Most agent harnesses either try to be an all-in-one platform or a thin wrapper; this project sits between: a focused external runner that gives an LLM a small, auditable workflow surface so work is repeatable, reviewable, and tmux-native. Its design emphasizes explicit interview→plan→execute loops and evidence-centric runs rather than opaque end-to-end automation.
What Sets It Apart
- Intent-to-software workflow: a compact default loop (deep-interview → ralplan → ultragoal) forces requirement elicitation, plan review, and goal-tracked execution so changes are deliberate and documented. This reduces blind edits by an agent and creates persistent evidence of decisions.
- TMUX-native execution and durable worktrees: runs in a chosen repository or worktree with optional tmux leader/team coordination for parallel workers. Execution is observable in-terminal and produces artifacts (session logs, notebooks, reports) for postmortem verification.
- Managed delivery & integrations: includes a configure-once SDK plus a managed Telegram reference daemon and Discord/Slack hooks so sessions can surface pending actions and accept replies without terminal scraping. The Telegram flow supports threaded per-session UIs, inline actions, and double-check acknowledgements.
- Research & desktop-control primitives: opt-in rlm (a shared Python REPL notebook mode that synthesizes reports) for iterative experiments, and a gated computer-use surface to optionally see the screen and drive mouse/keyboard for local desktop tasks.
Who It's For & Trade-offs
Great fit if you want an external, auditable harness around any LLM-based coding assistant: teams that need interview-before-change workflows, reproducible evidence of agent actions, or tmux-native parallel execution will benefit. It’s practical for interview-style coding tasks, reviewable refactors, and research loops where traceability matters.
Look elsewhere if you need a lightweight single-command code-completion plugin, a turnkey hosted agent platform, or a tightly integrated IDE extension. GJC intentionally avoids becoming a hidden plugin for specific agent runtimes; some features are beta/opt-in (desktop-control, native bindings) and require platform-specific native binaries or a Bun/npm install path.
Where It Fits
Positioned as a middle layer between raw LLM clients and full agent platforms: use it alongside Codex/Claude/OpenCode workflows when you want structured planning, persistent evidence, and tmux-backed execution without replacing your underlying model providers. The repo shows active development (2,242 stars at the snapshot) and maintains multi-platform release binaries plus an npm/Bun package path.
Practical notes
Expect experimental edges: many advanced surfaces (native @gajae-code/natives, computer-use) are gated and documented; default skills are small and opinionated (interview, plan, execute, team). License: MIT.
