Private data across machines

The harness is public, but a working installation holds private data: your accounts, your machines, your own instructions, the memory of your sessions. This page says what is private, how the harness keeps it out of the public repository, and how it reaches each of your machines, physical or virtual.

What is private

data	where it is	what it holds
the profile and its files	~/.config/research-harness/	profile.toml, every value that names a person, an institution or a machine; models.toml, the model of each tier; providers.toml, network.toml and repositories.toml. Adapting the profile names each key
the tree instructions	the directory that the profile key tree_agents names	your own instruction text about your research tree, and your own rules and skills
the memory and the plans	~/.claude/projects/<memory_project>/memory/ and ~/.claude/plans/	what Claude Code keeps from one session for the next. Research memory describes it
the credentials	~/.ssh/, the logins of gh and glab, the API keys	the keys that let a machine reach your accounts

The first three are text that the harness reads and installs. The credentials are not: the harness never reads, copies or installs a credential. The settings template denies a session a read of ~/.ssh/, ~/.gnupg/, ~/.config/api-keys/ and the token file of gh, both to the file tools and to the sandbox.

Keep private data out of the repository

A source of the harness names no person, no machine and no repository of your tree. Where a source needs such a value, it is a template that names a key of the profile as {key}, and a subcommand renders it with your profile before it uses it. So the repository holds the keys, and your profile holds the values.

Two checks keep the values out:

• harness leaks searches every tracked file, its path, and the templates rendered with the example profile. It looks for each string of the profile's leak list, each account of org, and the name of each directory below repository_roots.
• gitleaks looks for secrets and home paths.

The pre-push hook of the harness runs both before a push to main, and refuses the push on a hit. CI runs gitleaks on every push. CI has no profile, so it cannot search for your private strings: only the hook on your machine does. The security model gives the limits of this gate, and Development gives the two checks in detail.

One profile for every machine

One profile serves every machine of one user. A value that starts with ~ means the home directory of the machine that reads it. The key template_host names one machine, the template host, and the value of autocommit_branch then follows from the host name of each machine. So you write the profile once and copy it to each machine unedited. One profile for every machine gives the rules.

Synchronise across machines

Three mechanisms carry the private data to another machine. Each one is git, and each one is optional.

The dotfiles kit

The dotfiles kit, examples/dotfiles/, is a chezmoi source with placeholder data. You copy it into a private repository of your own with harness dotfiles-kit, and chezmoi applies that repository on each machine. The copy carries:

• the files of ~/.config/research-harness/: the profile and the model, provider, network and repository tables;
• the public key that signs, in ~/.config/git/allowed_signers;
• the table of your machines, .chezmoidata/machines.toml, with the commit of the harness that every machine runs;
• the setup scripts: the tool check; the clones of the remotes of machines.toml, such as ~/.claude, the harness and, in the example, the repository of the tree instructions; Julia; the harness at its commit, with harness install --apply; and the background jobs;
• the command dotfiles and its scripts.

One machine signs. Every other machine applies a change only when each new commit carries a good signature of a key that it already trusts, so a change that nobody signed never reaches a machine. The harness moves only when the commit of the harness in machines.toml changes, so each machine runs a reviewed commit, not the newest one. The dotfiles command gives the update path, and Start a private copy the steps.

The memory and the plans

~/.claude is a git repository with a private remote. The job claude-autocommit commits the plans and the memories of ~/.claude every 15 minutes, and pushes them to the branch of autocommit_branch. That branch is main on the template host, and memory/<short host name> on each other machine. So no two machines push to one branch, and the memory of the template host is the memory that the others start from. Jobs and configuration describes the job.

The tree instructions and the repositories

The tree instructions are a directory of a private repository, in the example ~/Research/Environment. The repositories of the research tree each have their own remote. harness clone clones them on a new machine, and harness repos sync pulls and pushes them. The research tree describes both.

The limits

• The credentials stay on each machine. A new machine needs its own SSH key and its own logins of gh and glab before its first clone. A new machine gives the steps.
• The leak check knows only the strings that you give it. A private string that is not in the leak list, and not a name of your tree, passes it.
• The kit's tool check installs no tool. It names a missing tool and stops the first apply; later, its harness step runs harness setup --yes, which installs the tools of [setup].
• The installed copies are not synchronised. Each machine writes its own with harness install --apply, from its own checkout of the harness. The kit runs the install when the commit of the harness in machines.toml changes.