overcast-stakeout

name	overcast-stakeout
description	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.

overcast-stakeout

Use this skill when the task is to sit on a target and catch new media as it is published (a "stakeout"), rather than a one-shot recon pass. Use the broad overcast skill and overcast/reference/verbs.md for exact flags. Only start a continuous loop when the user asks for ongoing monitoring.

Workflow

1. Set the standing scope. As new media arrives, matches auto-suggest leads — score/text triggers (a face ≥75 match, an image RANSAC hit, similar ≥85, cluster ≥70, or a text target the watch/listen text mentions) emit status:"suggested" findings that queue for triage (out of ask/brief until reviewed). --findings suggest is the default; pair it with --auto-sense watch:

overcast doctor --sources --json
overcast case init --json
overcast case setup --name stakeout --target "<name / plate / phrase>" --source "x:@handle,youtube:@channel" --auto-sense watch --findings suggest --yes --json

2. Sanity pass — one diff cycle, scheduler-friendly, to confirm sources resolve before you leave a loop running:

overcast monitor --once --pipe watch --json

3. Stand the watch up. Run the continuous loop under tmux; --alert mirrors new records to a sink and --brief summarizes each batch:

overcast monitor --every 15m --limit 5 --pipe watch --alert ./stakeout.jsonl --brief --json

4. Work the triage queue as leads accrue — finding list --state triage shows the suggested leads awaiting review; accept promotes a lead to evidence (it enters ask/brief), dismiss blocks it (never re-suggested for that match, but still auditable):

overcast finding list --state triage --json
overcast finding accept <finding-id> --json
overcast finding dismiss <finding-id> --json

5. Keep the control-room wall up as the visual surface — every case video muted and looping at its best evidence moment, freshest first, auto-restarting:

overcast wall --refresh 60 --theme csi --json
overcast wall --source x --since 24h --theme csi --json    # scope to one feed / window
overcast brief --export ./stakeout.html --json             # periodic cited report

Face-forward stakeout. A face/image suspect is an image target, which is excluded from text auto-findings. Pipe detection instead (monitor --pipe face) to surface faces in new media, then escalate the flagged clips manually with overcast face <clip> --match ./suspect.jpg --json, or keep a face-analysis index and search it (face --match ./suspect.jpg --index <id>).

Page-watch stakeout. To sit on a WEB PAGE instead of a feed, register the rendered-page source: overcast source add browser:<url> (no key; playwright optional dep), then monitor --every 30m — each pass re-renders the current page state to a PNG that flows into image auto-sense.

Live watch surface. wall is the static option (a written HTML snapshot you regenerate). For a LIVE self-updating page, an operator serves the situation room — overcast situation in its own pane, or /situation on in the TUI — with --every letting the page own the monitor cadence. Opening the listener is an operator action; the agent never runs serve. The agent then retunes it with situation set, reads situation status, and halts it with situation stop as leads accrue. Full drill: overcast-situation-room.

Output

A standing case that accrues cited findings over time: accepted matches with their source URL, record.id, and media.at; the alert-sink JSONL of new records; the freshness-overlaid wall; and periodic brief exports. State the cadence and which sources are live.

Caveats

Hard processing failures are marked seen (no infinite retry); credential/pending gaps stay retryable — run doctor --sources when a feed goes quiet. Apify-backed sources (x, tiktok, lens) bill per result, so keep --limit low on a frequent loop. The wall decodes real video — ~25 tiles is a practical ceiling (--limit); use --source/--since to scope it. Auto-findings are keyword matches on sensed text, so they shortlist — confirm before acting.