Custom Words: import from other dictation apps (PR-4a/b/c, epic #1619)

[saurabhav88]

Part of #1619 (Smart Import, PR-4a/4b/4c in the adopted v5 plan).

Tier: MEDIUM | Test: required (PostProcessing) | Modules: PostProcessing, AppKit

What

Read the user's vocabulary out of another dictation app they already use, and fold it through the same review-and-commit pipeline as every other source.

Schemas were grep-verified directly on this machine (bible §5.8) and re-confirmed tonight against live data.

Scope: three apps, not four

App	Storage	Live data here	v1
Wispr Flow	SQLite, Dictionary table	9 rows, 6 importable	✅
FluidVoice	JSON, terms[]	1 term, 2 aliases	✅
Superwhisper	JSON, vocabulary[] + replacements[]	2 words, 1 replacement	✅
TypeWhisper	CoreData SQLite	0 rows	❌ deferred

TypeWhisper is deferred deliberately. Its schema is confirmed but the table is empty on the only machine available, so an adapter would be written against a shape nobody has ever seen populated — and its WAL sidecar files add a real correctness question that cannot be tested without real content. Writing it blind and calling it done would be exactly the kind of unverified work this epic has been avoiding.

Read constraints each adapter must honour

• Wispr Flow: filter isDeleted = 0 (soft-delete — an unfiltered read resurrects words the user deliberately removed) and isSnippet = 0 (text expansions are a different feature). Open read-only/immutable: it is another app's live database.
• Superwhisper: probe BOTH ~/Documents/superwhisper/ (legacy, present here) and ~/superwhisper/ (current default). Both keys may be absent entirely on a fresh install, not present-but-empty.
• FluidVoice: read only the terms array; the sibling top-level keys are that app's own ASR tuning parameters, not vocabulary.

Synonyms are not imported

Consistent with the founder's v1 rule: competitor adapters supply the main word only, with every authority field .unspecified. A Wispr Flow entry like btw → by the way imports as the word "by the way"; the mapping itself is not brought across. Existing words are skipped and never modified.

This is an honest partial migration, and the screen says so rather than implying a full one.

Nothing is read until asked

No directory is touched at launch or on opening the sheet. Detection and reading happen only after the user picks an app, so an installed competitor is never quietly inspected in the background.

DoD

Each adapter tested against a fixture of that app's real shape · absent app reports "not found" rather than an error · present-but-empty reports nothing to import · another app's live database is never opened for writing · adds no second entry point · full suite + Debug/Release green.

[saurabhav88]

Ready to merge, blocked on the cloud review gate

PR #1693 is open, rebased onto the shipped import contract, and all four CI checks are green. Local Codex is clean over two rounds.

Not merged, because the GitHub Codex review never arrived: two @codex review requests, two poll-codex-review.sh runs of ~30 minutes each, both exiting no_review. The bot is not misconfigured — it reviewed #1687 earlier the same evening. GitHub's API was returning 503s during the first window, so the likely cause is that outage or a backlog behind it.

Per github.md RULE: codex-cloud-review-gate, no_review is not a merge-go. Leaving it for a review rather than merging on a silent gate, since this PR reads files belonging to other applications.

Same situation for #1694 (#1690's scratch-file fix).

To merge once a review lands: gh pr merge 1693 --squash --auto --delete-branch.

[chatgpt-codex-connector]

Review Result

• I checked out PR feat(user): import words from another dictation app (PR-4a/b/c, #1686) #1693 locally and reviewed the smart-import implementation. I did not find a code issue that needs a follow-up patch.
• No repository files were changed, so I did not create a commit and did not call the PR-creation tool.
• The import adapters are gated so disk access happens only after the user enters the app picker / starts the selected import path: adapter installation checks are driven from the smart app picker task, and loading happens through SmartImportSource.loadRawCandidates().
• Wispr Flow reads use an immutable SQLite URI after refusing any WAL/SHM sidecar presence, then re-check sidecars after the read to avoid presenting stale or incomplete data as successful.
• Wispr Flow SQL filters soft-deleted entries and snippets, and reads one past the import limit so over-limit imports can be rejected before deduplication.
• FluidVoice and Superwhisper stay within their intended vocabulary fields, with Superwhisper probing the current default before the legacy path.
• Export hardening for the live custom-words file and atomic 0600 publication remains in place.

Testing

• ⚠️ swift test --filter SmartImportSourceTests — attempted, but this Linux container cannot compile the macOS-only dependency path because MachTaskSelf.c includes mach/mach.h, which is unavailable here.
• ✅ git status --short --branch — working tree is clean on pr-1693.

View task →