Piik/docs/project-memory.md

Project Memory

Last updated: 2026-08-19

Confirmed Intent

• Build low-latency game screen sharing for one broadcaster and a small trusted group; public or large broadcasts belong on OBS/Twitch-class services.
• Viewers join from normal desktop/mobile browsers. A later native sender may share encoding and improve capture/audio without changing the Web viewer requirement.
• Use a small central service for access, rooms, signaling, deterministic topology, STUN, observability, and bounded SFU-root capacity; TURN is optional extreme-network transport.
• Minimize server bandwidth. The accepted invisible ladder is direct/peer UDP, then SFU/UDP roots, then optional TURN for a selected exceptional edge, then bounded failure. TURN is transport, not topology.
• Non-server nodes have at most two downstream edges; browser relays stay at one until resource gates pass. SFU normally feeds one or two roots that retain peer descendants; separately capped server edges may serve exceptional viewers that cannot attach behind a healthy root.
• Two-tree packet/layer striping may reduce endpoint upload toward one stream bitrate, but needs a bounded multi-parent, loss, sync, churn, and latency experiment; it is not in the current full-stream chains.
• Flagship native media encodes each active representation once and reuses it across direct, peer, and SFU transports; per-connection packetization, pacing, encryption, feedback, and upload remain independent. Healthy rooms have one shared HIGH; verified weak paths may add one shared on-demand LOW, never per-viewer encoders. Test simulcast, LiveKit/Dynacast, then SVC and stop at the first standard path that meets on-demand/resource gates; SVC has no software fallback.
• Keep decisions, snapshots, research, code, AGENTS.md, and .codex/ in Git; rewrite memory/status in place. Research current primary sources before material work and reject speculative machinery.
• Migrate client, server, and deployment atomically. After a canary, delete superseded config/wire/parsers/tests; do not retain compatibility layers, dual writes, or a second architecture without a current consumer. Git history owns the old implementation.

Current Recommendation

• Keep desktop Chrome/Edge and the responsive Web viewer as the baseline. Validate Android Chrome and iOS Safari as leaves.
• Use direct host P2P for one or two viewers. Keep ADR-0004 off broadly; use an isolated exact-room candidate until resource, quality, recovery, SFU/UDP, bounded-failure, and browser/mobile gates pass.
• Browser relays resend remote MediaStreamTrack values and re-encode at each hop; WebRTC does not guarantee a shared encoder across peer connections, so measure the cost.
• Keep experiments bounded: the standard representation sequence is below; native RTP relay, encoded-object striping, and FEC remain separate.
• ADR-0006's fixed-HIGH canary reached host setup and one encoder output but retained no downstream checkpoint before the first-viewer timeout. It is no-go-unclassified; no native product code is accepted, and any revisit starts at its staged evidence gate.
• ADR-0005 accepts SFU roots as the primary central fallback after direct/peer UDP. The candidate has no TURN consumer or credential wire; any future exceptional-edge grant is a complete separate change. PR #12 is superseded.
• Keep the failure-only controller default-off until the isolated STUN/SFU exact-room gates pass. Preserve break-before-make, sticky healthy subtrees, mobile leaves, and PR #20 prewarm; public transport/load remains open.
• Local reparenting is a later candidate: start with unassigned relay admission rescue; do not block current work.
• Flagship media is UDP; HTTPS/WSS stays TLS/TCP. Old production retains coturn TURN for rollback. The candidate uses self-hosted STUN-only ordinary ICE and separate LiveKit SFU/UDP, with no TURN/media-TCP wire.
• Treat settings as ceilings and deployed degradation as unclassified. Use correlated Host A+B/Viewer C to compare exact 769de201f7cc against current main for Host-only recovery, game FPS/load, preview cost, actual codec/encoder and one-variable rebuilds; then test standard simulcast/LiveKit/SVC before custom LOW. Never use UA or a composite score.
• Production reports poor film audio and self-echo when system capture includes voice software. Diagnose audio A/B/C and sync; Web cannot isolate arbitrary processes and maxBitrate is not quality-up. A Windows 11 native candidate defaults to game-process-tree audio and never widens silently; Windows 10 remains unresolved/unsupported. See docs/research/browser-screen-audio-quality.md.
• Do not add scene detection, dynamic-FPS control, or forced AV1 until stats and target hardware prove the need. Codec acceptance requires actual negotiation, power-efficient candidate evidence, interval encode cost, game FPS, CPU/GPU and sender count; Discord's native capture/hardware tuning is comparison evidence, not proof of server re-encoding or a reusable preset.
• Keep room policy deployment-driven: public/password-only rooms are random and temporary; password plus SQLite enables sequential persistent rooms and reusable stopped links.
• Deferred architecture audit: docs/maintenance.md.

Current Implementation

• The repository is one npm package using Node.js 24, React, TypeScript, Vite, native WebRTC, ws, Zod, Vitest, and separate coturn.
• Production 769de201f7cc at https://share.bonfire.icu has equal entry actions, default-closed details, bounded clarity-first quality, live source/quality changes, pause, and sender warnings.
• Production still uses one host RTCPeerConnection per viewer because peer assistance and LiveKit are unconfigured; above two viewers it still exceeds the target host-edge budget. Its real game-capture quality/pause cycle remains unverified.
• Access uses optional ACCESS_PASSWORD, a 12-hour stateless HMAC HttpOnly Strict cookie, internal host token, role-bound signaling, Origin/payload checks, and no accounts/JWT/session map.
• With ROOM_DATABASE_PATH and the password, SQLite stores room ID and host-token digest; links persist and room 1 survived deployment. Without it, rooms are temporary.
• Production uses nginx, Node.js 24.19.0, and authenticated coturn 4.17.2 at turn.bonfire.icu:3478; public STUN, TURN/UDP, TURN/TCP, and relay-only traffic pass. TURN/TLS is off.
• Candidate migration is process-wide: ordinary ICE is STUN-only, coturn disables TCP/TLS/DTLS, LiveKit 1.13.5 uses separate UDP participant ICE, and old TURN parser/signer/refresh/UI/SNI artifacts are gone. Old production is an isolated rollback.
• The default-off ADR-0005 controller has no production peer/SFU configuration. PEER_ASSISTED_ROOM_IDS restricts exact canary rooms; unlisted rooms stay P2P and empty/missing means all.
• When enabled, peer recovery precedes allowlisted SFU; session revisions prepare/commit/abort and active SFU gets one token refresh before failback. Browser relays re-encode; native sharing is outside product code.
• Chrome 151/LiveKit localhost A/B cut failure-to-active/render from 1.481/2.257 seconds to 0.200/0.320; 31 new frames and 25 ms sampling kept host edges at two. It is headless synthetic 720p30 and includes SDK/network prewarm, not public-network evidence.
• Draft native ladder #16/#18/#22/#23/#25/#28 has one bounded Chrome 720p30 loop: one WebCodecs object feeds two Pion/browser legs, applies an experiment-only minimum stock-GCC target, and recovers one isolated loss. It does not prove one hardware encode.
• Stock Pion GCC plus negotiated RTX is a no-go. The no-RTX candidate can distort loss statistics and stays outside product code until audio, heterogeneous estimates, broader loss, TURN/reconnect, and browser-diversity gates pass; do not add custom congestion control to bypass them.
• Chrome 151 short synthetic 1/3/5/8 runs kept host edges at two, relay edges at one, and all viewers decoding; a three-viewer relay close recovered in 5.32 seconds. This proves topology/control only, not quality, load, endurance, or silent partitions.
• The current quality stack uses one strict, non-persistent QualitySettings value with bounded ceilings and clarity/balanced/fluid preference. Sender readback and limitation warnings remain manual diagnostics. ADR-0007's automatic per-path HIGH/FALLBACK and unique on-demand shared LOW are accepted mainline behavior, but the state controller and LOW runtime are unimplemented.
• Chrome 151 synthetic 720p30 propagated balanced/clarity to one host and three viewers without changing peer identities; all viewers kept decoding and fanout stayed host two/relay one. This proves control continuity only.
• Local Host A+B aligns capture with one unique outbound RTP/remote/transport pair and rejects ambiguous or stale generations. Authenticated Viewer C now adds a nullable, sanitized two-second receive window of at most 2 KiB for each current ordinary or peer-assisted P2P hop; the server derives identity and current connection/revision, and the parent correlates only matching local B. It retains no raw fmtp/SDP/stats, performs no media action, and fails closed for SFU-fed roots. Automatic control and LOW remain absent.
• Release 769de201f7cc passed CI/hygiene/typecheck, 22 files/274 tests, both builds, zero production dependency vulnerabilities, atomic activation, public health/access, exact bundle, and service/log checks.

Provisional Quality Targets

• Controlled direct path at RTT no more than 40 ms and loss no more than 1%: glass-to-glass p50 no more than 150 ms and p95 no more than 250 ms.
• Regional SFU/UDP roots under the same endpoint conditions: glass-to-glass p95 no more than 350 ms; optional TURN is measured separately.
• First picture within 3 seconds after a watch request, excluding explicit permission time.
• A 60 fps profile must not remain below 50 decoded fps for more than 5 seconds without a visible downgrade or limiting reason.

These are measurement gates, not performance claims.

Open Decisions

• Sustainable viewer count by broadcaster and relay hardware, quality profile, network class, and media route; use instrumented 1/3/5/8 comparisons rather than a 1:8 claim.
• Whether ADR-0004 passes fanout, re-encoding, depth-four latency, reparenting, silent-partition, and mobile-leaf gates.
• Whether measured restrictive-network demand justifies a future selected-edge TURN grant at all, and which UDP transport/port it would use. The current candidate does not retain dormant TURN or media-TCP configuration.
• Exact mobile lifecycle behavior and whether Windows per-application audio is required for the first release.
• Project license and distribution model, which determines whether GPL/AGPL sources can move beyond study-only use.
• Initial deployment regions and expected mainland China, Hong Kong, and overseas network mix.
• Whether voice chat ever enters scope or Screener stays complementary to an existing voice application.
• Exact evidence thresholds/windows and the native hardware matrix for ADR-0007; #28's minimum-of-two rule is not a candidate.

Source Of Truth

• Documentation index and current phase: docs/README.md and docs/status.md
• Requirements and design: docs/需求理解.md and docs/方案设计.md
• Media research: docs/research/, indexed by docs/README.md
• Architecture: ADR-0001/0002, proposed ADR-0004, accepted-but-unverified ADR-0005, no-go proposed ADR-0006, accepted ADR-0007, and rejected/superseded ADR-0003
• Deployment and maintenance: docs/deployment.md, docs/maintenance.md, AGENTS.md, and .codex/