# ADR-003: Session-Rollup Unified Activity Feed on `/chronik`

## Status

Accepted

## Context

The app had two disconnected ways to see what was happening in the archive:

1. `/notifications` — personal mentions/replies only, delivered via the `notifications` table and a Bell dropdown.
2. Dashboard activity feed — ambient events (uploads, transcription, annotations, comments, mentions) via `/api/dashboard/activity`, which deduplicated using `DISTINCT ON (actor_id, document_id, kind, date_trunc('hour', happened_at))`.

Two separate lists was a poor mental model (personal vs. ambient feel the same to the user), the `/notifications` page wasted horizontal space, the dashboard's "Alle anzeigen" pointed to `/documents` (dead-end), and the hour-trunc dedupe produced ugly splits on natural sessions — saving 20 transcription blocks at 08:58, 08:59, 09:01 yielded two rows.

We needed one page that merges both streams, keeps personal mentions visually loud, and aggregates ambient noise coherently.

## Decision

**One page `/chronik` backed by two endpoints.** The SvelteKit `+page.server.ts` composes data from `/api/dashboard/activity` (for the ambient timeline) and `/api/notifications` (for the "Für dich" box). No new `/api/chronik` orchestrator — the frontend load function is the composition seam.

**Session-style rollup replaces hour-trunc dedupe everywhere.** `AuditLogQueryRepository.findDedupedActivityFeed` is renamed to `findRolledUpActivityFeed` and rewritten using a `LAG()`-based session algorithm:

```
LAG(happened_at) OVER (PARTITION BY actor_id, document_id, kind ORDER BY happened_at)
  → is_new_session = gap > 7200s (or first event in partition, or kind ∈ {COMMENT_ADDED, MENTION_CREATED})
  → SUM(is_new_session) OVER (... ROWS UNBOUNDED PRECEDING) = session_id
  → GROUP BY (actor_id, document_id, kind, session_id) → MIN(happened_at), MAX(...), COUNT(*)
```

Events within 120 min on the same `(actor, document, kind)` become one row with `count` and `happenedAtUntil` fields. `COMMENT_ADDED` and `MENTION_CREATED` always start a new session — these kinds never roll up. No hard cap on total session span (a 4-hour transcription sitting is one row). The hour-trunc dedupe SQL is **deleted**, not kept alongside — one aggregation strategy per query.

**URL is universal German `/chronik` across all locales**, matching the existing convention (`/dokumente`, `/personen`, `/briefwechsel`). Content is translated via Paraglide; the URL is a stable German identifier, not a translatable route.

**DTO extended, not replaced.** `ActivityFeedItemDTO` gains `count: int` (required, `1` for singletons) and `happenedAtUntil: OffsetDateTime?` (null for singletons, end-of-session for rollups). One DTO shape serves both the Chronik timeline and the dashboard side-rail.

**`/notifications` route is deleted outright.** The app is pre-production — no 301 redirect, no zombie page.

## Alternatives Considered

| Alternative | Why rejected |
|---|---|
| Fixed 2-hour wall-clock buckets (`date_trunc('hour', happened_at / 2)`) | Splits natural sessions at bucket boundaries (e.g. events at 13:58 / 13:59 / 14:01 land in two rollup rows) |
| Keep `DISTINCT ON hour-trunc` alongside new rollup query | Two aggregation strategies = zombie logic; dashboard and Chronik would drift |
| New `/api/chronik` endpoint that merges both streams | Couples two domains (notifications + audit) at the API layer; composition belongs in `+page.server.ts` |
| Localized URL slugs (`/chronik` / `/chronicle` / `/crónica`) | Breaks the project's existing German-URL convention and adds Paraglide routing overhead for zero UX value |
| Per-locale rollup in the SQL (e.g. align to local-day boundaries) | Timezone-aware SQL is brittle; rollup is a time-gap concept, not a calendar-day concept |

## Consequences

**Easier:**
- One hot path — `/api/dashboard/activity` is backed by a single partial covering index (`V49__add_audit_log_rollup_index.sql`) that matches the rollup query's WHERE clause exactly.
- Dashboard side-rail gets rollup for free — 20 block-saves appear as one "Papa transkribierte 20 Blöcke" row with a time range, not 20 dedup'd hour buckets.
- Component reuse — `ChronikRow.svelte` renders both singleton and rollup variants via a `$derived` discriminator; `DashboardActivityFeed.svelte` consumes the same DTO shape.

**Harder:**
- The session SQL is ~15 lines longer than `DISTINCT ON`. That's the price for not splitting natural sessions at fixed boundaries — worth it on day one.
- Historical `/api/dashboard/activity` consumers now see `count` and `happenedAtUntil`. No breaking change — `count` defaults to `1`, `happenedAtUntil` is nullable — but pre-existing tests needed updating.
- Rollup is load-bearing for the UX — if the index is missing or the query regresses, the page either runs slow or returns duplicate rows. Covered by the rolledUp integration tests and the partial covering index; worth a follow-up Grafana panel on `/api/dashboard/activity` p95 latency.