The user edits files in a normal git repo on their machine.
How the system stays in sync without asking the user to think like Git
BRB turns a local coding repo into a continuously mirrored Cloudflare-hosted project. The user keeps working locally. BRB handles sync, commit creation, context capture, restore points, and visibility in the UI.
Each repo maps to a Cloudflare durable agent that stores code and history.
The dashboard shows sync state, activity, conversation, commits, and restore points.
/presentations/how-it-works.html
Designed as a simple internal presentation deck
The whole product in one flow
Files change in a normal repo like ~/github/sample.
The CLI watch mode or daemon observes filesystem changes.
After a short pause, BRB batches the delta and syncs it.
If Codex is working in that repo, the latest session transcript is also synced.
The remote repo agent writes files, creates a git commit, and records a checkpoint.
The project dashboard and detail page show the new state.
What runs where
- Local git repo
- BRB CLI
watchor background daemon- Codex local session logs under
~/.codex/sessions/ - Local registry under
~/.brb/registry.json
- TanStack Start Worker for the app and API
- One durable repo agent per project
- Repo storage, commit history, sync status, checkpoints
- Workers AI for high-signal commit messages
The user does not need to manually commit their local repo for BRB to stay useful. The remote repo agent becomes the continuous, agent-friendly project record.
We wait for a pause so one burst of work becomes one meaningful commit
Sync after edits stop for a configured period, for example 5 seconds.
Force a sync during long editing bursts so the system does not wait forever.
Prevents noisy back-to-back commits when tools save repeatedly.
Instead of a commit for every file save, the user gets one commit per meaningful pause in the agent’s work.
Commit messages use more than filenames
- Changed file paths and content previews
- Recent synced conversation window
- Latest user request
- Latest assistant response
- Recent commit titles to avoid repetition
- Name the outcome, not the sync action
- Lead with the primary change, not incidental churn
- Write natural titles across frontend, auth, API, DB, tooling, and bug-fix work
- Fallback cleanly if AI output is unavailable
Why syncing agent conversation matters
A project often starts in one coding agent, then moves to another when credits run out or the user changes tools.
BRB mirrors the coding conversation into the project so another agent can see what was asked, what was done, and what the recent commits mean.
Conversation only attaches when the Codex session cwd matches the repo
path being watched. That keeps project context scoped correctly.
Restore points make agent work reversible
The system captures the visible repo snapshot and links it to the commit.
History stays auditable. Nothing silently rewinds in place.
The user sees restore points in the project detail page and can restore in one click.
One dashboard for scanning, one detail page for understanding
- All projects
- Sync-state badges
- Summary counts for watching, syncing, and errors
- Filters for all, watching, syncing, errors, and idle
- Activity and sync state
- Last commit and last message
- Conversation transcript
- Restore points
- Recent sync activity timeline
Simple local workflow
npm install -g brbgit
brb init --repo sample --source ~/github/sample
brb sync --source ~/github/sample
brb watch --source ~/github/sample
or install the macOS daemon to watch every registered repo automatically.
How we know it holds up
CLI and web app both build cleanly.
Commit-message quality is tested across frontend, auth, API, DB, tooling, and bug-fix scenarios.
A real sample app is pushed, updated, deleted, and verified through the live API.