@gotcos/glasses-server 6.1.0

package/server/routes/transcribe.ts

@@ -0,0 +1,55 @@
+// POST /api/transcribe — Whisper transcription endpoint (message-query / one-shot path).
+//
+// The actual Whisper + cleanup pipeline lives in server/lib/transcribe-audio.ts so
+// prompt-draft recovery and the legacy one-shot path share exactly the same behavior.
+
+import { Router } from 'express'
+import {
+  transcribeAudioBuffer,
+  resolveTranscribeMode,
+  NoSpeechDetectedError,
+  OpenAIWhisperBudgetExhaustedError,
+} from '../lib/transcribe-audio.js'
+
+export const transcribeRouter = Router()
+
+function resolveMode(req: { body?: { mode?: string }; query?: { mode?: string | string[] } }) {
+  return resolveTranscribeMode(
+    (req.body && typeof req.body.mode === 'string' ? req.body.mode : undefined) ??
+    (req.query && typeof req.query.mode === 'string' ? req.query.mode : undefined)
+  )
+}
+
+// Accept raw binary body up to 25MB (Whisper limit).
+transcribeRouter.post('/transcribe', async (req, res) => {
+  try {
+    const chunks: Buffer[] = []
+    for await (const chunk of req) {
+      chunks.push(Buffer.isBuffer(chunk) ? chunk : Buffer.from(chunk))
+    }
+    const audioBuffer = Buffer.concat(chunks)
+
+    if (audioBuffer.length < 100) {
+      return res.status(400).json({ error: 'audio too short' })
+    }
+
+    const result = await transcribeAudioBuffer(audioBuffer, { mode: resolveMode(req) })
+    console.log(`[perf] /transcribe: ${result.elapsedMs.toFixed(1)}ms | mode=${result.mode} | ${result.backend} | ${result.audioBytes}b | ${result.text.length} chars`)
+    res.json({ text: result.text, backend: result.backend, mode: result.mode })
+  } catch (err: any) {
+    if (err instanceof NoSpeechDetectedError) {
+      console.log(`[perf] /transcribe: DROPPED (hallucination or empty): ${err.rawText.length} chars`)
+      return res.status(204).send()
+    }
+    if (err instanceof OpenAIWhisperBudgetExhaustedError) {
+      console.error(`[transcribe] ${err.message}`)
+      return res.status(503).json({
+        error: err.message,
+        reason: 'openai_whisper_budget_exhausted',
+        spent_today_usd: err.spentTodayUsd,
+        cap_usd: err.capUsd,
+      })
+    }
+    res.status(500).json({ error: err.message })
+  }
+})

package/shared/model-preference.ts

@@ -0,0 +1,81 @@
+export type ClaudeModelPreference = 'opus' | 'sonnet' | 'haiku'
+export type CodexModelPreference = 'codex-high'
+export type ModelPreference = ClaudeModelPreference | CodexModelPreference
+
+export const DEFAULT_MODEL = 'opus' as const
+export const CODEX_HIGH_MODEL: CodexModelPreference = 'codex-high'
+
+// Optional codex model passed to `codex exec --model`. Empty (the default) means
+// "use whatever model your codex CLI is configured for" — so the public server
+// never pins a specific (possibly unreleased) model id. Set COS_CODEX_MODEL to
+// pin one. COS_CODEX_REASONING_EFFORT tunes the reasoning level (default high).
+export const CODEX_MODEL_ID = process.env.COS_CODEX_MODEL ?? ''
+export const CODEX_HIGH_REASONING_EFFORT = process.env.COS_CODEX_REASONING_EFFORT ?? 'high'
+
+export const MODEL_OPTIONS: ModelPreference[] = ['opus', 'sonnet', 'haiku', 'codex-high']
+
+const MODEL_SET = new Set<ModelPreference>(['opus', 'sonnet', 'haiku', 'codex-high'])
+
+export function isModelPreference(value: unknown): value is ModelPreference {
+  return typeof value === 'string' && MODEL_SET.has(value as ModelPreference)
+}
+
+export function normalizeModelPreference(value: unknown): ModelPreference | undefined {
+  return isModelPreference(value) ? value : undefined
+}
+
+export function isClaudeModel(model: ModelPreference): model is ClaudeModelPreference {
+  return model === 'opus' || model === 'sonnet' || model === 'haiku'
+}
+
+export function isCodexModel(model: ModelPreference): model is CodexModelPreference {
+  return model === 'codex-high'
+}
+
+export function modelLabel(model: ModelPreference): string {
+  switch (model) {
+    case 'sonnet': return 'Sonnet'
+    case 'haiku': return 'Haiku'
+    case 'codex-high': return 'Codex High'
+    case 'opus':
+    default:
+      return 'Opus'
+  }
+}
+
+export function modelShortLabel(model: ModelPreference): string {
+  switch (model) {
+    case 'sonnet': return 'Sonnet'
+    case 'haiku': return 'Haiku'
+    case 'codex-high': return 'Codex H'
+    case 'opus':
+    default:
+      return 'Opus'
+  }
+}
+
+export function modelButtonLabel(model: ModelPreference): string {
+  switch (model) {
+    case 'sonnet': return 'SNNT'
+    case 'haiku': return 'HAIKU'
+    case 'codex-high': return 'CODEX H'
+    case 'opus':
+    default:
+      return 'OPUS'
+  }
+}
+
+export function modelTag(model: ModelPreference): string {
+  switch (model) {
+    case 'sonnet': return 'S'
+    case 'haiku': return 'H'
+    case 'codex-high': return 'CH'
+    case 'opus':
+    default:
+      return 'O'
+  }
+}
+
+export function modelBracketTag(model: ModelPreference): string {
+  return ` [${modelTag(model)}]`
+}