## What
The admin Discord channel is Spanish, but the four scheduled `stats:*`
reports posted English. They now post Spanish, and the three
cohort/experiment reports open with a short AI-written summary so a
reader understands the situation without decoding the table.
### 1. Spanish reports
`stats:daily-report`, `stats:subscription-funnel`,
`stats:experiment-funnel` and `stats:ai-cohort-report` now post Spanish
embed titles, field names, ASCII table headers, legends and disclaimers,
with Spanish dates (`sáb., 13 jun. 2026`). Table headers and cells stay
ASCII (`Semana`, `Variante`, `UMad`, `pdte`, …) so `sprintf`'s byte
padding keeps the columns aligned inside the code block.
Hardcoded, not `__()`: this is an internal channel, not user-facing UI,
so it never needs a second language and doesn't belong in
`lang/es.json`. Everything else — code, comments, PHPDoc, command
descriptions, `$this->info()` — stays English.
### 2. AI summary (best-effort)
New `ReportSummarizer` + `ReportSummaryAgent` (laravel/ai, same
Gemini-Flash pattern and `AI_PROVIDER` switch as the other AI features,
config in `config/ai_reports.php`). It prepends a few sentences to the
embed description that compare the current period against the previous
one (week over week for the two weekly reports, month over month for the
monthly cohort one), say what got better or worse, and call out
explicitly when a figure isn't conclusive — small sample, immature
cohort, signup surge week, or no previous period.
- **Data**: only what each collector already computes. The two cohort
reports pass their weekly series with the per-metric maturity flags; the
experiment report passes the per-variant figures plus the
already-rendered significance verdict (one source of truth), with money
figures nulled exactly where the table renders `—`, so the summary can't
report a zero where the reader sees no data.
- **Previous period**: the experiment report has no time series, so each
run caches its figures (with a capture timestamp) as the next run's
baseline. A same-day manual re-run doesn't overwrite it, and the model
is told to flag a gap that isn't roughly one period.
- **Degrades safely**: no API key, a provider outage, a bad provider
name or a slow response (30s timeout) → the report is posted unchanged,
without the summary. Transient provider errors are logged; anything else
is reported, matching `CategorizeTransactions` /
`LaravelAiRuleSuggestionGenerator`.
### 3. Fix found while reviewing: Discord's embed limits
The translated "Cómo leerlo" field came out at 1152 characters, past
Discord's 1024-character cap on a field value — Discord rejects the
**whole** payload, so the experiment funnel would have silently stopped
appearing in the channel every Monday (`DiscordWebhook` only logs the
400). Both long fields are now tighter (745 and 898 chars),
`DiscordWebhook` trims anything still over the limit instead of losing
the report, and a test measures every scheduled report's embed so
growing copy fails in CI rather than in production.
## Testing
- `tests/Feature/Ai/ReportSummarizerTest.php`: baseline in/out, same-day
re-run, dry run, backtick stripping, truncation, empty answer,
transient-vs-reported failures.
- `tests/Feature/DiscordReportEmbedLimitsTest.php`: all four embeds
inside Discord's limits, plus the trimming fallback.
- The four command tests: Spanish assertions, summary is the first thing
in the description, and the report is still posted when the AI throws.
- 64 tests / 338 assertions pass locally, plus PHPStan and Pint.
## QA
Ran all four commands against the local DB with `--no-discord` (and
dumped the real webhook payloads with the Discord call faked). Real
Gemini output, e.g. the experiment funnel on the second run:
> No se ha producido ningún cambio en las métricas del experimento
respecto a la ejecución previa del 10 de agosto de 2026. La tasa de
conversión sobre usuarios maduros se mantiene en el 4,7 % para la
variante reducida, el 4,0 % para pay_now y el 3,7 % para el control. Las
cifras no son conclusivas ya que las diferencias entre variantes no son
estadísticamente significativas, con un p-valor de 0,593 frente al
umbral α de 0,017.
Also verified with an unreachable provider: the report renders in full,
without the summary.
## Not included
- `stats:stuck-cohort-report` is still English. It posts to the same
webhook but isn't scheduled in `routes/console.php`, so it was out of
the four active reports; worth its own decision.
- The `—` and `⚡` cells are still a couple of bytes off inside the code
block (multibyte in `sprintf`). Pre-existing on `main`, unchanged here.
## Problem
The production image runs the queue workers via supervisor but **nothing
runs the Laravel scheduler** (there is no `schedule:work` program and no
cron running `schedule:run`). As a result, none of the commands in
`routes/console.php` ever fire automatically — most notably
`banking:sync` (every 6h), so self-hosted users have to trigger bank
synchronization by hand. The manual "sync" button works because it
dispatches the same job to the queue, which the existing workers pick
up; only the scheduled trigger was missing.
## Changes
- **`docker/supervisor/supervisord.conf`**: add a `[program:scheduler]`
running `php artisan schedule:work`, mirroring the existing worker
programs (autostart/autorestart, dedicated log). This is what makes
every scheduled command fire on its own cadence.
- **Gate the operator-only stats reports** (`stats:daily-report`,
`stats:subscription-funnel`, `stats:ai-cohort-report`,
`stats:experiment-funnel`) behind `config('subscriptions.enabled')` with
an early return. These are subscription-conversion telemetry destined
for the admin Discord; on a self-hosted instance
(`subscriptions.enabled=false`) `stats:daily-report` would otherwise
make an unnecessary live Stripe API call, and the others would build
reports for an unconfigured Discord channel. On the SaaS (subscriptions
enabled) behavior is unchanged.
The guard lives **inside each command** (not in `routes/console.php`) so
it stays unit-testable. `banking:cancel-free-enablebanking` was left
alone (already a safe no-op via `hasProPlan()` when subscriptions are
off), as were the drip-email commands (already gated by
`mail.drip_emails_enabled`).
## Testing
- `vendor/bin/pint --test` passes.
- New Pest test per gated command asserting it skips and makes no
external call when subscriptions are disabled; existing tests updated to
enable subscriptions so they still exercise the real path.
- Full suite green except 5 unrelated pre-existing environment failures
(timezone alias, dashboard 409, Passport OAuth key) that fail
identically on `main`.
- Verified end-to-end with a real supervisor run: the `scheduler`
program reaches `RUNNING`, `schedule:work` boots, and autorestart
resurrects it; `php artisan schedule:test --name=banking:sync`
dispatches successfully.
> Takes effect after rebuilding the production image.
## Summary
Audited `stats:experiment-funnel` end-to-end (data acquisition, queries,
per-row counting, financial math, and statistics) and fixed the issues
that made it misleading for the win/no-win decision. The report used to
rank variants by an **absolute** contribution-margin total that
mechanically favoured whichever variant matured fastest, and it declared
statistical significance with a normal approximation that is invalid at
the small conversion counts this experiment has.
## What changed
**Presentation & comparability**
- Replace `A2P%` (numerator not a subset of the denominator → could
exceed 100%) with `Conv%` = conversions ÷ matured-assigned — always
≤100% and comparable across variants.
- Print `MatU` (matured cohort size) so the rate denominators are
visible and the table reconciles; revive `ARPU`.
- Reframe the guidance: compare on per-user `Conv%`/`ARPU`, not the
absolute `MRR`/`Cost`/`Burn`/`CM` totals (which scale with `MatU`).
**Data correctness**
- Count soft-deleted users (`withTrashed`) — they were assigned a
variant and their connections incurred real cost.
- Attribute by the deterministic `SubscriptionExperiment::bucket()`
instead of the resolved Pennant flag, so setting `force_variant` (the
winner rollout switch) no longer collapses the whole report onto one
variant. Also stops writing Pennant rows as a side effect.
- Resolve `MRR` for subscriptions on rotated/archived Stripe price ids
(fetch prices by product), warn on any net-active sub whose price is
unmapped, round yearly ÷ 12, and skip foreign-currency prices.
- `Burn` counts only users who never earned net revenue (no
subscription, or paid-then-refunded) — a paid-then-churned user is no
longer booked as connect-and-leave leak.
**Statistics**
- Measure conversion as "ever charged, net of refund" (time-invariant)
instead of a live active-now snapshot that biases older cohorts (which
have had longer to churn).
- Decide significance with **Fisher's exact test** (exact at any sample
size), Bonferroni-corrected over the three arms, instead of the
normal-approx z that overstates evidence when expected cell counts fall
below 5. Add a Newcombe difference-of-proportions CI and a small-sample
caveat.
- Extract the inference into `App\Services\Stats\ProportionSignificance`
(+ a `BinomialProportion` value object), with unit tests that pin the
exact interval and p-value numbers.
## Validation
Reconstructed every column in raw SQL against a production dump — all
reconcile **to the cent / to the row**. Independent recomputation of the
statistics (Wilson, Fisher exact, Newcombe, z) matches the command's
output.
## Testing
26 tests (20 feature + 6 unit, 94 assertions). `pint` and
`phpstan`/larastan green.
## Note
This branch also carries two small pre-existing commits unrelated to the
funnel (`fix(categories): fall back to gray…`, `chore(schedule): stop
scheduling the stuck cohort report`). Happy to split them into their own
PR if preferred.
## What
Reworks `stats:experiment-funnel` from a pure conversion funnel into a
**contribution-margin** one, so the 3-way trial/pricing experiment can
be judged on margin — not just conversion. Signups aren't free: each
bank connection costs money per month, so trials that connect banks and
never pay are burned cash, which the old report was blind to.
## Funnel
Per variant: **assigned → activated → carded → net-paying**
- **activated** = connected ≥1 bank connection **OR** gave AI consent
(triggered paid infrastructure), whether or not they paid.
- **carded** = completed Stripe Checkout (card on file).
- **net-paying** = live, non-refunded, mature subscription.
The **activated → carded** gap is where a user connects a bank and walks
away without paying — the exact leak the report now puts a number on.
## Metrics
| Column | Meaning |
|---|---|
| `A2P%` | net-paying ÷ activated (mature) |
| `Cost` | mature-cohort connections × cost/connection |
| `Burn` | connection cost of mature non-payers (money lost) |
| `CM` | MRR − Cost (the decision metric) |
- New `--cost-per-connection` option, default **0.40** per connection.
- Connection count includes soft-deleted/revoked connections (they still
cost money).
- All money/rate columns are gated on each variant's maturity window;
`Cost`/`Burn`/`CM` print `—` until a variant has mature volume. Legend
notes that raw `Assg`/`Actd`/`Card` are lifetime counts while
rates/money are mature-cohort only.
- Existing MRR/ARPU/Net% fields stay on the collector; dropped from the
printed table.
## Tests
Added coverage for activation (bank OR AI),
cost/burn/contribution-margin math, and the cost-per-connection
argument. Full file green (12 passed).
## Why
In the experiment funnel report, variants with a trial (control,
reduced) show `0` under Actv/Cncl/Rfnd while their signups are still
mid-trial — so the table looked empty even though many of those trials
have already been canceled by the user (Cashier keeps serving the trial
until it ends, then simply doesn't charge). That "already lost, just not
settled yet" cohort was invisible.
On prod right now: 28 subscriptions in `trialing`, of which **10 are
already scheduled to cancel** — a leading churn signal the report was
hiding.
## What
- `ExperimentFunnelCollector`: new per-variant `trialingCanceling`
counter = `stripe_status === 'trialing'` **and** `ends_at !== null`.
- Report table: two new columns — `Trl` (currently in trial, previously
collected but never printed) and `TrlX` (of those, already scheduled to
cancel and won't convert). Discord legend updated.
- Test covering the new counter.
`Trl`/`TrlX` are orthogonal leading indicators; the mature Net%/MRR/ARPU
metrics are unchanged.
## Testing
`php artisan test
tests/Feature/SendExperimentFunnelReportCommandTest.php` — 9 passed.
## What
Adds a `--no-discord` flag to `stats:experiment-funnel` so the
per-variant report can be run **on demand (e.g. in production) and only
printed to the console**, without posting to the Discord channel.
```bash
php artisan stats:experiment-funnel --no-discord
```
## Why
Handy for ad-hoc checks of the live experiment without spamming the
ops/Stripe Discord channel. The scheduled weekly run (and any run
without the flag) still posts to Discord exactly as before — default
behaviour is unchanged.
## Notes
- The console table already prints before the Discord step; the flag
just short-circuits the `->send()` and prints `Skipped Discord
(--no-discord).`
- Test added: with `--no-discord`, the command succeeds, prints the
table, and `Http::assertNothingSent()`.
- Pint + Larastan + the command's Pest suite green.
> Tip: the command queries the default DB connection, so to see
**production** numbers run it inside the prod container (e.g. via the
Coolify terminal).
## What
A 3-way experiment on how the paid plan is offered, plus per-variant
measurement. New signups (on/after `SUBSCRIPTION_EXPERIMENT_STARTED_AT`)
are split evenly into:
- **control** — current 15-day trial.
- **reduced_trial** — shorter trial: 3 days monthly, 7 days yearly.
- **pay_now** — charged immediately (no trial), with a self-service
money-back guarantee for the first 3 days.
Earlier users stay **legacy** and keep the 15-day trial. **While
`started_at` is null the experiment is off and everyone behaves like
control — inert until activated via env.**
## How it works
- **Assignment** — `App\Features\SubscriptionExperiment` (Pennant),
deterministic even split by a stable hash of the user id. QA can force a
variant with `feature:enable`.
- **Offer policy** — `ExperimentOffer` is the single source of truth for
trial days per plan, the pay-now flag, the refund window and refund
eligibility; shared by checkout, paywall and billing.
- **Checkout** — trial length comes from the variant (`trialDays(0)` for
pay_now → immediate charge).
- **Onboarding clarity** — the paywall states the exact terms above the
CTA: trial length for the selected plan, or "charged €X today + 3-day
money-back guarantee" for pay_now.
- **Self-service refund (pay_now)** — Settings → Billing, within the
window: refunds the upfront charge, `cancelNow`, revokes bank
connections keeping imported data. `refunded_at` records it and blocks a
second refund. Crash-safe ordering: the refund is stamped before
cancel/disconnect, which run best-effort in a try/catch.
## Measurement
`stats:experiment-funnel` (weekly → Discord): per-variant funnel
(assigned, subscribed, status breakdown, refunds) with a **net-active
rate** gated by each variant's decision window (control 15d / reduced 7d
/ pay_now 3d) so cohorts are read at equal age. Attribution reads the
variant Pennant actually served each user, so the report can't drift
from what users experienced. It also reports **MRR** (monthly run-rate
of mature net-active subs, yearly normalised ÷12) and **ARPU** (MRR ÷
assigned) per variant — ARPU is the revenue metric for the winner
decision. Plus a winner can be pinned org-wide with
`SUBSCRIPTION_EXPERIMENT_FORCE_VARIANT` (env, no deploy).
## Config (env)
- `SUBSCRIPTION_EXPERIMENT_STARTED_AT` — activates the experiment
(launch date). Null = off.
- `SUBSCRIPTION_EXPERIMENT_REDUCED_TRIAL_MONTHLY` (3), `..._YEARLY` (7),
`..._REFUND_WINDOW_DAYS` (3)
## Tests
- **Feature/unit:** assignment, offer policy, checkout wiring, refund
eligibility, the refund action incl. idempotency + crash-safe ordering
(Stripe mocked), and the funnel collector/command. ES + FR translations.
- **Browser** (`tests/Browser/SubscriptionRefundTest.php`): the
self-service refund UX end to end — card visibility + deadline, two-step
confirm, back-out, the refund control disappearing after confirming, and
gating (window passed / non-pay_now hidden). The `RefundSelfServe`
action is doubled so it never hits Stripe but applies the same DB
effect. Screenshots: `refund-card-visible`, `refund-confirm-step`,
`refund-completed`.
- Full non-browser suite green (the one failing `DashboardTest` is
pre-existing on `main` — Inertia 409 from the unbuilt local manifest).
Pint + ESLint + tsc (changed files) clean.
## Two independent reviews — acted on
**Fixed:** refund atomicity/idempotency (major) · funnel attribution now
reads Pennant's served value instead of recomputing, killing
report-vs-runtime drift (major) · pay_now copy shows the exact amount
charged · throttle + block-demo on the refund route · `resolve(?User)`
nullable · French translations.
**Reviewer notes (deferred, low value):**
- `refunded_at` is not cast to Carbon on Cashier's `Subscription` (safe
today — only null-compared; would need a custom Cashier model).
- `ExperimentFunnelCollector` walks users in PHP via `chunkById`; fine
at current volume, can move to grouped SQL if it grows.
## Confidence: 85 / 100
The critical money path is now **verified live against the Stripe
sandbox** (see below), which removes the earlier cap. All gates are
green and the acceptance criteria are met. Held at 85 (not higher)
because the browser UI test runs in CI rather than locally, and the
pay_now *hosted-checkout + webhook* leg reuses the standard Cashier
checkout already proven by the control flow (only `trialDays(0)`
differs) but wasn't re-driven through the hosted page. Given it moves
money + disconnects accounts, a human glance is still warranted before
enabling.
## Sandbox verification (live Stripe test mode)
`php artisan stripe:verify-refund` creates a real immediately-charged
subscription with a test card, runs the actual `RefundSelfServe`, and
checks the Stripe API. Result:
```
PASS subscription active after immediate charge (pay_now, no trial)
PASS canSelfRefund is true before refund
PASS latestPayment() resolves a payment intent
PASS refunded_at is stamped
PASS subscription is canceled
PASS canSelfRefund is false after refund
PASS Stripe charge shows a full refund (refunded=true)
```
The command is committed and guarded to Stripe test keys /
non-production, so it can be re-run before each launch toggle.
## Launch checklist
1. Stripe-sandbox smoke: `php artisan stripe:verify-refund` (done —
passing). Optionally also drive the hosted pay_now checkout once for
monthly + yearly to confirm the webhook leg.
2. Set `SUBSCRIPTION_EXPERIMENT_STARTED_AT` to the launch date (set
once; don't backdate).
3. Watch `stats:experiment-funnel`; a clean cohort baseline lands once
each variant's window matures.
Víctor Falcón
Jesús Mejías Leiva