scythe/PROJECT_GUIDE.md

Scythe: Comprehensive Project & Architecture Guide

This document is the definitive technical reference for developers and AI agents working on the Scythe codebase. It outlines the project's design philosophy, unified binary architecture, hardware encoding pathways, audio subsystem, power management, HUD overlay, and development standards.

---

1. Project Overview & Current Status

Current Version: v0.2.18
License: Dual MIT / Apache 2.0

Repository Topology

Remote	URL	Role	Status
forgejo	https://git.oiupoyt.space/oiupoyt/scythe.git	Self-hosted primary; source of truth, release assets	Active
spearzificial	https://spearzificial@github.com/spearzificial/scythe.git	Active development mirror (currently maintained account)	Active
origin	https://github.com/oiupoyt/scythe.git	Legacy GitHub mirror	Read-only / stale token
gitlab	https://gitlab.com/oiupoyt.cat/scythe.git	Optional mirror	Optional

GitHub account policy: GitHub pushes now go to the spearzificial account. The oiupoyt GitHub OAuth token is expired, so origin is not a valid push target. Always name the account in the remote URL (https://spearzificial@github.com/...) so Git selects the correct stored credential when multiple GitHub accounts exist in the credential store.

Dual-Push Rule

Every push must reach both active remotes:

git push spearzificial main
git push forgejo main

Scythe is a cross-platform, ultra-low-overhead desktop screen recorder, instant replay buffer, and livestreamer written in Rust. It delivers functionality comparable to NVIDIA ShadowPlay and GPU Screen Recorder with minimal CPU and GPU overhead.

Core Capabilities:

1. Unified Multicall Binary Architecture: UI overlay, CLI client, and headless daemon engine are consolidated into a single executable (src/main.rs) with subcommands (scythe daemon, scythe record, scythe save, etc.) and backward-compatible symlinks (scythe-daemon, scythe-ui).
2. Zero-Copy Hardware Video Encoding:
   • Linux: FFmpeg VA-API (h264_vaapi, hevc_vaapi, av1_vaapi) with direct DRM DMA-BUF import for Intel and AMD GPUs. Automatic low-power hardware fallback handles drivers lacking VAEntrypointEncSliceLP.
   • Windows: Native NVENC, AMD AMF, Intel QuickSync (QSV), and multithreaded software fallbacks via DirectX Graphics Infrastructure (DXGI) desktop duplication.
3. High-Fidelity Audio Engine:
   • Multi-source routing: System Audio, Microphone, Both, or Muted.
   • Dual codec support: Native Opus (AV_CODEC_ID_OPUS, 48 kHz stereo, 160 kbps, AV_SAMPLE_FMT_FLT) as recommended default, with native AAC (AV_CODEC_ID_AAC, 48 kHz stereo, 192 kbps, AV_SAMPLE_FMT_FLTP) as universal fallback.
   • Interactive sound verification suite with live dB / percentage VU meters and test tone playback.
4. Lock-Free Circular RAM Replay Buffer:
   • Continuous lock-free packet ring buffer in system RAM (configurable from 5 seconds to 30 minutes).
   • Keyframe-aware draining ensures saved MP4 clips begin cleanly on IDR boundaries with zero decoder corruption.
5. Smart Power Management:
   • AC power connectivity monitoring via Linux /sys/class/power_supply and Windows Win32 GetSystemPowerStatus.
   • Dedicated 1-second background power poller thread.
   • "Power-only" replay mode automatically drains and pauses the replay buffer on battery power to conserve battery life, and seamlessly resumes with keyframe forced resync upon AC connection.
6. KAT Square Minimalist Interface:
   • Strictly conforms to DESIGN.md guidelines: border-radius: 0 !important; across all windows, inputs, buttons, and badges.
   • Obsidian dark palette (#0b0b0b / #111111) with hairline #222222 framing and restrained functional accents.
   • Borderless, shadowless terminal-style notification toasts.
   • Monospace typography option with integrated Adwaita Mono font.
7. Screenshot Resolution Presets:
   • Intuitive presets: Native (Source), 1080p (FHD), 1440p (QHD), 2160p (4K UHD), 720p (HD) with hardware-accelerated Lanczos scaling via FFmpeg.
8. In-App Auto-Updater:
   • Synchronizes against Forgejo releases API (https://git.oiupoyt.space/api/v1/repos/oiupoyt/scythe/releases/latest) with automated binary replacement.

---

2. High-Level Architecture & Process Model

Scythe uses a unified multicall binary dispatch model:

                         +-----------------------------------+
                         |           scythe binary           |
                         +-----------------+-----------------+
                                           |
               Inspects argv[0] and first CLI argument
                                           |
         +---------------------------------+---------------------------------+
         |                                                                   |
         v                                                                   v
   "scythe daemon"                                                   "scythe [command]"
  or argv[0] == "scythe-daemon"                                     or argv[0] == "scythe-ui"
         |                                                                   |
+--------v-------------------------+                               +---------v-----------------------+
|    Headless Daemon Engine        |                               |      Client CLI & Overlay       |
|      (src/daemon.rs)             | <--- Unix Domain Socket / --- |       (src/overlay_egui.rs)     |
| - Screen Capture (PipeWire/DXGI) |      Local TCP IPC Loop       | - egui HUD Overlay (Alt+Z)      |
| - Hardware Video Encoder         |                               | - Audio Diagnostics & VU Meters |
| - Native Opus/AAC Audio Encoder  |                               | - Clip Trimmer & Gallery        |
| - Lock-Free Circular Ring Buffer |                               | - Terminal Notification Toasts  |
| - AC Power Watcher Thread        |                               | - Global Hotkey Listener Pump   |
| - MP4 Muxer & RTMP Streamer      |                               +---------------------------------+
+----------------------------------+

Multicall Dispatching:

• Daemon Mode: Invoked when args[1] == "daemon", args[1] == "--mock", or when executed via symlink scythe-daemon / vrec-daemon. Runs the headless capture and recording engine (src/daemon.rs).
• Client / UI Mode: Invoked when executed as scythe, scythe-ui, or with client arguments (save, record, screenshot, status, menu, etc.). Runs src/main.rs and launches src/overlay_egui.rs.

---

3. Inter-Process Communication (IPC)

The protocol is implemented in src/ipc.rs.

Transport Layer:

• Linux: Unix Domain Socket at $XDG_RUNTIME_DIR/scythe.sock (with legacy fallback to /run/user/<UID>/vrec.sock).
• Windows: Local TCP socket on 127.0.0.1 binding to candidate ports 42069–42072 and persisting the active port in %TEMP%\scythe-ipc.port.

Message Payload:

Packets are framed with a 4-byte little-endian length prefix followed by JSON-serialized structs:

• Command (Command):
  • SaveReplay: Drains the circular buffer and muxes an MP4 replay.
  • StartRecording / StopRecording / ToggleRecording: Controls manual recording.
  • StartStreaming / StopStreaming / ToggleStreaming: Controls RTMP broadcasting.
  • ToggleAudio: Toggles microphone/system audio mute state.
  • CycleAudioMode: Cycles through System -> Mic -> Both -> Muted.
  • ToggleCursor: Inverts cursor visibility state and triggers capture synchronization.
  • ShowOverlay: Opens the HUD overlay via the daemon if it is not already running (scythe --show-overlay). Distinct from --menu, which toggles an already-open overlay closed.
  • ReloadConfig: Live-reloads configuration changes without restarting the daemon unless video parameters change.
  • PowerStateChanged(bool): Internal notification dispatched by the power poller when AC state transitions.
  • GetStatus: Requests the current DaemonStatus.
  • StopDaemon: Triggers clean daemon shutdown.
• DaemonStatus (DaemonStatus):
  • is_recording: bool
  • recording_duration_sec: u64
  • is_replay_active: bool
  • audio_muted: bool
  • audio_mode: String ("system", "mic", "both", "muted")
  • show_cursor: bool
  • mic_volume / system_volume: f32
  • mic_level_peak / system_level_peak: f32 (0.0 to 1.0 real-time peak values)
  • is_streaming: bool
  • streaming_duration_sec: u64

---

4. Video Capture Subsystem

Capture sources implement the FrameSource trait located in src/capture/mod.rs.

Frame Representation:

pub enum Frame {
    Raw {
        width: u32,
        height: u32,
        stride: u32,
        data: Arc<Vec<u8>>,
    },
    DmaBuf {
        fd: std::os::fd::RawFd,
        width: u32,
        height: u32,
        stride: u32,
        format: u32,
        modifier: u64,
        offset: u32,
    },
    D3D11Texture {
        handle: usize,
        width: u32,
        height: u32,
    },
}

Platform Implementations:

1. Wayland (src/capture/wayland.rs & src/capture/wayland_stream.rs):
   • Uses xdg-desktop-portal and PipeWire screencast stream.
   • Negotiates DMA-BUF memory buffers (SPA_DATA_DmaBuf) for zero-copy GPU video encoding.
   • Falls back to shared memory buffers (SPA_DATA_MemPtr) when DMA-BUF is unavailable.
2. X11 (src/capture/x11.rs):
   • Uses MIT-SHM (xcb::shm) for direct shared-memory frame transfer from the X server.
3. Windows (src/capture/windows.rs):
   • Uses the DirectX Graphics Infrastructure (DXGI) Desktop Duplication API (IDXGIOutputDuplication).
   • Iterates through all display adapters and outputs to probe and select the active desktop.
   • Uses a single Direct3D 11 staging texture (D3D11_USAGE_STAGING) with D3D11_CPU_ACCESS_READ for low-latency GPU readback.
   • Tightly packs rows into BGRA format if GPU pitch padding is detected.
   • Ensures fallback / idle frames generate opaque black ([0, 0, 0, 255]) to prevent uninitialized green chroma artifacts.
4. Mock Capture (src/capture/mock.rs):
   • Generates simulated 1080p frames for headless unit tests and CI runners where no display server is attached.

---

5. Video Encoding Pipeline

The video encoder module is located in src/encoder.rs.

Linux Encoder (VaapiEncoder):

• Uses FFmpeg libavcodec with VA-API hardware acceleration (h264_vaapi, hevc_vaapi, av1_vaapi).
• Direct DRM DMA-BUF Import: Zero-copy surface creation via vaCreateSurfacesFromProperties for Intel and AMD GPUs.
• Color Standards: ITU-R BT.709 color primaries (AVCOL_PRI_BT709), transfer characteristics (AVCOL_TRC_BT709), and MPEG limited range (AVCOL_RANGE_MPEG).
• SIMD Color Space Conversion: When software fallback is required, sws_scale operates with SWS_FAST_BILINEAR and explicit ITU-709 coefficient mapping.

Windows Encoder (WindowsHwEncoder):

• Hardware Probing Priority:
  • H.264: h264_nvenc -> h264_amf -> h264_qsv -> libx264
  • HEVC: hevc_nvenc -> hevc_amf -> hevc_qsv -> libx265
  • AV1: av1_nvenc -> av1_amf -> av1_qsv -> libsvtav1
• Standardized Pixel Format: NV12 (AV_PIX_FMT_NV12) across hardware encoders, with YUV420P fallback.
• Color Conversion: Converts captured BGRA to NV12 using sws_scale configured with ITU-R BT.709 matrix and full-range to limited-range mapping (sws_setColorspaceDetails), eliminating washed-out colors and green screen bugs.
• Low Latency Settings: Rate control VBR with burst ceiling, zero B-frames (max_b_frames = 0), ultrafast tuning, and 1-second GOP size (gop_size = fps).

---

6. Audio Subsystem

Audio capture and conversion is implemented in src/capture/audio.rs and src/encoder.rs.

Pipeline:

1. Capture:
   • Linux: Uses PulseAudio/PipeWire parec targeting @DEFAULT_MONITOR@ (for system audio loopback) and @DEFAULT_SOURCE@ (for microphone input) with a 20ms low-latency buffer.
   • Windows: Direct WASAPI loopback capture using the cpal crate.
2. Channel Downmixing & Resampling:
   • Downmixes 5.1 and 7.1 surround sound to stereo using center and surround coefficients:

     L = FL + 0.707 * C + 0.707 * SL + 0.5 * LFE
     R = FR + 0.707 * C + 0.707 * SR + 0.5 * LFE
     

   • Resamples arbitrary input sample rates to 48,000 Hz using linear interpolation and soft-clipping limiting (tanh curve).
3. Encoding (AudioEncoder):
   • Opus (AV_CODEC_ID_OPUS): Native Opus stereo at 160 kbps, 48 kHz with float sample format (AV_SAMPLE_FMT_FLT). Primary recommended standard.
   • AAC (AV_CODEC_ID_AAC): Native AAC stereo at 192 kbps, 48 kHz with planar float format (AV_SAMPLE_FMT_FLTP). Universal fallback standard.
   • Microsecond-accurate PTS rebasing aligns audio packets against video IDR keyframes.
4. Interactive Verification & Live VU Meters:
   • The UI includes live decibel / percentage VU meters for both system sound and microphone.
   • Integrated test suite (test mic and test sys with automated chime playback) allows users to immediately verify sound capture without running external applications.

---

7. Lock-Free RAM Ring Buffer & Power Management

Implemented in src/ring.rs, src/daemon.rs, and src/config.rs.

Buffer Mechanics:

• Keeps encoded video and audio packets in RAM without writing to disk during active buffer mode.
• Fixed Capacity: Capacity is calculated dynamically from duration (seconds) * target framerate * safety factor. On low-power hardware, capacity is tightly bounded to conserve memory.
• Keyframe-Aware Pruning:
  • Replay clips must always begin with an IDR keyframe to prevent visual corruption.
  • When the circular buffer reaches maximum capacity, packets are dropped starting from the oldest non-keyframe sequence up to the next IDR keyframe boundary.
• Drain Operation: When Command::SaveReplay is received, the buffer packets are cloned into an asynchronous channel without blocking the capture and encoding threads.

AC Power Management:

• Connectivity Detection: ScytheConfig::is_ac_power_connected() inspects /sys/class/power_supply/*/online on Linux and calls GetSystemPowerStatus on Windows.
• Poller Thread: A dedicated 1-second background thread continuously monitors AC power changes.
• Automatic Buffer Gating: When autostart_power_only is enabled:
  • On battery power loss: The replay buffer is automatically paused, drained of existing frames, and a toast notification (power lost, replay paused) is dispatched.
  • On AC connection restore: Replay buffering automatically resumes, forces an encoder keyframe, and dispatches a notification (power restored, replay resumed).

---

8. Muxer & Livestreaming

Implemented in src/muxer.rs.

MP4 Container Muxing (Muxer):

• Wraps FFmpeg libavformat to produce universal MP4 files (.mp4) containing H.264/HEVC/AV1 video and Opus/AAC audio.
• Emits faststart headers (+faststart) to place the moov atom at the beginning of the file for instant streaming and playback.
• Rebases PTS and DTS timestamps to zero for both video and audio streams at file initialization.
• Automatic Collision Handling: Formats filenames as Replay-HH-MM-SS_DD-MM-YYYY.mp4 or Recording-HH-MM-SS_DD-MM-YYYY.mp4 and resolves conflicts via incremental suffixes (_1, _2).

RTMP Livestreaming (spawn_stream_worker):

• Connects to RTMP endpoints (Twitch, YouTube, Kick, Custom RTMP).
• Encapsulates video and audio NAL units into FLV/RTMP transport on a dedicated worker thread with non-blocking channel dispatch.

---

9. In-Game HUD Overlay & Notifications

Implemented in src/overlay_egui.rs.

Design System: KAT Square Minimalist:

• Strictly adheres to DESIGN.md:
  • border-radius: 0 !important;: Zero rounded corners or pill badges anywhere in the interface.
  • Obsidian Dark Palette: #0b0b0b root background, #111111 surface panels, #222222 hairline borders.
  • Restrained Accents: Functional emerald (#10b981), red (#ef4444), cyan (#38bdf8), and vendor themes (AMD, NVIDIA, Intel).
  • Monospace Typography: Full support for Adwaita Mono / JetBrains Mono with an interface toggle.

Views:

• Main HUD:
  • Rolldown / swipe animated dropdown menus (ui.ctx().animate_bool(...) over 180ms) for Replay, Record, and Broadcast.
  • Quick actions with drop shadow framing.
• Settings View:
  • File Information: Output directory with native folder pickers.
  • Video Parameters: Resolution scaling, FPS (15 to 360), bitrate (5 to 150 Mbps), codec (H.264, HEVC, AV1), color range (Full vs Limited), low-power recording mode.
  • Audio: Mode routing, volume sliders (0-150%), Opus/AAC codec choice, and live verification test suite.
  • Screenshot: Resolution presets (Native, 1080p, 1440p, 2160p, 720p), format (PNG vs JPEG), and cursor capture toggle.
  • General: Theme accents, 8 global keybinds with interactive recording listener, compact Reset Hotkeys button, and typography toggle.
• Notification Toasts (ShadowPlayToastApp):
  • Borderless, shadowless terminal-style card.
  • Obsidian #111111 fill with 2px vertical accent rule and 6px status dot.
  • Uppercase monospace title, lowercase metadata subtitle, and 2px hairline progress countdown line.

---

10. Global Hotkeys & Window Hooks

Implemented in src/hotkey.rs and src/hyprland_binds.rs.

Keybind Registry:

1. Menu Overlay: Alt+Z
2. Save Instant Replay: Ctrl+Shift+R
3. Start / Stop Recording: Ctrl+Shift+F9
4. Pause / Resume Recording: Ctrl+Shift+F8
5. Take Screenshot: Ctrl+Shift+S
6. Toggle Mouse Cursor: Ctrl+Shift+F10
7. Mute / Unmute Microphone: Ctrl+Shift+M
8. Mute / Unmute System Audio: Ctrl+Shift+N

Dispatch Strategies:

1. Hyprland Direct IPC: Registers binds dynamically using hyprctl keyword bind to bypass Wayland global hotkey restrictions. Watches hyprland.conf modifications and auto-reapplies.
2. Universal Desktop: Uses the global-hotkey crate for standard Windows and X11 hotkey registration.

---

11. Configuration Management

Implemented in src/config.rs.

File Locations:

• Linux: ~/.config/scythe/config.json
• Windows: %APPDATA%\scythe\config.json

Autostart Integration:

ScytheConfig::apply_autostart_system(enable_replay, enable_overlay) manages:

1. Systemd user service: systemctl --user enable/disable scythe-daemon.service.
2. Standard XDG autostart desktop entries in ~/.config/autostart/scythe-daemon.desktop and scythe-overlay.desktop.

Retired Settings (v0.2.18 cleanup):

These keys were accepted by the settings UI and persisted to disk, but nothing in the capture, encoder, or daemon paths ever read them. They were removed from ScytheConfig and the corresponding UI controls were deleted, so the overlay no longer presents options that do nothing:

Retired key	Why it was removed
capture_source	Capture always selected the focused monitor; window/region modes were never implemented
replay_storage	Replay buffer is always RAM-backed; the disk path was never built
change_resolution, resolution_w, resolution_h	Encoder always used native capture dimensions
container_format	Muxer always writes MP4
show_scroll_lock_led	No LED driver integration existed
autostart_on_game	No game-detection layer existed
save_folder_by_game, name_by_game	No game detection; filenames are timestamp-based
language, ui_color_theme	No translation layer or theme engine; accent_color drives the real palette
record_enabled	Dead field superseded by runtime is_recording state

Unknown keys in an existing config.json are ignored by serde, so upgrading does not break saved configurations. scythe rewrites the file without these keys on the next save.

---

12. In-App Auto-Updater

Implemented in src/updater.rs.

Workflow:

1. Queries the self-hosted Forgejo release API: GET https://git.oiupoyt.space/api/v1/repos/oiupoyt/scythe/releases/latest (Falls back to GitHub API if Forgejo is unreachable).
2. Parses SemVer strings and checks if is_newer(latest, CURRENT_VERSION).
3. Downloads the singular platform package:
   • Linux: scythe-v<version>-linux-x86_64.tar.gz
   • Windows: scythe-v<version>-windows-x86_64.zip (includes standalone portable FFmpeg DLLs).
4. Atomically swaps the executable in ~/.local/bin/scythe or %USERPROFILE%\.local\bin\scythe.exe.

---

13. Verification, Testing & Cross-Compilation

Local Linux Commands:

# Type check all targets
cargo check --all-targets

# Run the complete unit test suite (27 tests)
cargo test

# Strict clippy lint verification
cargo clippy --all-targets -- -D warnings

# Build optimized Linux release binary
cargo build --release

# Install release binary and symlinks to ~/.local/bin
bash install.sh

# Run end-to-end integration test suite
bash tests/e2e_test.sh

tests/e2e_test.sh refuses to unlink a socket owned by a running daemon. Unix sockets cannot be re-linked: removing the path entry of a live daemon leaves the kernel socket bound but unreachable, and the user's session stays broken until it restarts. The script therefore sends StopDaemon and waits for the socket to disappear first, aborting if the daemon will not stop.

Windows Cross-Compilation (from Linux Host):

FFMPEG_DIR="$HOME/.local/share/ffmpeg-win64/ffmpeg-n9.0-latest-win64-gpl-shared-9.0" \
PKG_CONFIG_PATH="$HOME/.local/share/ffmpeg-win64/ffmpeg-n9.0-latest-win64-gpl-shared-9.0/lib/pkgconfig" \
PKG_CONFIG_ALLOW_CROSS=1 \
cargo build --release --target x86_64-pc-windows-gnu

---

14. Behavioral Rules & Coding Standards

1. ABSOLUTELY ZERO EMOJIS: Under no circumstances should emojis be added to code, comments, log outputs, commit messages, or communication.
2. File & Symbol Links: Every file path or code symbol mentioned in communication must be formatted as a clickable markdown link using the file:// scheme (e.g., src/encoder.rs).
3. KAT Square Minimalist: Strict border-radius: 0 !important; on all visual elements.
4. Dual-Push Rule: When pushing commits or tags, always push to both active remotes (forgejo and spearzificial):

   git push spearzificial <branch>
   git push forgejo <branch>
   

5. Memory Safety & Process Isolation: Avoid blocking the audio or video capture loops. Use unbounded or high-capacity crossbeam channels for frame dispatch.