finds.dev← search

// the find

Yeachan-Heo/gajae-code

★ 2,289 · TypeScript · MIT · updated Jul 2026

Gajae Code MVP

Gajae-Code is an external harness for coding agents (Claude Code, Codex, OpenCode) that enforces a structured workflow before code changes: requirements interview, plan review, then goal-tracked execution. It runs as a separate CLI beside your existing agent, not as a plugin that patches another runtime. Aimed at developers who want the agent to plan before mutating, with tmux-backed parallel workers and git worktree isolation for riskier tasks.

The interview-before-guessing, plan-before-mutation sequence (deep-interview → ralplan → ultragoal) is enforced mechanically by the tool, not left to the developer's discipline. The Rust native bindings package (pi-natives) handles grep, glob, diff, PTY, filesystem isolation, and computer-use input rather than shelling out from TypeScript — the right call for anything on the hot path. The WebSocket loopback SDK and gjc daemon protocol give external tools a real machine interface for session control rather than terminal scraping. Worktree isolation (--worktree flag) keeps agent mutations in a separate git worktree so a destructive run doesn't touch the main checkout.

The entire artifacts/ directory — JSON QA reports, redteam receipts, cli-replays, PTY captures — is committed to the main branch. This is internal dogfood evidence that belongs in CI artifacts or a private wiki, not the published tree. Standalone release binaries cannot dynamically load @gajae-code/natives due to Bun's compile limitations, and the README explicitly tells you not to use compiled binaries for development; it's unclear how production standalone binaries actually ship the natives module. The core parallel execution feature (--tmux + team) is described as 'not fully supported' on native Windows — the platform that can't run real tmux natively gets the worst experience for the tool's main differentiator. The repo also includes a geobench spec for measuring 'LLM hit rate, MRR, and share of voice', which is an AI-visibility marketing benchmark dressed up as tooling; it has no place in a coding harness.

View on GitHub →

// want more like this?

We dig through GitHub every week and send a few repos picked for what you actually care about — each with an honest take like this one.

Get finds in your inbox → Search again →