asi se fue xd

This commit is contained in:
2026-02-18 12:13:22 -06:00
parent d27da30494
commit d0fdd04132
17 changed files with 612 additions and 735 deletions

View File

@@ -50,7 +50,7 @@ export async function handleWhisperRoutes(req: Request): Promise<Response | null
return Response.json({
...result,
...state,
message: state.enabled ? 'Whisper enabled (GPU)' : 'Whisper disabled (using Web Speech API)'
message: state.running ? 'Whisper GPU running' : 'Whisper GPU starting...'
})
}

View File

@@ -1,6 +1,7 @@
/**
* Whisper Service - Manages the Python Whisper server process
* Provides GPU-accelerated speech-to-text as an alternative to Web Speech API
* Whisper Service - Singleton persistent GPU speech-to-text server
* Auto-starts with the system, auto-restarts on crash.
* Single instance processes all client requests.
*/
import { join } from 'path'
@@ -8,18 +9,19 @@ import { Subprocess } from 'bun'
const WHISPER_PORT = 4104
const WHISPER_SCRIPT = join(import.meta.dir, '..', 'whisper_server.py')
const RESTART_DELAY_MS = 3000 // Wait before auto-restart after crash
interface WhisperState {
enabled: boolean
running: boolean
starting: boolean // Prevents multiple simultaneous start attempts
starting: boolean
process: Subprocess | null
model: string
device: string
}
const state: WhisperState = {
enabled: false,
enabled: true, // Always enabled by default
running: false,
starting: false,
process: null,
@@ -32,104 +34,16 @@ const state: WhisperState = {
*/
async function killProcessOnPort(port: number): Promise<void> {
try {
// Use PowerShell to find and kill process on port
const proc = Bun.spawn(['powershell', '-Command',
`Get-NetTCPConnection -LocalPort ${port} -ErrorAction SilentlyContinue | ForEach-Object { Stop-Process -Id $_.OwningProcess -Force -ErrorAction SilentlyContinue }`
], { stdout: 'ignore', stderr: 'ignore' })
await proc.exited
// Wait a moment for port to be released
await new Promise(resolve => setTimeout(resolve, 1000))
} catch {
// Ignore errors
}
}
/**
* Start the Whisper Python server
*/
export async function startWhisperServer(): Promise<boolean> {
// Prevent multiple simultaneous start attempts
if (state.starting) {
return false
}
if (state.running && state.process) {
return true
}
state.starting = true
console.log(`[Whisper] Starting (${state.model})...`)
// Kill any existing process on the port
await killProcessOnPort(WHISPER_PORT)
try {
// Use Bun.spawn with inherit to show logs directly in console
// -u flag disables Python output buffering for real-time logs
const proc = Bun.spawn(['python', '-u', WHISPER_SCRIPT], {
cwd: join(import.meta.dir, '..'),
stdout: 'inherit',
stderr: 'inherit',
env: { ...process.env, PYTHONUNBUFFERED: '1' }
})
state.process = proc
// Wait a bit for the server to start
await new Promise(resolve => setTimeout(resolve, 2000))
// Check if process is still running
if (proc.exitCode !== null) {
console.error('[Whisper] Process exited with code:', proc.exitCode)
state.process = null
state.starting = false
return false
}
// Check if WebSocket is ready
const isListening = await checkPort(WHISPER_PORT)
if (isListening) {
console.log('[Whisper] Ready')
state.running = true
state.enabled = true
state.starting = false
return true
}
// Wait more if model is still loading (up to 120 seconds total for large models)
for (let i = 0; i < 40; i++) {
await new Promise(resolve => setTimeout(resolve, 3000))
if (proc.exitCode !== null) {
console.error('[Whisper] Process died')
state.process = null
state.starting = false
return false
}
const ready = await checkPort(WHISPER_PORT)
if (ready) {
console.log('[Whisper] Ready')
state.running = true
state.enabled = true
state.starting = false
return true
}
}
console.error('[Whisper] Timeout (120s)')
state.starting = false
return false
} catch (err: any) {
console.error('[Whisper] Error:', err.message)
state.process = null
state.starting = false
return false
}
}
/**
* Check if Whisper WebSocket is ready using PowerShell
*/
@@ -152,7 +66,125 @@ async function checkPort(port: number): Promise<boolean> {
}
/**
* Stop the Whisper server
* Monitor the Whisper process and auto-restart on crash
*/
function monitorProcess(proc: Subprocess) {
proc.exited.then((exitCode) => {
console.error(`[Whisper] Process exited with code ${exitCode}`)
state.process = null
state.running = false
state.starting = false
// Auto-restart after delay
console.log(`[Whisper] Auto-restarting in ${RESTART_DELAY_MS / 1000}s...`)
setTimeout(() => {
startWhisperServer().catch(err => {
console.error('[Whisper] Auto-restart failed:', err)
})
}, RESTART_DELAY_MS)
})
}
/**
* Start the Whisper Python server (singleton - only one instance)
*/
export async function startWhisperServer(): Promise<boolean> {
// Prevent multiple simultaneous start attempts
if (state.starting) {
console.log('[Whisper] Already starting, skipping')
return false
}
// Already running
if (state.running && state.process) {
console.log('[Whisper] Already running')
return true
}
// Check if an external instance is already listening
const alreadyListening = await checkPort(WHISPER_PORT)
if (alreadyListening) {
console.log('[Whisper] External instance already running on port', WHISPER_PORT)
state.running = true
state.enabled = true
return true
}
state.starting = true
console.log(`[Whisper] Starting singleton server (${state.model})...`)
// Kill any orphan process on the port
await killProcessOnPort(WHISPER_PORT)
try {
const proc = Bun.spawn(['python', '-u', WHISPER_SCRIPT], {
cwd: join(import.meta.dir, '..'),
stdout: 'inherit',
stderr: 'inherit',
env: { ...process.env, PYTHONUNBUFFERED: '1' }
})
state.process = proc
// Monitor for crashes and auto-restart
monitorProcess(proc)
// Wait for initial startup
await new Promise(resolve => setTimeout(resolve, 2000))
// Check if process died immediately
if (proc.exitCode !== null) {
console.error('[Whisper] Process exited immediately with code:', proc.exitCode)
state.process = null
state.starting = false
return false
}
// Check if WebSocket is ready
const isListening = await checkPort(WHISPER_PORT)
if (isListening) {
console.log('[Whisper] Server ready (GPU)')
state.running = true
state.enabled = true
state.starting = false
return true
}
// Wait for model loading (up to 120 seconds for large-v3)
for (let i = 0; i < 40; i++) {
await new Promise(resolve => setTimeout(resolve, 3000))
if (proc.exitCode !== null) {
console.error('[Whisper] Process died during model loading')
state.process = null
state.starting = false
return false
}
const ready = await checkPort(WHISPER_PORT)
if (ready) {
console.log('[Whisper] Server ready (GPU)')
state.running = true
state.enabled = true
state.starting = false
return true
}
}
console.error('[Whisper] Timeout waiting for server (120s)')
state.starting = false
return false
} catch (err: any) {
console.error('[Whisper] Start error:', err.message)
state.process = null
state.starting = false
return false
}
}
/**
* Stop the Whisper server (only for manual override, not used in normal flow)
*/
export function stopWhisperServer(): boolean {
if (!state.process) {
@@ -163,8 +195,7 @@ export function stopWhisperServer(): boolean {
state.process.kill()
state.process = null
state.running = false
state.enabled = false
console.log('[Whisper] Stopped')
console.log('[Whisper] Stopped manually')
return true
} catch (err) {
console.error('[Whisper] Stop error:', err)
@@ -173,27 +204,26 @@ export function stopWhisperServer(): boolean {
}
/**
* Toggle Whisper server on/off (async - returns immediately when starting)
* Toggle is now a no-op for stop - Whisper always stays on.
* If not running, triggers a start.
*/
export async function toggleWhisperServer(): Promise<{ enabled: boolean; success: boolean; starting: boolean }> {
// Prevent toggle while starting
if (state.starting) {
return { enabled: false, success: false, starting: true }
return { enabled: true, success: false, starting: true }
}
if (state.enabled && state.running) {
const success = stopWhisperServer()
return { enabled: false, success, starting: false }
} else {
// Start server in background - don't await
startWhisperServer().catch(err => {
console.error('[Whisper] Start error:', err)
state.starting = false
})
// Return immediately - frontend will poll for status
return { enabled: false, success: true, starting: true }
if (state.running) {
// Already running - just confirm it's on
return { enabled: true, success: true, starting: false }
}
// Not running - start it
startWhisperServer().catch(err => {
console.error('[Whisper] Start error:', err)
state.starting = false
})
return { enabled: true, success: true, starting: true }
}
/**
@@ -217,13 +247,12 @@ export async function getWhisperState(): Promise<{
state.enabled = true
} else if (!isListening && state.running) {
state.running = false
state.enabled = false
state.process = null
// Keep enabled=true since we auto-restart
}
}
return {
enabled: state.enabled,
enabled: true, // Always enabled
running: state.running,
starting: state.starting,
port: WHISPER_PORT,
@@ -233,15 +262,12 @@ export async function getWhisperState(): Promise<{
}
/**
* Check if Whisper is enabled
* Check if Whisper is running
*/
export function isWhisperEnabled(): boolean {
return state.enabled && state.running
return state.running
}
// WebSocket server for Whisper (proxies to Python server or handles directly)
let whisperWsServer: any = null
export function getWhisperPort(): number {
return WHISPER_PORT
}

View File

@@ -3,10 +3,12 @@
* Terminal Server - Independent process
* This runs separately from the main server to maintain stable Claude Code sessions
* even when the main server restarts due to code changes.
* Also manages the Whisper GPU server (singleton, persistent).
*/
import { startTerminalServer } from './services/terminal'
import { startSyncServer } from './services/sync-server'
import { startWhisperServer } from './services/whisper'
import { WORKING_DIR } from './config'
console.log('')
@@ -14,6 +16,7 @@ console.log('='.repeat(50))
console.log('Terminal Server (Independent Process)')
console.log(` Terminal WebSocket: ws://localhost:4103`)
console.log(` Sync WebSocket (Git + Torch): ws://localhost:4105`)
console.log(` Whisper GPU: ws://localhost:4104 (auto-start)`)
console.log(` Working Dir: ${WORKING_DIR}`)
console.log('')
console.log('This process is stable and won\'t restart')
@@ -23,3 +26,14 @@ console.log('')
startTerminalServer()
startSyncServer()
// Auto-start Whisper GPU server (singleton, persistent, auto-restart on crash)
startWhisperServer().then(ok => {
if (ok) {
console.log('[Whisper] GPU server started successfully')
} else {
console.warn('[Whisper] Failed initial start (will auto-retry)')
}
}).catch(err => {
console.error('[Whisper] Boot error:', err)
})

View File

@@ -26,35 +26,25 @@ except ImportError as e:
def convert_audio_to_wav(input_data: bytes, input_format: str = "webm") -> bytes:
"""
Convert audio data to WAV format using ffmpeg.
Whisper requires WAV/PCM format, but browsers typically record in WebM/Opus.
Uses stdin/stdout pipes so ffmpeg probes the actual data format
instead of relying on file extensions.
"""
# Create temp files for input and output
with tempfile.NamedTemporaryFile(suffix=f".{input_format}", delete=False) as in_file:
in_file.write(input_data)
input_path = in_file.name
output_path = input_path.replace(f".{input_format}", ".wav")
try:
# Use ffmpeg to convert to WAV (16kHz mono, which Whisper prefers)
result = subprocess.run([
"ffmpeg", "-y", # Overwrite output
"-i", input_path, # Input file
"-ar", "16000", # Sample rate 16kHz
"-ac", "1", # Mono
"ffmpeg", "-y",
"-i", "pipe:0", # Read from stdin (auto-detect format)
"-ar", "16000", # Sample rate 16kHz
"-ac", "1", # Mono
"-c:a", "pcm_s16le", # PCM 16-bit little-endian
output_path
], capture_output=True, text=True, timeout=30)
"-f", "wav", # Output format
"pipe:1" # Write to stdout
], input=input_data, capture_output=True, timeout=30)
if result.returncode != 0:
print(f"[Whisper] ffmpeg error: {result.stderr}")
print(f"[Whisper] ffmpeg error: {result.stderr.decode('utf-8', errors='replace')}")
return None
# Read the converted WAV file
with open(output_path, "rb") as f:
wav_data = f.read()
return wav_data
return result.stdout
except subprocess.TimeoutExpired:
print("[Whisper] ffmpeg conversion timed out")
@@ -65,16 +55,6 @@ def convert_audio_to_wav(input_data: bytes, input_format: str = "webm") -> bytes
except Exception as e:
print(f"[Whisper] Conversion error: {e}")
return None
finally:
# Cleanup temp files
try:
os.unlink(input_path)
except:
pass
try:
os.unlink(output_path)
except:
pass
# Configuration
HOST = "0.0.0.0" # Listen on all interfaces (needed for Traefik proxy)