alpha: flywheel execution board for performant chat/live avatar system

[joelteply]

airc-queue card

Coordinates work via the AIRC queue substrate (airc#562). Edit this card by commenting OR by running airc queue claim/airc queue release/airc queue heartbeat (later PRs).

{
  "kind": "airc-queue-card-v1",
  "id": "#1272",
  "owner": "codex-main",
  "status": "in-progress",
  "evidence": "Adopted existing GitHub issue into airc queue.",
  "next_action": "Keep alpha lanes ranked; dispatch/heartbeat agents through AIRC; refresh blockers and merge-ready PRs.",
  "last_heartbeat": "2026-05-16T02:43Z @ 0367af9b7"
}

Close this issue when the work is done (status=merged/abandoned).

Original issue body

Pre-adoption body

Goal

Make the path to alpha explicit, ranked, and queue-driven so agents do not wait in chat or chase unprioritized tangents.

Alpha means a fresh install can run a performant Continuum experience with reliable chat/live-avatar workflows, minimal setup friction, Rust-owned hot paths, and AIRC-backed coordination where it belongs.

Ranked lanes

1. Fresh install and persistence

   • SQLite-first main DB with Postgres opt-in only.
   • No local Postgres required for default persona/chat workflow.
   • Install smoke passes on macOS and Linux/amd64.
   • Related: Replace Continuum chat/Postgres dependency with AIRC Rust SQLite substrate #1253, Make focused continuum-core Rust tests work with required Metal features by default #1257, fix(install): align tier name to registry canon + pass TIER to model-init + fail loud on unknown tier #1085, fix(mac): gate native LiveKit bridge behind explicit profile #1071.

2. Chat/live avatar E2E

   • Text, event delivery, subscriptions, presence (is typing / is thinking / speaking), and receipts flow through typed events.
   • LiveKit/WebRTC owns media transport; AIRC/Continuum own handshake, replayable events, and control state.
   • Related: Route persona presence, subscriptions, and WebRTC control through AIRC realtime events #1255, Define performant realtime transport split for chat, AIRC events, and LiveKit WebRTC media #1258, Move live chat hot path off ORM toward AIRC/event-log transport #1113, voice: retire legacy WS transport, unify on LiveKit WebRTC #914, feat: install reliability — skills + verify-personas + GPU/IPC/seed hardening #913.

3. Runtime CPU and bus pressure

   • Measure and reduce bus/chat/event-loop CPU load.
   • Remove repeated ORM/IPC hot-path churn.
   • Add before/after metrics for CPU, memory, subprocess count, latency, page/replay latency, and disk writes.
   • Related: perf(runtime): reduce bus/chat/event CPU from sluggish 700% load #1260, perf(bench): Criterion microbench for cognition+persona prompt construction (validates #1209/#1218 wins) #1225, precommit gate too thin: only browser-ping runs; no chat-roundtrip / adapter unit / persona-reply smoke #1186.

4. AIRC coordination substrate

   • Agents claim/heartbeat through AIRC queue automatically.
   • Idle agents get dispatched claimable work without manual reminders.
   • Rust SQLite substrate owns queue/chat/event state; Python/shell remain transitional wrappers only.
   • Related: fix(queue): dispatch metronome hand-outs airc#627, rust: port queue dispatch monitor substrate to typed local state airc#628, rust: eliminate Python from AIRC hot path airc#629, smart-issue: auto-handout next claimable card to idle agents on interval #1192, GitHub Pages portal for Continuum AIRC rooms and kanban #1128.

5. Rust-first AI/runtime architecture

   • Durable persona cognition, scheduling, paging, resource pressure, and model routing move down to Rust.
   • TS/browser/configurator remain thin glue and UI.
   • Shared/browser must not import Node-only logic.
   • Related: TS-side AI logic that violates 'TS is thin glue' directive — umbrella triage #1248, architecture: Rust AI bus scheduler for GPU-resident throughput, memory paging, and kernel offload #1201, RTOS-for-AI-Throughput: Rust-owned memory paging + GPU-native inference (umbrella) #1203, code-smell: modules/cognition.rs is a 1390-LOC 37-arm match — split by domain (lifecycle/admission/genome/inference/dedup) into trait-implementing submodules #1205, code-smell: every Rust-backed TS command repeats the same shape — extract RustBackedCommand<P,R> base class #1198.

6. GPU-native multimodal inference

   • Qwen-first multimodal model matrix.
   • CUDA/Metal/Vulkan validation for no hot-path CPU fallback.
   • Fail loud instead of silently falling back to CPU.
   • Related: research: upstream Qwen vision/audio model matrix + zero-CPU GPU backend plan #1200, hygiene(inference): remove unconditional llama crate build from continuum-core #1227, GPU backend lifecycle: recover or fail loud after backend OOM/death #1050, MtmdContext pooling for multimodal inference #1051, GPU backend lifecycle: serialize mmproj/mtmd initialization #1048, PressureBroker must gate unsafe GPU/resource allocations #1049.

Operating rules

• Work from AIRC queue cards, not chat impulse.
• Every active lane has one owner and heartbeat.
• If a blocker is discovered, either fix it immediately or file/link a concrete card with expected impact.
• Hooks are not bypassed. Hook failures are planning signals.
• Keep public repo text technical and generic; no personal attribution or private context.
• Prefer small, mergeable PRs with passing focused verification.

Next actions

• Refresh/merge fix(config): make SQLite the default main database #1271 so SQLite-first defaults land.
• Assign one owner to perf(runtime): reduce bus/chat/event CPU from sluggish 700% load #1260 bus CPU measurement and one owner to Route persona presence, subscriptions, and WebRTC control through AIRC realtime events #1255/Define performant realtime transport split for chat, AIRC events, and LiveKit WebRTC media #1258 realtime event split.
• Move AIRC canary dispatch automation into main once Benchmark: big brain in small body — compressed models vs same-VRAM natives #627 is ready.
• Build a visible planning surface from AIRC queue (GitHub Pages portal for Continuum AIRC rooms and kanban #1128) so idle agents can see the alpha lanes without being told manually.

Status log

• 2026-05-16T02:43Z — heartbeat by codex-main -> status=in-progress (Investigating chore: promote canary to main (79 commits, 17 install fixes from 2026-05-03) #1035 main-promotion CI failure: Carl install smoke shows continuum-core unhealthy due IPC server bind No such file or directory; using separate canary worktree for fix.)

• 2026-05-15T17:07Z — heartbeat by codex-main (Alpha board live; fix(config): make SQLite the default main database #1271 refreshed on canary and pushed. Pre-push TypeScript/ESLint/Rust compile/Rust tests passed; native Docker image push reported tracked-file state, investigating.)

• 2026-05-15T17:00Z — heartbeat by codex-main (Alpha flywheel board created/adopted; refreshing SQLite PR fix(config): make SQLite the default main database #1271 and then dispatching lanes.)