compare_app_release

Purpose

Read descriptive aligned release cohorts.

Access and effects

Owner workspace credential only. Member, partner and app-bound affiliate credentials cannot use this tool. App IDs must belong to the authenticated workspace. record_app_release persists a marker; the other tools read evidence. None activates flows or sends email.

Request

Fictional IDs; replace with owned app and saved records.

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "compare_app_release",
    "arguments": {
      "appId": "11111111-1111-4111-8111-111111111111",
      "releaseId": "22222222-2222-4222-8222-222222222222",
      "funnelId": "33333333-3333-4333-8333-333333333333",
      "version": 1,
      "windowDays": 7
    }
  }
}

Evidence and limits

Health is rule-based prioritization with visible weights and skipped unknown signals, not a calibrated churn probability. Validation requires actual persisted decision-time snapshots and distinct episodes with mature30day outcomes and3day lead. The provisional gate needs100 episodes,20 uninstall outcomes, recall at least50%, precision at least25% and F1 above a known trial-only baseline on the same population;95% intervals remain descriptive.

Daily actions use current associated USD MRR times score for paying merchants and separate unpaid priorities, top10 combined: paying first, then unpaid filling remaining slots. Handling suppresses7days unless comparable measured risk worsens. Targeted flows are separate Drafts. Digest opt-in is disabled by default, uses saved owner recipient/sender, timezone08:00 and durable local-date reservations; uncertain acceptance is never silently retried.

Release comparisons require exact app, releaseId, funnelId and version; windowDays is1–90 (default7). Equal exclusive UTC windows surround the marker. Optional currentPlan, country and installKind filters apply to both sides. Feature and retention cohorts use earliest installation per merchant in each window, mature outcomes and coverage; funnel entry uses the saved definition. These are descriptive comparisons, not causality or release lift. Marker idempotency keys cannot be reused for different content. Read status/reasons before interpreting any totals.

Health evidence policy v1 permits at most30minutes of lifecycle watermark lag. Rolling windows and observed counts end at the covered endpoint; asOf and snapshot persistence remain actual decision time. Responses expose evidence (policyVersion, coveredThrough, lagMs, maximumLagMs, fresh, uncoveredTail) and each signal’s exact window. Stale, failed, incomplete or missing lifecycle coverage remains unknown. Events after the covered endpoint do not enter rolling counts; current known installation closures still remove closed relationships.

Release failure and attempt totals are independently null unless their configured event pin and complete window coverage are known. Reasons distinguish unconfigured events from uncovered windows. A covered zero-event window is zero; a rate also needs a positive, covered attempt denominator. Digest SMTP and HTTP sends share the absolute20second cancellation deadline; uncertain acceptance is never retried or sent through a fallback.