Skip to main content

Documentation Index

Fetch the complete documentation index at: https://gump.build/docs/llms.txt

Use this file to discover all available pages before exploring further.

Git Worktrees

Every run executes in an isolated Git worktree. Your main branch stays clean — no half-finished code, no broken builds, no risk of corruption from a misbehaving agent.

How it works

When you run gump run, Gump creates a worktree in .gump/worktrees/run-<uuid> attached to branch gump/run-<uuid>, runs all steps inside it, and commits a snapshot after every agent step.

Worktree modes

Each step type has a default worktree mode, overridable:
ModeDescription
read-writeThe agent can read and modify the worktree (default for code)
read-onlyThe agent can read; guard enforces no writes (default for split, validate)
noneNo worktree — temp directory. The state is the only I/O channel
read-only is enforced via the no_write guard (post-step check). OS-level sandboxing is planned for day 2.

Parallel worktrees

When parallel: true is set on a split or a group, each branch gets its own worktree. Merge is sequential by declaration order. Same file modified by two branches = fatal.

Apply

gump apply merges the run’s branch into your current branch with a Gump-Run: <uuid> trailer. gump apply --run <uuid> targets a specific run.

Retry behavior

In retry, worktree: reset means git reset --hard + git clean -fd. Without the override, the worktree is kept as-is (keep).

Cleanup

gump gc --keep-last 10
Removes old worktrees, branches, and run artifacts. Unapplied runs prompt for confirmation.

Dirty check

Before creating a worktree, Gump checks that your working directory is clean. .gump/ is excluded from this check.