Wiki
The vocabulary, defined once. These entries are the terms this whole product is built out of: what each one means in general, and what changes about it when several agents are working the same repository at the same time.
Several agents, one repository
CLI coding agent
A coding agent that runs as a command-line process on your machine (Claude Code, Codex, OpenCode, Gemini CLI) rather than inside an editor or behind a web app. It reads and writes your real files with your real permissions, which is both the appeal and the risk.
Dispatch
Dispatch is handing a task to an agent along with the environment it needs (working directory, model, effort level and an isolated checkout) so that starting work does not require a human to set the stage first.
Kanban agent task
A unit of work on a board that an agent can pick up on its own, work in an isolated checkout, and hand back with evidence. It differs from an ordinary ticket in that the acceptance criteria have to be machine-checkable, not just human-readable.
Topic
A topic is one unit of work that holds everything belonging to it (the agent conversation, a real terminal, the project files and their diffs, a browser, and what the turns cost) instead of scattering them across four windows that only your memory connects.
Worktrees and landing
Git worktree per agent
A git worktree is a second working copy of the same repository, checked out to its own branch in its own directory. Giving each coding agent one is what lets several of them edit the same project at once without touching each other's files.
Landing vs approving
Approving accepts an agent's work; landing merges it into your main branch; publishing pushes it to a remote. Collapsing the three into one button is how code you have not read ends up somewhere you cannot easily take it back from.
Squash-landed branch
A branch whose changes are all in main but whose commits are not, because landing squashed them into one new commit. Git's "is it merged" check answers about ancestry, so it says no, and any cleanup that trusts that answer either keeps everything or deletes the wrong thing.
Worktree reaping
Removing a git worktree once its work has landed. Skipping it leaves branches, disk and (worse) anything still pointing at that directory in a state that looks broken rather than finished.
What agents cost
Agent turn
A turn is one complete cycle of request, reasoning, tool calls and reply. It is the unit that cost, latency and context growth are all measured in, which makes it the only honest thing to count when you want to know what an agent is spending.
Auto-compaction
The automatic summarising of older context when a session approaches the model's window limit. It keeps long sessions alive, and it is also the moment a watchdog is most likely to mistake a working agent for a dead one.
BYOK (bring your own key)
A model where the software holds no credential of its own and proxies nothing. It uses your API key or your signed-in CLI, so your prompts go straight to the model vendor and the tool never sees them, never resells inference and cannot take a margin on it.
Context window
The ceiling on how much text a model can see in one turn. Now that a million tokens is an ordinary size, the binding question has changed from "will it fit" to "what does it cost to keep it there", because everything in the window is re-read and re-billed every turn.
Effort level
Effort is how hard a model is asked to think before it answers. It is set per session rather than per account, it changes both the quality and the price of a turn, and a session started from a bare shell can silently get the cheapest tier without saying so.
Prompt caching
Prompt caching lets a provider reuse the already-processed prefix of a prompt instead of reading it again. It is the difference between paying for your system prompt once and paying for it on every single turn, and it breaks the moment anything near the front of the context changes.
Subscription vs API billing
Two different ways to pay for the same model, and they are not interchangeable. A subscription authorises the vendor's own CLI or app; the API bills per token against a key. Nothing bridges them, which is why a tool that wants to use your plan must run the real binary.
The substrate
Checkpoint
A saved point in a session you can return to, taking the working tree back with you. It is the answer to an agent that went somewhere wrong twenty minutes ago and has been building on it since.
Local-first
An architecture where your data lives on your disk and the network is an optimisation rather than a requirement. It is not the same as offline support: the difference is which copy is authoritative when the two disagree.
Lossless reattach
Reconnecting to a terminal session and getting back the whole scrollback, not just the last screen. Almost every embedded terminal gets this wrong, and the difference shows up exactly when it matters: when you want to know what happened while you were not looking.
LWW (last-writer-wins)
A conflict rule that keeps the most recent write. It is the simplest thing that works for synchronising state across devices, and it is only correct if "most recent" is decided causally rather than by reading two clocks that disagree.
PTY (pseudo-terminal)
A pseudo-terminal is a pair of kernel devices that makes a program believe it is talking to a real terminal. It is why `vim`, `htop` and every interactive CLI behave differently when you pipe them, and it is the only channel through which a paid CLI subscription can be driven.
PTY bridge
A process that owns the terminal sessions independently of the window showing them. Because a PTY dies with whatever holds it, keeping sessions alive across a reload, a crash or a server restart means moving ownership somewhere longer-lived than the interface.
Scrollback
The lines a terminal has already printed and still keeps. It is the part almost every embedded terminal discards, and the only part that answers the question you actually have when an agent finishes: what did it do while I was away.
Tombstone
A record that something was deleted, kept instead of simply removing the row. Without one, a sync cannot tell "you deleted this" from "I have something you have not seen yet", so deleted things come back.
What the shell costs to run
Pane
A pane is one live surface inside a topic: a chat, a terminal, a browser, a diff. Unlike a tab in a text editor, a pane usually has a process behind it, which means it has a lifecycle and a measurable cost even when you are not looking at it.
Split
A split divides a pane in two, producing a tree of rows and columns rather than a flat list. It is what lets a terminal, the file it is editing and the page it renders sit on screen together instead of taking turns.
Window (group)
A group is a set of tabs you live in: the unit a window represents. Confusing it with a cell of a split layout is the most common mistake in tiling interfaces, because the two look identical on screen and behave nothing alike.
Protocols and the toolchain
ACP (Agent Client Protocol)
A protocol that lets one interface talk to different coding agents without hard-wiring any of them. Where MCP standardises what an agent can reach, ACP standardises how a client and an agent talk to each other.
MCP (Model Context Protocol)
A protocol for exposing tools, data and prompts to an AI agent over a standard interface, so a capability written once can be used by any client that speaks it. The interesting question it raises is not transport. It is who decides which tools an agent can see.
MCP scoping
Restricting which MCP servers a given session can see. It matters because every tool an agent is offered costs context on every turn and adds one more way to pick the wrong thing, and because inherited configuration is the usual reason an agent has tools nobody meant to give it.