Squash integration of the reviewed Go core, frontend audit fixes, release tooling, and deployment hygiene. CI passed all Server/Client jobs.
6.7 KiB
Server Consolidation
- Reviewed: 2026-09-06
- Status: implemented in the candidate; the acceptance list below is the remaining boundary
- Scope: one shared Go backend for Hosted and self-contained Client operation
Current Ownership
cmd/screener-server (Hosted) and cmd/screener-client (Client) compose the
same application from internal/server and own only their defaults,
reachability, and lifecycle policy. Within that scope protocol owns wire types,
strict decoding, and shared scalars; config owns environment validation plus
the Hosted and Local compositions; room owns room authority, participants, and
the SQLite database; route decides graph transitions without I/O; signal
executes signaling and SFU effects under one lock; sfu owns admission, LiveKit
room control, and token issue; app owns HTTP composition, site access, static
assets, and lifecycle; and webassets carries the embedded Browser bundle.
internal/client keeps the native, launcher, and platform boundaries and
supervises no server process. These are useful boundaries, not duplication to
remove.
Structural Simplification
- Room expiry and abandonment share one
terminateRoompath insignal. Preserve the distinct semantics of stopping a share, ending a room, and restarting a persistent server. - The Hybrid connection-ID mirror is gone. The viewer-quality evidence path and the active media-edge lookup read the connection identity from the committed edge, and the evidence gate keeps its own last-observation identity; router tests cover in-place connection adoption during a rebuild and exact direct/peer-relayed evidence sources.
These are structural simplifications, not reproduced functional failures. The audit reconciliation separates verified fixes from disputed recommendations. Its external staged plan is not an instruction to perform extractions for their own sake.
Product Choice
The graph controller is now universal. The former rollout switch and ordinary Host-star signaling path were removed together, so Hosted and Local share one route authority and LiveKit remains an optional fallback of that graph.
Target Composition
one Go module
cmd/screener-server -> shared server application
cmd/screener-client -> shared server application + launcher + native media
internal/server -> rooms, routing, HTTP/signaling, persistence, SFU adapters
internal/client -> current native and platform boundaries
one React/Vite static artifact -> embedded in both binaries
Use packages for cohesive owners and files for local concerns. Do not introduce a framework layer per file, a second product API, or a multi-module build. Hosted and Local construct the same validated application options; entry points own their defaults, reachability and lifecycle policy.
Each room keeps one serialized state owner. Go HTTP/WebSocket handlers must not mutate room, share and route maps independently. Run password derivation and LiveKit I/O outside that owner, then apply results only while the exact session, share and operation identities still match. Preserve the existing distinction between decisions and effects rather than replacing Node serialization with unrelated locks around each map.
Reusable Components
- Standard
net/httpandembedserve HTTP and the built assets. - Signaling reuses the Client's
coder/websocket. - LiveKit room control and tokens use the standard library only: HS256 JWTs and Twirp JSON requests. The official LiveKit Go SDK adds a large module and license surface for a handful of call shapes and is not used.
- Room persistence keeps the current schema and transactions on the pure-Go
modernc.org/sqlitedriver, which builds without cgo and holds one exclusive connection; no ORM is needed. - The layout follows Go's shared
internaland multiplecmdguidance. Pion, capture sidecars and control-tunnel ownership are unchanged.
Minimum Acceptance
Preserve exact Host/session replacement, grant revocation, atomic room replacement, SQLite restart continuity and Local shutdown semantics. Exercise one graph and one operation through Peer, SFU, same-edge regeneration, bounded overlap and rollback, including stale asynchronous completions. Preserve short-lived SFU authority and release resources only after physical drain confirmation.
Use the same Browser contract scenarios against Hosted and Local, plus one physical mixed Browser/Client media flow. Package size, startup and idle memory are measured below. Keep one strict wire contract and one schema owner; avoid hand-maintained parallel TypeScript/Go validation rules. The Node server, bundled runtime, supervisor and old packaging paths retire in this same accepted integration boundary, preserving ADR-0005 and ADR-0010 as amended by ADR-0012.
Measured Before And After
One machine, 2026-09-06: Windows 11, Node 24.15.0, Go 1.26.6. A script spawned
each server with the same production configuration (PORT=18787, HTTPS public
origin, site-access password, one STUN URL, loopback listen host), polled
/healthz every 10 ms until {"status":"ok"}, waited 5 s idle, sampled
resident memory through PowerShell WorkingSet64, then killed the process.
Five runs per side; the median is reported. Before is the Node baseline at tag
baseline-b20fd88; after is screener-server under the same script.
| metric | before | after |
|---|---|---|
| cold start to healthy | 160 ms | 81 ms |
| idle resident memory | 77,242,368 B | 10,928,128 B |
| shipped Hosted bytes | ~106.7 MB | ~14.0 MB |
The before total is the Node runtime (91,694,408 B), the production
node_modules tree (12,640,469 B in 1,513 files), dist/client (1,492,721 B)
and dist/server (856,898 B). The after total is the linux/amd64
screener-server built with -s -w and embedded assets (13,983,906 B) plus its
notices and REVISION. For reference, the windows/amd64 binaries are
14,276,608 B (screener-server) and 22,703,104 B (screener-client, which also
carries Pion and the terminal UI); the Client package additionally drops the
Node runtime and dependency tree it used to ship beside its Go executable.
The first two starts of a freshly built binary took 469 ms and 336 ms, which is consistent with on-access antivirus scanning of a new executable; the median above is the steady state. Not measured: throughput or latency under load, which this port does not change, and the Linux deployment host.