CoqPi

A private desktop copilot for multilingual calls, interview practice, meeting transcription, and reviewed follow-up.

MVP: CoqPi is useful for local preparation and transcription, but it has not yet completed repeated real-call latency and speaker-attribution validation.

Current source and locally verified app: 0.3.11. GitHub Release downloads may lag behind source; this update is published as source, without a new release artifact.

Compact communication workspace

The current source opens with three modes: Conversation, Recorder and Learning. Conversation setup explicitly chooses headphones plus a separate call-audio input, or a single-input loudspeaker conversation. The active call has no navigation, branding or timers: small Russian meaning, a large answer and Stop. Manual ME/Space control is available only for the single-input route. Preparation, knowledge and settings remain reachable before starting, without Finder/mail in the main navigation.

The default window is 640x320 (minimum 400x260). Settings separate Audio, Conversation, Recording, My data and Connection; technical tools are collapsed. The compact build was installed locally after short real BlackHole/OpenAI EN/FR, language-switch and persistence checks. Long-call and physical-headset acceptance remain open. See the real-user acceptance report for actual audio evidence, synthetic regressions and outstanding scenarios.

Live brief replies request only the displayed translation and answer. The scheduler uses 250 ms settling and overlaps it with cooldown, without bypassing cooldown or stale-response guards. Short real EN/FR probes measured roughly 3-5 seconds to a complete answer (about 4-5 seconds in the clean-input probes); this is not a long-call latency guarantee. OpenAI brief replies now stream translation and answer text into the current request's UI before final completion. Stop/context changes cancel that request and hide its partial text. Only validated final output is stored as an assistant result. First-text timing and full-answer timing are distinct; model comparisons remain open.

Interview answer safety

Learning now includes a complete bounded virtual interview: five spoken EN/FR questions, real Live hints, microphone answers, independent WAV backup, incremental transcript and saved timings. Fictional context is isolated from personal CVs and approved replies. After Stop, Voice answer review can send one explicitly selected microphone clip to OpenAI for tentative audio-backed feedback, replay the recording and speak a correct example. Confirm/Dismiss stays local. Repeat & compare saves a separate short attempt, rechecks the original recommendation and compares audio only after explicit consent. Identical audio, unconfirmed advice or insufficient evidence cannot count as improvement; model feedback remains experimental. No automatic mistake promotion or pronunciation score. Personal interview rehearsal now freezes the exact submitted CV, vacancy and reviewed company information for adaptive EN/FR questions, the same Live hints, voice recording and an evidence-backed post-Stop review. General biography and approved replies cannot override that CV. See audio review.

Rehearse a specific interview

1. Open Learning → Personal interview and choose English or French.
2. Add the exact CV sent to that employer, the vacancy and reviewed company text. Paste text or choose a file; public links can load an excerpt for review.
3. Confirm the materials and OpenAI processing, then Freeze reviewed materials.
4. Start a 3–10-question rehearsal. Listen, read or paraphrase the large hint, answer aloud, then use Next or Stop.
5. After Stop, request Review interview. Recording, timings and exports are under Recording & timing; pronunciation review is a separate opt-in action.

The submitted CV controls personal facts. Other CVs, global biography, memory and the approved reply bank are excluded from this mode. Missing specific experience must lead to a conditional approach, not an invented past event.

The final review is draft coaching: exact citations do not prove that an assessment is correct. Generic language advice, overinterpretation and occasional English commentary remain known limitations. Advice is not automatically promoted into lessons or verified personal facts. PDF/Office conversion requires a local MarkItDown/Python runtime; reviewed pasted text works without it.

The short real synthetic-speech EN/FR check saved four answers and both WAVs; first hint text took 0.8–1.9 s and complete hints 1.1–2.3 s, excluding interviewer question generation and speech playback. See the acceptance evidence and limits.

EN/FR brief responses distinguish owner facts from interviewer assertions and proposed commitments. Recognized high-risk questions (work authorization, availability, accepting terms, asserted achievements and missing behavioral examples) receive a conservative local clarification instead of a fabricated answer. This is deliberately not a universal semantic safety guarantee.

If a model answer takes more than 1.8 seconds, the existing answer area can show a rotating EN/FR thinking phrase. It disappears on answer, Stop, error or ME speech and is never persisted as an assistant answer. Typical interview structures guide generation. A local EN/FR personal answer bank in Prepare now reuses explicitly reviewed answers for exact approved question variants, without a model request. Profile or selected-context changes invalidate approval; new constraints fall back to normal analysis. Personal drafts remain outside Git. See personal answer bank and content safety checks and limits.

Conversation transcription uses a 500 ms end-of-speech pause; recorder mode and the separate ME channel retain 800 ms. Explicit system audio bypasses microphone echo/noise/gain processing. The transcription model, auto language switching and independent audio backups are unchanged. See the STT latency checks; shorter VAD pauses do not guarantee bounded cloud recognition latency.

Recording reliability

• Stop/Reset invalidates pending assistant responses; delayed answers cannot reappear in the next conversation.
• Audio capture starts before cloud transcription and continues through STT failures. Transcribe also records locally without an API key; use Recover later for text.
• An explicitly selected loopback input now feeds Live as OTHER; the separate microphone feeds ME, which is recorded but excluded from suggestions. With both sources enabled, two cloud transcription streams are used.
• Each acknowledged PCM chunk is synced locally with a usable WAV header. Recover can repair interrupted recordings and saves recovered text chunk by chunk.

See recording reliability and setup. Hardware routing, recognition accuracy and long-call latency still require real-call checks. These changes are included in v0.3.4.

What it does

CoqPi keeps the communication loop in one desktop app:

• Live transcribes an active microphone, follows EN/FR/RU/UK language changes, and prepares one short response.
• Transcribe records a plain local meeting transcript without assistant analysis.
• Training offers a fictional five-question interview and a personal 3–10-question rehearsal using the exact submitted CV, vacancy and company materials. Both reuse Live hints, voice recording and timings. Personal follow-ups adapt to saved answers; the final review is requested after Stop. Optional audio review remains experimental. Tutor continuity is the next step in the training plan.
• Prepare builds a reviewed session brief from explicitly selected context.

Finder, Gmail/Calendar and Monitor integrations remain legacy/optional backend capabilities, not primary modes. No stored data or approval boundaries have been removed; this release focuses on communication, recording and learning.

With separate microphone and system inputs, your microphone is ME and call audio is OTHER. For shared speakerphone input, hold Space in the focused Live window or toggle Говорю я. ME remains recorded but excluded from suggestions.

Privacy and authority

CoqPi is local-first, but OpenAI-backed transcription and assistant features send the selected audio or prompt context to OpenAI.

• Secrets stay in the Electron main process and may be stored with macOS secure storage.
• Runtime profiles, transcripts, context packs, receipts, and OAuth tokens stay outside the Git repository.
• Assistant requests pass through a local privacy gate that redacts recognized contact/tracking data and blocks secret-like material.
• Recovered transcript chunks are visibly marked before export; weak chunks can be excluded/restored for Markdown/TXT or merged into the previous segment without deleting the local session record.
• Gmail sending and Calendar creation require explicit, hash-bound user approval.
• The Monitor bridge is disabled by default, requires explicit client-speech permission, keeps its token in macOS secure storage, and never sends ME, partial, or system transcript items.
• CoqPi cannot autonomously join calls, control other apps, send messages, or publish content.
• CortexABV handoff accepts compact, reviewed artifacts only; raw transcripts and cross-tenant promotion remain denied.

See Cortex context contract and agent operations contract.

Install the latest macOS build

The current public build targets Apple Silicon Macs.

1. Open the latest release.
2. Download CoqPi-<version>-arm64.zip and SHA256SUMS.txt.
3. Verify the checksum, extract the archive, and move CoqPi.app to Applications.
4. Because the app is unsigned, open it with right click → Open when macOS asks.

Replacing the app does not remove local data under ~/Library/Application Support/CoqPi.

Run from source

Requirements:

• Node.js 22.12 or newer;
• pnpm 11;
• macOS for the packaged desktop app;
• an OpenAI API key for realtime transcription and OpenAI assistant responses.

git clone https://github.com/markoblogo/CoqPi.git
cd CoqPi
pnpm install
cp .env.example .env
pnpm dev

Set OPENAI_API_KEY in .env, or save it through Settings. Never commit the key.

The source default fast-response model is gpt-5.6-luna; realtime transcription defaults to gpt-4o-transcribe. Backend environment overrides can select different models; recorded measurements apply only to the actual tested configuration.

First useful test

1. Open Learning -> Virtual interview, choose EN/FR and start the fictional five-question exercise.
2. Check the answer, language, model, and latency.
3. Open Transcribe, select the microphone and language, and record a short local sample.
4. Stop the sample. If finalized text is missing but the audio backup completed, use Recover before export.
5. Export it and confirm that the Markdown/TXT file contains the expected finalized text.
6. Only then try a short Live call.

For the real-provider checklist, use Realtime smoke test. For storage and recovery, use Meeting transcription mode.

Speaker controls

• Hold Space before speaking while CoqPi is focused; release it when finished.
• Use Говорю я for longer answers or while switching to Meet/Zoom.
• Stop the session before reviewing or exporting it.

These controls apply to shared-input conversations, not the separate-source mode. There is no voice biometrics. If both speakers share one uninterrupted item, CoqPi conservatively excludes that item from suggestions. See Manual speaker mode.

Architecture

flowchart LR
  Mic["Microphone / explicit system input"] --> RT["OpenAI realtime transcription"]
  Mic --> Backup["Independent local WAV backup"]
  RT --> Log["Local durable transcript"]
  Log --> Gate["Language + privacy + speaker gates"]
  Gate --> Assistant["Short assistant response"]
  Context["Selected local context"] --> Gate
  Log --> Export["Markdown / TXT export"]
  Context -. "compact admitted handoff" .-> Cortex["CortexABV-private"]

Loading

Electron main owns secrets, filesystem access, provider calls, and narrow IPC handlers. React renders the operator surfaces. Simple Assistant uses only the selected Markdown profile, selected scenario, and recent transcript; the legacy retrieval path remains available separately.

Read Architecture for the full data flow and Conversation delivery order for current acceptance gaps.

Verification

pnpm check

The gate checks:

• no runtime/private data is tracked;
• TypeScript and ESLint;
• the complete deterministic Node test suite;
• renderer and Electron builds;
• high-severity dependency advisories.

The Electron UI harness and real microphone test are separate because they require a graphical macOS session:

pnpm test:conversation-ui
pnpm test:pass2-live-smoke-readiness

Current limits

• Apple Silicon macOS build only; the app is unsigned and not notarized.
• Microphone input is the default capture path. A separate System backup input can be selected when macOS exposes a virtual/routed call-audio device such as BlackHole or Loopback; CoqPi stores that source as system.wav.
• Raw-audio backup writes local WAV/PCM files next to the transcript manifest. Stopped sessions can be manually recovered into timestamped journal chunks without duplicating a previous recovery, and recovery reports show recovered, failed, skipped, and empty chunks.
• Speaker attribution follows explicit sources; shared-input marking cannot split one mixed speech item.
• Monitor live preview depends on a selected client, an authenticated Monitor account, and correct manual ME/OTHER marking; it never creates or sends an operational BID/OFFER/TRADE.
• Live quality and p50/p95 response latency still need repeated real-call measurement.
• Local Ollama can provide controlled text fallback, but there is no offline realtime STT.

Documentation

• Architecture
• Conversation delivery order
• Realtime smoke test
• Meeting transcription
• CortexABV boundary
• v0.3.4 release and verification
• Security policy
• Contributing
• Changelog

License

MIT

ABVX ecosystem

• AGENTS.md_generator — Keeps repository guidance and machine-readable context current. Current release: v0.5.1.
• abvx-agent-skills — Uses shared, reviewable agent capabilities during maintenance. Current release: v0.16.1.

This block is generated from the reviewed ABVX ecosystem registry.