AI coding without the vicious spiral. Infil into the codebase, brief the task, ship in splashes, exfil with the diff — under your command, not the model's.
Ship every change through a written, military-style OP cycle that stays disciplined, recoverable, and accountable. A web app plans the ops, a CLI gives agents every tool they need in your codebase, and AI skills drive the cycle.
Two landscapes, one loop. Intel lives in Mission HQ. Code lives in the battlefield. AI skills sync between them or operate on the ground.
┌────────────────────────────────────────────────────────┐ │ Mission HQ: Naholo web app │ │ │ │ [Ideation: Drop intel as logs or notes] │ └────────────────────────────────────────────────────────┘ │ /infil ^ /sitrep ^ /exfil v (pull OP data) │ (push OP data) │ (push + close) ┌────────────────────────────────────────────────────────┐ │ Battlefield: Your codebase │ │ │ │ /warno (write architecture decisions) │ │ ↓ │ │ [Review architecture decisions] │ │ ↓ │ │ /opord (cut into single-commit tasks) │ │ ↓ │ │ [Review specs] │ │ ↓ │ │ /splash (ship one task) ◂┐ │ │ ↓ │ [Repeat] │ │ [Review diff] │ (/opord if revising tasks) │ │ └────────────────┘ │ └────────────────────────────────────────────────────────┘
/infilPull notes, tasks, and logs from the server into a local working dir.
/warnoResearch the codebase and lock-in the architecture decisions before any tasks are cut.
/opordCut the WARNO into single-commit-sized tasks. Each task ships in one splash.
/splashShip one taskl implement the change, then write an After-Action Report that flags any deviations from the plan.
/sitrepCheckpoint mid-mission: your tasks, notes, and a summary log sync to the server without closing the operation.
/exfilClose out the operation and push all local context to the server: notes, tasks, and agent usage stats.
The web app is an issue tracker for your codebase: operations are the issues, logs are the comments, and notes are markdown scratchpads.
The OP is the unit of work — the cycle revolves around it. Drop context as logs (comments) and freeform notes (markdown) onto each operation; the CLI infils both into the local working dir when you start work.
Tokens spent, model breakdown, and session timing are all recorded to the OP on exfil. Review the usage and run the next op more efficiently.
A local CLI wires your AI agent into the OP: it exposes MCP tools and commands for Mission HQ, tracks each session per OP, and stays worktree-safe.
/fob starts an OP from the codebase, /recon talks one out, /raid is the small-OP shortcut, and /chop splits an OP.The loop is the same either way; what changes is how naholo lives in your codebase. Classify your project by how much control you have over it.
You deploy into territory you don't control, so no base goes up: nothing is committed and no trace is left behind. Your kit stays in your ruck (~/.naholo). Run naholo covert init and go dark.
Choose this when you're working on a team project but don't want to, or can't, introduce naholo to the rest of the team. You'll need to run naholo covert init each time you create a new worktree.
This is your territory, so you set the rules. Deploy a full base at .naholo/ in the project root with config checked into source, so every operator runs the same loop. Run naholo init and push forward with your team.
Choose this when you own the repo or have the authority to introduce naholo to your team. Config lives in source, so new worktrees work out of the box; no extra setup required.
What you need before boots hit the ground.
Claude CodeOpenAI Codex will be supported later.
macOS
Node.js v22+May work on other Unix systems(not yet confirmed). Linux and Windows will be supported later.
Workflow involves reviewing and editing markdown docs.
Start your first operation, or fork it and spin up your own instance. Available under MIT as open source.
View on GitHub