Local-only strip: audit

Status: done. The strip landed on 2026-09-19. The plan below is kept as the record of what was removed and why; see Outcome for what it actually cost.

Dictate is no longer a commercial product. It is a local-first dictation app worked on when there is time. Everything that exists to support accounts, subscriptions, hosted transcription, cloud sync and cloud deployment is dead weight and comes out.

The goal after the strip: a nice working package for local dictation, so a user installs one thing and it works, instead of assembling a UI and a model runtime themselves. No account. No API key. No network dependency in the transcription path.

Outcome #

The strip removed 27,064 lines across 107 files, deleting 53 files outright — more than the 20,020 lines across 50 files estimated below.

The estimate was low for three reasons the audit did not anticipate:

1. The front-end was not counted. ui/ contributed 2,201 lines, dominated by AccountDialog and its tests.
2. Deleting a module means fixing every test that touched it. The tests row is nearly as large as src/, and most of that was editing surviving tests rather than deleting cloud ones.
3. CI referenced the deleted code. Three workflows ran the cloud-sync audit and volume smoke, invoked deleted test modules by name, and carried a whole linux-user-sync-smoke job that package, publish and release gated on.

Two things had to be kept that the audit listed for deletion. The update UI — version rows, the Stable/Beta channel selector, the update actions — lived inside AccountDialog, so that dialog was replaced by a smaller AboutDialog rather than deleted. And scripts/msstore-submit.py stayed once Microsoft Store distribution was confirmed as staying.

One audit claim was wrong: whisper-cpp was listed as a surviving local backend, but whisper_cpp_backend.py has no BACKEND_REGISTRY entry, no DEFAULT_MODELS entry and is absent from the SttBackend literal. It is not selectable, and making it so would be new work, not preservation.

After the strip: 589 Python tests and 85 UI tests pass.

Headline (as estimated, before the work) #

Removing the cloud surface deletes roughly 20,000 lines across 50 files, about 36% of src + tests combined (55,257 lines today). A further ~10,500 lines across 12 files need editing rather than deleting.

There are no existing users on the cloud side, so nothing needs migrating, deprecating or keeping alive for compatibility. The code can just come out.

This is not only "delete the cloud code". A large part of the win is that machinery which exists because the cloud can fail also stops being needed.

Tier 1 — delete wholesale #

These have no local-only purpose. Nothing survives them.

scripts/msstore-submit.py (307 lines) is not in this list. Microsoft Store submission stays — see Tier 3.

Why provider_supervisor.py goes #

Its own docstring says it "handles graceful degradation when the remote fails, and schedules automatic recovery probes using class-aware exponential backoff", with on-device as "the always-available floor". With no remote, there is no degradation to manage and no probe to schedule — the floor becomes the whole building. 387 lines of supervisor plus 1,188 lines of tests exist purely to survive a network that local-only never touches.

This is the shape of the whole change: the cloud did not just add features, it added failure modes, and the code handling those failure modes is larger than the features were.

src/dictate/pro/ detail #

store.py alone is 80 KB. With client.py (51 KB), service.py (36 KB) and server.py (31 KB), four files carry most of the weight. stripe_handler.py, plans.py, browser_auth.py, email_delivery.py and signing.py are subscription and identity plumbing with no local equivalent.

Note dictate-pro-server is a console entry point in pyproject.toml — a hosted server shipped inside a desktop app. That goes with the package.

Tier 2 — edit, do not delete #

These are load-bearing for local operation but currently reference cloud paths.

ui/src/App.jsx is the front-end equivalent — 185 lines match account / sync / billing / API-key identifiers, and ui/src/ipc.js has 51. productDestinations.js goes entirely.

Tier 3 — docs to retire #

Move to docs/archive/ rather than delete, so the history stays readable:

• DICTATE_PRO_CLOUD_SYNC_PLAN.md
• dictate-pro-subscription-architecture.md
• desktop-browser-signin-architecture.md
• dictate-pro-terms.md
• msstore-in-app-subscriptions.md — the subscription mechanism only; the Store channel itself stays
• deck-sections-and-dictate-space-spec.md
• platform/dictate-platform-inventory-v1.md
• platform/arc-forge-backend-handoff-v1.md
• contracts/dictate-platform-v1.json (2,048 lines)

dictate-privacy-policy.md needs a rewrite, not an archive: with no account and no network transcription path, the policy becomes short and much stronger.

What stays #

The local transcription path is already first-class, which is why this strip is feasible rather than a rebuild:

• stt/parakeet_backend.py, parakeet_pyannote_backend.py, parakeet_speaker_backend.py — local Parakeet and diarisation
• stt/faster_whisper_backend.py, whisperx_backend.py (whisper_cpp_backend.py exists but is not registered — see Outcome)
• Audio capture, hotkey, tray, history, lexicon, note chunking, outputs
• Installers, updater, packaging
• resolve_default_local_backend / resolve_default_local_model in stt/factory.py — the local default already exists

Runtime dependencies in pyproject.toml need no change: they are already local-ML (faster-whisper, onnx-asr, onnxruntime, numpy). There is no Stripe or web-framework dependency to drop, because the hosted server was built on the standard library. cryptography was there only for sync record encryption and has been dropped: after the strip nothing in the repository imports it.

Decisions #

Microsoft Store stays. It remains the intended public Windows channel. The listing is overdue for an update and gets one once the first stable local-only build is done. msstore-publish-msix.yml, msstore-api-smoke.yml, windows-msix-store-bundle.yml, scripts/msstore-submit.py, msstore-automation.md and msstore-listing.md all stay; the listing copy needs rewriting for local-only positioning when that update happens. Only the in-app subscription content goes, since there is nothing to subscribe to.

No migration needed. There are no existing users on the cloud side, so there is no saved stt_backend: openai|gemini|xai config in the wild to fall back gracefully from, and no stored API keys in an OS secret store to clear. Both of these would otherwise have been required work; neither is. The cloud code can be deleted outright rather than deprecated.

Open question #

history/note_store outbox. Both take SyncOutbox. Confirm removing it leaves local history intact — it should, but the write path needs reading before cutting. This is answerable from the code and does not block starting.

Suggested sequence #

Each step should land green on its own.

1. Delete pro/ and its tests; drop the dictate-pro-server entry point.
2. Delete cloud STT backends; collapse factory.py onto the local default.
3. Delete provider_supervisor.py; unwire from daemon, ui_server, ui_launcher, __main__.
4. Delete sync; unwire SyncOutbox from history.py and note_store.py.
5. Delete API-key management; strip the tray, control panel and UI surfaces.
6. Archive docs; rewrite the privacy policy; drop the in-app subscription doc.