T3 Code just got a rulebook that can lock what coding agents are allowed to do

A new test build lets one file set locked rules for every agent on a computer, and a separate fix waits for agents that keep opening extra threads.

Listen to this story
2:02

On Nov. 1, Ping, the team behind the open-source app T3 Code, put out a nightly build with a new feature called managed policy. A nightly is a daily test build, ahead of the stable version most people install. T3 Code runs coding agents such as Claude Code and Codex side by side, each in its own thread: one running conversation with one agent. The release notes and the merged change, pull request 27197, describe a policy as one file of rules for a whole computer: approvals, AI providers, updates and remote access.

The rules come in three strengths, according to the managed policy guide. A default fills in any setting the user has not picked; the user's saved choice beats it. A lock sets a value that neither users nor agents can change. A ceiling caps how much freedom a thread gets; a thread above it cannot take its next turn until someone lowers it. The app's server, the background program the windows connect to, reads the policy only when it starts, so reconnecting a phone or browser changes nothing. No policy means the app runs unmanaged, and a policy with mistakes stops it from starting.

On Nov. 1, Jon decided that every long-running agent in his fleet, and every new working session, would move to T3 Code. These controls are about to become the controls on his agents.

A second change is still waiting. Issue 28986, filed Nov. 2, reports three times an agent handed a job by another agent read the brief as a fresh request and opened yet another thread. A pull request, or PR, is a proposed code change. In the open fix, PR 28999, the author reports Claude Sonnet 5.5 opened an extra thread in 4 of 4 runs before the fix and 0 of 7 after. That is the author's own small test, and the PR is not merged.

AI Parabellum tested it on one Mac on stable 0.1.6. It reports that new threads start in Full access, where commands and edits run without asking, and that a five-minute pairing link gives a device a 30-day session. The remote access guide says ending a device's access means revoking its session, separately from the link.

How to use it

  1. Write the policy as JSON with "schemaVersion": 1 and an organization name. Run t3 policy check policy.json.
  2. Put the file where the guide says: /Library/Application Support/T3Code/policy.json on macOS, /etc/t3code/policy.json on Linux and C:\Program Files\T3Code\policy.json on Windows. Watch the Windows path: PR 27197's description still says C:\ProgramData\T3Code\policy.json, while the pinned guide and code use Program Files. Only administrators should be able to change the file.
  3. Restart the server: quit and reopen the desktop app, or run t3 service restart.
  4. Run t3 policy show. It says whether the running server loaded the installed policy. In Settings, locked controls read "Managed by"; try changing one to confirm the lock holds.
  5. When one agent hands a job to another, check the work in the receiving thread, not just the new-thread link.

The catch: the policy governs T3 Code only, so Claude Code started outside the app follows its own settings. And managed policy is only in nightlies so far. The newest nightly today is 0.1.7-nightly.20261102.4080. The newest stable, 0.1.6 from Oct. 31, has no policy guide.

What nobody knows yet

When a stable release will carry managed policy. Whether PR 28999 merges, and whether its fix holds beyond Claude: the author could not make the bug happen in two Codex tries. Nobody outside the project has published a test of whether the locks and ceilings actually hold.

What to watch

The releases page for a stable 0.1.7, and whether PR 28999 merges.

Sources