gstack 1.102.1.0 closes a gap where a secret written after a slash could be printed unhidden, locks down gstack's usage reporting and makes it more useful for people who opt in, fixes a set of bugs in how gstack picks the right skill for a prompt, and makes a gstack skill's start step run as Claude Code loads the skill, so it can't be skipped.
Usage reporting (telemetry, opt-in)
- The usage-reporting backend now accepts data only through its own checked endpoint: the public key in the repo can no longer read or write the tables, and a signed-in Supabase user is denied everything. gstack's own test runs never send usage data, test traffic is left out of
gstack-community-dashboard, and its weekly active count now counts real installs instead of update checks. gstack versions older than v0.11.16.0 (March 2026) stop sending usage data, because they wrote straight to the tables. (#4537) - If you opt in, each skill run now also sends nine new fields, each a pick from a fixed list: which coding agent ran it, the model family, whether you, the model or another skill started it, why it ended, whether its own checks passed, whether gstack's routing section and hooks are set up, a random session id, and for upgrades the version you came from and why an upgrade failed. Runs from CI and gstack's own evals label themselves, and a duration gstack couldn't measure is sent empty instead of 0. The opt-in question lists the new fields, and anonymous mode now says it sends "no device ID; a random id per agent session that is never reused". No field carries code, file paths, prompts or repo names; every field is listed in docs/privacy.md. (#4542)
Skills and skill suggestions
- In Claude Code, each gstack skill's start step now runs while the skill loads, so the model can no longer skip it (in test runs, models had followed a loaded skill's start step in only 45% to 62% of sessions). If loading fails, the skill says so in one line and the model runs the start step itself, as before. Codex and other agents keep the old instructions. (#4509) gstack's paid test that checks this start step now counts one that ran during loading. (#4543)
- When a prompt reads like a bug report, the skill-suggestion hook now names your installed bug-fix skill alone and never suggests /verify ahead of it. On 60 real bug-fix prompts it now names the bug-fix skill on all 46 bug reports. (#4539)
- The skill-suggestion hook now ignores words inside backticks and code blocks, so a quoted command such as "Run
gh pr review 123" no longer triggers /review, and a--fixflag in a pasted log no longer makes a prompt look like a bug report. Thanks to @vilano93, who found and first fixed it. (#4566) - gstack now tells you in one line what it's about to add before it installs a background job (
gbrain autopilot) or the gstack browser extension, including how to remove it. A new test fails if a future install step leaves that line out. (#4538)
Safety and security
- The destructive-command check (/careful) closes more ways around its disk-wipe and force-push blocks:
wipefs -ais blocked on every target and in more spellings, a write into a disk is blocked when its path is quoted or goes through a link to/dev, andgit push -fn --no-dry-runcounts as a real force-push. Nothing that was blocked or asked about before is allowed now. (#4534) - /cso can now audit repositories that contain symbolic links. It lists each link and its target in the report and never follows it, where before a single link (even
CLAUDE.md -> AGENTS.md) stopped the audit. Thanks to @nathanshox for the report and first fix. (#4567) - gstack's secret hider now also hides a password, API key or token written as
name=valueafter a slash (such ascfg/api_key=…orhttps://host/cb/token=…), where before it caught them only at the start of a line. /cso's history report had printed such values; it and every other place gstack hides secrets (review material, usage data, /ship logs) now hide them. (#4571)
/ship and code review
- A green CI run still counts toward /ship and merging when GitHub refuses to store the run's files because the repository's storage is full: gstack can now read the same record from the job's log, under the same checks. (#4536)
- /ship now stops and tells you what to do when CI's earlier test run can't be reused, instead of starting the whole test suite on your machine. (#4565)
- When the owner rules a stopped review's findings acceptable as open notes, those rulings now let /ship finish (#4552), the review record says which findings are still waiting for a ruling (#4553), and rulings also work when the review's findings were saved in a separate file (#4568).
- A pull request that adds only an eval's plan and its plain data files (a preregistration) now gets the lighter one-reviewer review meant for write-ups, and rows of data no longer count as lines of code when gstack sizes a review. (#4544)
- Run ids in /ship's approval records, such as
507-7192673, are no longer mistaken for phone numbers and blocked, and a team whose name starts with "GStack" can record rulings. (#4549)
Browser and tests
- The browser helper no longer picks port 10080, which Chrome refuses to open, so about one start in 39,000 no longer fails to load its own pages; one flaky test was fixed too. (#4554)
- gstack's documentation-sync eval no longer fails when /ship repairs a reviewer's result line in the allowed way. (#4555)
- For people working on gstack: every test that failed and then passed on its retry in gstack's free test runs is now recorded, including runs started by hand on a branch, so flaky tests get counted and fixed. (#4572)
- For people working on gstack: test runs on pull requests start one step sooner, because the job that splits the tests now runs alongside the earlier checks instead of after them. (#4570)