[FlyM1ss]

DO NOT MERGE until 2026-10-23 at the earliest (30 complete fact days) AND the prod read-only gate below has been re-run and passes. Merging deploys prod and drops two tables irreversibly.

Admin layer redesign PR B (design docs/superpowers/specs/2026-09-15-admin-layer-redesign-design.md, plan docs/superpowers/plans/2026-09-15-admin-layer-redesign-prB-read-paths-delete.md). It moves every /api/admin/analytics/* read path off the legacy snapshot tables and onto user_daily_facts / user_activity / rollups. It deletes states.py, lifecycle_backfill.py, maintenance.py and the reaper sweep. It drops user_analytics_snapshots and user_lifecycle_daily_snapshots at boot.

Gate (run read-only on prod before lifting draft)

• At least 30 complete fact days exist. The first complete day is 2026-09-23.
• Legacy lifecycle days with no fact row: 18 on 2026-10-15 (07-11..07-28, 283 rows). These get rolled into anonymous lifecycle_segment_count/lifecycle_transition rollups before the drop. Re-count them, and confirm that every legacy segment is one of the six lifecycle segments.
• user_analytics_snapshots.activated_at exists, so the seed can read it.
• user_activity sanity: the users without a row are never-active users (123/271 on 2026-10-15).
• Fact days with no lifecycle_transitions row (the PR A history copy, 07-29..09-22, plus no-movement days). The first boot derives their log from consecutive fact days before the drop; the count is informational.

Deploy

• Roll forward only. A rollback to main recreates empty legacy tables.
• The migration runs on the daily-facts worker thread, not the startup hook. Order: seed user_activity, legacy lifecycle rollup, history transition backfill, drop. Any of those failing skips the drop. The activation_users rollup backfill runs regardless of that chain. A failure (e.g. the drop's 5 s lock_timeout) is retried before each worker tick until it succeeds; the boot log prints the report.
• The activation_users backfill fills every rolled-up day inside the 180-day raw-event retention from one grouped query. No operator step is needed: backfill_analytics would not do it (it rebuilds rollups only for days where it inserts new events).
• The outgoing instance logs snapshot_maintenance_failed every 60 s until it exits. This is harmless.

Value changes

• Retention, /groups and /lifecycle read the PR A history copy, which holds zero activity before about 2026-09-23. Retention marks those cells incomplete, and /groups now reports partial. Recomputing those days from raw events is a follow-up (admin: recompute PR A's copied fact days from raw events before they expire #645).
• The "ATL platform model cost" figure comes from agent_runs.est_cost_usd and now excludes BYOK runs. Facts written since 2026-09-23 still carry BYOK estimates. Runs with no owner_user_id are excluded. "ATL Credits debited" now sums ledger consumption.
• include_internal=true no longer widens retention, /groups runs/cost, lifecycle segments, state counts, attention, the overview's funnel/failures/tokens/active/conversion/repeat, or /commercial platform cost (today's raw slice is internal-free like the rollups).
• Excluding an account or promoting it to admin removes its existing fact rows from every fact read (state counts, attention, history, transitions, /groups, expiry rollups) on the next request.
• Before tonight's facts land (or while the job fails), /lifecycle current counts and /operational state counts read the latest landed day (stale=true, status partial, coverage = that day) instead of zeros. Overview attention reads the day before yesterday at most, else it is unavailable. /groups is partial when any past day in the window has no fact rows.
• The activation funnel is unavailable while any rolled-up day in the window lacks activation_users rows.
• Retention counts a ledger movement (purchase, grant, debit) as activity, as on main.
• /users:
  • The filter and the priority placement use yesterday's segment, while each row shows the live segment.
  • Before tonight's facts land, the page falls back to the latest landed day, and as_of says which day that is.
  • A within-tier tie is no longer broken by lifetime net purchased.
• /overview drops user_state_counts, and /users?status= returns 422. No frontend reads either.
• The activation funnel and active/conversion/repeat ignore the billing/provider/model filters. The rollups have no such dimensions. The funnel sums per-day distinct users, so a user who repeats a step on two days counts twice. That is kept because an exact count needs a raw read on every load.
• top_failure_categories[].affected_users is renamed event_count and counts failure events, filtered or not. On main it counted distinct users from a raw read. The page column now reads "Events". rollup_day stops writing affected_users rows. Closes admin: decide what the overview's failure count means #647.
• Attention recent_runs = the sum of runs_requested over the 30 days through yesterday.
• Profile:
  • The state follows the live axes.
  • Run/cost/spend figures are as of yesterday.
  • Tokens, provider mix and lane mix come from the newest 100 events. Milestones are first occurrences in the selected range, from one grouped MIN() (≤ 2 rows): a second raw-event read beside the timeline page, recorded as an amendment to design §6.12.
  • Sessions: a session crossing the 30-day cutoff is read whole.
  • evidence_event_ids is always [].
  • Transitions: one row per day.

Follow-ups

• admin: recompute PR A's copied fact days from raw events before they expire #645 recompute the PR A copied fact days from raw events (unrecoverable after ~2027-01-25).
• admin: top_operational_reasons has no state, but the page renders one #646 top_operational_reasons needs state before PR D wires it.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
agentic-trading-lab	Ready	Preview	Oct 17, 2026 12:24am UTC