AICoder

Appearance

Core concepts ​

A few ideas explain how everything in AICoder fits together. Learn these once and the rest of the app is obvious.

Task → Agent → Worktree ​

This is the spine of the app.

Task           a unit of work; owns a branch off your base branch
 └─ Agent      an AI agent working the task in an isolated git worktree
     └─ Session one run of that agent (coding or review)
  • Task — a feature, bug, or investigation. It owns a git branch created off your base branch (e.g. main).
  • Agent — the AI (claude or codex) doing the work. It runs in an isolated git worktree under .aicoder/worktrees/, so its changes never touch your main checkout. A task can also run several agents in parallel (see Run agents in parallel), each in its own worktree.
  • Session — a single run of an agent against a task. A task can start, stop and resume many sessions over its life.
  • Worktree — a real git worktree: a separate working directory sharing the repo's history. This is what lets agents edit code without disturbing your working copy — and lets several run at once.

Parallel agents

Use Split into parallel agents in the task drawer to run more than one agent on a task at the same time — for example, to race two approaches. Each agent gets its own worktree.

The review pipeline ​

When a coding session exits cleanly, AICoder automatically runs a review pipeline over the change:

  1. DRY — flags repeated / copy-pasted code.
  2. Docs — flags missing or stale documentation.
  3. Security — a dedicated validated security pass.
  4. Plus any custom checks you define in Settings.

If everything passes, the work is marked approved. If not, it's marked needs work — and Fix with AI re-runs the agent with the findings injected so it can address them. See Reviews.

Model routing ​

Agents are the real CLIs, so they spend your subscription quota. To keep cost down, AICoder routes work to the cheapest capable model by default:

  • Coding → Sonnet
  • Reviews, judge, explain, Azure → Haiku
  • Opus / Fable → opt-in per task, with a warning before they run.

You can override the model per task in the task drawer. Full detail in Model routing.

Attention, not noise ​

AICoder assumes you're running several agents at once. Its job is to route decisions to you and keep the rest out of the way:

  • The Attention Queue is the home view — a list of sessions that need a human (a question, a review, a ship decision).
  • The Workbench is the opposite: a live grid of everything running, for when you want the firehose.
  • The Dashboard is the top-level summary across all projects.

Two project types ​

  • Software projects pair with a git repo and run the full loop above.
  • Azure projects need no repo. They run read-only az investigations and save findings to project docs — no worktree, no review, no PR. See Azure investigations.