Concepts & glossary
A short reference for the words this documentation uses as terms of art. You do not need to read it front to back — look something up when a page uses a word you haven't met.
How work is organized
A project is one git repository that Cairn manages. It holds issues and threads.
An issue is a piece of work. Running it creates an execution, which works through the steps a recipe defines — plan, build, open a pull request. Each step is a node, and the execution panel lists them with their status as they run. Running the same issue again creates a second execution alongside the first; both stay readable.
Glossary
Agent — a role an execution can run: a prompt plus the settings that decide which model it uses and how much freedom it has. A planner, a builder, a reviewer. See Agents and skills.
Arc — a thread's living document: where its topic stands and what is coming. It appears as a tab on the thread, like any other artifact.
Artifact — a node's output, shown as a tab you can read and edit: a plan, a pull-request description, a thread's arc. Some artifacts are gated, meaning the execution waits for you to confirm one before continuing.
Backend — the service that runs the model: Claude, Codex, OpenRouter, or Ollama. See Backends.
Check — a test, lint, or build command your project declares, which Cairn runs automatically against an agent's commits. See Checks.
Coordinator — a long-running agent that drives one feature by filing child issues, reviewing what comes back, and deciding what happens next. It writes no code itself. See Coordinators.
Execution — one run of a recipe against an issue, numbered per issue. It takes a copy of the recipe and its agents at launch, so editing a recipe does not change a run already in progress.
Fence — how much freedom an agent has outside its own project: ask pauses
for your approval, deny refuses, allow removes the boundary.
Issue — a piece of work and everything that accumulated around it: its description, comments, executions, transcripts, diffs, and pull request.
Memory — something an agent learned while working and wrote down, which is reviewed before it becomes durable guidance. See Memories.
Node — one step in an execution, shown as a row in the execution panel with its own transcript, diff, and artifacts.
Recipe — the definition of a workflow: which agents run, in what order, and where it pauses for you. The default is plan-then-build. See Recipes.
Skill — a reusable instruction package an agent can draw on for a specific domain. See Agents and skills.
Task — work an agent hands to a sub-agent while it keeps its own place. Tasks appear nested under the node that spawned them, with their own transcripts.
Thread — a standing topic, named rather than numbered, that stays open across many pieces of work and files issues to do them. See Threads.
Tier — a capability level (sm, md, lg) rather than a specific model, so
an agent works the same under any backend. See Backends.
Wake — what starts an idle agent when something needs it: a failing check, a child issue's question, a merged pull request, a message from you.
Workflow — a script that orchestrates agent calls and ordinary code, for automation that a recipe does not fit. See Scripting.