← Zulu Assistant · Release · v1.2.0 · November 10, 2026
Zulu Assistant 1.2: the trust release, and a security upgrade for 1.0.x
Update: 1.3.2 is out. 1.3 closes more ways an agent could run a held action without asking you: a held call behind a launcher's flag that takes a value (timeout -s KILL 5 …, caffeinate -t 5 …) and in-place edits inside a guarded folder (sed -i after a cd there, plutil -replace), present in 1.2.4 and earlier releases and found by our own review; and the dry run the gate asks for before a write is no longer taken on trust. 1.3.1 lets the installer find a Node in /usr/local/bin without nvm and no longer pins an approved merge written with &>, which ran unpinned under dash. 1.3.2 makes an approved call in a Codex session run when you approve it, with a Codex merge pinned to its head, and a slow GitHub read no longer uses up an approval. From 1.2.x, use the update card's button or run zulu upgrade, then follow the short list in Upgrading to 1.3. From 1.0.x, follow the steps below; they bring you to 1.3.2. What is new in 1.3 → · Release notes →
Earlier update, November 14: 1.2.4 is out. It closes two ways an agent could run a held action without asking you: by calling its interpreter or shell by a versioned name (python3.13, node22, bash5) or through a shell the gate did not know (tcsh), and by handing a held program that is a script to its interpreter by its file (node …/vercel deploy --prod, bash bin/zulu upgrade), present in 1.2.3 and earlier releases and found by our own review. From 1.2.x, use the update card's button or run zulu upgrade; there is nothing to reconfigure. From 1.0.x, follow the steps below; they bring you to 1.2.4. Release notes →
Earlier update, November 12: 1.2.3 is out. It closes a way an agent could push to main without asking you, by switching to it earlier in the same call (git switch main && git push origin), present in 1.2.2 and every earlier release and found by our own review. Release notes →
Earlier update, November 11: 1.2.2 is out. It closes three ways an agent's push to a public repository could run without asking you (a Codex call's folder, a folder change that may not apply, and the remote of a push told another way), present in 1.2.1 and every earlier release and found by our own review. Release notes →
Earlier update, November 11: 1.2.1 is out. It closes two ways a gated script could start without your approval (behind an interpreter flag, or handed to the interpreter on its input), present in 1.0.x and 1.2.0 and found by our own review. On 1.2.0, the update card's button takes you there once the board has seen the release (it looks every six hours), or run zulu upgrade now; there is nothing to reconfigure. From 1.0.x, follow the steps below. Release notes →
Security: if you run 1.0.0 or 1.0.1 with Guardrails, upgrade to 1.2.0
The gate hook in both releases had a flaw. A call that should have waited for your approval could run without it, if the call's own text contained the name of the hook's after-call event. The gate was still filed on the board, but the call went ahead.
We found this ourselves. We have no report of it being used against anyone.
To check your own history: after upgrading, node scripts/gate-replay.mjs --mentions PostToolUse --list lists past checks whose text named that event (time, agent, event and answer, never the call's text).
After upgrading from 1.0.x:
- Follow the install guide, §7b.
- Run
scripts/guardrails-install.sh --applyonce more. That adds the settings-change hook, which an older install does not have. - Check the guardrails card. An upgrade does not change your rules, so the card says how many shipped rules are not in yours. §7b has the step to add them.
If you export MESH_BOARD or MC_PORT in your shell and it does not name the board the hook belongs to, unset it. The hook now denies gated calls in that case and says why.
The trust release
Guardrails reads more of what an agent can do and protects itself, an approval is something only you can give, receipts are chained, and the gate hook no longer trusts the session it runs in. Also new: one-click upgrade, a Guardrails-only install, push notices to your phone, GitHub intake, the evidence panel and a daily ledger backup. There was no public 1.1.0. This release holds everything since 1.0.1.
An approval is something only you can give
When an agent records that you said yes in its pane, the record is now a proposal. Your row on the board quotes it and names who typed it, and you click CONFIRM, DECLINE or NOT WHAT I SAID. Until you confirm, the gate stays pending and nothing runs. Before, the record was the approval, and it rested on a name the agent declared for itself.
The board also looks at the process behind a decision before it takes one. An approve, decline or snooze that an agent session is behind is refused, with the decision token or without it, and Mission Control is told which agent.
Receipts are chained
Every change to a gate appends one line to a receipts file, and each line carries the hash of the line before. task.sh audit N --verify recomputes the chain and compares it with the task's gates. The guardrails card does the same over every task at start and every six hours, and goes red on a difference. An approved call runs only when the chain holds the approval the running board recorded, so an approval written into the files by hand is refused.
The gate checks and protects itself
- It doesn't believe the session's environment. The hook puts the system folders first on
PATH, takes the board's address from its own checkout, and answers "deny" itself before its time limit runs out. - A settings-change hook. A project settings file can switch off a hook that your user file registered. The new hook blocks such a change for the running session, and the board says so on the ticker.
- Gate integrity on the guardrails card (off by default). The board compares the gate's own files with what has landed on the remote, or with the release tag on a release install, and checks that the hook is registered as it should be. It only reads.
- Self-protection (ships in observe). Once you set it to enforce, writes to the gate's own files, to hook settings and to shell start files are held, and so is stopping the board. In observe, the log says what it would have held.
- It reads more. Code handed to an interpreter on the command line, a renamed program whose first words are a gated action, a gated command behind a shell keyword such as
then, and a file named in another letter case.
The Guardrails guide lists each form and what is still not read.
Guardrails on their own
./install.sh --guardrails installs the board, the ledger, the gate hook and the page with the waiting rows, and nothing else: no cmux, no Homebrew packages, no Mission Control, no scheduled triage job, no brain. It shows the hook it would add to your Claude settings and asks before writing it. The Guardrails quickstart has the steps.
One-click upgrade
1.2.0 is the first release you can upgrade from with one click. A release install gets an update card with an update button when a newer release is out. The upgrade takes a backup, moves to the release tag, runs the migrations and the tests, restarts the board and checks its health. If anything fails after the code moved, it rolls back by itself to the exact previous commit. Your ledger is never rolled back. UPGRADE.md has the details.
Coming from 1.0.x there is no button. Use the three steps at the top of this post, once.
Also new
- Push notices to your phone (prototype, off by default). When a gate, an approval or a plan review waits on you, your phone buzzes through a self-hosted ntfy server on your tailnet. A push carries ids and labels only, never the command. The tap opens the board with that item highlighted, and you still decide there. RELAY.md
- GitHub intake (off by default). New issues and discussions on your public repo can become ledger tasks, one per thread, addressed to Mission Control. It reads with your
ghlogin and never writes to GitHub. GITHUB-INTAKE.md - The evidence panel. A read-only page at
/?evidence=1: one row per task with what its close rests on (the request, the PR and its head commit, test stamps, each gate's outcome) and its gate receipts. Gate command text is never shown. - A daily backup of the ledger.
zulu backup --statewrites a small archive of the ledger, the board's state files and your config, and keeps its newest 14. It never removes an archive it did not create. You install the daily job yourself. - The attention ring. WAITING ON YOU can show one row per task. Gates, approvals and plan reviews come first, with their buttons showing. With snooze switched on (off by default), a row can wait 15 minutes, 1 hour, 4 hours or until tomorrow morning. A snooze never decides anything and never holds an agent.
- A weekly operating reading (off by default). Once a week the board reads the ledger and reports how the workforce ran against the week before: closes, reopens, how long tasks waited on you, gates by outcome, and the time from approval to action.
Install, or upgrade
curl -fsSL https://zuluassistant.com/install | bash
Just looking? ./install.sh --demo runs a synthetic board, gate included, with Node only. Nothing on your Mac is touched. Tested versions are in VERSIONS.md.
Full notes: the v1.2.0 release. Questions and install reports are welcome in Discussions.