analyse-target

name	analyse-target
description	Analyse a single tracked target's WhatsApp probe data. Pulls live from the GCP server's SQLite (WAL-safe — does not interrupt collection), separates target-side acks (delivery type) from our own collector account's echo acks (sender type), classifies the target's actual physical devices into fast/phone bands by RTT signature, decomposes phone RTTs into behavioural states (foreground/screen-off/etc.), and emits a clean human-readable presence timeline in NL local time. Triggers include "analyse <name>", "show timeline for <name>", "what was <name> doing", "decompose <name>'s activity", "/analyse-target <name>".
argument-hint	<target_name_or_jid> [--since 24h|3d] [--mode timeline|devices|phone-states|all]
user-invocable	true
allowed-tools	Bash, Read, Write

analyse-target

Project-scoped skill for inspecting one target's collected probe data and rendering a human-readable behavioural timeline. The single most important methodological point is in the section below — read it before invoking.

The ack_type distinction (load-bearing methodology)

A WhatsApp delete-probe receipt comes back from two distinct sources:

• ack_type=sender — from OUR collector account's other linked devices (typically the warm laptop where you scanned the original QR code). When we send a delete, WhatsApp's multi-device sync notifies our other devices so they can mirror the deletion; they ack as 'sender'. This receipt happens regardless of whether the target's phone is on, off, or anywhere. It is signal about us, not them.
• ack_type=delivery — from the target's actual physical device(s). This is the only ack that tells us anything about the target.

The same @lid JID will appear in EVERY target's data with ack_type=sender because every probe we send is mirrored back to our own laptop. An earlier version of this analysis treated those as if they were the target's "always-on laptop" — a mistake that produced believable but completely wrong narratives ("they were at their desk for 4h" actually meaning "we had our laptop on for 4h"). Don't repeat it. The current analyze.py filters by ack_type; if you reimplement any logic, do the same.

When to use

• The user asks "what's been doing", "show me 's timeline", "decompose 's activity", or types /analyse-target <name>
• Any time the user wants to look at a specific target rather than account-level stats

Do NOT use for:

• Account-level health (use /api/stats)
• Cross-target comparisons (this is single-target)
• Live monitoring (one-shot snapshot; for live, point them at the UI tunnel)

How it works

gcloud-ssh to wa-collector
   -> sqlite3 -csv probe_events for target_id
   -> Python analysis pipeline:
        1. classify_devices()       -- segregate by ack_type:
                                       delivery -> target's actual physical devices
                                       sender   -> our own account's linked devices (IGNORED)
                                     Within target devices, by RTT signature:
                                       fast  : median < 800ms   (rare phone-foreground or
                                                                 primary-JID device session)
                                       phone : median > 1500ms  (typical mobile push-wake)
        2. phone_state_breakdown()  -- map delivery-acked phone RTT to states
                                       baseline = phone p10
                                       foreground   = baseline + 0..200ms
                                       minimized    = baseline + 200..500ms
                                       screen-on-bg = baseline + 500..1500ms
                                       screen-off   = baseline + 1500..2500ms
                                       deep-sleep   = baseline + >2500ms
        3. per_minute_classify()    -- bucket each UTC minute (sender acks ignored):
                                       PHONE_ACTIVE     -- >=2 fast delivery acks
                                       PHONE_REACHABLE  -- >=2 slow phone delivery acks
                                       SILENT           -- 0 delivery acks despite probes
        4. smooth_and_coalesce()    -- absorb 1-2 min flips, merge contiguous runs
        5. render_timeline()        -- per-day local-time output with totals

The implementation lives in scripts/analyze.py. Don't reimplement — call it.

Invocation

python3 scripts/analyze.py <target>                       # default: timeline, NL local
python3 scripts/analyze.py <target> --since 24h           # last 24h only
python3 scripts/analyze.py <target> --mode all            # devices + phone-states + timeline
python3 scripts/analyze.py <target> --tz -5               # different TZ
python3 scripts/analyze.py <target> --source file --file exports/foo.csv  # offline CSV

<target> resolves against display_name, full jid, or phone-number prefix.

Interpretation guide (what the output means and what it doesn't)

The timeline emits three states, all derived from delivery-type acks only:

• "phone active" — multiple low-RTT delivery acks in one minute. Strong evidence that the target's phone is in screen-on / WA-foregrounded state. Likely WhatsApp open or notification just consumed.
• "phone reachable" — slow but consistent delivery acks. Phone is on the network, but in screen-off / push-wake state. Reachable, not in active use.
• "phone silent" — probes were sent but no delivery acks came back. Phone is powered off, in airplane mode, in deep sleep beyond push reach, or out of cellular coverage. Cannot distinguish "off" from "in poor signal".

What the timeline can NOT tell you:

• Location. Phone reachability is decoupled from physical location (especially with wifi-roaming). Never claim "she was at home/at her desk/outside" from this signal alone.
• Sleep vs awake. Without multi-day diurnal patterns, "phone silent overnight" looks identical to "phone in basement".
• App-specific use. The target may be using a non-WhatsApp app while phone is foreground-active; our delete-probe RTT only drops if WhatsApp itself is foregrounded.
• Whose laptop is on. The 'fast'-RTT @lid you see in ack_type=sender is OUR laptop. The target's laptop, if any, doesn't show up in delete-probe traffic.

For confident "they were actively in WhatsApp" claims, look for runs of 3+ consecutive phone active minutes. Single isolated phone active minutes are likely individual notification glances, not sustained engagement.

Caveats

• Stalled-collection windows are invisible. If our collector stalled (e.g. WS disconnect bug), gaps in probe_events look identical to "target was silent". Always cross-check account_health for ws_disconnects and probe-rate dips before claiming long silences are real.
• Timezone: defaults to NL (UTC+2 CEST). Override with --tz <hours>.
• Multiple target devices: a target may have several ack_type=delivery JIDs (e.g. primary phone JID <num>:<sess>@s.whatsapp.net plus an @lid). The script treats faster-RTT delivery JIDs as the "fast" path (correlating with PHONE_ACTIVE) and slower-RTT JIDs as the "phone" path (correlating with PHONE_REACHABLE).
• A :[e-mail] JID with N>0 IS a primary-number device session, not just a multi-device endpoint. These sometimes ack faster than the target's @lid and are more diagnostic of "phone is at its most awake".
• Don't read 'silent' as 'asleep'. Phone silent over a minute means probes did not produce delivery acks; it does not mean the person is asleep, off the grid, or anything specific about their state — only that the device wasn't reachable enough to ack.

Example narrative shape

When the user asks "what was X doing this afternoon", lead with what the data shows, not what you'd like it to show:

"From 16:48-17:55 NL her phone was reachable but in background state — on the network, not in active use. 17:55-18:37 her phone went silent; could be off, in DND, or in poor signal — can't say which. Two short reachable bursts at 18:37 and 18:59 suggest the phone was pulled out briefly. From 19:10-21:07 the phone was completely silent for nearly 2 hours — the strongest signal in her timeline. Total 'actively in WhatsApp' time today: 14 minutes, all between 23:01-23:15 yesterday evening."

Never claim location, posture, or social context — the most you can say is what the phone's network behaviour shows.