gstack’s new Claude Code hooks pick the right skill and check it finished

A new gstack hook points Claude Code at the right skill. It scores each prompt against every skill's triggers, with no model call, and names the one that fits. It is off by default.

Slide text

gstack’s new Claude Code hooks pick the right skill and check it finished

./setup --skill-routing

UserPromptSubmit hook: names the gstack skill your prompt fits

Off by default. Scored against each skill’s triggers, with no model call.

Skills Browser Reviews Security Shipping

garrytan/gstack · CHANGELOG.md · 1.94.6.0

A second hook runs when Claude Code stops. If a skill ends without the outputs its contract requires, like a test after the last edit, it blocks once. This one is off by default too.

/investigate loses its five-phase procedure, which Opus 5.5 never followed. It keeps its Iron Law, scope lock and debug report, and shrinks by about a third.

Opus 5.5 now gets a lite copy of every skill, minus ten coaching sections. Capy sessions pick it themselves, and Claude Code can opt in at setup. Every other model keeps the full skills.

The browser.

When several agents drive gstack's browser at once, each one now gets its own tab and stays on it. Each tab is separate, while cookies and storage stay shared.

/qa can now make a tab act like an iPhone 15 and send real taps, swipes and pinches. It is Chromium emulating a phone, so Safari still needs a real device.

Reviews.

/review and /ship used to triage only Greptile's comments. Now they triage every review bot you list, so CodeRabbit and Macroscope get the same treatment.

/review now names the kind of pull request first, so a CI or script change gets checks that fit it. It also flags a script that doesn't fit the repo, like a Windows batch file in a bash project.

On a pull request clone, /ship could compare your branch with itself, find zero changes and pass. Now it refuses. A stale local main also stops standing in for the real base.

Security.

On Capy, a feature plan gets one lean reviewer that reads your draft code and sends it to an outside model. That draft now leaves out every env and credentials file, and redacts keys in ordinary files.

On a public repo, /ship used to paste the owner's exact words into the pull request, where GitHub keeps every edit for good. Now it writes one line: approved, with a record id.

Shipping.

When CI already passed on the exact same clean tree, /ship now uses that run instead of running your tests again. The repo has to map its lanes first, and fork runs never count.

A flaky test can now go into quarantine for a week, with an owner. It still runs every time and never counts as a pass. With the owner's grant, named flakes outside your change can clear the pregate.

The new gstack-clean command trims old scratch files and browse logs, and changes nothing until you pass apply. QA in ship and review stops writing reports into your repo. And Codex reviews now print their token usage.

If you use gstack with Claude Code, I would turn on skill routing first. It costs no model call and sends every prompt to the right skill. Every change is in the gstack changelog.