Repository navigation
gstack 1.102.2.0 fixes two things that broke skills on Codex, stops /plan-tune sending your typed answers to the model without a record, lets large repositories use /ship, makes /office-hours start from ground truth, and moves skill routing out of your CLAUDE.md into the hook that now does that job.
Codex
- Skills no longer stop on Codex 0.170 when they clean up temporary files: gstack's Codex skills stopped using the
rm -fform that Codex refuses (67 refusals before, 0 after). (#4612) - /review on Codex with OpenAI models no longer fails a review that really ran its security pass. Codex encrypts the message sent to that reviewer, so gstack now recognises the pass from the reviewer it started instead of from the message text. (#4622)
Privacy and telemetry
- /plan-tune's answer summariser now writes the off-machine send record the README promises before it sends your typed answers to the model, and sends nothing if that record can't be written. A new test fails on any model call that writes no record. Thanks to @robbiesierra for the report. (#4621)
- On a Capy install, only the owner's opt-in sends usage data, reviewers that only read code send none, and error reports carry an error class. (#4597)
Skills
- /office-hours is rewritten: it asks for ground truth first, and every startup session ends with a dated test. (#4468)
- gstack no longer offers to add a
## Skill routingsection to your project's CLAUDE.md. The skill-suggestion hook now does that job: with it, Claude Code loaded the right bug-fix skill in 45 of 60 and 48 of 60 test sessions, against 0 of 60 with no gstack. Existing CLAUDE.md files are left alone. (#4629) - Plan reviews no longer read a /design-review audit summary as the branch's design doc. Thanks to @robbiesierra for the report. (#4623)
- /cso checks every folder above a source file before reading the file, so it never reads through a symlinked folder. (#4592)
/ship and review on large repositories
GSTACK_REVIEW_PLAN_TIMEOUT_MSnow raises the review-plan time limit everywhere /ship uses it, so repositories with hundreds of thousands of files can finish planning a review. The timeout message names the setting. (#4632, #4637)- When the base branch's review policy is invalid only in its exclude lists (for example more than 64 patterns), /ship now reviews every file with a loud warning instead of refusing, so the pull request that fixes the policy can itself be reviewed. A pull request still can't change its own review policy. (#4637)
- Node's built-in test files now run with
node --testin repositories that use vitest or jest. Thanks to @eoin-obrien for the report. (#4626) - /ship no longer stalls on review notes a reviewer already cleared, counts review rounds the way the review policy reads them, and keeps security findings out of the pull request body. (#4595)
- /ship's review no longer stops when Capy's drive sync database grows past 100 MB. (#4606)
- /ship's review readers are told never to stop processes they didn't start. (#4607)
- /ship no longer asks for a fresh review because of a stacked parent that main already had (#4580), keeps a reviewed head current across merges that only bring in the base (#4577), honours your repository's time limit for rerunning tests CI skipped and says when a fresh worktree needs its packages installed (#4608).
- Landing a ready pull request no longer depends on a hand-copied gate decision, and the ready note, release PR and squash message say the right thing. (#4596)
- Results pull requests with large captured agent sessions can be reviewed. (#4574)
Evals
/eval-planrefuses gpt-4o and any superseded model as an arm, judge, reader or simulator and names the replacement; a competitor tested with its vendor's default model is exempt when the plan says so. Plans frozen from 2026-11-22 need 100 human-labelled outputs, up from 40, before a model judge's numbers count. (#4630)- gstack's own eval helpers moved from Claude Haiku 4.5 to Haiku 5.5, which restarts every paid case's pass-rate history. One scripted test actor stays on Haiku 4.5, because Haiku 5.5 acts before the event that test waits for. (#4630, #4639)
/eval-planwrites plans that waste less money. (#4591)
Docs
- The README and docs no longer claim gstack makes your agent fix more bugs; they say what it does and doesn't do. (#4598)
- New page: working across agent systems. (#4583)
For people working on gstack
- Free Tests runs its shards on 16-core runners (median 154 s, down from 222 s over the first 5 runs) and the shards finish closer together. (#4600, #4578)
- A quarantined test that the runner skipped now fails the run instead of passing it. Thanks to @Huyvux12, whose #4551 found and first fixed it. (#4590)
- Flaky fixes: the spawnSync overflow check (#4613); the weekly test-time recording can land again (#4581) and the recorded test times are current (#4633); the job-log sleep is long enough for real runners (#4589).
- The slowest free test file is split into six (#4588), Windows setup tests finish sooner on install changes (#4575), and pull requests no longer wait on the flaky-test upload (#4579).
- Habits from nine teams that run evals, written into gstack's own working rules (#4585, #4602); gstack's own use on Capy is counted separately from real use (#4576); token check-in lines for agent threads on Capy are halved to match Capy's new counting (#4582).
- An experiment on whether /ship needs its helper agent got its plan updates: an uncounted pilot per arm, one planned look at half the cases, a stop at the final remote-gate refusal, and one more labelled attempt for its judge. (#4587, #4605, #4609, #4611)