Files
GitHop/wiki/Getting-Started.md
T
Hop 7139b9e1a7 Automatically push every accepted Hop transition
Hop-State: A_06FN6FF8HFSFD9ANGK6BY4G
Hop-Proposal: R_06FN6FDTK9G17RXFFT82ZYR
Hop-Task: T_06FN6CFVXVVAP1KZAJ2X62R
Hop-Attempt: AT_06FN6CFVXTRZ8ZDG28QR5Z0
2026-07-11 15:13:13 -07:00

2.5 KiB

Getting started

Use Hop with an agent integration

  1. Install Hop.
  2. Select an existing Git project in a compatible agent client, or make it the controller's working directory.
  3. Ask the agent to make a normal change.

That is the full user workflow. Do not manually create .hop, route the prompt through a terminal, or tell the agent to work inside .hop/workspaces. A Hop integration does that coordination for the agent.

Without --path, hop skill install writes the same Hop-managed skill files to ~/.agents/skills/hop and ${CODEX_HOME:-~/.codex}/skills/hop. Compatible runtimes can use the shared bundle. An explicit --path installs only to the requested skills directory.

Codex Desktop example

Restart Codex Desktop after installing or upgrading the skill, select a Git project, and ask Codex to work normally. The skill is eligible for implicit activation on every repository task; mention $hop for deterministic explicit activation.

What happens on the first task

Before reading or changing project files, the agent runs hop begin. Hop then:

  • initializes local state without moving the Git branch or index;
  • stores the prompt after redacting detected credentials;
  • creates an isolated attempt workspace;
  • returns the state and workspace to the agent; and
  • keeps all project-changing work inside that workspace.

The agent validates, proposes, and lands the result. A successful hop land updates the visible project folder to the accepted tree.

Confirm the result

From the selected project directory:

hop status
hop history
hop doctor

A normal interactive result reports Root: synchronized.

If the active Git branch has an upstream—or the repository has one unambiguous origin/single-remote destination—landing also fast-forward pushes the accepted commit automatically. Hop never force-pushes. Repositories without a remote remain local without treating that as an error.

Ask for review before landing

Automatic landing is the default because the original task authorizes the local code change. To stop at a proposal, say one of the following in the task:

  • review first
  • proposal only
  • do not land

Connect another agent runtime

If a compatible runtime does not read ~/.agents/skills, install the embedded bundle into that runtime's skills directory:

hop skill install --path /path/to/agent/skills --force

Controllers that can persist a prompt before model delivery should use hop start, hop env, and hop prompt; see Agent workflow.