Built, running, logged.

notes

• The job queue is a table in SQLite (WAL mode) worked by a Tokio task. No broker to run or babysit.
• Reply duties carry a four-working-hour timer and a "never silent" rule: ask once, for exactly the missing items.
• Password plus OTP login with Argon2id. 895 Rust tests and 306 web tests at last count. Large module splits (agents/tools, jobs/handlers) done with no behaviour change.
• LLM reasoning mode stays disabled after one model spent its whole token budget thinking.

notes

• Hybrid topology: 2–4 peers talk over a WebRTC P2P mesh, so there is no server media cost; at 5+ the room switches to an SFU, automatically as people join and leave.
• Two SFU backends behind one flag: mediasoup by default, or a Go/Pion server. 1080p with HD fallback and bitrate caps (2.5 Mbps video, 64 kbps audio); ICE restart on failure; TURN relay with HMAC credentials.
• Load tested to 1500 concurrent connections across 500 rooms with zero errors. Per-IP and per-socket rate limits, strict CSP, vendor libraries self-hosted, Prometheus metrics with a Grafana dashboard, optional Telegram alerts.
• Dropped better-sqlite3 for Node's built-in sqlite and removed an admin UI that crashed on Node 24.

notes

• SPEC.md is the source of truth and every rule has an ID with a test named after it. P-01: raw events are never rewritten. P-04: unknown is shown as unknown. P-06: inaccurate samples are stored and filtered only at display time.
• Poison-event handling: each event in a batch is validated on its own and bad rows go to a rejected_events table, so one bad event never blocks a device's upload queue. Ingest is idempotent by event id and ordered by a per-install sequence, not the device clock.
• trackcore is a set of small Go packages — cleanpath, gaps, filter, drift, snap, livedistance, visits — with a 41-page plain-language guide on GPS error, snapping, gaps and battery, written while learning.
• The Kotlin lab app runs a foreground service with a boot keeper, motion profiles, backoff and an upload mirror, to generate real traffic for experiments.

notes

• The device is the client. The server never connects in; to send it work you queue a command and wait for the next poll.
• Fail-safe storage: if the raw request cannot be persisted the gateway answers 503, so the device keeps its buffer and retries. Never acknowledge what you did not store.
• A research log labels every protocol fact DOCUMENTED / OBSERVED ON DEVICE / ASSUMED / UNKNOWN and is updated from captured traffic. A simulator script replays sample requests in tests.
• SSE console for live punches, HMAC-signed webhooks with retries, and plain-sentence alerts such as late or absent.

notes

• Cloud sync through Supabase and AI spending summaries through OpenRouter are optional layers that never block core tracking.
• Built in nine documented phases, with ARCHITECTURE and DATABASE docs written ahead of the code.

notes

• Built for local search: TravelAgency / LocalBusiness structured data, area-served markup, and descriptive metadata on every page.
• Next.js 16 App Router with Tailwind v4, deployed on Vercel.

notes

• Three install profiles — minimal (4 plugins), standard (7), full (14) — chosen by token budget. Every combination is tested for conflicts in CI. MIT licensed, versioned releases.
• Hooks do the work without commands to remember: load past learnings on session start, snapshot state before compaction, capture wiki pages from significant sessions, block prompt injection and read-before-edit violations.
• Two siblings: ai-setup-generic applies the same setup across Claude Code, Codex, Gemini, Cursor, Windsurf, Aider, Continue and Copilot; ai-setup-creative does it for writers, designers, researchers and marketers.