Dvir Gross

Toolbox

How this site runs

1. FastPaste

   A macOS desktop app and fp CLI that turn a pasted path fragment into the right file, read local coding-agent sessions, and dispatch cloud agent runs.

   A macOS utility for developers running many coding agents across many repos. Paste a path fragment — or any text containing one — and FastPaste opens the file it refers to, even from an agent transcript, a GitHub link, or a branch that isn't checked out. A companion CLI, fp, resolves the same way from a terminal. It is a Tauri v2 desktop app summoned by a global hotkey that opens a black-and-white reader, plus a bundled fp command-line tool; both surfaces link one Rust engine, fp-core, which resolves a query over roughly 120k files in under 4 ms at p95.

   Key capabilities

   • Files — resolve path fragments and pasted references, rank ambiguous matches, read markdown and other file kinds, navigate folders.
   • Sessions — read local Claude Code, Codex, Pi, omp, fx, and DeepSeek Harness transcripts read-only; inspect conversation projections, tool calls, subagents, and context windows.
   • Dispatch — start, steer, interrupt, and stop cloud agent runs, answering the agent's questions from cards in the pane.
   • CLI — fp resolve, fp find (content search), fp sessions (index/list/show/drift), fp pin, fp bench.

   Architecture

   fp-core owns indexing, git resolution, permissions, and ranking; the Tauri app serves its warm engine over a local socket while the fp CLI links the same library. Provider adapters (fp-sessions) normalize agent transcripts into one shared model; fp-dispatch gates every outbound request behind an explicit network switch. Sessions index on demand, never at launch.

   Tech stack

   • Tauri v2 (Rust shell)
   • React + TypeScript (TanStack)
   • Rust workspace (fp-core, fp-sessions, fp-dispatch, fp)
   • local-first with a network off switch

   Experience

   Dvir built FastPaste end-to-end as its sole author — git history attributes over six thousand commits to him, against a handful from bots. He wrote the product and design briefs (PRODUCT.md, DESIGN.md), defined the shared fp-core resolver and its ranking, the provider-agnostic session model behind fp-sessions, and the explicit network switch that gates every fp-dispatch outbound request.

   Management

   Work is tracked through a Keep-a-Changelog / SemVer changelog, a large dated research-and-tasks corpus (docs/research, docs/tasks), and postmortems for lessons learned. CI is the executable contract: just ci plus GitHub Actions run Clippy with warnings denied, deny.toml dependency advisory and licence gates, and the Rust and frontend test suites.

2. Navada Platform

   An AI agent orchestration platform: a Tauri desktop control surface, a Go API daemon, an execution controller, and a terminal client.

   Navada is an AI agent orchestration platform. Operators create and configure agents, dispatch work, watch iterations and tool calls stream in, and reconstruct runs afterwards. Agents are data (MongoDB documents, not code); pluggable runtimes execute them behind one operator surface. Four cooperating services: Navada Desktop (Frontend, Tauri v2 + React) is the operator control surface; the Daemon (Go, Gin + Huma) owns the canonical agent/session/execution API; harness-controller (Go, ConnectRPC) runs the iteration loop around pluggable harness adapters; nvd-cli is a single-binary Go terminal client (Bubble Tea) driving the Daemon's remote API for multi-agent, multi-session work.

   Key capabilities

   • Agent create/monitor/execute, with SSE streaming of iterations and tool calls as they happen.
   • Session and execution history with search, status/recency filters, and persisted run/trace viewing.
   • Pluggable harness adapters (C4mula, Jarvis, Reacher, Revi, Zhu-Li) behind one Harness contract.
   • Keyboard-first, multi-tenant (Stytch B2B) UI, a loopback-only local coding-agent session inspector, and a terminal workflow via nvd-cli.

   Architecture

   The Frontend calls the Daemon over HTTPS; the Daemon reaches harness-controller over ConnectRPC/h2c, which iterates over registered harness adapters toward LLM providers (OpenRouter, Anthropic). Streaming follows a canonical SSE event schema, and execution IDs are stable and retry-safe. Tenant scoping is enforced in the data layer through scoped repositories.

   Tech stack

   • Tauri v2 + React + TypeScript + Vite (Frontend)
   • Go, Gin + Huma, MongoDB, Stytch, OTel (Daemon)
   • Go + ConnectRPC (harness-controller)
   • Go, Bubble Tea + Lipgloss + Cobra (nvd-cli)

   Experience

   Dvir authored the platform's four services — Frontend, Daemon, harness-controller, and nvd-cli — as their primary committer; git history across all four repos shows the work under his own accounts. He defined the Harness adapter contract, the canonical SSE streaming event schema, and the scoped-repository tenant model, and made the binding decisions recorded as owner rulings in the workspace register.

   Management

   Navada is run through a formal decision register: workspace-level D-records in the Atlas decisions register, per-project registers, and ADRs, alongside roadmap and adoption notes per service. Each service carries its own CI workflow, and internal RPC contracts are distributed by publication plus local generation.