Add long-press-to-talk on session search box (uses configured dictation provider)

[heavygee]

Context

Session search (web/src/components/SessionList.tsx) already got a clear/X button via #1544/#1545. A prior local attempt at long-press-to-dismiss on the same control stalled before ever reaching a PR (worktree session-search-longpress-dismiss, uncommitted, superseded by the X button) but its useLongPress wiring on the collapsed search trigger is reusable groundwork.

Ask

Repurpose that wiring: long-press the collapsed search trigger should start realtime dictation into the search query, using whichever dictation provider the operator currently has configured in Settings (useVoiceInputPreferences.ts / realtimeTranscription.ts) — cloud or Browser on-device, whatever is selected. Do not force or special-case any specific provider.

Scope

• Reuse worktree ~/coding/hapi/worktrees/session-search-longpress-dismiss (branch fix/session-search-longpress-dismiss) — has existing uncommitted useLongPress wiring on the search trigger button to build from.
• Change the long-press action from "clear query" to "start dictation" (short tap still expands to text search as today).
• Long-press should be enabled regardless of whether there is already a query (unlike the old dismiss-only gating).
• Reuse the existing realtime dictation hooks (useDictation/useRealtimeDictation) rather than building new STT plumbing — this is a UI wiring task, not a new STT backend.
• Standard gates apply: typecheck, tests, Playwright + visual evidence before operator dogfood.

Not in scope

• The mobile fail-closed STT restriction (tracked separately in Re-examine mobile fail-closed restriction on Browser on-device (native SpeechRecognition) STT #1593) — do not change browserLocalSpeech.ts gating as part of this work.

[github-actions]

Confirmed: this can reuse the existing dictation stack without adding another STT backend or branching on individual providers.

The provider-neutral path is useVoiceInputPreferences(api) → useDictation(...). It resolves the stored provider and transcription mode, includes Browser on-device when supported, and appends the final transcript to the current text. The collapsed search button can use the native-button long-press interaction so a short tap keeps its current expand behavior.

One verified constraint: configured providers do not all support realtime mode. OpenAI, ElevenLabs, and Deepgram support standard + realtime; Browser on-device is realtime-only; Groq and OpenAI-compatible are standard-only. Calling useRealtimeDictation directly therefore cannot support every selected provider. useDictation supports all of them by honoring the configured transcription mode.

The shared SessionListSearch is also used by the share-target picker, so the dictation affordance should be enabled explicitly for the main session list unless both surfaces are intended to change.

Relevant code:

• web/src/components/SessionList.tsx:700 - reusable session-search component and props
• web/src/components/SessionList.tsx:793 - collapsed search trigger; short click currently expands search
• web/src/components/SessionList.tsx:1162 - main session list already receives the API client and owns search state
• web/src/hooks/useLongPress.ts:238 - native-button mode preserves normal click behavior and suppresses the post-long-press compatibility click
• web/src/hooks/useVoiceInputPreferences.ts:42 - resolves stored provider/mode and Browser on-device availability
• web/src/hooks/useDictation.ts:30 - provider-neutral dictation wiring and transcript append
• web/src/hooks/useDictation.ts:168 - selects realtime or standard implementation from the configured mode
• web/src/hooks/useRealtimeDictation.ts:27 - existing realtime implementations for OpenAI, ElevenLabs, Deepgram, and Browser on-device
• shared/src/voice.ts:266 - provider capability/mode definitions
• web/src/routes/share/index.tsx:327 - second consumer of SessionListSearch

Need more info:

• Should releasing the long press stop and commit transcription (push-to-talk), or should long press toggle recording until another action stops it? The current long-press hook reports the threshold action but has no release callback.
• Should search honor the configured transcription mode via useDictation (supports every provider), or must this be realtime-only when a standard-only provider is selected?

---

HAPI Bot