Overcast Bolo

overcast-bolo

Use this skill for a be-on-the-lookout / all-points-bulletin: "here is a face, alert me on any match across incoming media." It is a standing watch keyed on a VISUAL reference (a face or image), not a one-shot lineup search. Use the broad overcast skill and overcast/reference/verbs.md for exact flags. Only leave a continuous loop running when the user asks for ongoing monitoring.

This packages existing machinery: a target --image is the reference line hits attach to, monitor --pull --pipe captures + senses new media, and the persist hook auto-runs the score triggers on every evidence record — a face --match ≥75 (or similar match ≥85) auto-emits a suggested finding linked to that line. The BOLO flow names and sequences those pieces. It is the close cousin of overcast-stakeout (a general standing watch, often keyed on a text target) and overcast-lineup (a one-shot identify against a local face DB); reach for those when the watch is text-keyed or the job is a single database lookup.

Prerequisites

Pick the matcher up front — it decides the backend the watch needs:

• tinycloud face (default face provider): CLOUDGLUE_API_KEY set, tinycloud CLI on PATH. No local Python. face <clip> --match <ref> runs against the clip.
• deepface-local (offline, no cloud): the uv-managed visual-DB Python (scripts/visual-db-uv.sh --face, then OC_VISUAL_DB_PY). Reference faces live in a local deepface-local index; similar match uses a local basic-clip index instead. Nothing leaves the case.

overcast doctor --json                 # confirm the face/visual-db backend is ready
overcast case init --json

Workflow

1. Register the watchlist reference. The reference face becomes an image target — the line of investigation every hit attaches to. Add one target per person/object you are looking out for:

overcast target add ./suspect.jpg --image --question "Does the watchlisted person appear in incoming media?" --json
overcast target list --json                                  # confirm the image line (kind: image)

For LOCAL matching, also stand up a face/image index and register the reference(s) in it, so you can match case-wide without re-passing the file:

overcast index create bolo-faces --type deepface-local --local --json    # local face DB
overcast index add ./suspect.jpg --to <index-id> --json                  # enroll the reference face
# semantic image alternative: overcast index create bolo-seen --type basic-clip --local --json

2. Point the watch at incoming feeds and stand it up. Reuse any configured source (x / youtube / tiktok / instagram / telegram / browser / …). Recording the reference as a case --face-ref with --findings suggest + --auto-sense watch makes new captures auto-sense and auto-suggest without a manual pass:

overcast source add "x:<query>" --json
overcast case setup --name bolo --face-ref ./suspect.jpg --findings suggest --auto-sense watch --yes --json
overcast monitor --once --pull --pipe watch --json           # one diff pass: confirm sources resolve
overcast monitor --every 30m --pull --pipe watch --json      # the standing loop (run under tmux)

3. Auto-match incoming media against the reference — the BOLO core. Because face --match / similar match fire the suggested-finding trigger on run, run the reference match over each newly captured item. A hit at/above the threshold (face ≥75, similar ≥85) auto-emits a suggested finding linked to the image target line:

overcast face <new-clip> --match ./suspect.jpg --json                      # tinycloud/deepface: find the face IN the clip
overcast face <new-clip> --match ./suspect.jpg --index <index-id> --json   # deepface-local matcher over the clip
overcast similar match ./suspect.jpg --index <index-id> --json             # basic-clip semantic image match

4. The alert / triage queue — the BOLO board. New hits queue as suggested leads (quarantined from ask/brief until reviewed). Confirm a real hit with accept (stamping it onto the reference line); reject a false positive with dismiss (never re-fires for that match):

overcast finding list --state triage --json                          # the BOLO board (leads awaiting review)
overcast finding accept <finding-id> --target <target-id> --json     # confirm the hit onto the reference line
overcast finding dismiss <finding-id> --note "wrong person" --json   # reject a false positive

5. The wall / live surface (optional). Keep a control-room view of the standing watch — every case video muted and looping at its best evidence moment, freshest first:

overcast wall --refresh 60 --theme csi --json          # static HTML monitor wall
overcast brief --export ./bolo.html --json             # periodic cited report

For a LIVE self-updating page, an operator serves the situation room (overcast situation in its own pane, or /situation on in the TUI) — never the agent. Full drill: overcast-situation-room.

Output

A standing case that alerts on visual matches over time. For each confirmed hit return: the reference (image target) it matched, the source clip's record.id + media.at, the match score (face similarity 0–100 / similar 0–100), the suggested→accepted finding id, and the reference target line it lands on. State the cadence, which sources are live, and which matcher (tinycloud vs deepface-local) the watch runs on.

Caveats

• A face/image target is EXCLUDED from text auto-findings — the visual match is the trigger, so you must run face --match / similar match on new media (or wire it via case setup), not rely on a text phrase.
• Auto-suggested visual hits are LEADS, not identifications — face embeddings degrade on poor lighting, small faces, and heavy angles. Corroborate a single borderline match (a second clip, overcast-lineup) before naming anyone.
• Apify-backed sources (x, tiktok, instagram, lens) bill per result; keep --limit low on a frequent loop. The wall decodes real video (~25 tiles is a practical ceiling — use --source/--since to scope it).
• Hard processing failures are marked seen (no infinite retry); credential/pending gaps stay retryable — run overcast doctor --sources when a feed goes quiet.

Deferred follow-up (not in v1)

• Alert on a text description ("a person in a red hoodie carrying a backpack"). This needs a standing open-vocab detector (see --detect) on every incoming frame — noisier and costlier than a reference match, so it is out of scope for v1. Today, key a BOLO on a face/image reference. For a text-KEYWORD watch (a name/phrase in sensed transcripts), use overcast-stakeout with a non-image text target instead.
• A first-class bolo state verb wrapping steps 1–4 (register → watch → auto-match → triage) is a possible future convenience. v1 is skill-only over the existing verbs; no new verb was added.