Glossary

Terms

Coding agent

An AI tool that reads, edits and tests code from your terminal, such as Claude Code or Codex. Treelab organises its work; access to the agent is not included.

Worktree

An additional working folder attached to the same git repository, with its own branch. Treelab opens one per task, so that each agent works in isolation. See the worktrees guide.

Main checkout

The repository’s original folder. With Treelab, it never receives a commit: every change goes through a worktree and its branch.

Lane

The lab CLI’s word for a task and its worktree: its messages say “lane” where this site says “worktree”. Older batch files and versions say « couloir », its former French name; lab help glossary lists each of lab’s words with its former French name.

Scoping

A short interview at the start of a task: the grill-me skill asks its questions in rounds, then summarises the decisions, which you confirm.

Lot

The smallest useful delivery of a task: one branch, one agent session, one measurable objective. lab’s messages call it a batch. See the lot method.

Lot document

A lot’s document: objective, scope, completion criteria, steps, risks, log and handoff. It is the only source of the lot’s state. lab calls it the batch file.

Objective (/goal)

The first line of the document: an observable result, which lab launch hands to the agent as written. The agent does not guess it; it receives it. lab calls it the Goal.

Completion criteria

The evidence that a lot is done. Each one is checked by a command or an observation, never by an impression. lab calls them exit criteria.

Handoff

A section of the document written by the agent before passing on: where it stands, what remains, the next action. The next session starts from there.

Independent review

A review of a lot’s diff by an agent that did not write it, against the document: only correctness or scope gaps are reported.

Proof report

The document stored with the lot that keeps commands, exit codes, measurements and screenshots. The lot document points to it instead of copying. lab calls it the evidence report.

Acceptance trial

Your trial of the result, after the last lot. lab acceptance records your approval, tied to the revision tried; lab close delivers only after that.

Guardrail

A default refusal of a common destructive operation, with the safe alternative. A closed list, not a sandbox. See the security page.

Hook

A script the coding agent runs at a precise moment: when a session starts, when it ends or before each command. Treelab uses hooks for context and guardrails.

Skill

Reusable instructions the agent applies on request: scope, plan a lot, run it, close it. The same for Claude Code and Codex.

cmux

A tabbed terminal application for macOS. Treelab opens one tab per worktree in it, with its agent; it has no affiliation with cmux.