voice-cat/CLAUDE.md

CLAUDE.md — agent hub for VoiceCat

Auto-loaded each session. This is the map: build commands, architecture at a glance, and where everything is. For the working method read AGENTS.md; for what's done and what's next read PROGRESS.md; for design read docs/.

One-line status: M3 (multi-stream, per-channel tuning) is complete, plus a follow-up pass closing the device enumeration / VAD-PTT input gate / stereo playback / WASAPI loopback gaps it left open (ctest --test-dir build/m1-dev green — 12/12 tests, including test_vad_ptt_devices, and vccli --voice/--list-devices manually verified live). Real webrtc-audio-processing (AEC/NS/AGC) is still unbuilt — no working Windows/MSVC port upstream — so v1 ships a lightweight energy/RMS VAD instead. Next up is M4 (native clients). See PROGRESS.md.

VoiceCat = self-hosted native voice & text chat (TeamSpeak/Mumble-style). Plain TCP (control)

• UDP (media), no WebRTC, encrypted by default. A shared C++ core (libvoicecat) drives native clients (Swift on macOS/iOS, C# on Windows) and the server.

---

Build & test commands

The M0 skeleton builds with no third-party dependencies — just CMake + Ninja + a C++20 compiler. Deps (vcpkg) are off until a subsystem needs them.

# Configure + build the skeleton (default; no vcpkg needed)
cmake --preset dev
cmake --build --preset dev

# Run the tests (behavior smoke test today; grows per milestone)
ctest --preset dev                      # or: ctest --test-dir build/dev --output-on-failure

# Run the binaries (Windows adds .exe; Linux/macOS no extension)
./build/dev/bin/vccli                    # headless test client
./build/dev/bin/voicecat-server --help
./build/dev/bin/voicecat-server --name "My Server"

# Build a single target / be verbose
cmake --build --preset dev --target vccli
cmake --build --preset dev --verbose

# Clean
rm -rf build/dev                         # nuke; or:
cmake --build --preset dev --target clean

When you start a subsystem that needs real libraries (mbedTLS, libsodium, opus, protobuf, …), turn vcpkg deps on:

# one-time: git clone https://github.com/microsoft/vcpkg && ./vcpkg/bootstrap-vcpkg.sh  (.bat on Windows)
export VCPKG_ROOT=/path/to/vcpkg         # works on Linux / macOS / Windows
cmake --preset server-release            # auto-installs deps pinned in vcpkg.json
cmake --build --preset server-release

Other useful toggles (pass with -D at configure time):

cmake --preset dev -DVOICECAT_BUILD_SHARED=ON     # build libvoicecat as a .dll/.so/.dylib (for the C# client)
cmake --preset dev -DVOICECAT_BUILD_SERVER=OFF     # core + tools only
cmake --preset dev -DVOICECAT_BUILD_TESTS=OFF

Formatting: clang-format config is .clang-format (Google base, 100 cols, 4-space).

git ls-files '*.cpp' '*.h' | xargs clang-format -i

---

Architecture at a glance

Full detail: docs/architecture.md. The short version:

 Swift (macOS/iOS)  ─┐                                    ┌─ C# (Windows)
                     ├──▶  libvoicecat  (C ABI: voicecat.h) ◀──┤
 voicecat-server  ───┘   net · crypto · codec · protocol ·     └─ all UIs are thin
   (links core)          session · audio                          the core owns audio

• One core, many faces. Protocol, Opus, crypto, networking, jitter buffer, and mixing live once in C++. Clients call the C ABI (core/include/voicecat.h); the server links the same core, so framing/crypto never drift between ends.
• Two transports. TCP + TLS 1.3 (control, protobuf Envelope) and UDP + exported-key ChaCha20-Poly1305 AEAD (media, fixed binary voice frame). Encryption is mandatory.
• Threading. Real-time audio threads never allocate/lock/block; they exchange data with the net thread via lock-free ring buffers; a worker pool absorbs blocking work.

Subsystem map (code ↔ design doc)

Path	Subsystem	Design
core/include/voicecat.h	The C ABI (client/server contract)	architecture.md §4
core/proto/voicecat.proto	Control-plane wire format (source of truth)	protocol.md
core/src/net/	Asio TCP/UDP transport, [u32 len][payload] framing	protocol.md §1, voice.md §2
core/src/crypto/	TLS 1.3 (mbedTLS), media AEAD (libsodium), anti-replay	security.md
core/src/codec/	Opus encode/decode, FEC/DTX	voice.md §3–4
core/src/protocol/	Envelope (de)serialize, request/response, dispatch	protocol.md
core/src/session/	Channels, users, streams, permissions, ephemeral text	protocol.md §5
core/src/audio/	miniaudio I/O, APM DSP, jitter buffer, mixer	voice.md §8–11
core/src/core/	vc_client — the handle behind the C ABI	architecture.md §4
server/	Connection mgr, session registry, SFU relay, SQLite	architecture.md §5
tools/vccli/	Headless client that drives/verifies the protocol	—
clients/apple/, clients/windows/	Native GUIs (M4)	architecture.md §4

---

Documentation index (source of truth)

Read docs/ before changing behavior. Order:

1. docs/README.md — overview, locked decisions, glossary
2. docs/architecture.md — core, C ABI, threading, server
3. docs/protocol.md — control plane, Envelope, message catalog
4. docs/voice.md — UDP media, Opus, multi-stream, two-sided NR, VAD/PTT
5. docs/security.md — mandatory encryption, TLS+AEAD, accounts, threat model
6. docs/tech-stack.md — libraries, permissive-license rule, tooling
7. docs/deployment.md — zero-config self-host (Docker / binary / source)
8. docs/roadmap.md — milestones + resolved decisions
9. docs/building.md — what each CMake preset is for + manual server/vccli testing

---

Keeping track of progress

PROGRESS.md is the living status file. When you finish a task, check it off there and note the next step, so the next agent can pick up instantly. Treat it as part of the work, not an afterthought — update it in the same commit as the code.

---

House rules (hard constraints)

• A clean compile is the floor, not the goal. "Done" = the milestone's observable exit criterion in docs/roadmap.md passes (e.g. M1 = two vccli actually chat over TLS). Encode it as a test. See AGENTS.md.
• No GPL/LGPL dependencies, ever (closed-source redistribution is a goal). docs/tech-stack.md §5.
• Encryption is mandatory — never add a plaintext transport path. docs/security.md.
• Real-time audio threads never allocate, lock, or block. docs/architecture.md §3.
• Keep docs + code in sync. Changing a wire format (voicecat.proto) or the C ABI (voicecat.h) is a deliberate, versioned act — update the doc in the same commit (protocol.md §8).
• Every commit must build (cmake --build --preset dev) and pass ctest --preset dev.