feat(voice): add send/edit/cancel confirmation buttons for voice transcription

[DovyLive173]

Summary

Replace the auto-send flow for voice transcriptions with an interactive confirmation step. After transcription, the user now sees a message with inline buttons to Send, Edit, or Cancel the recognized text before it is sent to OpenCode.

Changes

New file: src/bot/handlers/stt-confirm.ts

• showSttConfirmation() — displays the recognized text with Send/Edit/Cancel inline keyboard
• handleSttConfirmCallback() — handles button callbacks:
  • Send — applies the optional STT_NOTE_PROMPT and forwards to processUserPrompt
  • Edit — transitions the interaction to expectedInput: "mixed" (accepts both text and a new voice recording) and prompts the user to send corrections
  • Cancel — clears the interaction and marks the voice message as cancelled
• handleSttEditText() — receives corrected text (or a re-recording routed through the voice handler) and sends it to OpenCode via processUserPrompt

Modified: src/bot/handlers/voice.ts

• Simplified: removed the inline processPrompt call and processPrompt dep injection in favor of showSttConfirmation
• Removed unused FilePartInput import and processPrompt option from VoiceMessageDeps

Modified: src/interaction/guard.ts

• Inside the expectedInput === "mixed" block: added an exception to allow inputType === "other" (voice/audio messages) when the active interaction is a custom STT interaction (kind === "custom" with sttTranscript metadata). This enables re-recording during edit without being blocked by the guard.

Modified: src/bot/index.ts

• Registered handleSttConfirmCallback in the callback query handler chain
• Registered handleSttEditText before processUserPrompt in the text handler to intercept edited text

i18n

• Added 9 new keys (stt.confirm_message, stt.confirm_send, stt.confirm_edit, stt.confirm_cancel, stt.confirm_sending, stt.confirm_edit_prompt, stt.confirm_edit_sending, stt.confirm_cancelled, stt.confirm_inactive) to all supported languages: en, de, es, fr, ru, zh

Tests

• Updated tests/bot/handlers/voice.test.ts:
  • Removed tests for the old auto-send flow
  • Added verification that the confirmation interaction is created after transcription
  • Updated existing tests to check for confirmation interaction state instead of direct processPrompt calls

How it works

User sends voice message
  → Voice handler transcribes audio
  → Status message updated with recognized text + inline buttons
  → User chooses:
     • Send → prompt sent to OpenCode (respects STT_NOTE_PROMPT)
     • Edit → user can send corrected text OR a new voice recording
               → New recording is re-transcribed and shows confirmation again
               → Text is sent directly to OpenCode
     • Cancel → transcription discarded

Testing

• npm run build — passes
• npm run lint — passes (0 warnings)
• npm test — passes (942 tests, 111 files, all passing)
• Manual testing with a Telegram dev bot verified:
  • Full confirmation flow (Send / Edit / Cancel)
  • Text correction during edit
  • Voice re-recording during edit (new transcription shows confirmation again)
  • Interaction expiry and inactive confirmation handling