← Zulu Assistant · Release · v1.3.0 and v1.3.1 · November 15, 2026
Zulu Assistant 1.3: approvals bound to what you saw, and a lighter first hour
Security: if you run Guardrails, upgrade to 1.3.2
In 1.2.4 and every earlier release, a held action could run without asking you in other spellings:
- A word in front of the held call. A launcher's flag that takes a value (
timeout -s KILL 5 …,caffeinate -t 5 …,sudo -D <dir>), zsh'snoglob, and plain launchers such asarchandstdbufwere read as the program, and the gated call behind them passed. They are looked through now. - In-place edits inside a guarded folder.
sed -iafter acdthere was not held, andplutil -replace,-insert,-remove,-convertand-extractwere held nowhere. Both are held now;plutil -pand-lintstay free. - The dry run was taken on trust. For a script run with a write flag, the gate asked for the same call without the flag first. A script that pays or sends without its flag would have run as "the dry run". Now the dry form is read by the same rules before it is offered; if it is held too, the gate comes to you as it stands, and its row says
dry run: NONEand why.
1.3.1, a patch the same day, closes one more: an approved merge written with &> ran without its head pin under dash, which reads & as "run in the background". Only a session whose shell is dash was affected; on macOS, /bin/sh runs bash unless its owner changed it. Such a merge now has its head read again before it is let through.
We found all of these ourselves. None is known to have been used.
Two more spellings, an interpreter named with its version (python3.13, node22) and a held program run through its interpreter by its file, were patched in 1.2.4 the day they were found. 1.3 reads more of them than that patch could.
An approval is bound to what you saw
A merge gate records the pull request's head commit and shows it on your row. The approved merge runs pinned to that head, so GitHub refuses it if the pull request moved. A push gate records the commit every pushed ref points at. And the content of the scripts a call runs is part of what you approve: if any of them changed after you said yes, nothing runs, the gate is closed as superseded, and the call is asked again, saying what moved.
When a call could move the commit itself before the push or merge (it switches branch, commits or runs a script first), the row says it is not bound and names what runs before it. A held call that changes a script and then runs it is refused, with the step to take: make the change as its own call first.
The gate reads more, in observe first
New checks read a session an agent starts for itself, programs that run other programs (npx vercel --prod, uv run gh pr merge 3), MCP tools nobody has classified, writes through gh and SQL, typing into another agent's pane, and the desktop. They ship in observe: they report what they would hold and change no answer until you turn them on. The Guardrails guide lists each form and what is still not read.
- A gate hook that cannot start stops the call. The installer now writes the hook entries with
"onFailure": "block". Before, a hook whose script was gone let every call run. Before you move or delete the checkout, runscripts/guardrails-install.sh --remove --apply. - An approval nobody used no longer stays approved. It is marked expired, and a grant runs only while the task's own record says the gate is approved and still in time.
- An agent running the installer is held, and naming the product in a note or a folder holds nothing.
- Settings that point git at another repository turn the guardrails card red, and the settings-change hook refuses them in a running session.
- Anchor the hook (optional).
scripts/guardrails-anchor.shprints one line for you to type with your password. It puts the hook into Claude Code's managed settings, which a session cannot leave out.
A smaller scheduled triage pass
The pass that runs four times a day now has two tools, no free curl, one small reading of the board, and a short list of exact restarts built from your registry. It writes only its own files and reads only your user settings. A new install has it ask Mission Control for a worker instead of starting one; an install made before 1.3 keeps starting workers until you change one setting (below).
A lighter first hour
./install.sh (and the one-line install) now does four things: a Node check, your config, the board as one LaunchAgent, and it opens the board. It installs no Homebrew package, writes nothing to your Claude settings and loads nothing that costs money. A next-steps card offers the rest, one command each. ./install.sh --full is the old behaviour. On a Mac where cmux has never run, the board shows your live Claude Code sessions, read-only.
Known limitations
Guardrails in a Codex session
- An approved call still stops at Codex's own prompt. Codex (checked on 0.162.1) accepts only a "deny" from the gate hook, so after you approve, Codex shows its own "Would you like to run the following command?" prompt. Answer "Yes, proceed", soon after you approved. Never answer "Yes, and don't ask again": Codex then never asks for that command again.
- The approval is checked when it is released, not when it runs. The call runs only when someone answers the Codex prompt, which may be much later; a change made in between is not seen, and the head pin of a merge does not apply in Codex.
Fixed in 1.3.2: after scripts/guardrails-install.sh --apply --codex and a restart of each Codex session, an approved call runs when you approve it, with no Codex prompt, a Codex merge must carry its head pin, and the row says whether the call ran (checked on Codex 0.162.1 and 0.170.0). A held MCP tool call in a Codex pane can still meet Codex's prompt. Claude Code sessions are not affected. 1.3.2 notes · Troubleshooting
Names enforced, with a scheduled job mapped
If you map a scheduled job in callerEvidence.jobs (the triage pass included), keep callerEvidence.mode at 'observe', the shipped setting, rather than 'on'. A mapped job that starts within 30 seconds of the board last finding it not running is not recognised, so now and then its messages go out marked UNVERIFIED and its ledger steps are refused. In 'observe' this only adds a [caller mismatch] line to the log. 1.3.1 does not change this either.
How to upgrade
From 1.2.x, use the update card's button or run zulu upgrade. It brings the code, runs the migrations and the tests, and restarts the board. Then do the steps it does not do, because each touches something that is yours:
- Your rules file. An upgrade never changes
overlay/policy.json. Copy the rules 1.3 adds fromexamples/policy.json; the guardrails card says how many are missing. - The hook entries. Run
scripts/guardrails-install.sh --applyonce more. It needs Claude Code 2.1.295 or later. - Triage. Add
triage: { dispatch: 'ask' }toSETTINGSinoverlay/config/jobs.mjs, and if you keep your ownoverlay/config/TRIAGE.md, copy the example's steps.
Upgrading to 1.3 has the full list. In a checkout reached through a symbolic link, use the button or zulu upgrade, or run cd "$(pwd -P)" && ./install.sh upgrade. From 1.0.x there is no button: follow the install guide, §7b.
Install
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: v1.3.0 and v1.3.1. Questions and install reports are welcome in Discussions.