# Roadmap

Where Rollup is headed, relative to the [product contract](CONTRACT.md) and
[competitive position](COMPARISON.md). Guidance for contributors, not a
commitment calendar.

Rollup reduces a chosen information environment into a **bounded reading
object**. The next work should deepen that briefing — not add another stream
to triage.

Existing sources already fit the model. Thunderbird owns the ongoing
newsletter stream; Reddit, LinkedIn, and article capture supply extra
material at digest time; Rollup takes a window and produces the rollup.

## Product sequence

**Toward 1.0 — make the briefing the product, then stop expanding surface.**

1. **Documentation consolidation.** README is the product entry; `docs/` is the
   reference. The contract stays short and authoritative. Remaining work is
   keeping the split honest as features land.
2. **Attention / relevance / skip-safely.** Explicit interest profiles (simple
   prose, not dozens of weights) that rank or annotate entries: high relevance /
   worth scanning / background / safely skip. **Preserve deterministic
   inclusion** — ranking changes attention, never silently drops material.
   Later, existing Quality / read / save / dismiss signals can inform the
   model.
3. **Cross-source consolidation.** Detect redundancy, overlapping coverage,
   and recurring themes across newsletters, LinkedIn, Reddit, and articles in
   the same window. The briefing should say “this showed up three ways” rather
   than reprint it three times.
4. **Briefing-level synthesis.** Beyond per-item summaries and final-review
   QA: a “what actually mattered this week?” layer over the collection window.
   The output stays a finished reading object, not a ranked feed.
5. **Search the accumulated corpus.** SQLite FTS over cached reader bodies,
   titles, and summaries. Then optionally a tightly grounded “Ask Rollup” over
   that corpus. Makes the archive useful after the weekly reading event without
   turning Rollup into a notes app. Reader bodies stay a convenience cache.
6. **Feedback → source intelligence.** Turn existing Quality / rating data into
   suggestions: unused for eight weeks, mostly duplicate, overlapping coverage,
   consider lowering priority. Suggestions only; never auto-unsubscribe or
   mutate Thunderbird. The source registry is the right home.
7. **Reading destinations.** The writer seam is already the right extension
   point. OPDS for e-readers; then optional send-to-Kindle / Calibre-style
   delivery as plugins. Reinforces “produce something worth reading away from
   the computer” rather than another infinite-scroll app.

**Then 1.0.**

**After 1.0 — Gmail ingestion.** Optional Gmail API (OAuth) as a read-only
alternative to local mbox, aligned with the [product contract](CONTRACT.md).
Thunderbird mbox remains primary through 1.0. Gmail expands the addressable
audience; doing it earlier risks becoming another generic inbox reader.

## Shipped (recent)

- Optional TOML sticky config, run profiles, effort presets, folder themes ([CONFIG.md](CONFIG.md))
- Output-writer plugin seam + builtins (xteink, txt, json, epub) ([OUTPUT_WRITERS.md](OUTPUT_WRITERS.md))
- Loopback web UI: Archive, Quality, Registry, Admin, reader bodies ([WEB.md](WEB.md))
- Configuration Centre (`/settings`) and Run Studio (`/run`) on the real TOML + CLI digest path
- Shared `sticky_flags` registry; `cli_parser` extraction; `run_digest` phase helpers
- Optional LinkedIn `fromMember` folders, Reddit subreddits, webpage article queue
- **Reddit + LinkedIn listing cache** (`rollup.db` schema v15): persist network listings; `--reddit-refresh` / `--linkedin-refresh` force a live pull
- **Firefox Add to Rollup**: toolbar/context-menu capture into `webpage_queue` via `POST /articles/capture` (Bearer token in `{state_dir}/extension_token`). Ingest-only; `rollup web` must be running.
- **Google Scholar detailed mode**: `[scholar].mode = "detailed"` / `--scholar-mode detailed` fetches paper landing pages from Scholar alert emails and summarises each paper (schema v16 cache). Default mode keeps alerts as `item_list` mail.

## Standing rule (network sources)

Persist fetched payloads in `rollup.db` and avoid re-calling expensive or
rate-limited APIs when a fresh-enough snapshot exists.

Network sources are **ingest transports** for the collection window: fetch at
digest time, fold into the briefing, do not grow an inbox.

## Parked (not the 1.0 arc)

These expand surface. They wait until the briefing itself is distinct.

- **RSS as ingest transport** — “include this publication in my weekly
  rollup”; fetch the lookback window at digest time; never an RSS inbox,
  unread state, or continuous feed. Same conceptual slot as LinkedIn / Reddit
  today. An RSS *reader* is a [non-goal](#non-goals-still).
- **Richer LinkedIn sources** beyond faceted `fromMember` search — keyword SRP
  if LinkedIn exposes a working content-search query again; company/org posts;
  follows/mentions. v1 stays author-list URLs mapped through Voyager
  `profileUpdatesV2`.
- Engineering hygiene when the next feature forces the edit: `state.py` /
  `source_registry.py` splits; `render.py` MD/HTML shared structure.

## Non-goals (still)

From [CONTRACT.md](CONTRACT.md). Do not prioritise these; existing non-goals
are helping the project. Output artifacts are a cheaper path to cross-device
reading than a mobile app.

- IMAP / Maildir backends (Gmail API → post-1.0 above)
- Thunderbird add-on (XPI)
- Multi-user or non-loopback web UI
- In-app digest scheduler (Run Studio stays a synchronous guided runner; use [CRON.md](CRON.md) / launchd)
- Exposing classifier thresholds as user knobs
- Built-in workflow engine or generic plugin-everything infrastructure
- Richer email-client operations (Rollup is not a mail client)
- Official mobile app
- An RSS **reader** (unread counts, folders, stars, sync, mark-all-read,
  scrolling feeds). That is Miniflux / FreshRSS / Readeck territory.

## Open product questions (parked)

- Richer onboarding when mail paths are undiscoverable
- Optional remote Ollama remains explicit (`--allow-remote-ollama`); optional LiteLLM via `rollup[llm]` + `--llm-provider litellm` (API keys from env only)
- Deeper Admin failure history when manifests are missing (incomplete-history disclaimer stays)
- LinkedIn HTML/API fragility — isolated fetch module; may break when LinkedIn changes

## Related docs

| Doc | Role |
|-----|------|
| [CONTRACT.md](CONTRACT.md) | Product shape and publication integrity |
| [COMPARISON.md](COMPARISON.md) | Competitive position |
| [CONFIG.md](CONFIG.md) | TOML, profiles, sticky ↔ CLI |
| [WEB.md](WEB.md) | Loopback UI surfaces |
| [DOCKER.md](DOCKER.md) | Default Compose deploy |
| [EXAMPLES.md](EXAMPLES.md) | Runnable recipes |
| [CHANGELOG.md](../CHANGELOG.md) | What shipped |
| [design/firefox-capture.md](design/firefox-capture.md) | Firefox Add to Rollup (packages, SLOC, reload, known issues, review) |