Repository navigation
Description
IO-2 parallel integration handoff — G-MV + N-MOD + Multi-Device I/O
Coordination record, 2026-11-19 Asia/Jakarta. This is an executable handoff contract, not a request to wait for MVVM completion and not a second authority or a promise of background work. Refresh actual main, PR heads, source, and checks on every execution; source below is snapshot evidence.
Why this exists
User explicitly wants multi-device capture to move in parallel with GUI MVVM and native modularization, without losing JUCE realtime reliability or causing merge collisions. The closed #397 control-tower protocol was superseded; this issue does not restore its 4-lane admission bureaucracy. Follow current AGENTS.md, Multi-I/O ARD v2, PRD, and MVVM ADR.
Truth: what has really shipped vs pending
- IO-0 #770 MERGED: pinned-JUCE dual vendor-ASIO opt-in physical compatibility probe and synthetic policy CTest. It is NOT actual dual-vendor hardware FAT and does not route extra audio into a mixer channel.
- IO-1 #771 DRAFT / OPEN at
a0af6448c19d20d35830271c07803f41fe9ea85b: secondary ASIO input-only producer andSecondaryCaptureClockStreamreuse original bounded SPSC/ASRC. Exact-head Windows MSVC/CTest workflow 44224277959 completed SUCCESS. Branch is substantially behind main and GitHub marks it dirty; recheck exact diff and reconcile before merge. The adapter is NOT wired to productionLiveAudioEngine,ChannelId,.smix, native ACK/readback or /preview/. No physical dual-ASIO FAT. - G-MV27 #811 active and owns native GUI scalar/control authority work in
web/approved-native-bridge.js,web/approved-channel-strip-view-model.jsand their JS contracts. - N-MOD #799 active and owns native Player/control code including
src/Main.cpp,CMakeLists.txtand its tests. The absence of IO in Player code means no reason to halt either lane. - Primary single-device direct ASIO, existing
LiveSecondaryOutputDeviceWASAPI MAIN copy, and same-interface multichannel physical outputs are distinct existing capabilities. Separate physical BUS/CH→device B sink is a later milestone, not IO-2.
Parallel development contract — no waiting for G-MV/N-MOD 100%
| Track | Safe autonomous deliverable now | Exclusive or high-conflict paths that require a single integrator |
|---|---|---|
| IO-A (first) | Reconcile #771 onto latest main without modifying audio behavior; repair actual code conflicts, review specific ASIO class/lease and callback RT correctness; exact-SHA Fast Native, MSVC/CTest and required gate; merge only when clean | CMakeLists.txt is shared with N-MOD #799. Coordinate its final integration; NEVER overwrite newer Player targets/tests. IO adapter headers/tests remain IO-owned. |
| IO-B (IO-2) | Develop a small pure device-qualified SourceBinding / virtual input mapping contract + independent C++ tests using the accepted current ChannelId graph/port masks, with fail-closed captureGeneration/primaryEpoch and stale readback handling. Keep default OFF, no new output arm | No IO edit to src/Main.cpp or web/approved-native-bridge.js while #799 / #811 own them. Do not create another native command router, primary manager, mixer graph or JS accepted store. |
| G-MV | Continue pure GUI ViewModels + existing ACK/readback, scalar/MAIN authority and exact /preview/ parity | Own web/approved-native-bridge.js until G-MV PR merges. Do not implement dual-ASIO device opens in JS or create IO-specific PCM bridge. Future SOURCE picker consumes one native-qualified input inventory/read model once IO backend proves it. |
| N-MOD | Continue Player/setlist services, durable transactions, Native control modularization under existing native application/session owners | Own src/Main.cpp and shared CMakeLists.txt during #799; no speculative extra capture callback or second native graph; expose or preserve a clean native application boundary for later IO-2. |
| Single integration slice | After shared-file owners merge or a negotiated explicit handoff, attach verified IO-1 B stream into ONE primary render callback and the existing accepted ChannelId graph, add one serialized native ADD/REMOVE/ASSIGN capture transaction and schema-v6 readback, then add source picker projection in existing GUI ViewModel | ONE owning integration PR, exact updated main, no bulk branch merging or duplicate CMake targets/transport, no lost G-MV/N-MOD changes. |
Define the IO-2 acceptance before editing production renderer
- A.Input1 → CH01 and B.Input1 → CH05 uniquely identified (same physical pin name does not collide), with actual independent PCM and channel meters.
- B runs its own JUCE capture-only callback, writes a bounded single-producer FIFO/ASRC; only primary output render callback consumes; ONE native ChannelId/FX/VST3 graph (never per-device FX).
- B capture dropout/unplug -> only B bound source mute/anti-click; A primary playback keeps running; no forced global session, MAIN arm, FX/routing restart. Stale epochs/generations fail silently.
- Existing SOURCE ONLY / WIRE / BUS / MAIN and ChannelId/EdgeId semantics preserved; .smix offline intended physical binding and deterministic remap/restore, no friendly-name guessing.
- One native requestId / expected epoch & graphRevision / accepted ACK / later matching installed readback; no JS-local LIVE claim; existing /preview/ keeps appearance and input workflow.
- Regression gate: preallocated RT (no heap/lock/COM/driver reopen from callback), 44.1/48/96 + drift/jitter/underflow tests, zero-new-FIFO primary path, no second VST3 invocation, no output pin automatically armed.
- CI: fast tests every meaningful source change, one exact-head Windows MSVC/CTest when native slice ready, Chromium only if GUI actually changes; real two-vendor ASIO FAT and audible playback remain separate explicit external proof. Don't label simulated/native CI as FAT.
Research-backed multi-agent working method
- Short-lived PRs into current main, not long-running mutually dependent MVVM and IO megabranches; branch by abstraction + default-OFF capability/keystone interface for unfinished device I/O (AWS, Fowler CI, Atlassian).
- One writer per shared integration boundary, multiple independent domain implementation PRs, explicit producer/consumer contract and evidence. Don't mistake chat threads for communicating shared memory: each new assistant/thread must fetch this GitHub issue + active PR heads to participate. A comment is a handoff, not an automatic agent task trigger.
- Negative compatibility and rollback are first-class. No blind retry loops, unsupported ASIO fallback speaker routing, fake green statuses, force push, or merge-gate bypass. Existing exact-head required gate stays authoritative. Optional GitHub Merge Queue would require separate repository policy/CI
merge_groupsupport; do not enable it merely for this feature.
Current next action (concrete)
IO owner: refresh main/#771/#799/#811; reconcile IO-1 headers and CMake without losing N-MOD; maintain IO-1's source contract; obtain current exact-head native CI; merge after proper checks. In parallel IO-2 can implement device-qualified input binding pure module and tests on fresh main. IO-2 production native integration is not blocked by finishing ALL MVVM; only shared-file changes need synchronization. G-MV and N-MOD stay on their own bounded work.
Handoff format for any related PR
MAIN_SHA | PR_HEAD | OWNED_FILES | PRODUCER_CONSUMER_CONTRACT | CHANGED_SHARED_FILES | FAST_TESTS | WINDOWS_CTEST | PHYSICAL_FAT | MERGE_BLOCKER | NEXT_VERTICAL_SLICE.
No user action required for coordination. No claim of future autonomous execution merely by this issue.
Activity