ENGINEERING VERSION LOG · 2026.10.04
AI editing settings: choice, route, and recent result
The 2.0.38 source candidate separates the editing switch from style, explains when a request can run, and records what happened recently. It has not been installed or formally released.
01 · USER PROBLEM
“Raw” hid two different decisions.
Previously, Raw represented both “AI editing off” and a text style. A temporary block could leave the selector showing Raw and hide the user’s saved choice. A fixed model label could also imply that a particular provider handled a request, even when no request was made. The candidate gives the user an independent on/off switch and a separate preferred style.
02 · PREFERENCE AND CONSENT
Turning editing off does not erase the preferred style.
The existing execution and roaming setting still saves aiMode=raw when editing is off. A separate preferred style is remembered on this device for the current account; it is not synced through the API. An older non-Raw choice carries forward. If an old Raw configuration contains no style memory, the candidate uses its default style without claiming to have recovered an unknown choice.
Offline, signed-out, or quota-blocked users can still turn editing off or choose a future style while idle. Changing this preference does not grant text-upload permission. Recording and processing hold the session’s effective settings steady; later account or quota refreshes do not rewrite the meaning of that recording.
03 · REAL PROCESSING ROUTE
The screen describes when editing may run.
| Recognition path | AI editing time | Boundary |
|---|---|---|
| Local recognition | Optional API text editing after recording stops | No separate local editing model; recognition audio is not sent for cloud ASR |
| Cloud Live · clean / message | May edit stable segments as they arrive | Shown as a possible route, not proof that a request succeeded |
| Cloud batch · any editing style | Wait for the complete transcript | Actual execution still depends on prerequisites |
| Live instant insertion | No AI editing | Already inserted text is not rewritten |
API text editing still requires sign-in, network access, explicit upload permission, and available AI-edit quota. Text, vocabulary, correction rules, and custom instructions may be sent when an authorized request runs; recognition audio is not sent by the local recognition path. Vocabulary and explicit correction rules are separate from semantic AI editing. The candidate freezes those inputs and the effective style at recording start so a later settings edit cannot change an in-flight request.
04 · WHAT THE HISTORY CAN SAY
Recent result is account-scoped; old history stays unknown.
The current account can see its most recently recorded AI result on this device, including success, failure, or skip where that result was actually captured. Legacy entries without an AI result display unknown. The UI does not infer an AI provider, model, or skip reason from an ASR model, a current preference, or old timing data. Switching accounts clears the previous account’s displayed result. Rapid recordings retain each queued request’s own intent and style, with an account-generation guard against stale updates.
05 · VERIFICATION AND RELEASE
Source checks and isolated layouts pass; real-user acceptance remains open.
The 2.0.38 candidate passed 192 of 194 tests with no failures; two Windows-only checks were skipped on the test host. The TypeScript and Vite build passed. Regression checks cover the separate switch, local account memory, blocked-state editing, legacy unknown status, frozen session inputs, rapid setting changes, and account transitions.
Isolated native Electron checks covered 36 synthetic screens each on Mac and Linux: nine settings states in English and Simplified Chinese at 720 and 1200 pixels. Layout checks passed with no horizontal overflow, labelled visible controls, and switch targets of at least 44 pixels. Selected Mac wide and narrow screenshots were visually reviewed. These synthetic checks do not establish interactive behavior with a real account, microphone, target-app paste, or an installed-user session.
The Mac arm64 candidate passed Developer ID identity, architecture, and strict-signature checks, but it remains unnotarized and uninstalled. The Mac arm64 packaged public-WAV check passed five languages and six pipeline scenarios in a serial recheck. Its initial model startup timed out at 60 seconds; that failure remains recorded, and this is not a claim of stability in every environment. Other platforms and real-user sessions are accepted separately. The layout probes ran in isolated test contexts; a Mac Electron window-lifecycle fix was confined to the probe.