Codex task report - 2026-06-16

W2 - Per-peer EQ

Added src/audio/eq.rs: a 3-band listener-side RBJ biquad EQ (low shelf, mid peaking, high shelf) with per-peer state and flat bypass.
Added local config persistence in AppConfig.peer_eq, keyed by peer node id string.
Added local CoreCommand::SetPeerEq and mixer-side per-peer Eq state. EQ is applied after local volume and before pan/mix; raw multitrack stems remain pre-volume/pre-EQ.
Added participant-card controls for Low/Mid/High gain sliders (-12 dB to +12 dB). Changes apply live and persist on slider release.
Tests added for flat identity, low/high boost energy, coefficient finiteness, clamping, and hot-signal processing.

Unverified: subjective voice quality and zipper/noise behavior on real devices.

W1 - Per-listener pan / stereo playback

Added src/audio/pan.rs: constant-power pan_gains() with tests, plus playback gains that preserve the legacy default dual-mono center.
Converted playback mix to interleaved stereo in src/core/mod.rs.
Switched PipeWire playback output to 2-channel S16LE and adjusted ring target/capacity/stride accounting in src/audio/pipewire_impl.rs.
Kept capture, Opus encode/decode, jitter buffers, and network audio mono.
Limiter now receives the interleaved stereo bus; shared limiter gain ducks both channels consistently.
Mixed WAV and multitrack convenience mix fold the listener stereo mix back to mono before writing. Per-peer stems remain raw mono.
Updated audio_probe to send dual-mono stereo frames.
Added tests for exact center dual-mono behavior, hard-left pan contribution, and stereo fold-down.

Decision for senior sanity-check: pure pan law is constant-power, but playback scales it by sqrt(2) so pan=0 is exactly the old mono signal in both ears. This satisfies the "default behavior unchanged" guardrail at the cost of louder hard-panned extremes, which the existing limiter catches.

Unverified: real PipeWire stereo playback, underrun behavior on actual hardware, and recorded WAV listening checks.

Added src/hotkeys.rs: serializable KeyBinding, HotkeyAction, HotkeyMap, parse/format/lookup, tier checks, and duplicate conflict detection.
Added AppConfig.hotkeys with defaults: F9 mute, F10 deafen, F2 Settings, Space push-to-talk, Leave unset.
Replaced the hard-coded PTT key capture with config-backed binding capture.
Added Settings hotkey editor with Set/Clear per action and live conflict warnings.
Added top-right hotkey info popup that lists every action and current binding, showing unset for unbound actions.
Routed focused iced key events through the map. App-wide actions can fire from any screen while focused; room-only actions require an active call. PTT press/release still uses SetPttActive.
Tests added for unset formatting, duplicate detection, room-tier lookup, defaults, and character parse/format.

Unverified: manual keyboard interaction in the GUI. No OS-global hooks were added.

W3 - PipeWire pro-routing plan (not implemented)

I stopped at design for W3. The current backend already supports simple target-node routing through PipeWire stream property node.target, but true "pro routing" (explicit ports / manual graph links / no-autoconnect patching) would require backend changes that are not safely verifiable offline.

Proposed future scope:

Expose two advanced route targets: capture source node and playback sink node, with optional future per-port routing.
Enumerate available nodes with the existing pw-cli list-objects Node parser. For port-level routing, add a separate parser for pw-cli list-objects Port collecting object.id, node.id, port.name, direction, and channel position.
For node-level routing, continue using PipeWire stream property node.target on stream creation. This is the low-risk path and matches current backend behavior.
For explicit port routing, do not use AUTOCONNECT; instead capture the created PeerSpeak stream node/port ids from the PipeWire registry, then link with PipeWire-native APIs or pw-link <source-port-id> <sink-port-id>. Degrade by falling back to node.target autoconnect if any selected node/port is missing.
Offline tests should cover pure routing-plan decisions: selected node exists/missing, selected port exists/missing, capture/playback direction mismatch, and fallback choice. Real-device tests still need a PipeWire graph.

Reason for not implementing: the current run_playback / run_capture code does not retain stream node or port ids, and changing AUTOCONNECT behavior plus adding manual pw-link calls could destabilize the working audio path. That matches the assignment's "bail if risky" instruction.

Backlog A21/A22 - correctness fixes

Fixed A21 in src/core/jitter.rs: implausibly large sequence discontinuities now reset the per-peer jitter stream instead of being treated as ordinary late packets or packet loss.
The reset threshold is 500 frames, about 10 seconds at 20 ms/frame. That covers both same-identity sender restart back to sequence 0 and a faulty/malicious jump far ahead that would otherwise force a long PLC run.
Added jitter regression tests for both far-behind restart and far-ahead jump cases.
Fixed A22 in src/audio/recorder.rs: WavWriter now tracks data bytes as u64, checks additions before writing, and rejects data that cannot fit both the RIFF size field and the data chunk size field.
Added a WAV overflow regression test that exercises the limit without creating a huge file.

Unverified: the same-identity peer restart has not been exercised in a live 2-machine call; the WAV fix is counter/size-field tested, not a real >12h recording.

Backlog A14 - orderly window-close shutdown

Added CoreCommand::Shutdown and UiEvent::ShutdownComplete.
Window close now saves config, marks the GUI as closing, asynchronously queues Shutdown, and exits only after the core acknowledges completion or after a 5-second fallback timeout.
Core shutdown finalizes active mixed/multitrack recordings before session teardown, stops the standalone mic monitor, runs ActiveSession::shutdown() for active calls, clears room presence/routing, closes the persistent network stack, sends ShutdownComplete, and ends the core loop.
The shutdown command is queued with an awaited mpsc::Sender::send task instead of the best-effort try_send, so a full command queue does not immediately drop the close command.

Unverified: actual GUI window-close behavior during a live call/recording still needs a manual run; tests/builds only prove the path compiles and existing unit coverage still passes.

Verification

cargo check passed.
cargo test --lib passed: 288 passed, 0 failed, 2 ignored.
cargo clippy --all-targets passed.
cargo build --release passed.
Formatted the touched Rust files with rustfmt --edition 2024; I did not run repo-wide cargo fmt to avoid unrelated formatting churn.

No new dependencies were added. Runtime/manual/field verification is still pending for audio-device and 2-machine behavior.

7.0 KiB Raw Blame History