Anvil

Docs · Threads

Threads

A thread is an agent at work on one thing. You read it as a conversation as it goes, answer it when it asks, and its work becomes a change.

Agents

A thread runs Claude Code, Codex, Copilot CLI or any agent an extension adds. A project picks its default in its config; a thread can pick another when it starts:

# .aide/config.toml
[threads]
agent  = "claude"     # or "codex", "copilot"
model  = "opus"
effort = "high"
anvil thread start --project harbor --agent codex --prompt "…"

Its branch

Every thread works on a branch of its own, in a worktree of its own, so threads in one repository never share files with each other or with you. The branch is the change: when the thread finishes, the change opens from it (see Changes and review).

Reading one

A thread's conversation shows what it was asked, what it says, and every command it ran and file it changed, with the diff a click away. Its terminal is a real one; switch to it to watch, or type into it.

When it asks

A thread that needs you asks once, with the answers it sees, and goes on with whatever else it can. The ask comes to your list of what needs you, wherever you are. A permission it needs (a command, a file outside its branch) can be allowed, allowed for good, or denied with a note saying what to do instead. Anything you write to it meanwhile waits until you've answered.

The Lead

Each project has a Lead, a thread that knows the project: its code, its issues, its other threads. Ask it what's going on, or give it work; it starts a thread for anything that changes files, and keeps track of the threads it started. Above all your projects, the orchestrator answers the same questions across all of them.

Where it runs

A thread runs on the machine it was started on, or wherever the queue places it, in a process of its own: closing Anvil doesn't stop a turn, and the next launch picks the thread up where it is. Move a thread to another machine between turns and it takes its conversation and its branch along (see Machines).