Teknium
50ea107d9d
With `voice.auto_tts` on and HUD mode active, a reply could be spoken twice: the HUD renderer and the hidden app window both claim `speak:<messageId>` through `hermes:ambient:claim`, and main collapsed the two claims with the same 1 s deduper it uses for the turn-end beep. The app window under the HUD is throttled by Chromium, so its transcript subscription fires well past 1 s, finds the key pruned, and is told it owns the cue. Fixes #99717. The unit was wrong, not the window: a spoken reply is minutes of audio keyed by a durable message id. `createAmbientClaimArbiter` keeps the tick-sized window for `sound:*` and holds `speak:*` claims for a 10-minute TTL. No HUD special-casing: the same race exists between any two windows showing the same chat, and the arbiter fixes them all. Diagnosis credit: #99810 (@liuhao1024) and #100289 (@shivamjg101), whose TTL framing this follows.
51 lines
2.0 KiB
TypeScript
51 lines
2.0 KiB
TypeScript
// Cross-window de-dupe for one-shot side-effects (OS notifications, the turn-end
|
|
// sound, spoken replies). Every desktop window is its own renderer process, so N
|
|
// open windows each independently react to the same backend event. The main
|
|
// process is the one place they all share and it handles IPC serially, so it's
|
|
// the race-free owner: the first window to claim a key within the interval wins;
|
|
// peers see it's taken and stay quiet. Pure + injectable clock, so it's
|
|
// unit-testable without Electron.
|
|
|
|
const DEDUPE_INTERVAL_MS = 1000
|
|
|
|
// Returns true when `key` was already claimed within the interval (caller drops
|
|
// this one). Self-evicting: stale keys are pruned on every call, so the map
|
|
// can't grow unbounded.
|
|
export function createEventDeduper(intervalMs = DEDUPE_INTERVAL_MS) {
|
|
const lastSeenAt = new Map<string, number>()
|
|
|
|
return function isDuplicate(key: string, now = Date.now()): boolean {
|
|
for (const [k, at] of lastSeenAt) {
|
|
if (now - at >= intervalMs) {
|
|
lastSeenAt.delete(k)
|
|
}
|
|
}
|
|
|
|
if (lastSeenAt.has(key)) {
|
|
return true
|
|
}
|
|
|
|
lastSeenAt.set(key, now)
|
|
|
|
return false
|
|
}
|
|
}
|
|
|
|
// A `speak:<messageId>` cue is seconds of audio keyed by a durable backend
|
|
// message id, not an instant beep: the peer's claim can arrive well past the
|
|
// 1 s window (the app window hidden under an open HUD is throttled by Chromium
|
|
// and its transcript subscription fires late), and the reply was read twice
|
|
// (#99717). One reply is one claim for as long as a reply can plausibly play.
|
|
export const SPEECH_CLAIM_TTL_MS = 10 * 60_000
|
|
|
|
// Cross-window arbiter for every ambient cue: `speak:*` keys hold for the
|
|
// speech TTL, everything else keeps the tick-sized window.
|
|
export function createAmbientClaimArbiter(intervalMs = DEDUPE_INTERVAL_MS, speechTtlMs = SPEECH_CLAIM_TTL_MS) {
|
|
const cues = createEventDeduper(intervalMs)
|
|
const speech = createEventDeduper(speechTtlMs)
|
|
|
|
return function owns(key: string, now = Date.now()): boolean {
|
|
return !(key.startsWith('speak:') ? speech(key, now) : cues(key, now))
|
|
}
|
|
}
|