read foreign-language text off an image or video frame, translate it, and re-search in the SOURCE language — OCR a sign/screen/poster with see --ocr, translate it yourself, then scan the open web (and dork) with native-language queries, and cite what the text revealed.

Save media into global archive buckets reusable across cases, stand up bucket indexes (face/semantic/image/audio/ask) with the archive setup wizard, and run "does this match anything I've archived?" checks from inside any case.

Map a target's internet-exposed hosts and services with the `shodan` source, capture host reports, brief the exposure, and optionally stand up a monitor for newly exposed services.

Google-dork a target domain for exposed documents, directory listings, login portals, and misconfigurations using the `dork` source (real Google operators), then capture the result pages and produce a cited exposure brief.

Build a persistent local face database out of case media — the mugshot book — then run a suspect photo through it to identify who they are and where else they appear, with cited similarity scores.

Reconstruct a single event from multiple clips — sense each one, cross-anchor shared moments, surface corroborations and contradictions, and produce one cited chronological brief.

Hunt re-uploads and reskins of original video content across X / YouTube / TikTok — escalate from cheap metadata triage to frame/face/transcript matching and produce citable copycat findings.

Analyze screen recordings, product demos, customer support videos, and audio notes into actionable, cited bug reports for coding agents.

Scan or monitor public sources for a target, capture relevant hits, sense media, and produce cited investigation briefs.

Create small, installable agent skills that wrap focused Overcast workflows. Use when the user asks to make an Overcast skill for a specific investigation, media analysis, recon, monitoring, or case-memory workflow.