Reasonix Voice Input

Voice-input mic for Electron chat apps: Gemini transcribes, inserts and sends (auto or manual), with attachments, queue and a hidden right-click menu. Ported to Reasonix Studio, Cline Desktop, Open Cowork and AionUi.

Start here: Tutorial · Tutoriel (FR).

Mockup — no personal data.

Unofficial project. Not affiliated with, endorsed by, or supported by Reasonix Studio or DeepSeek. It patches a local copy of the app's static UI, is provided as is, without warranty, and may break on any app update. You must provide your own Gemini API key; no key is included.

Portable method. The same module and the same 8-step method apply to other Electron chat apps. Reasonix Studio is the current target; AionUi 2.2.2 is supported by AionUi-Voice-Input using the same method.

App	Reasonix Studio 2.40.0 (Electron + local host)
Patched UI	resources\frontend-next\dist — plain files, no asar repack
Status	runtime-validated on 2.40.0: mic, Gemini key, auto-send, queue, multi-pane anchoring; see §Status
Module version	1 (src/voice-input.reasonix.js)

Signature — what makes it different

The mic lives in the chat, not in another app. Left click records; right-click opens the hidden settings menu; the rest is about sending exactly how you want.

Feature	Behavior
Left click	starts recording; click again to stop and transcribe
Right-click	opens the hidden menu: Gemini key, model, language, auto-insert, auto-send, send keyword
Manual mode	text is inserted; you review, edit, then send yourself
Auto-send	at stop, text is inserted and sent automatically; an attached file goes with it
Attachments	attach the file first, then dictate; text + file are sent together
Queue	if the agent is working, the message is queued; several messages can wait
Force / fallback	the module retries ~2 s; if the send button stays disabled, it shows an explicit error and copies the text; you can always click send/queue
No separate app	no server, no account, no extra window; the module only adds the mic

The strong point: Gemini transcription is significantly better than Windows dictation (Win+H), and the auto-send/queue/attachment workflow means you can dictate several messages without touching the mouse — while keeping a manual mode when you want to review.

Files

File	Role
src/voice-input.reasonix.js	voice module adapted to Reasonix selectors and storage
scripts/apply-patch.ps1	patcher: verified backup, copy, script tag, check, reapply, rollback
PREPARER-SAUVEGARDE-MIC-REASONIX.bat	creates a verified backup without patching
INSTALLER-MIC-REASONIX.bat	first install
REAPPLIQUER-MIC-REASONIX.bat	reapply after a Reasonix update
RESTAURER-MIC-REASONIX.bat	rollback (works even if the app no longer starts)
DESINSTALLER-MIC-REASONIX.bat	alias for rollback
VERIFIER-MIC-REASONIX.bat	read-only check: tag, module hash, backup
docs/	inspection, method, journal, problems/solutions, version fingerprints
docs/PROBLEMS-SOLUTIONS*.md	problem → solution table

Install

1. Close or leave Reasonix Studio open — the patcher never closes it.
2. Recommended: PREPARER-SAUVEGARDE-MIC-REASONIX.bat — creates and verifies index.html.bak + .sha256 without modifying the app.
3. INSTALLER-MIC-REASONIX.bat — the patcher re-verifies the backup before any write.
4. Reload (Ctrl+R) or relaunch Reasonix Studio.
5. Right-click the mic → paste your Gemini key (AIza…) → Save → Test the Gemini key.

Command line:

.\scripts\apply-patch.ps1 -ScriptPath .\src\voice-input.reasonix.js -Force

Rollback — guaranteed before any write

The patcher verifies the backup before modifying anything; if the copy does not match, it aborts. A checksum file is written: index.html.bak.sha256.

Step	Command
Prepare a verified backup without patching	PREPARER-SAUVEGARDE-MIC-REASONIX.bat
Roll back	RESTAURER-MIC-REASONIX.bat (or apply-patch.ps1 -Rollback)
App no longer starts	close it (Task Manager if needed), run RESTAURER-MIC-REASONIX.bat, relaunch
Package lost	copy index.html.bak over index.html, delete voice-input.js

Manual fallback:

$fe = "$env:LOCALAPPDATA\Programs\Reasonix Studio\resources\frontend-next\dist"
Copy-Item "$fe\index.html.bak" "$fe\index.html" -Force
Remove-Item "$fe\voice-input.js" -Force -ErrorAction SilentlyContinue

The patch never touches app.asar or the host binary: only index.html and voice-input.js in the static frontend. The app reference index.html is not shipped in this repository (third-party file); see docs/reference/README.md to recreate it locally.

Reapply after an app update

frontend-next\dist is replaced by the update → the patch disappears.

1. REAPPLIQUER-MIC-REASONIX.bat (or PREPARER... then INSTALLER...).
2. Reload / relaunch.
3. VERIFIER-MIC-REASONIX.bat should report: html tagged = True, voice-input.js present = True, hash matches source = True, backup present = True, backup sha256 = True.

Exact changes

File	Change
…\frontend-next\dist\voice-input.js	added
…\frontend-next\dist\index.html	one <script src="./voice-input.js"></script> before </body>
…\frontend-next\dist\index.html.bak	backup of the original
…\frontend-next\dist\index.html.bak.sha256	checksum of the backup

Reasonix selectors (2.40.0)

• Composer: textarea[data-action-keydown="session.send"][role="combobox"]
• Send: button.btn.send[data-action="session.send"]
• Stop (anchoring only, never clicked): button.btn.stop[data-action="session.stop"]
• Queue: queue.send-now, queue.cancel, queue.edit, queue.pause, queue.retry
• Storage keys use the rx- prefix so Reasonix's preload persists them across launches.

Portable method — other apps

The module is app-agnostic; only the selectors, storage keys and preload APIs change.

App	UI location	Patch type	State
Reasonix Studio 2.40.0	frontend-next\dist (plain)	copy + one script tag	ported and validated
Cline Desktop	Electron app	same module + selectors	ported; anchoring to adjust
Open Cowork	Electron app	same module + selectors	ported (base)
AionUi 2.2.2	resources\app.asar	asar repack with --unpack	ported and validated

The 8 generic steps are in docs/METHODOLOGY.md (EN) and docs/DEMARCHE.md (FR):

1. Locate the composer, send and stop controls; record stable selectors.
2. Determine whether the UI is plain or inside an asar.
3. Plain UI: copy the module + one script tag; asar: extract, preserve app.asar.unpacked, repack with --unpack-dir / --unpack, verify the manifest.
4. Adapt storage keys to the app's persistence (e.g. rx- in Reasonix).
5. Adapt preload APIs (external links, config/logs availability).
6. Anchor the mic in the real action row; never anchor onto itself; float above when narrow.
7. Verify: no renderer freeze, clickable mic, key test, recording, insertion, auto-send, queue.
8. Keep a verified rollback (.bak + checksum, or app.asar.bak + unpacked manifest).

Why Gemini? Gemini transcription is significantly more accurate and better punctuated than Windows dictation (Win+H), especially for technical terms and long messages. That is the module's main advantage; Windows remains an offline/no-key fallback.

For an agent: AGENT-PLAYBOOK.md (EN) / FICHE-AGENT.fr.md (FR) is a copy-paste roadmap. It covers ~95% of the port; the last 5% are exact selectors and anchoring measurements — small adjustments, not a constraint.

Status / to validate

Point	State
Mic button, anchoring, resize	validated on 2.40.0
Gemini key, test button, transcription	validated
Auto-send, queue while the agent works	validated
CSP / host headers, sandboxed fetch	validated at runtime
UI language of the module	French (v1); docs are FR/EN

Traceability

File	Role
docs/INSPECTION-2026-10-18.md / .en.md	inspection, selectors, risks — FR/EN
docs/DEMARCHE.md / docs/METHODOLOGY.md	method — FR/EN
docs/JOURNAL.md / docs/JOURNAL.en.md	append-only journal — FR (authoritative) / EN mirror
docs/PROBLEMS-SOLUTIONS.md / .en.md	problems/solutions — FR/EN
docs/VERSION-REASONIX.md / .en.md	2.40.0 fingerprints — FR/EN
CHANGELOG.md / CHANGELOG.fr.md	versions — EN/FR
docs/PUBLICATION.md	publication checklist — FR/EN
SECURITY.md	key handling and reporting — FR/EN

Tags

voice-input · speech-to-text · dictation · gemini · electron · reasonix-studio · cline-desktop · open-cowork · aionui · productivity · accessibility · windows

License

MIT — see LICENSE. The app itself is not distributed here; its name and trademarks belong to their owners.