Skip to content
Security

Built so an agent can't go rogue.

What Fixroom refuses to do is enforced in code, not left to a prompt. Here are the guardrails, the architecture behind them, and an honest list of what isn't built yet.

Guardrails

What it refuses to do, enforced in code.

A prompt is not a guarantee, and neither is a comment. These limits live in the code. Several are checked in tests that read the source.

  • Run a shell command for the model

    There is no bash tool. Approval can never make a raw shell string safe, so it was removed instead of gated.

  • Merge, squash, or auto-merge

    The part that opens pull requests has no merge operation, and a test reads the source to make sure it never gains one.

  • Force-push over your work

    If someone else pushed, Fixroom rebases or stops. A branch already under review is never rebased, because that would need a force push.

  • Move the goalposts after the fact

    The change contract is written before the patch and can't be changed afterwards. Every read checks its content hash. A patch that touches files the contract didn't name is reverted.

  • Judge how risky its own change is

    Change class and blast radius come from the file paths, not from the model. Migrations, auth, payments, crypto and CI config can only be proposed, never applied automatically.

  • Silence the failing test

    A change that edits only tests is refused, because it hides the symptom instead of fixing it.

  • Open a PR that cites something imaginary

    Every cited file, line, symbol and commit is checked against the tree before the push. One unresolved citation stops the pull request.

  • Keep your prompts by default

    Model inputs contain your code and production logs, so prompt capture is off. Turning it off again deletes every stored payload at once.

Architecture

Everything runs on a server you control.

Your team works in the browser. The work that has to keep going runs on your server, even when nobody has the page open, and the browser only watches it. Two things leave: the model calls you configure, and a branch and pull request on GitHub, where a person merges.

Your teamIn the browser, on aphone or a tabletReportsHelp desks, email, chat,logs and signals01WorkspaceYour teamTickets, fix sessions,review, roles and theaudit trail.02Agent roomThe agentReads your code andproposes changes.No shell, no merge.03RecordsOne databaseEvery run, ticket anddecision, backed upbefore each upgrade.04Test benchYour checksRun the change withoutFixroom's keys or settings.Your serverYour AI modelOnly the callsyou configureGitHubA branch and a pullrequest. A personmerges it.
People decideThe agent proposesYour server runs it

Fixroom · Sheet A-2 · Where everything stays · Business view, not to scale

Three ordering rules carry the guarantees

  1. 01

    Write the stage before running it

    A process that dies mid-stage picks up at that stage, not at the beginning.

  2. 02

    Write the final status before releasing the lock

    No other run can claim the repository while this one is still settling.

  3. 03

    Record the event before anyone hears of it

    A viewer who opens the run late replays exactly what someone watching from the start saw.

Roles are enforced on the server, and a route without a rule is refused, even for an owner. The roles are owner, admin, developer and reviewer. People join by single-use invite links that expire after seven days.
Project status

What's built, and what isn't yet.

Built
  • Users, roles, invites, per-project assignment, and authorisation checked on the server
  • GitHub-connected projects through the shared Fixroom App
  • Fix sessions with review, verification, commit, push and pull request
  • Tickets Room with help desks, a support mailbox, Slack, Discord and in-app feedback
  • Pull request review with a required-check gate, and Code Scanner
  • Investigate and Solve from Slack and Discord
  • Logs with SDKs for five languages, and AI observability
  • Signals, log watchers, autofix in worktrees, and the ops loop
  • Fixroom OS, and a platform admin view of the install itself
Not built yet
  • A second organisation per install. For now there's one
  • Sandboxed host access. Until then, give developer only to people you'd trust with a CI runner
  • A real deploy target. The ops loop judges releases but deploys through a no-op
  • Preview environments, and a CI adapter
  • Outlook and Microsoft 365 mailboxes, which need OAuth
  • A viewer role