petal/MULTIUSER_PLAN.md

# Petal multi-user plan

**Status:** Phase 0 (identity plumbing) landed 2026-07-26 in `6901cdb`. Everything
below it is proposed, not agreed.

This document exists to be argued with. Sections marked **OPEN** are real
decisions with real trade-offs, not rhetorical questions — a reviewer should
push back on them. Where I have a recommendation I say so and say why.

---

## 1. Where Petal is today

Single user, by construction. `db.Open` seeds one row (`users.id = 'local'`) and,
until this week, every query named that constant directly.

What was already right: **the schema has been multi-user-shaped from day one.**
`documents`, `tags`, and `vocab_words` all carry `user_id`; `document_versions`
and `suggestions` scope through their parent document. No migration is needed to
support a second user — only a way to know which user is asking.

### What Phase 0 changed

- New `internal/auth`: `Middleware(Resolver)` resolves the caller once per API
  request and stores the id in the request context. Handlers read
  `auth.UserID(r.Context())`.
- `Resolver` is a one-method interface — `Resolve(*http.Request) (string, error)`
  — and is the only thing a real identity provider has to implement.
- `StaticResolver(db.LocalUserID)` supplies today's single user, so behavior is
  unchanged.
- `/api` is split into a public group (`/health`, `/version`) and an
  authenticated group (everything else).
- Two pre-existing access-control gaps fixed: `setStatus` (accept/dismiss) had
  **no ownership check at all**, and `fetchPending` read suggestions by `doc_id`
  alone — which leaked the quoted source sentences.
- Two-user isolation test suites (`docs`, `suggestions`) mount the same routers
  twice behind two resolvers over one database.

**The remaining work is not "make Petal multi-user."** It is "authenticate
someone, provision them, and clean up the three places where data is still
global."

---

## 2. Goals and non-goals

**Goals**
- Two or more people use one Petal instance without seeing each other's writing.
- The existing local user's data survives, attached to a real account.
- Adding a user is an operator action, not a code change.

**Non-goals (explicitly out, unless a reviewer argues otherwise)**
- Sharing, collaboration, or multi-author documents. Petal is a private writing
  space; every feature to date assumes one reader. Sharing would change the
  passport's meaning (authorship evidence) and is a product decision, not an
  auth one.
- Roles, permissions, or an admin UI.
- Public signup. Accounts are provisioned deliberately.

---

## 3. Phase A — authentication

### OPEN #1: forward-auth vs. in-app OIDC

This is the biggest fork and it should be settled first, because everything in
Phase B depends on it.

**Option A — Traefik forward-auth to an Authentik outpost.**
Traefik is already in the deferred deploy bucket. Authentik's outpost terminates
the login, and Petal receives a trusted header (`X-authentik-uid`, plus email and
name). The `Resolver` becomes ~20 lines: read the header, map to a user id.

- *For:* no OIDC library, no session store, no cookie handling, no redirect
  plumbing, no token refresh, no logout endpoint. Petal keeps zero auth code.
  Login/MFA/password reset are entirely Authentik's problem.
- *Against:* Petal is only secure if it is **never reachable except through
  Traefik**. Anyone who can hit the container directly can forge the header and
  become any user. That's a deployment invariant enforced by network config, not
  by code — and it is exactly the kind of invariant that quietly breaks. It also
  makes local development awkward (no proxy → no identity), though
  `StaticResolver` covers that.

**Option B — Petal is an OIDC client itself.**
`github.com/coreos/go-oidc` + `golang.org/x/oauth2`, a `/auth/callback` route,
and a signed session cookie. The config fields already exist
(`AUTHENTIK_URL`, `AUTHENTIK_CLIENT_ID`, `AUTHENTIK_CLIENT_SECRET`,
`SESSION_SECRET`).

- *For:* self-contained and safe to expose directly. No trust-the-proxy
  invariant. Works the same in dev and prod.
- *Against:* meaningfully more code — a session table or signed-cookie scheme,
  CSRF on the callback, token expiry, logout. Two new dependencies in a project
  that currently has exactly two.

**My recommendation: Option B**, but not confidently. The deciding factor for me
is that "must never be reachable directly" is a footgun that survives long after
whoever set it up has forgotten, and Petal already holds someone's private
journals. But if the deployment is definitively a single Traefik-fronted box on a
LAN and will stay that way, Option A is *much* less code and I'd not object.

A reviewer should weigh: how likely is this instance ever exposed beyond the LAN?
Does the operator want Petal to be independently deployable?

### Session handling (assumes Option B)

- Server-side sessions in a `sessions` table (id, user_id, expires_at,
  created_at, user_agent), cookie holds an opaque random id.
  Preferred over signed stateless cookies because it makes logout and
  revocation actually work — worth the one table.
- Cookie: `HttpOnly`, `SameSite=Lax`, `Secure` when `BASE_URL` is https.
- **OPEN #2:** session lifetime. This is a personal writing tool used daily on
  trusted devices; a 30-day sliding expiry is kind, an 8-hour one is
  conventional. I lean long — an editor that logs you out mid-draft is hostile,
  and the auto-save makes a surprise 401 genuinely costly. Needs a decision.

### The 401 problem

Every frontend fetch currently assumes success. Once a session can expire,
**any** call can return 401 mid-session — including the 1.5s auto-save, which is
the one that must not fail silently.

Proposal: a single interceptor in `web/src/api/client.ts` that, on 401, halts
auto-save, surfaces a warm bilingual "请重新登录 / Please sign in again" state
rather than a raw error, and preserves unsaved editor content across the
re-login (localStorage draft keyed by doc id). This is small but easy to forget,
and getting it wrong means lost writing.

---

## 4. Phase B — user provisioning and migration

### Provisioning

On first successful login, upsert a `users` row from the OIDC claims
(`sub` → `users.id`, plus email and name). No signup flow; whoever Authentik lets
in gets an account.

**OPEN #3:** should there be an allowlist? Authentik may host other applications
with a broader user set than Petal should have. A `PETAL_ALLOWED_SUBS` env var,
or an Authentik group check on a claim, would gate it. Probably yes, cheaply.

### Migrating the existing `local` user

The live database on millenia holds real writing under `user_id = 'local'`. That
data must end up owned by the wife's real account.

Recommended: a migration that **renames** rather than copies — update the
`users.id` and let `ON UPDATE CASCADE`… except SQLite FKs here are declared
without `ON UPDATE`, so this needs either a deliberate multi-table update inside
one transaction (`documents`, `tags`, `vocab_words` — versions and suggestions
follow their parents) with `PRAGMA foreign_keys=OFF` around it, or an explicit
one-off admin command.

I'd rather do this as a **documented one-off script run with the app stopped and
a copy of the DB taken first** than as an automatic startup migration, because it
depends on knowing the new OIDC subject id, which isn't available until that
person logs in once. Sequence: deploy auth → she logs in → new empty account is
created → stop app, back up, run script to move `local`'s rows onto her real id,
delete the empty row → restart.

**OPEN #4:** is that acceptable, or is a small admin endpoint preferable to a
script? The existing project convention (`scripts/`, hand-run Python) suggests a
script is in keeping.

---

## 5. Phase C — the data that is still global

Found during the Phase 0 audit. None of these break with two users; all of them
leak or bleed.

### Image store — the real one

`internal/images` is a flat content-addressed directory. There is no per-user
association and no database row at all. Any authenticated user who knows a
sha256 can fetch any other user's image.

That is capability-URL security. Hashes aren't guessable, so this is not an
emergency — but "unguessable filename" is not access control, and images pasted
into a private journal are exactly the content that shouldn't rely on it.

Proposal: an `images` table (hash, user_id, content_type, created_at, size) with
the fetch handler joining on the caller. Content addressing is kept — the same
image uploaded by two users is stored once on disk and simply has two rows, so
deduplication survives. Deleting the last row referencing a hash removes the
file.

**OPEN #5:** is this worth doing before real multi-user, or is it acceptable to
ship auth first and treat this as a known limitation? I lean toward doing it *in
the same phase as auth*, since the moment a second account exists the exposure is
real and the fix requires a migration either way.

### Frontend `localStorage`

`petal.spell.personal` (personal dictionary), `petal.companion` (chosen mascot),
plus sound and petal-effect preferences are all per-browser. Two users on one
device share them — and the personal dictionary is the one that matters, since
it's built from someone's own writing.

Cheapest fix: namespace every key by user id once the client knows who it is.
The honest fix for the dictionary is to move it server-side into a table, which
also means it follows a user between devices — arguably a feature.

### Not affected

`export-all` is correctly scoped. The TTS cache is content-addressed audio of
text the requester supplied, no cross-user inference. The lexicon is a static
dataset identical for everyone.

---

## 6. Phase D — per-user language (and DreamDict)

Tracked here because it lands on the same `users` row and shouldn't be designed
twice.

**English is always the target language.** What varies is the user's *native*
language — the one glosses and explanations are written in. Mandarin ships today;
European Portuguese (pt-PT, explicitly not pt-BR) is wanted; French is possible.

### OPEN #6 is answered: DreamDict

The original worry here was data sourcing — Petal's gloss comes from ECDICT
(English↔Chinese), and a pt-PT equivalent of comparable quality and license
looked like the blocker.

`~/git/dreamdict` already solves it, and more completely than expected. It
supports **en, fr, pt-PT, and zh** (~136k/56k/136k/121k words), and its shape maps
almost 1:1 onto `lexicon.Result`:

| Petal field | DreamDict |
|---|---|
| `Gloss` | `Translate(word, "en", L1)` |
| `Phonetic` | pronunciation (CMU + IPA for en, Wiktionary IPA elsewhere) |
| `Definitions` | `Define(word, lang)` — curated sources ranked above Wiktionary |
| `Synonyms` | `Synonyms(word, lang)` |

It also carries data Petal has no equivalent for and could use: `Antonyms`,
`Frequency`, `Difficulty`, and `Etymology`.

So Phase D stops being gated on data and becomes an integration decision.

### OPEN #6a (new): how to integrate

**Option 1 — HTTP client.** Petal calls DreamDict on localhost:7777, exactly the
pattern already used for Piper TTS (including graceful degradation when it's
down).
- *For:* zero coupling, DreamDict updates independently, all endpoints available.
- *Against:* a second service Petal now depends on at runtime, and the gloss is a
  350ms hover tooltip where "the dictionary service is down" is a visible
  regression from today's always-there embedded data.

**Option 2 — build-time extraction.** A script (sibling to the existing
`scripts/build_gloss.py`) generates Petal's embedded `.json.gz` datasets per
language from DreamDict's `dict.db`.
- *For:* preserves the embedded/offline property exactly; no runtime dependency;
  no architectural change at all.
- *Against:* every language multiplies the binary (the four current gz files are
  already ~11.6 MB); updating the dictionary means rebuilding and redeploying
  Petal; the richer fields are lost unless separately extracted.

**Option 3 — import the package, open `dict.db` read-only.** DreamDict's
`internal/dictionary` is a plain library with `NewReadOnly(dbPath)`, and its only
dependency is `modernc.org/sqlite` — the same CGO-free driver Petal already uses.
Petal opens `dict.db` as a second read-only handle beside `petal.db`.
- *For:* no service, no HTTP, no new dependency, lookups stay local-file fast,
  all four languages at once, and it deletes ~11.6 MB of embedded gz plus the
  ECDICT build scripts. One dictionary, maintained once, shared with GogoBee.
- *Against:* Petal stops being a self-contained binary in the "just run it" sense
  — `dict.db` has to be deployed alongside. In practice Petal already ships a
  data directory (`petal.db`, images, TTS cache), so this is a smaller loss than
  it first sounds.

**My recommendation: Option 3.** It is the only one that gets all four languages,
keeps lookups offline and instant, and *removes* code rather than adding a
subsystem. Option 1's runtime dependency buys flexibility Petal doesn't need for
a dictionary that changes a few times a year.

**Prerequisite:** DreamDict's module path is currently `module dreamdict`, which
isn't fetchable. Importing it needs the module renamed to something like
`gitea.parodia.dev/drwily/dreamdict` (or a local `replace` directive for
development). Small, but it must happen first.

### Migration caution

Whichever option wins, the zh path is **currently working and in daily use**. The
gloss quality difference between ECDICT and CC-CEDICT is unknown and matters more
than the architecture.

Proposal: introduce DreamDict behind Petal's existing lexicon interface as a
*provider*, wire pt-PT and fr to it first (nothing to regress — they don't exist
yet), and keep zh on ECDICT until the two have been compared on real lookups from
her actual documents. Converge only if quality holds. This also de-risks the whole
change: if DreamDict turns out to be a poor fit, only the unshipped languages are
affected.

### Still per-user regardless

Native language becomes a `users` column, and these become per-user lookups:
LLM prompt copy (`internal/llm/prompts.go`, currently Mandarin-first), companion
tips (`tips.ts`), the L1 Piper voice (Piper has pt-PT voices), and the CJK font
stacks (not needed for Latin-script L1). English-side machinery — nspell en-US,
the phonetic dataset, the EN voice — is unaffected and stays shared.

## 7. Suggested sequence

1. **Settle OPEN #1** (forward-auth vs. in-app OIDC). Everything else follows.
2. Deploy plumbing: Dockerfile, Traefik, real hostname, HTTPS. Auth needs a
   stable `BASE_URL` and a redirect URI regardless of which option wins.
3. Auth itself: `Resolver` implementation, sessions if applicable, frontend 401
   handling.
4. Image store table + migration (same phase, per OPEN #5).
5. Provision the second real account; migrate `local`'s data.
6. `localStorage` namespacing.
7. Per-user language. **No longer gated on data** — DreamDict covers all four
   languages. Sequence within it: rename DreamDict's module path → wire it in as
   a lexicon provider → pt-PT/fr first → compare zh quality → converge if it
   holds.

---

## 8. Risks

- **Silent unscoping.** Phase 0 hit this exactly once: `docs.fetch` took a
  `userID` parameter and kept binding `db.LocalUserID` in the query. Unused
  parameters are legal Go — it compiled, `vet` was silent, and every existing
  test passed while the lookup stayed unscoped. Only the two-user isolation test
  caught it. **Every new user-scoped endpoint should get an isolation case in the
  same commit**; the existing suites are the template.
- **Migrating live data.** The wife's real writing is the thing being moved.
  Back up first, run with the app stopped, verify counts before deleting
  anything.
- **A 401 mid-draft losing work.** See Phase A.
- **Scope creep into sharing.** Multi-user and collaboration are different
  products. Adding accounts should not quietly become adding sharing.

---

## 9. Questions for the reviewer

1. Forward-auth or in-app OIDC? (OPEN #1 — the one that matters most)
2. Session lifetime, given a daily-use editor with auto-save? (OPEN #2)
3. Allowlist Petal accounts separately from Authentik's user set? (OPEN #3)
4. Migration script vs. admin endpoint for moving `local`'s data? (OPEN #4)
5. Fix the image store alongside auth, or ship auth with it as a known
   limitation? (OPEN #5)
6. ~~Is there a usable open English↔European-Portuguese dictionary dataset?~~
   **Answered: DreamDict**, which covers en/fr/pt-PT/zh. The live question is now
   *how* to integrate it — HTTP service, build-time extraction, or importing the
   package and opening `dict.db` read-only. (OPEN #6a; I recommend the third)
7. Does replacing the zh gloss (ECDICT → CC-CEDICT) risk regressing a feature in
   daily use, and should zh stay on ECDICT until the two are compared?

Anything above that reads as settled but shouldn't be is also fair game.