path: root/tmp-decisions.md

# Client architecture — decisions

> Created 2026-08-13 after the second security review. D1/D2/D3 decided the same day;
> D4 (hub minimization) deferred. Fold into `docs/meshbay-draft-v5.md`.
> The analysis below is kept as the rationale behind the decisions, not as open questions.

---

## Status

| # | Decision | State |
|---|---|---|
| D1 | Does the hub keep serving the web UI? | ✅ **DECIDED 2026-08-13 — yes** |
| D2 | Browser extension, native desktop client, or both? | ✅ **DECIDED 2026-08-13 — native client, offered alongside the hub-served SPA.** Shell revised 2026-08-17: **Electron**, see `docs/desktop-client-v1.md` |
| D3 | Transport: aiortc primary, QUIC at parity, TCP+HTTP removed | ✅ Decided 2026-08-13. Unchanged for the **node**; the desktop client uses Chromium's WebRTC rather than aiortc, and QUIC via a Python sidecar |
| D4 | Hub minimization (old Phase 12) | ⏸️ **Deferred, may be dropped** |

> **2026-08-17 — the browser-extension options are no longer live.** 12.2 (signed digest
> of the served bundle) is what Option A needed, and it remains scheduled; nothing else
> from A or B is planned. The desktop client is the answer to T3 for users who want one.

**What was decided.** The hub keeps serving the web UI — that is the zero-install path
and it stays. A native desktop client is offered *in addition*, not as a replacement.
Hub minimization is off the critical path and may be dropped entirely.

**What that means, stated once and then respected.** Keeping the hub in the trusted path
is a legitimate product call, and this project is not obliged to defend against its own
operator. But two consequences should be carried deliberately rather than by accident:

1. **T3 is accepted permanently for browser users.** A hub that serves the code can
   exfiltrate keys from the page regardless of what the protocol does. The native client
   gives users who care an alternative; browser users are trusting meshbay.org, and the
   docs should say so plainly rather than claiming end-to-end integrity.
2. **H3 was the last open High finding and its only fix lived in the dropped phase.**
   The hub is the public key directory: substituting a key during an invite hands it the
   group key, silently, with no forgery and no code injection. So key transparency and
   safety numbers were kept and are now Phase 12.1 — everything else from hub
   minimization is dropped. If Phase 12 is later dropped too, H3 stays open by choice,
   and "unreadable by other parties, even the hub" stops being a claim the project can
   make about an adversarial hub.

The honest framing that survives all of this: **the hub cannot read your content unless
it actively attacks you.** That is still a strong property, and it is defensible.

## Rationale — why the native client is not a T3 fix

The second review recommended a native client and claimed *"T3 disappears — code integrity
stops depending on the hub."* **That claim was wrong and has been corrected** in
`second-review.md` §9.

If the hub operator is the adversary, a native client downloaded from `meshbay.org` and
signed with a key that operator holds does not remove the trust — it relocates it from "the
JS they serve" to "the binary they serve." What actually changes is **detectability**:

- **Browser:** an attack is one HTTP response, targeted at one user, leaving no artifact.
  Undetectable in principle.
- **Native:** an attack requires shipping a build. That build is an artifact — hashable,
  archivable, comparable between users, reversible. Targeting one user means giving them a
  different binary, which reproducible builds and published hashes make detectable.

That is a real improvement, but **the value lives in the verification machinery
(reproducible builds, published hashes, independent rebuilds — Phase 18.7), not in the
packaging format.** Without it, a native client from meshbay.org is only marginally more
trustworthy than the SPA from meshbay.org.

Native also has real costs that were under-weighted: loss of the browser sandbox (a Python
process with full user privileges vs a seccomp-confined renderer), ownership of patch
velocity for WebKitGTK and every bundled dependency, and new attack surface (loopback media
server, IPC bridge, update client).

**Conclusion recorded:** the native client is justified on *product* grounds — durable keys,
no browser tab, background connectivity, better video, hub-less `group://` access over QUIC.
It should not be justified as the fix for T3 unless 18.7 ships with it.

---

## D1 rationale — hub keeps serving the UI ✅

Keeping it is defensible. It is how anyone tries the platform without installing anything,
and it stays the fallback when a device has no client installed.

What must be true now that it stays (Phase 12.2/12.3):

- strict CSP and Subresource Integrity on the bundle
- the hub publishes a **signed digest** of the served bundle, so any third party — an
  extension, a native client, a curious user — can verify it
- `/app/` carries an explicit, visible "reduced trust: this hub serves this code" notice
- the docs never claim end-to-end integrity for the hub-served SPA path

The honest framing: hub-served SPA is a **convenience tier**, not the secure tier.

---

## D2 rationale — native chosen; extension not taken up

Three shapes, cheapest first:

**Option A — Extension as a verifier (hub still serves the UI)**
The extension does not ship the UI. It hashes the bundle the hub served and compares it
against a digest signed by the project. Mismatch → visible alarm, optionally block.
Converts a silent targeted injection into a loud one. Small effort, keeps today's
architecture, compatible with D1 = yes.

**Option B — Extension ships the UI (hub serves the API only)**
The UI lives in the extension, distributed and signed by Mozilla/Chrome — a channel **the
hub operator does not control**. Manifest V3 forbids remote code, which works in our favour:
the structure enforces exactly what we want. Keys live in extension storage, isolated from
page JS. Moderate effort.

**Option C — Native desktop client** ← **CHOSEN**
Phase 13. Full control, durable keys in an OS keystore, QUIC, hub-less access, best UX.
Highest effort, and the security argument depends on 18.7.

> **Shell revised 2026-08-17: Electron, not pywebview** (+ an optional Python sidecar for
> `group://` over QUIC). See `docs/desktop-client-v1.md` §2. The comparison table below
> was written against pywebview and **two of its rows are wrong for the chosen shell**:
>
> - *Browser sandbox* — Electron with `sandbox` and `contextIsolation` **keeps** the
>   Chromium renderer sandbox. The row reads "❌ None (partial under Flatpak)"; that was
>   WebKitGTK's property, not native's.
> - *Crypto available* — the client keeps WebCrypto **and** gains local Argon2id/ChaCha20
>   in the main process; the row implied WebCrypto is lost with the browser engine.
>
> The row that gets *worse*: **patch velocity**. "You own it" is more true of Chromium
> than of a distro-maintained WebKitGTK, and the renderer parses attacker-controlled
> content from nodes. That is the real cost and it is accepted deliberately.

### Comparison

| | Hub-served SPA (today) | A: extension verifies | B: extension ships UI | C: native desktop |
|---|---|---|---|---|
| Code distribution channel | Hub (the adversary) | Hub, but **verified** | Store (independent) | Hub download + own signing key |
| Silent targeted injection | Undetectable | **Detected** | Not possible | Detectable *if* 18.7 |
| Browser sandbox | ✅ Full | ✅ Full | ✅ Full | ❌ None (partial under Flatpak) |
| Patch velocity | Browser auto-updates | Browser auto-updates | Browser + store review latency | **You own it** (WebKitGTK, Python deps) |
| Key storage | IndexedDB, page-reachable | unchanged | Extension storage, page-isolated | **OS keystore** |
| Crypto available | WebCrypto only (no ChaCha20/Argon2id) | unchanged | unchanged | **Full** (ChaCha20, Argon2id 256 MB) |
| Transport | WebRTC | WebRTC | WebRTC | WebRTC **+ QUIC** |
| Large file → disk | Chrome only (FS Access API) | unchanged | unchanged | **Native, unlimited** |
| Hub-less `group://` | ❌ | ❌ | ❌ | ✅ |
| New attack surface | — | negligible | negligible | loopback server, IPC, updater |
| Platforms to maintain | 0 | 2 stores | 2 stores | 3 OSes |
| Effort | 0 | Low | Moderate | High |

### Observations for the decision

- **A and B are not exclusive with C.** A/B protect browser users; C serves users who want a
  real application. "Both" is coherent — just sequence them.
- **B gives most of C's security benefit at a fraction of the cost**, because the win was
  never the packaging format — it was getting the code off the adversary's distribution
  channel — and the extension keeps the browser sandbox while doing it.
- **Store review latency is the one place B is worse than C**: a critical fix waits on
  Mozilla/Google. Mitigate with a version-pinned kill switch.
- **For node operators specifically, Phase 14 (CLI) beats all three.** The operator is the
  highest-value target — holds the GEK, is the content authority — and today must use
  hub-served JS to initialize GEKs and invite members. The CLI removes that dependency at a
  fraction of any client's cost. If only one thing gets built for T3, it should be the CLI.
- If **D1 = yes** (hub keeps serving the UI), Option A is the natural companion and is nearly
  free once 12.6 publishes the signed digest.

### Not yet investigated

- Whether AMO/Chrome Web Store policy accepts an extension whose purpose is P2P file sharing
- Whether MV3 service-worker lifetimes can hold a long-lived WebRTC DataChannel (may need an
  offscreen document — worth a spike before committing to B)
- Safari/iOS: no extension route comparable to AMO; likely out of scope either way

---

## Impact on the roadmap

| Decision | If yes | If no |
|---|---|---|
| D1 hub serves UI | Phase 12.6 as written (CSP, SRI, signed digest, reduced-trust notice) | 12.6 shrinks to removing `/app/`; hub becomes API-only |
| D2 = A | Small new phase; 12.6 is a prerequisite | — |
| D2 = B | New phase, ~Phase 13-sized; 13.1 platform split is reused directly | — |
| D2 = C | Phase 13 as written (13.2–13.11), + 18.7 for the security claim to hold | Phase 13 reduces to 13.1 only |

**13.1 (platform adapter split) is worth doing regardless of D2.** It is pure refactoring
whose acceptance criterion is "the browser SPA behaves identically," and it is the
prerequisite for A, B and C alike.