Twenty-three specialists turn Claude Code into a process
gstack packages 23 role-based specialists and eight power tools around coding agents. The useful idea is sequencing: /office-hours questions the product, planning commands force decisions into writing, /review examines the diff, /qa drives a browser, and /ship handles the release path. That gives a technical founder or staff engineer named checkpoints instead of hoping one prompt covers product judgment, implementation, testing, and deployment.
The current gstack README lists support for 10 agent hosts, including Claude Code, Codex, Cursor, OpenCode, and Factory Droid. Some hosts receive the full install while others get generated methodology files or pointers. Claude Code remains the deepest path: hooks, browser control, review records, test-evidence ledgers, project memory, and team auto-update all meet there. If you use only one or two commands, much of the installed system will sit idle.
The 50.9 MB checkout becomes a 964 MB working install
commit 65bfb0c contained 2,826 files, about 439,890 source lines, and occupied 50.9 MB when checked out. Our install pulled 242 packages and expanded the environment to 964 MB. That is a reasonable price for a browser engine and compiled helpers on a developer workstation, but it is far beyond a folder of Markdown prompts. Review disk use and generated binaries before rolling it into disposable workspaces.
The base requirements are Git and Bun 1.0 or newer, with Node.js added on Windows. Browser commands prefer Aside on macOS and otherwise build bundled Chromium. The /cso security officer has stricter requirements: a compatible Bun release plus a static-capable compiler on Linux, Xcode tools on macOS, or Visual Studio C++ tools on Windows. Setup can leave CSO unavailable while installing the rest, which is better than hiding a partial install.
What happened when we ran it
Our sandbox installed commit 65bfb0c in 43 seconds and built it in 13 seconds. The container had 3 CPUs, 8 GB of RAM, Node 22, no secrets, and no elevated privileges. Installation and compilation completed successfully. The repository had 19 CI workflow files and a tests directory, but no Dockerfile. Those facts describe repository mechanics, not the quality of code an agent produces with the skills.
The test command did not finish inside the 900-second cap. Its final lines still listed failures in trust-policy handling, office-hours child environments, security and community dashboard fallbacks, and a setup-gbrain remote-caller race. Several dashboard cases concerned stale or incomplete data being presented safely. The log tail does not identify one cause that explains the group, so blaming the container, network, or a single regression would go beyond the evidence. The defensible result is that our complete gate stayed red.
Release automation has an open fail-open report
Open issue #2995 describes /land-and-deploy using only GitHub checks marked as required. On a repository with no required checks, the report says the workflow may skip ordinary red or pending CI and proceed to a direct merge. A related pull request was active when we checked, but an open fix is not a shipped guarantee. Teams on plans or branch setups without required checks should keep this command away from merge authority until they verify the installed revision.
That concern matters because gstack is designed to do more than advise. Team mode can add project instructions and an update hook, while shipping commands act on pull requests and deployments. Start with planning, review, and read-only QA. Add write authority after a trial repository proves the hooks, CI interpretation, and rollback behavior. The privacy section is unusually direct about opt-in telemetry and egress receipts, but a receipt records an attempted send rather than blocking every network path.
CSO stops at a 64 MiB source snapshot
Open issue #2993 records a CSO audit stopping before it received a run ID because the source exceeded a fixed 64 MiB admission limit. The report also identifies separate 1 MiB limits for individual payloads and the serialized manifest. This affects a specific security workflow, not ordinary review or QA, but it rules out treating CSO as complete coverage for every large monorepo.
Windows deserves its own trial. Open issue #1675 reports six manual workarounds during /setup-gbrain on Windows 11, including a Bun shell build failure, a misleading version-check failure, and an undocumented jq dependency. The reporter eventually produced a working brain. That outcome is encouraging for a patient individual and unsuitable for a company-wide unattended installer.
October 1 activity shows speed and unfinished edges
GitHub showed 134,616 stars and 949 combined issues and pull requests, and the repository was pushed on October 1, 2026. Issues #2993, #2995, #3004, and #3007 were all created or active during the final days of September. That combination says more than an old release tag would: maintainers and users are working quickly, while important behavior is still being discovered around security coverage, restored context, planning hooks, and CI gates.
gstack makes the most sense when your bottleneck is supervising whole agent-driven changes rather than generating code. Its 23 specialists give review and release work a shared vocabulary, and the documentation explains many sharp edges plainly. Our 900-second timeout prevents a clean recommendation for immediate organization-wide rollout. Install it in one noncritical repository, keep deploy authority off, and require a green run of the checks that match your host before expanding its reach.