llm-workbench

Turns AI coding conversations into isolated, reproducible Git workspaces.

llm-workbench is a standalone, provider-neutral Git/Bash/npm chat harness for blank or existing repos.

Run AI coding sessions in isolated git worktrees without dirtying main.

Why use it?

AI coding tools are useful, but they make it easy to lose context, mix unrelated changes, or pollute your main branch.

llm-workbench gives every chat session its own branch, worktree, and session log so you can experiment safely and merge only when ready.

Before LLM-Workbench

• one long AI conversation trying to solve multiple problems
• random commits
• forgot why code changed
• dirty main
• abandoned branches

After LLM-Workbench

Main
    │.                         │
    ▼.                         ▼
Start chat #1.            Start chat #n
    │.                         │
    ▼.                         ▼
Create worktree/branch.       ...
    │
    ▼
Code with AI
    │
    ▼
Checkpoint
    │
    ▼
Continue tomorrow
    │
    ▼
Merge back to main

• one branch per AI session
• session logs
• isolated worktrees
• merge when ready

Quick start

Public npm install is the intended user path:

cd /path/to/your/repo

npx llm-wb init --dry-run
npx llm-wb init

If your repo already has workbench files from an earlier manual copy or bootstrap, start with adoption instead:

npx llm-wb@latest adopt --dry-run
npx llm-wb@latest adopt --apply

Then start and inspect work from that repo:

npx llm-wb new "your prompt here"  // creates new chat session with prompt; happens automatically if you start a chat in code-assistant
npx llm-wb sessions list // lists open chat sessions
npx llm-wb commit -m "Describe the completed work" // ask chat to commit - runs gates and checks
npx llm-wb merge-main  // merges to local main, handles conflict

llm-wb list lists installed workbench commands. To list active chat sessions, use llm-wb sessions list.

This installs into your target repo:

• Adds the llm-wb CLI
• Configures AGENTS.md
• Installs workbench scripts
• Enables safe uninstall
• Writes .llm-workbench/lock.json and .llm-workbench/manifest.json
• Creates session logs that preserve decisions, prompts and commit history.

Update to the latest published workbench with a dry run first:

npx llm-wb@latest update --dry-run
npx llm-wb@latest update --apply

Roll back or test an older published version by pinning the package:

npx [email protected] update --dry-run
npx [email protected] update --apply

Updates only replace files recorded as workbench-managed in .llm-workbench/manifest.json. Files such as commitLogs/, product code, branches, worktrees, and repo-specific layers are left alone. If a managed file was changed locally, update stops with a conflict instead of overwriting it.

Each new llm-chat session then creates:

• a chat-owned branch
• a git worktree for that branch
• a session log/checkpoint record
• merge-readiness checks before promotion back to main

Works with existing repositories. No migration required. Uninstall removes only files owned by LLM Workbench.

Who is this for?

Developers using Claude Code, Codex, Cursor or Copilot coding assistants who want safer multi-session repo workflows.

Security Boundaries

• does not push to remote
• does not rewrite history
• does not delete branches without approval
• does not overwrite locally changed managed files
• does not require a specific LLM provider