Gorka Molero

Case · agent IDE

Koda

A private review-first agent IDE concept for steering multiple coding agents through visible sessions, workspaces, checkpoints, sidecars, and visual diffs.

Working desktop build History scrubbed, publication-ready Paused after Claude access change

Problem

Modern coding agents can move quickly, but they often leave reviewers with a terminal transcript, a pile of edits, and weak control over competing attempts. The missing product layer is a workbench where the human can see parallel agent lanes, compare outputs, checkpoint progress, and decide what is allowed into the real codebase.

What I Built

Koda is positioned as a desktop review workbench for agent-era development. The private project combines a Tauri desktop shell, React UI, Rust boundary, Node sidecar, agent sessions, PTY management, workspaces, worktrees, SQLite state, checkpoints, and visual diffs.

The project was stopped when Claude removed the third-party access path Koda relied on. The useful signal is still the product and systems work: a real attempt to make agent coding reviewable, stateful, and controllable before the platform boundary changed.

Human control Review-first workflow instead of invisible agent execution.
Agent lanes Claude, Codex, Kimi, and similar runners can be treated as inspectable work sessions.
Trust layer Worktrees, checkpoints, state, visual diffs, and handoff packets make changes reviewable.
Koda architecture diagram showing operator surface, Tauri and React core, Node sidecar, SQLite state, agent sessions, checkpoints, worktrees, visual diffs, and handoff packets.
Public-safe architecture diagram grounded in the portfolio migration docs. It intentionally avoids private repository internals, local paths, session transcripts, prompts, and workspace data.
Koda desktop app running: workspace sidebar with an active main session, ASCII Koda terminal panel, and a right-side review panel listing three changed files with additions and deletions.
The real Tauri desktop app running against the Koda repo itself: agent chat surface, session tracking, and the right-side review panel showing changed files. Captured from a clean workspace with no private data; the visible shell-connection error is the honest live state at capture time.
Koda initializing splash screen with pixel-art Koda wordmark on a black background.
Startup splash from the Koda renderer. It captures the product tone without exposing private workspace state.

System Architecture

Technical Decisions

  • Treat agent work as something to inspect, compare, and checkpoint before accepting.
  • Use a desktop shell for local filesystem, PTY, and workspace affordances that browser-only tools struggle to handle cleanly.
  • Separate the UI, desktop boundary, sidecar execution, and durable state so each responsibility stays legible.
  • Make visual diffs and handoff packets the main review surface, not a secondary afterthought.

What This Proves

Koda is evidence for agent-product judgment: understanding where coding agents are heading, what makes them hard to trust, and how to design the operator layer around review, isolation, and control.