From a0534ce6db39a8da1f3c23ec070aac2479e44be9 Mon Sep 17 00:00:00 2001 From: Garry Tan Date: Thu, 19 Nov 2026 10:42:28 -0800 Subject: [PATCH] v1.102.0.0 release (#4535) - `gstack-artifact validate` no longer passes a run.json that gstack-autoplan never wrote. (#4494) - /ship's changed-number check no longer stops on tests that only share a word with the change (#4478). (#4496) - On Capy, /ship no longer edits or comments on pull requests as you. (#4518) - The destructive-command check stops asking about scratch deletes and catches the real damage it used to miss. (#4513) - The destructive-command check reads commands more carefully, closing gaps found after the last release. (#4525) - The Codex readiness check and the second-model review now pick the same model for a custom Codex provider (GST-3089). (#4491) - gstack's Codex hooks work when your Codex folder's path has a `(`, `&` or other shell character in it. (#4528) - Uninstalling gstack keeps your own Codex hooks trusted. (#4528) - Codex hook setup matches Codex more closely. (#4528) - On Windows, setup copies the Claude Code hook code that gstack's Codex hooks import (#4528) - /review's checklist on Codex reads shell commands more carefully. (#4528) - On Codex, gstack now suggests the right skill for a prompt and stops destructive commands before they run. (#4512) - On Codex, /review is now held once when it tries to finish without its adversarial pass or its report, as it already is on Claude Code. (#4522) - On Codex, gstack's skill suggestion now names the skill the way Codex lists it. (#4523) - On Codex, gstack's suggested prompts and docs now name skills the way Codex lists them. (#4524) - /review on Codex with gpt-6.1-sol is no longer held back for a second-model pass it already ran. (#4526) - On Codex, office-hours, plan-design-review and gstack-upgrade work again when you name them. (#4510) - gstack knows when Codex is running unattended. (#4510) - Codex reviewers start with a clean context. (#4510) - Docs no longer say Codex has no hook support. (#4510) - A command whose output was cut off no longer counts as a finished read. (#4492) - A folder marked as published data can no longer hide source code or agent files from /ship's review. (#4516) - Correction to the on-by-default helper hooks entry: the hook still routes only to /review, /qa and /verify, so skills you've installed yourself aren't suggested yet. With gstack's skills it routes 14 of 60 real bug reports (was 17), and long prompts are no longer skipped. (#4505) - /sync-gbrain's read check now works when the first page in your code source is over 64 KB. (#4497) - Free-suite CI no longer fails before any test runs when Ubuntu's package mirror is slow. (#4530) - Syncing a second repo no longer makes the first one ask for /sync-gbrain again. (#4532) - gstack's helper hooks are now on by default in Claude Code, OpenClaw and Hermes, and the prompt-routing hook fires on real bug reports. (#4503) - `gstack-pregate clear --receipt` accepts a remote gate run from before a merged release when the release only regenerated a declared release output. (#4483) - The second-model review now keeps the reviewer's own result, and reports a used-up budget as a budget stop. (#4515) - When the reviewer wrote its review to the result file it was asked for, Codex then replaced that file with its short closing message ("Done, I wrote the review"), so the review was thrown away and reported as missing its markers. (#4515) - When the provider refuses the call with HTTP 402 Payment Required (usually a metering proxy that hit its allowance, or an account out of credits or quota), gstack used to record only "exit 1". (#4515) - The pre-commit /verify check now fails a check that times out, and its checks share one time budget (GST-3201). (#4527) - /ship's review collector no longer reads a reviewer's own scope list or list of passed checks as findings (GST-3201). (#4527) - Merge-queue stack tests fail when git cannot list a PR's added test files (GST-3201). (#4527) - /ship can now review results PRs that publish raw session data. (#4511) - /ship's review pass works when the checkout you ship isn't the project's default checkout. (#4507) - A second review round no longer reports done when nothing was reviewed. (#4520) - The prompt-routing hook now suggests skills you install yourself, and no longer sends ordinary bug reports to /qa (GST-3123). (#4506) - gstack's script runner and merge-queue checks now also refuse cut-off output. (#4502) - A failed second-model review reports the cause that actually stopped it, and the runner never deletes its own prompt. (#4519) - `gstack-ship-queue ready` now fills in the trial merges and the second-model trail read instead of printing lines to fill. (#4529) - Five small ship-tool fixes. (#4529) - A ruling the owner gives while a `/ship` run is open can be relayed. (#4529) - /ship no longer kills a working test run as stalled when the test command starts its workers in their own session. (#4533) - /ship no longer stops a review because the state-machine reader started its reply the way its instructions say. (#4504) - When Claude Code runs /verify right before a commit, it now gets a quick check instead of the full fix loop (GST-3150). (#4508) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> --- CHANGELOG.md | 88 +++++++++++++++++++ VERSION | 2 +- agents-digest/gstack-AGENTS.md | 2 +- ...tifact-validate-refuses-hand-built-runs.md | 1 - .../call-site-literals-need-the-module.md | 1 - .../capy-ship-writes-through-host-pr-tool.md | 4 - .../careful-fewer-false-asks-real-catches.md | 4 - changelog.d/careful-parser-gaps.md | 4 - ...x-custom-provider-one-plan-review-model.md | 1 - changelog.d/codex-hooks-hardening.md | 5 -- .../codex-hooks-setup-routing-guard.md | 4 - changelog.d/codex-review-contract-hook.md | 1 - ...dex-routing-hint-uses-codex-skill-names.md | 1 - .../codex-skill-mentions-use-listed-names.md | 1 - changelog.d/codex-sol-spawn-dispatch.md | 1 - ...ex-unattended-runs-and-duplicate-skills.md | 4 - .../cut-off-child-output-fails-closed.md | 1 - .../declared-content-folders-keep-code.md | 4 - changelog.d/docs-hooks-default-followup.md | 1 - .../fix-sync-gbrain-large-page-probe.md | 1 - changelog.d/flaky-fix-2026-11-19.md | 1 - .../gbrain-read-check-per-repo-sync.md | 1 - changelog.d/helper-hooks-on-by-default.md | 8 -- ...ear-accepts-regenerated-release-outputs.md | 1 - .../report-review-budget-stop-not-login.md | 4 - ...view-collector-and-verify-timeout-fixes.md | 3 - ...iew-content-patterns-for-published-data.md | 6 -- ...eview-readers-start-in-another-checkout.md | 1 - .../round-2-review-runs-after-a-gate-fix.md | 1 - ...ng-hook-stops-sending-bug-reports-to-qa.md | 1 - ...r-and-merge-queue-refuse-cut-off-output.md | 1 - ...cond-model-review-final-failure-decides.md | 1 - changelog.d/ship-tooling-small-fixes.md | 3 - .../stall-check-counts-own-session-workers.md | 1 - ...tate-matrix-check-accepts-reader-header.md | 4 - .../verify-quick-check-before-commit.md | 1 - package.json | 2 +- 37 files changed, 91 insertions(+), 80 deletions(-) delete mode 100644 changelog.d/artifact-validate-refuses-hand-built-runs.md delete mode 100644 changelog.d/call-site-literals-need-the-module.md delete mode 100644 changelog.d/capy-ship-writes-through-host-pr-tool.md delete mode 100644 changelog.d/careful-fewer-false-asks-real-catches.md delete mode 100644 changelog.d/careful-parser-gaps.md delete mode 100644 changelog.d/codex-custom-provider-one-plan-review-model.md delete mode 100644 changelog.d/codex-hooks-hardening.md delete mode 100644 changelog.d/codex-hooks-setup-routing-guard.md delete mode 100644 changelog.d/codex-review-contract-hook.md delete mode 100644 changelog.d/codex-routing-hint-uses-codex-skill-names.md delete mode 100644 changelog.d/codex-skill-mentions-use-listed-names.md delete mode 100644 changelog.d/codex-sol-spawn-dispatch.md delete mode 100644 changelog.d/codex-unattended-runs-and-duplicate-skills.md delete mode 100644 changelog.d/cut-off-child-output-fails-closed.md delete mode 100644 changelog.d/declared-content-folders-keep-code.md delete mode 100644 changelog.d/docs-hooks-default-followup.md delete mode 100644 changelog.d/fix-sync-gbrain-large-page-probe.md delete mode 100644 changelog.d/flaky-fix-2026-11-19.md delete mode 100644 changelog.d/gbrain-read-check-per-repo-sync.md delete mode 100644 changelog.d/helper-hooks-on-by-default.md delete mode 100644 changelog.d/pregate-clear-accepts-regenerated-release-outputs.md delete mode 100644 changelog.d/report-review-budget-stop-not-login.md delete mode 100644 changelog.d/review-collector-and-verify-timeout-fixes.md delete mode 100644 changelog.d/review-content-patterns-for-published-data.md delete mode 100644 changelog.d/review-readers-start-in-another-checkout.md delete mode 100644 changelog.d/round-2-review-runs-after-a-gate-fix.md delete mode 100644 changelog.d/routing-hook-stops-sending-bug-reports-to-qa.md delete mode 100644 changelog.d/script-runner-and-merge-queue-refuse-cut-off-output.md delete mode 100644 changelog.d/second-model-review-final-failure-decides.md delete mode 100644 changelog.d/ship-tooling-small-fixes.md delete mode 100644 changelog.d/stall-check-counts-own-session-workers.md delete mode 100644 changelog.d/state-matrix-check-accepts-reader-header.md delete mode 100644 changelog.d/verify-quick-check-before-commit.md diff --git a/CHANGELOG.md b/CHANGELOG.md index 4b9ba11e7..296af3da0 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,93 @@ # Changelog +## [1.102.0.0] - 2026-11-19 + +- **`gstack-artifact validate` no longer passes a run.json that gstack-autoplan never wrote.** A session that skipped `gstack-autoplan next` and built run.json with a script, copying the reference fixture, got `ARTIFACT_VALID`. The validator now also checks that the run came from gstack-autoplan: run.json must carry `host: gstack-autoplan`, `plan_input` and the other fields it writes on every run, and must agree with its `attempts.jsonl` journal, its bound review outputs and its checkpoint. A run that fails says `ARTIFACT_NOT_AUTOPLAN_RUN` or `ARTIFACT_JOURNAL_MISMATCH`, and `gstack-artifact ack` refuses it too. Public copies (`validate --public`) are unchanged. + +- **/ship's changed-number check no longer stops on tests that only share a word with the change (#4478).** When an ordinary property such as `limit: 120` changed to `30`, every test that used the word "limit" and held the number 120 was reported. A test now counts only when it also names the changed module, as `docs/pregate.md` already said; tests that import the module are still reported. Thanks to @dtoelke for the report and first fix. + +- **On Capy, /ship no longer edits or comments on pull requests as you.** On a Capy machine, `gh` is signed in as the machine owner's own GitHub account, so when /ship recorded its review rounds in a PR body (`gstack-ship-receipt stub`, `stop`, `claim` and `carry`) or posted a body, title, comment or new PR through `gstack-post`, the change showed up under the owner's name, even on someone else's PR. Under `GSTACK_CAPY=1` these commands now run every check they ran before, write the exact text they would have sent to a private file, and print a `HOST_PUBLISH:` line naming it and the repository (and host, when it isn't github.com) it was meant for, so the agent publishes it through Capy's own PR tool instead. Nothing changes outside Capy. + +- **The destructive-command check stops asking about scratch deletes and catches the real damage it used to miss.** Deleting anything under `/tmp`, a `mktemp` folder, a cache or build output no longer asks; in a replay of 254,322 agent commands those were 95% of its false asks. It now also asks before deletes outside the session's folder (including `rsync --delete` and `find -delete`), git commands that would throw away uncommitted work (only when the tree really has some), history rewrites, remote branch deletes, database resets and cloud deletes, and it reads every command in a chain on its own. A recursive delete of `/`, your home folder or a system folder, a force-push to `main` or the default branch, and wiping a whole disk are now blocked however the command is written. `DROP TABLE` and `TRUNCATE` ask only when a database client runs them, so the words in a grep or a test file no longer do. Agent SDK programs now get a warning instead of an ask, because the SDK refuses an ask outright when no permission callback answers it; an SDK app where a person does answer can set `GSTACK_SESSION_KIND=interactive` to keep the ask, and other harnesses with nobody watching can set `GSTACK_SESSION_KIND=unattended`. + +- **The destructive-command check reads commands more carefully, closing gaps found after the last release.** A force-push to `main` is blocked again when written with `git push --repo=origin +main`, after a `cd`, or with `--all` or `--mirror`. A `cd` inside a pipeline no longer makes a later delete look like scratch, and `git reset --hard "$REV"` asks again. It now also asks for quiet `git clean -q`, `--work-tree` and `GIT_DIR` forms, `find -L … -delete`, `rm dir -rf` with flags at the end, `..` paths, a `mktemp` variable pointed somewhere else, a discard after `git stash pop`, and a `terraform destroy` that only had a dry run earlier on the line. It stops blocking `git push -nf` (a dry run) and `wipefs` on image files it names on the command line (any `wipefs -a` fed by `xargs`, or aimed at `/dev`, stays blocked), and stops asking about a grep for SQL text or a folder named `psql`. `GSTACK_SESSION_KIND=interactive` now keeps the ask even when `CI=1` is set. Long commands are checked about three times faster on macOS's bash 3.2. + +- **The Codex readiness check and the second-model review now pick the same model for a custom Codex provider (GST-3089).** `gstack-codex-probe role-ready` refused a custom provider even when its `~/.codex/config.toml` named `gpt-6-astra`, while `gstack-outside-voice run --runner codex-cli` ran whatever that file named, so a session that trusted the check skipped a review that would have run. Both now use one selection. A custom provider's own `model` counts for plan review when it is exactly the selected tier's catalog ID, any other name is refused as `custom_provider_model_not_tier`, and no name stays `custom_provider_requires_model`. Each of these policy refusals prints a `CODEX_MODEL_POLICY: ` line. For the `autoplan` and `spec` gates the runner now also follows the plan-review tier with the stock OpenAI provider, instead of `config.toml`'s model. + +- **gstack's Codex hooks work when your Codex folder's path has a `(`, `&` or other shell character in it.** An unquoted `(` made the hook command a shell syntax error, which Codex reads as "block", so every prompt and Bash call was blocked once you trusted the hooks; an `&` split the command in two. Both Claude Code and Codex hook commands now quote those paths. A path made only of letters, digits and `_ @ % + = : , . / -` keeps the command it had; any other path is quoted once, so Codex asks you to trust those hooks again. +- **Uninstalling gstack keeps your own Codex hooks trusted.** When your hook groups come after gstack's in `hooks.json`, uninstall (and turning a hook off) now leaves an empty group where gstack's was, so your hooks keep the positions Codex trusts them by. If your handler sits inside gstack's group, gstack leaves the file alone and names the one handler to delete. +- **Codex hook setup matches Codex more closely.** `description: null` is accepted, MCP hook input with a null is refused, `rollback` won't restore a backup Codex would skip, a gstack hook already declared in `config.toml` under any matcher that covers Bash is caught (Rust's `(?i)`, `\A` and `\z` included), and `gstack-doctor` reads hook trust from the same key Codex uses when `CODEX_HOME` is a symlink. +- **On Windows, setup copies the Claude Code hook code that gstack's Codex hooks import** into Codex's gstack folder, so the hook scripts are complete. Native Windows Codex still starts hooks through `cmd.exe`, which can't run these Bash scripts, so the hooks don't run there yet. +- **/review's checklist on Codex reads shell commands more carefully.** A `;` inside quotes no longer splits a command, a read that never ran or failed isn't counted, a read error that names another file no longer drops a good read, PowerShell's "Cannot find path" counts as an error, and a message with only an image starts a new turn. + +- **On Codex, gstack now suggests the right skill for a prompt and stops destructive commands before they run.** `./setup --host codex` (Codex 0.170 or newer) adds two hooks to `~/.codex/hooks.json`: one adds a line naming the gstack skill a prompt fits (review, QA, verify), and one runs /careful's check before every shell command. Codex can't ask you about a single command, so in an interactive session a force-push, `git reset --hard`, `DROP TABLE` or the like is blocked, and the model is told to ask you first; if you want it, type `!` and the command to run it yourself. Under `codex exec` it's a warning instead, and deleting `/` or your home folder stays blocked everywhere. Codex runs new hooks only after you trust them, so pick "Trust all and continue" the next time Codex starts (setup prints a reminder, and `gstack-doctor` shows whether they're trusted). Your own hooks in that file are kept as they are, the same `gstack-config set skill_routing off` and `gstack-config set destructive_guard off` switches work on both hosts, and `gstack-uninstall` or `./setup --no-team` removes only gstack's entries. Setup also suggests `codex -c model_reasoning_effort=high` for gstack work and leaves your `config.toml` alone. + +- **On Codex, /review is now held once when it tries to finish without its adversarial pass or its report, as it already is on Claude Code.** `./setup --host codex` adds the skill-contract Stop hook to `~/.codex/hooks.json` next to the routing hint and the destructive-command check, and Codex runs it once you trust it at its startup prompt or in `/hooks`. The hook reads Codex's rollout file (it tells a Codex rollout from a Claude Code transcript by what the file holds), counts a skill as loaded when Codex injects or the model reads a `gstack-*` SKILL.md, and counts a `spawn_agent` call that started an agent as a dispatch. /review's specialist-reviewer item is skipped on Codex, which doesn't run those reviewers yet. `gstack-skill-contract check` takes a Codex rollout too. The same `skill_contracts` switch turns it off on both hosts. Details: docs/skill-routing-and-contracts.md#codex. + +- **On Codex, gstack's skill suggestion now names the skill the way Codex lists it.** The routing hint used to say "gstack-review", a name Codex doesn't know, so mentioning it as `$gstack-review` loaded nothing. It now reads each skill's name from its installed `SKILL.md`, so the hint says `review` and `$review` loads the skill it suggests. + +- **On Codex, gstack's suggested prompts and docs now name skills the way Codex lists them.** The picker prompts for nine skills (office-hours, review, ship, qa, autoplan, land-and-deploy, modernize, runaway-stopper and write-skill) said `$gstack-office-hours` and so on, which Codex doesn't recognize, so choosing one loaded nothing. They now say `$office-hours`, `$review` and so on, which load the skill. The README's Codex row and the Claude outside-voice note in docs/skills.md say the same. + +- **/review on Codex with gpt-6.1-sol is no longer held back for a second-model pass it already ran.** With Sol, Codex starts subagents through its `collaboration` tools, whose result names the new agent by `task_name` instead of `agent_id`, so the end-of-session check missed the adversarial pass and blocked the session once. It now counts a started subagent either way; a spawn that failed still doesn't count. + +- **On Codex, office-hours, plan-design-review and gstack-upgrade work again when you name them.** `./setup --host codex` left a second copy of each of these three skills inside gstack's own folder, so Codex listed them twice and loaded nothing when you asked for one (`$office-hours` started without its instructions). On macOS and Linux, setup now links those files instead of copying them, so Codex lists each skill once; Windows setup still copies them, because it can't rely on symlinks there. Re-run `./setup --host codex`, or upgrade gstack, to fix an existing install. +- **gstack knows when Codex is running unattended.** A skill started under `codex exec` (or `codex review`) now reports `SESSION_KIND: headless` instead of `interactive`, the same as an unattended Claude Code run, so it stops with a clear BLOCKED line instead of asking a question nobody will answer. The Codex TUI, and Codex inside Conductor, still count as interactive. +- **Codex reviewers start with a clean context.** The design reviewers, /autoplan's review phases and /context-restore's summary helper now start their Codex subagents with `fork_turns: "none"`, so they no longer inherit the parent's whole conversation (GST-3091). +- **Docs no longer say Codex has no hook support.** Codex CLI 0.170 runs hooks; gstack's routing, contract and destructive-command hooks just aren't ported to it yet. + +- **A command whose output was cut off no longer counts as a finished read.** Bun can report exit status 0 for a `spawnSync` child whose output passed the read limit (ENOBUFS), handing back only part of the output. The credential pre-push hook, the pregate secret and literal scans, the sealed-output guard, the ship receipt, review scoping and the other gstack commands that read large git or gh output now treat that case as a failure, named `SPAWN_OUTPUT_INCOMPLETE`, instead of judging part of a diff or file list. Before this, a push whose diff passed 64 MiB could carry a credential past the pre-push hook. + +- **A folder marked as published data can no longer hide source code or agent files from /ship's review.** In a folder listed under `review_content_patterns`, source code in any language and letter case (`.java`, `.c`, `.ps1`, `Makefile` and the like, binary or not) still gets reviewed, and so do agent instruction files, config, executables and any data file that kept code names by its full path. Before this, only nine source extensions were kept, a binary `AGENTS.md` was left out, and a file named by a kept script could leave review. No base policy sets the key yet, so nothing was exposed. Details: docs/ship-policy.md#review-content-patterns. + +- **Correction to the on-by-default helper hooks entry: the hook still routes only to /review, /qa and /verify, so skills you've installed yourself aren't suggested yet. With gstack's skills it routes 14 of 60 real bug reports (was 17), and long prompts are no longer skipped.** The OpenClaw and Hermes start hook stays off on upgrade whenever it was off, and the one-line "turn it on" hint prints only for an off saved before this release; an off set with `GSTACK_HARNESS_START_HOOK` or `harness_start_hook` stays off without a hint. docs/skill-routing-and-contracts.md (an unreadable config turns all three hooks off; the host table lists the destructive-command check) and docs/hosts.md (only an explicit start-hook request exits 4) now match the code. + +- **/sync-gbrain's read check now works when the first page in your code source is over 64 KB.** The check read the whole page just to confirm its source and slug, and a large first page (a lockfile is the usual one) hit its 64 KiB output limit, so every run reported the read as `unknown`. It now asks gbrain for a 4 KB window of the page (`--body window --max-bytes 4096`), which keeps the response small however big the page is. A gbrain from before body windows (0.60.231.0) gets the old whole-page read with the same limit, and output cut off at the limit still counts as `unknown`, never `ready`. Thanks to @nfitzsimmons for the report and the fix (#4396). + +- **Free-suite CI no longer fails before any test runs when Ubuntu's package mirror is slow.** Playwright's system packages and the gate tools are cached per runner image and Playwright version, so a run on a known image installs them with dpkg instead of downloading about 38 MB from archive.ubuntu.com, and apt waits for a timed-out attempt's lock instead of failing the retry in seconds. On 2026-11-19, 27 free-suite and quick-tier jobs failed this way. Two flaky free tests are fixed too: `first-use` now lists several user drives in sorted order on every filesystem, and the gstack-lease lock test waits for the owner line instead of reading a file that was created but not yet written. Two eval checkers stop refusing correct reads. The read-only command grammar accepts `\"` inside double quotes, as bash and zsh do, and the coverage audit follows a `cd` into a subfolder followed by `cd ..` in the next call. + +- **Syncing a second repo no longer makes the first one ask for /sync-gbrain again.** The GBrain read check looked only at the most recent code sync on the machine, so once you synced another repo, every earlier repo reported itself as unsynced at session start. It now falls back to the per-repo sync record `/sync-gbrain` already keeps. A repo that never synced still asks, and so does one whose record is missing, malformed or points at another folder, or whose own sync is the most recent on the machine and failed. Reported by @SatriaSuriaAriffin (#4499). + +- **gstack's helper hooks are now on by default in Claude Code, OpenClaw and Hermes, and the prompt-routing hook fires on real bug reports.** `./setup` now registers the prompt-routing hook (one line naming the gstack skill a prompt fits), the skill-contract hook (holds /review once if it stops before its specialist reviewers ran), and a new destructive-command hook that runs /careful's check before every Bash and PowerShell command, so `rm -rf`, `DROP TABLE`, force-push and the like ask before they run even when nobody loaded /careful (in `claude -p` and other sessions nobody is watching, it prints a warning instead of asking, so nothing waits on an answer). New OpenClaw and Hermes registrations get the start hook (a one-line note on each turn telling the agent to run a skill's start command). Why: on Opus 5.5, unattended, a hook line naming a skill got it loaded in 58 of 60 sessions, where a skill relying on its description loaded in 0 of 60. The routing hook used to skip any prompt over 8,000 characters and miss bug reports that never say "bug"; it now scores a long prompt's first 4,000 and last 1,000 characters, reads words like error, exception, traceback, crash and fails as a bug, and keeps paths such as `site-packages` whole. On 60 real bug reports it now fires on 40 instead of 22, and on none of 60 plain prompts that are not bug reports, reviews, QA or checks. + + **Upgrading:** the first `./setup` or `/gstack-upgrade` after this release turns these hooks on once and prints what it turned on. An off you chose earlier (a `./setup --no-...` flag, `gstack-config set off` or an environment variable) is kept. An OpenClaw or Hermes start hook that is off stays off on upgrade, and the setup output prints one line saying how to turn it on: `gstack-harness register openclaw --start-hook` (or `hermes`). + + **Turning them off:** `gstack-config set skill_routing off`, `gstack-config set skill_contracts off` or `gstack-config set destructive_guard off` (each takes effect on the next prompt or command), or `./setup --no-skill-routing`, `--no-skill-contracts`, `--no-destructive-guard`. For OpenClaw or Hermes: `gstack-harness register openclaw --no-start-hook` (or `hermes`), or `gstack-config set harness_start_hook off` and `./setup`. Details: docs/skill-routing-and-contracts.md and docs/hosts.md#start-hook. + +- **`gstack-pregate clear --receipt` accepts a remote gate run from before a merged release when the release only regenerated a declared release output.** A PR that merged a release commit was refused because a generated file the policy lists in `release_outputs` was checked against the `stamp_files` list first, and `clear` never reran the release tool the way the tree receipt does. A `release_outputs` path is now judged by the release-output rule, and with `--allow-repo-commands` `clear` reruns `release_tool` on the head and accepts an output it regenerates byte for byte. Without the flag the receipt is refused as before, and the refusal now says to pass it. A hand edit to the generated file, or any path outside `release_outputs`, is still refused. + +**The second-model review now keeps the reviewer's own result, and reports a used-up budget as a budget stop.** Two fixes to the Codex review that /autoplan, /review, /ship and the other skills run: + +- When the reviewer wrote its review to the result file it was asked for, Codex then replaced that file with its short closing message ("Done, I wrote the review"), so the review was thrown away and reported as missing its markers. Codex's closing message now goes to a separate `.last-message` file, and gstack keeps whichever of the two carries the review. +- When the provider refuses the call with HTTP 402 Payment Required (usually a metering proxy that hit its allowance, or an account out of credits or quota), gstack used to record only "exit 1". It now reads the cause from Codex's JSON error events and reports "the provider refused the review with HTTP 402 (payment required)", and says to check the proxy's allowance, or the account's quota and credits. It has its own reason code, `budget_exhausted`, in the run record and in docs/troubleshooting.md, and the skills' "authentication failed, run `codex login`" rule no longer fires on it. A 401 or 403 is still a login problem, and its record now keeps Codex's status line instead of "exit 1". + +- **The pre-commit /verify check now fails a check that times out, and its checks share one time budget (GST-3201).** A check that ran out of time used to be listed as "not run", so the commit still got `VERIFY PASS`. Now it is `VERIFY FAIL: TIMEOUT`. `gstack-checks quick --timeout` (540 seconds by default) is now one budget for the whole pass instead of a fresh allowance per check: each check gets the time that is left, and one that can't start in time is a TIMEOUT. The test stage starts with the time that is left, but its runner can still run past it, so a very slow test run can still outlast Claude Code's 10-minute command limit. +- **/ship's review collector no longer reads a reviewer's own scope list or list of passed checks as findings (GST-3201).** A bullet such as "a HIGH finding still refused" in the list of what the tests cover, or "`missing_verdict` can't occur alongside P0 or P1 findings" in the list of checks that held, collected as a HIGH or P0 blocker even though the reviewer declared `REVIEW_VERDICT: blocking=0`. When the reviewer declares no blockers, a plain bullet with no file:line, no Severity or Blocks field and no severity other than one naming findings as a kind keeps its label but is marked for the lead to lower with a written reason, so a bullet that really reports an open blocker still holds the merge until someone judges it. +- **Merge-queue stack tests fail when git cannot list a PR's added test files (GST-3201).** That failure used to read as "no new test files" and pass the slot. + +- **/ship can now review results PRs that publish raw session data.** A repo's ship policy can list folders under a new key, `review_content_patterns`, next to `review_exclude_paths`. A file both keys match counts as data whatever its name, so /ship's reviewers leave it out, the same way they already leave out JSON, CSV, logs and images. Source files in that folder (`.ts`, `.py`, `.sh` and the like, in any letter case) still get reviewed. Before this, files such as `.patch` diffs, a `MANIFEST.sha256` and licence texts with no extension always counted as code, so a folder of them could never be left out, and a PR like gstack-evals #352 (13.8 MB, of which 13.6 MB is published session data) could not get under /ship's 2 MB review limit. Nothing is guessed: only a pattern on the base branch counts, a content pattern on its own excludes nothing, and the secrets scan and the sealed-output guard still read every excluded file. Details: docs/ship-policy.md#review-content-patterns. + + **Also fixed:** when a `review_exclude_paths` pattern kept too many files for the reviewers' read command (more than 24,000 bytes of file names), /ship quietly dropped the whole pattern and sent the folder to the reviewers anyway. It now stops before review with `REVIEW_EXCLUDE_KEEP_TOO_LARGE`, naming the pattern, how many files it keeps and why, and the fix (docs/troubleshooting.md#review-exclude-keep-too-large). + +- **/ship's review pass works when the checkout you ship isn't the project's default checkout.** On Capy, review readers start in the thread's working directory, and /ship used to stop with `STOP: readers on the shared machine start in , not in ` whenever that was a different checkout, for example a project whose default checkout is another repository. Every reader's brief already opens by telling it to `cd` into the shipped checkout and report that checkout's tree on its first line, and a reply naming any other tree or path is still refused, so the pass now dispatches. It still stops before anything starts if readers can't reach the shipped checkout at all, or if they start in another worktree or clone of the same history, where a reader that skipped the `cd` would quietly review the wrong diff. (GST-3145) + +- **A second review round no longer reports done when nothing was reviewed.** After the first review pass, a fix that touched a gate-class path (CI, auth, merge checks) was planned as "the step applied no edits" unless another fix flag came with it, so /ship's round 2 printed done and no reviewer read the fix. That fix now gets its full review pass on the changed range. When the plan is told nothing was fixed but the branch has new commits since the reviewed head, it now refuses with `review_required_no_readers` instead of printing done. `gstack-review-dispatch` also stops reading its own flags out of the review-plan flags after `--`, so a plan's `--pass 1` can no longer stand in for the pass id. + +- **The prompt-routing hook now suggests skills you install yourself, and no longer sends ordinary bug reports to /qa (GST-3123).** Any installed skill that lists `triggers:` in its own SKILL.md can now be named by the hook next to gstack's /review, /qa and /verify, so a bug-fix skill with a "fix this bug" trigger is suggested for bug reports. One line names at most two different skills, gstack's own skill wins a tie, and a skill whose name or trigger text holds anything but letters, digits, spaces and simple punctuation is skipped, so its text can't break out of the quoted trigger or add markup to the line. gstack's own generated skill copies count as gstack's even without an install marker, so they still never route. Separately, /qa no longer takes bug reports on everyday words ("cannot find" plus "fails", or "call site" and "tests" far apart), while "please test the site" and "find bugs on my site" still route to it. `gstack-skill-route --explain` and the routing log show where each suggested skill came from. + +- **gstack's script runner and merge-queue checks now also refuse cut-off output.** The helper that runs gstack's own scripts and `bash -c` commands, and the merge queue's branch-ownership, stacked-test, release-version and merge-group record checks, now fail with `SPAWN_OUTPUT_INCOMPLETE` when Bun cuts a command's output off at its read limit, as the pre-push hook and the pre-merge scans already do. The earlier entry about cut-off output said every other gstack command that reads large git or gh output was covered; it was not, because these went straight to Node's `spawnSync`. A new test now stops any gstack command from calling Node's `spawnSync` directly. The 68 files that still do so mostly run short commands or already check for this error. The rest (the co-author trailer helper's commit-log read, the CI health report and the release-stamp precondition) never decide a security check, a ship record or a merge check. The test lists them, and the list can only shrink. The browser's process-list read is still not covered. + +- **A failed second-model review reports the cause that actually stopped it, and the runner never deletes its own prompt.** When Codex logged a 402 on an early attempt and then failed for good with a 401, gstack called it a used-up budget; the final failure now decides. A request id or a duration that happens to contain "402" no longer reads as a budget stop. Claude Code exiting 0 with `subtype: error_during_execution` and a 402 is now a budget stop instead of a generic provider error, and a Claude Code usage limit is reported as `budget_exhausted` instead of with Codex's usage-limit advice. `gstack-outside-voice run` refuses a `--prompt` that is one of the files the run writes for `--out` (such as `.last-message`) instead of deleting it. When the reviewer's own result file has no final verdict, or its first line names another plan snapshot, Codex's complete closing message is used instead. + +- **`gstack-ship-queue ready` now fills in the trial merges and the second-model trail read instead of printing lines to fill.** It lists the open PRs that change a file this PR changes and prints each one's `git merge-tree` result, and with `--trail ` it names the reader from the trail's newest `trail-read` row, under a train policy too; a gate, security, money, auth or data PR with no such row prints `MISSING: cross-family trail read`. +- **Five small ship-tool fixes.** `Plan: exempt:` accepts bug-fix wording such as "bug fix with failing-first tests", and ready uses the Plan line `/ship` finish resolved when the PR body has none. `gstack-approvals round-grant` refuses only gstack's own names, so "gstack-evals Captain" can grant a round. The Capy PR hand-off follows `GH_REPO` the way gh's own write does and refuses a PR URL that names another repository, and a pull request opened from a fork names its head as `owner:branch`. A targeted review read on a path with non-ASCII characters now carries across a later docs-only change. +- **A ruling the owner gives while a `/ship` run is open can be relayed.** `gstack-approvals rule --owner-ref :` records the owner's words with the message they came from and marks the ruling relayed in the record, the receipt and the PR body; without it the refusal now says `APPROVAL_RULING_RUN_OPEN`. `rule --not-security` lets the owner say a red-team finding is not a security finding, so its PR text shows the finding instead of "security finding, details recorded privately"; it is refused while a `/ship` run is open. + +- **/ship no longer kills a working test run as stalled when the test command starts its workers in their own session.** The stall check read CPU time from the test command's process group only, so a quiet suite run through a launcher that starts its workers in a new session (pnpm 11's `pnpm exec` running Vitest) looked idle and was killed as `PREGATE_STALLED`. It now counts every process descended from the command, and a stall or timeout kill reaches those workers too. A truly hung run is still killed at the same stall timeout. Reported by @dtoelke in #4477. + +- **/ship no longer stops a review because the state-machine reader started its reply the way its instructions say.** Every /ship reader is told to open its reply with a `REVIEWED_TREE: ` line and, for readers that can block, a `REVIEW_VERDICT: blocking=` line under it. The state-machine check read those two lines as unexplained text and reported the review as incomplete (`STATE_MATRIX_CHECK: incomplete gaps=2`), which stopped /ship before the adversarial review step on any change that touches a proposal, hold or approval flow. The check now skips exactly those two lines at the top of the reply, using the same parsers /ship's collector uses, so the two can't drift apart. Any other line that isn't a matrix record or a finding, including a misplaced or malformed header line, is still reported as a gap. + +- **When Claude Code runs /verify right before a commit, it now gets a quick check instead of the full fix loop (GST-3150).** Claude Code 2.1.286 and later tell the model to run any skill named `verify` before every commit that isn't docs-only or tests-only, so gstack's /verify was starting its whole build-and-fix workflow, first-run notices included, on ordinary commits. Now that call runs one command, `gstack-checks quick`: the repo's existing checks (typecheck, lint) and the tests /ship would pick for the change, with no code edits, no new tests and no questions, ending in one line: `VERIFY PASS`, `VERIFY FAIL`, or `VERIFY skipped: `. A repo with no `origin` remote, check commands you haven't approved yet (type /verify once in that repo to review and approve them), or nothing to run is skipped and the commit goes ahead; a failure makes the model stop and ask you before committing. Typing /verify yourself still runs the full skill. To turn the pre-commit check off, run `gstack-config set verify_before_commit off`; to stop Claude Code from asking for it at all, install with `./setup --prefix`, which names the skill /gstack-verify. + ## [1.101.0.0] - 2026-11-18 - **A test, or a scratch repository in the temp directory, can no longer write approval rows into your real Capy Drive (GST-3027).** On Capy, `gstack-approvals` mirrors every row to a Drive folder. When it runs under `bun test`, or its state root or repository sits in the temp directory, the mirror now has to resolve, links followed, inside that temp directory; otherwise the write fails with `APPROVAL_RECORD_WRITE_FAILED`, names the path it refused, keeps the local row and reads no rows from that Drive. A review probe on 2026-11-09 had written three QA rows into a real project-user Drive this way. diff --git a/VERSION b/VERSION index 78df83493..894b94ce1 100644 --- a/VERSION +++ b/VERSION @@ -1 +1 @@ -1.101.0.0 +1.102.0.0 diff --git a/agents-digest/gstack-AGENTS.md b/agents-digest/gstack-AGENTS.md index 403067e4a..1855124cc 100644 --- a/agents-digest/gstack-AGENTS.md +++ b/agents-digest/gstack-AGENTS.md @@ -1,4 +1,4 @@ -# gstack digest v1.101.0.0 — regenerate/re-copy after upgrading gstack +# gstack digest v1.102.0.0 — regenerate/re-copy after upgrading gstack Behavioral rules from gstack (https://github.com/garrytan/gstack), compressed for agent hosts without a full skill install. The full skills add workflows, diff --git a/changelog.d/artifact-validate-refuses-hand-built-runs.md b/changelog.d/artifact-validate-refuses-hand-built-runs.md deleted file mode 100644 index b6fa64e12..000000000 --- a/changelog.d/artifact-validate-refuses-hand-built-runs.md +++ /dev/null @@ -1 +0,0 @@ -- **`gstack-artifact validate` no longer passes a run.json that gstack-autoplan never wrote.** A session that skipped `gstack-autoplan next` and built run.json with a script, copying the reference fixture, got `ARTIFACT_VALID`. The validator now also checks that the run came from gstack-autoplan: run.json must carry `host: gstack-autoplan`, `plan_input` and the other fields it writes on every run, and must agree with its `attempts.jsonl` journal, its bound review outputs and its checkpoint. A run that fails says `ARTIFACT_NOT_AUTOPLAN_RUN` or `ARTIFACT_JOURNAL_MISMATCH`, and `gstack-artifact ack` refuses it too. Public copies (`validate --public`) are unchanged. diff --git a/changelog.d/call-site-literals-need-the-module.md b/changelog.d/call-site-literals-need-the-module.md deleted file mode 100644 index 35dc90cc8..000000000 --- a/changelog.d/call-site-literals-need-the-module.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship's changed-number check no longer stops on tests that only share a word with the change (#4478).** When an ordinary property such as `limit: 120` changed to `30`, every test that used the word "limit" and held the number 120 was reported. A test now counts only when it also names the changed module, as `docs/pregate.md` already said; tests that import the module are still reported. Thanks to @dtoelke for the report and first fix. diff --git a/changelog.d/capy-ship-writes-through-host-pr-tool.md b/changelog.d/capy-ship-writes-through-host-pr-tool.md deleted file mode 100644 index 33e8d4528..000000000 --- a/changelog.d/capy-ship-writes-through-host-pr-tool.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: security ---- -- **On Capy, /ship no longer edits or comments on pull requests as you.** On a Capy machine, `gh` is signed in as the machine owner's own GitHub account, so when /ship recorded its review rounds in a PR body (`gstack-ship-receipt stub`, `stop`, `claim` and `carry`) or posted a body, title, comment or new PR through `gstack-post`, the change showed up under the owner's name, even on someone else's PR. Under `GSTACK_CAPY=1` these commands now run every check they ran before, write the exact text they would have sent to a private file, and print a `HOST_PUBLISH:` line naming it and the repository (and host, when it isn't github.com) it was meant for, so the agent publishes it through Capy's own PR tool instead. Nothing changes outside Capy. diff --git a/changelog.d/careful-fewer-false-asks-real-catches.md b/changelog.d/careful-fewer-false-asks-real-catches.md deleted file mode 100644 index 48e9cee4f..000000000 --- a/changelog.d/careful-fewer-false-asks-real-catches.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -bump: minor ---- -- **The destructive-command check stops asking about scratch deletes and catches the real damage it used to miss.** Deleting anything under `/tmp`, a `mktemp` folder, a cache or build output no longer asks; in a replay of 254,322 agent commands those were 95% of its false asks. It now also asks before deletes outside the session's folder (including `rsync --delete` and `find -delete`), git commands that would throw away uncommitted work (only when the tree really has some), history rewrites, remote branch deletes, database resets and cloud deletes, and it reads every command in a chain on its own. A recursive delete of `/`, your home folder or a system folder, a force-push to `main` or the default branch, and wiping a whole disk are now blocked however the command is written. `DROP TABLE` and `TRUNCATE` ask only when a database client runs them, so the words in a grep or a test file no longer do. Agent SDK programs now get a warning instead of an ask, because the SDK refuses an ask outright when no permission callback answers it; an SDK app where a person does answer can set `GSTACK_SESSION_KIND=interactive` to keep the ask, and other harnesses with nobody watching can set `GSTACK_SESSION_KIND=unattended`. diff --git a/changelog.d/careful-parser-gaps.md b/changelog.d/careful-parser-gaps.md deleted file mode 100644 index f71efb010..000000000 --- a/changelog.d/careful-parser-gaps.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: regression ---- -- **The destructive-command check reads commands more carefully, closing gaps found after the last release.** A force-push to `main` is blocked again when written with `git push --repo=origin +main`, after a `cd`, or with `--all` or `--mirror`. A `cd` inside a pipeline no longer makes a later delete look like scratch, and `git reset --hard "$REV"` asks again. It now also asks for quiet `git clean -q`, `--work-tree` and `GIT_DIR` forms, `find -L … -delete`, `rm dir -rf` with flags at the end, `..` paths, a `mktemp` variable pointed somewhere else, a discard after `git stash pop`, and a `terraform destroy` that only had a dry run earlier on the line. It stops blocking `git push -nf` (a dry run) and `wipefs` on image files it names on the command line (any `wipefs -a` fed by `xargs`, or aimed at `/dev`, stays blocked), and stops asking about a grep for SQL text or a folder named `psql`. `GSTACK_SESSION_KIND=interactive` now keeps the ask even when `CI=1` is set. Long commands are checked about three times faster on macOS's bash 3.2. diff --git a/changelog.d/codex-custom-provider-one-plan-review-model.md b/changelog.d/codex-custom-provider-one-plan-review-model.md deleted file mode 100644 index b87db7f6e..000000000 --- a/changelog.d/codex-custom-provider-one-plan-review-model.md +++ /dev/null @@ -1 +0,0 @@ -- **The Codex readiness check and the second-model review now pick the same model for a custom Codex provider (GST-3089).** `gstack-codex-probe role-ready` refused a custom provider even when its `~/.codex/config.toml` named `gpt-6-astra`, while `gstack-outside-voice run --runner codex-cli` ran whatever that file named, so a session that trusted the check skipped a review that would have run. Both now use one selection. A custom provider's own `model` counts for plan review when it is exactly the selected tier's catalog ID, any other name is refused as `custom_provider_model_not_tier`, and no name stays `custom_provider_requires_model`. Each of these policy refusals prints a `CODEX_MODEL_POLICY: ` line. For the `autoplan` and `spec` gates the runner now also follows the plan-review tier with the stock OpenAI provider, instead of `config.toml`'s model. diff --git a/changelog.d/codex-hooks-hardening.md b/changelog.d/codex-hooks-hardening.md deleted file mode 100644 index a8f241a80..000000000 --- a/changelog.d/codex-hooks-hardening.md +++ /dev/null @@ -1,5 +0,0 @@ -- **gstack's Codex hooks work when your Codex folder's path has a `(`, `&` or other shell character in it.** An unquoted `(` made the hook command a shell syntax error, which Codex reads as "block", so every prompt and Bash call was blocked once you trusted the hooks; an `&` split the command in two. Both Claude Code and Codex hook commands now quote those paths. A path made only of letters, digits and `_ @ % + = : , . / -` keeps the command it had; any other path is quoted once, so Codex asks you to trust those hooks again. -- **Uninstalling gstack keeps your own Codex hooks trusted.** When your hook groups come after gstack's in `hooks.json`, uninstall (and turning a hook off) now leaves an empty group where gstack's was, so your hooks keep the positions Codex trusts them by. If your handler sits inside gstack's group, gstack leaves the file alone and names the one handler to delete. -- **Codex hook setup matches Codex more closely.** `description: null` is accepted, MCP hook input with a null is refused, `rollback` won't restore a backup Codex would skip, a gstack hook already declared in `config.toml` under any matcher that covers Bash is caught (Rust's `(?i)`, `\A` and `\z` included), and `gstack-doctor` reads hook trust from the same key Codex uses when `CODEX_HOME` is a symlink. -- **On Windows, setup copies the Claude Code hook code that gstack's Codex hooks import** into Codex's gstack folder, so the hook scripts are complete. Native Windows Codex still starts hooks through `cmd.exe`, which can't run these Bash scripts, so the hooks don't run there yet. -- **/review's checklist on Codex reads shell commands more carefully.** A `;` inside quotes no longer splits a command, a read that never ran or failed isn't counted, a read error that names another file no longer drops a good read, PowerShell's "Cannot find path" counts as an error, and a message with only an image starts a new turn. diff --git a/changelog.d/codex-hooks-setup-routing-guard.md b/changelog.d/codex-hooks-setup-routing-guard.md deleted file mode 100644 index 64110b69a..000000000 --- a/changelog.d/codex-hooks-setup-routing-guard.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -bump: minor ---- -- **On Codex, gstack now suggests the right skill for a prompt and stops destructive commands before they run.** `./setup --host codex` (Codex 0.170 or newer) adds two hooks to `~/.codex/hooks.json`: one adds a line naming the gstack skill a prompt fits (review, QA, verify), and one runs /careful's check before every shell command. Codex can't ask you about a single command, so in an interactive session a force-push, `git reset --hard`, `DROP TABLE` or the like is blocked, and the model is told to ask you first; if you want it, type `!` and the command to run it yourself. Under `codex exec` it's a warning instead, and deleting `/` or your home folder stays blocked everywhere. Codex runs new hooks only after you trust them, so pick "Trust all and continue" the next time Codex starts (setup prints a reminder, and `gstack-doctor` shows whether they're trusted). Your own hooks in that file are kept as they are, the same `gstack-config set skill_routing off` and `gstack-config set destructive_guard off` switches work on both hosts, and `gstack-uninstall` or `./setup --no-team` removes only gstack's entries. Setup also suggests `codex -c model_reasoning_effort=high` for gstack work and leaves your `config.toml` alone. diff --git a/changelog.d/codex-review-contract-hook.md b/changelog.d/codex-review-contract-hook.md deleted file mode 100644 index c7ec24f10..000000000 --- a/changelog.d/codex-review-contract-hook.md +++ /dev/null @@ -1 +0,0 @@ -- **On Codex, /review is now held once when it tries to finish without its adversarial pass or its report, as it already is on Claude Code.** `./setup --host codex` adds the skill-contract Stop hook to `~/.codex/hooks.json` next to the routing hint and the destructive-command check, and Codex runs it once you trust it at its startup prompt or in `/hooks`. The hook reads Codex's rollout file (it tells a Codex rollout from a Claude Code transcript by what the file holds), counts a skill as loaded when Codex injects or the model reads a `gstack-*` SKILL.md, and counts a `spawn_agent` call that started an agent as a dispatch. /review's specialist-reviewer item is skipped on Codex, which doesn't run those reviewers yet. `gstack-skill-contract check` takes a Codex rollout too. The same `skill_contracts` switch turns it off on both hosts. Details: docs/skill-routing-and-contracts.md#codex. diff --git a/changelog.d/codex-routing-hint-uses-codex-skill-names.md b/changelog.d/codex-routing-hint-uses-codex-skill-names.md deleted file mode 100644 index ca8e60cec..000000000 --- a/changelog.d/codex-routing-hint-uses-codex-skill-names.md +++ /dev/null @@ -1 +0,0 @@ -- **On Codex, gstack's skill suggestion now names the skill the way Codex lists it.** The routing hint used to say "gstack-review", a name Codex doesn't know, so mentioning it as `$gstack-review` loaded nothing. It now reads each skill's name from its installed `SKILL.md`, so the hint says `review` and `$review` loads the skill it suggests. diff --git a/changelog.d/codex-skill-mentions-use-listed-names.md b/changelog.d/codex-skill-mentions-use-listed-names.md deleted file mode 100644 index d90c8c633..000000000 --- a/changelog.d/codex-skill-mentions-use-listed-names.md +++ /dev/null @@ -1 +0,0 @@ -- **On Codex, gstack's suggested prompts and docs now name skills the way Codex lists them.** The picker prompts for nine skills (office-hours, review, ship, qa, autoplan, land-and-deploy, modernize, runaway-stopper and write-skill) said `$gstack-office-hours` and so on, which Codex doesn't recognize, so choosing one loaded nothing. They now say `$office-hours`, `$review` and so on, which load the skill. The README's Codex row and the Claude outside-voice note in docs/skills.md say the same. diff --git a/changelog.d/codex-sol-spawn-dispatch.md b/changelog.d/codex-sol-spawn-dispatch.md deleted file mode 100644 index 76ae98da9..000000000 --- a/changelog.d/codex-sol-spawn-dispatch.md +++ /dev/null @@ -1 +0,0 @@ -- **/review on Codex with gpt-6.1-sol is no longer held back for a second-model pass it already ran.** With Sol, Codex starts subagents through its `collaboration` tools, whose result names the new agent by `task_name` instead of `agent_id`, so the end-of-session check missed the adversarial pass and blocked the session once. It now counts a started subagent either way; a spawn that failed still doesn't count. diff --git a/changelog.d/codex-unattended-runs-and-duplicate-skills.md b/changelog.d/codex-unattended-runs-and-duplicate-skills.md deleted file mode 100644 index 08cb67bf0..000000000 --- a/changelog.d/codex-unattended-runs-and-duplicate-skills.md +++ /dev/null @@ -1,4 +0,0 @@ -- **On Codex, office-hours, plan-design-review and gstack-upgrade work again when you name them.** `./setup --host codex` left a second copy of each of these three skills inside gstack's own folder, so Codex listed them twice and loaded nothing when you asked for one (`$office-hours` started without its instructions). On macOS and Linux, setup now links those files instead of copying them, so Codex lists each skill once; Windows setup still copies them, because it can't rely on symlinks there. Re-run `./setup --host codex`, or upgrade gstack, to fix an existing install. -- **gstack knows when Codex is running unattended.** A skill started under `codex exec` (or `codex review`) now reports `SESSION_KIND: headless` instead of `interactive`, the same as an unattended Claude Code run, so it stops with a clear BLOCKED line instead of asking a question nobody will answer. The Codex TUI, and Codex inside Conductor, still count as interactive. -- **Codex reviewers start with a clean context.** The design reviewers, /autoplan's review phases and /context-restore's summary helper now start their Codex subagents with `fork_turns: "none"`, so they no longer inherit the parent's whole conversation (GST-3091). -- **Docs no longer say Codex has no hook support.** Codex CLI 0.170 runs hooks; gstack's routing, contract and destructive-command hooks just aren't ported to it yet. diff --git a/changelog.d/cut-off-child-output-fails-closed.md b/changelog.d/cut-off-child-output-fails-closed.md deleted file mode 100644 index f6b7fb0ec..000000000 --- a/changelog.d/cut-off-child-output-fails-closed.md +++ /dev/null @@ -1 +0,0 @@ -- **A command whose output was cut off no longer counts as a finished read.** Bun can report exit status 0 for a `spawnSync` child whose output passed the read limit (ENOBUFS), handing back only part of the output. The credential pre-push hook, the pregate secret and literal scans, the sealed-output guard, the ship receipt, review scoping and the other gstack commands that read large git or gh output now treat that case as a failure, named `SPAWN_OUTPUT_INCOMPLETE`, instead of judging part of a diff or file list. Before this, a push whose diff passed 64 MiB could carry a credential past the pre-push hook. diff --git a/changelog.d/declared-content-folders-keep-code.md b/changelog.d/declared-content-folders-keep-code.md deleted file mode 100644 index 30c406535..000000000 --- a/changelog.d/declared-content-folders-keep-code.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: blocker ---- -- **A folder marked as published data can no longer hide source code or agent files from /ship's review.** In a folder listed under `review_content_patterns`, source code in any language and letter case (`.java`, `.c`, `.ps1`, `Makefile` and the like, binary or not) still gets reviewed, and so do agent instruction files, config, executables and any data file that kept code names by its full path. Before this, only nine source extensions were kept, a binary `AGENTS.md` was left out, and a file named by a kept script could leave review. No base policy sets the key yet, so nothing was exposed. Details: docs/ship-policy.md#review-content-patterns. diff --git a/changelog.d/docs-hooks-default-followup.md b/changelog.d/docs-hooks-default-followup.md deleted file mode 100644 index 1f5db076d..000000000 --- a/changelog.d/docs-hooks-default-followup.md +++ /dev/null @@ -1 +0,0 @@ -- **Correction to the on-by-default helper hooks entry: the hook still routes only to /review, /qa and /verify, so skills you've installed yourself aren't suggested yet. With gstack's skills it routes 14 of 60 real bug reports (was 17), and long prompts are no longer skipped.** The OpenClaw and Hermes start hook stays off on upgrade whenever it was off, and the one-line "turn it on" hint prints only for an off saved before this release; an off set with `GSTACK_HARNESS_START_HOOK` or `harness_start_hook` stays off without a hint. docs/skill-routing-and-contracts.md (an unreadable config turns all three hooks off; the host table lists the destructive-command check) and docs/hosts.md (only an explicit start-hook request exits 4) now match the code. diff --git a/changelog.d/fix-sync-gbrain-large-page-probe.md b/changelog.d/fix-sync-gbrain-large-page-probe.md deleted file mode 100644 index 2c777fe68..000000000 --- a/changelog.d/fix-sync-gbrain-large-page-probe.md +++ /dev/null @@ -1 +0,0 @@ -- **/sync-gbrain's read check now works when the first page in your code source is over 64 KB.** The check read the whole page just to confirm its source and slug, and a large first page (a lockfile is the usual one) hit its 64 KiB output limit, so every run reported the read as `unknown`. It now asks gbrain for a 4 KB window of the page (`--body window --max-bytes 4096`), which keeps the response small however big the page is. A gbrain from before body windows (0.60.231.0) gets the old whole-page read with the same limit, and output cut off at the limit still counts as `unknown`, never `ready`. Thanks to @nfitzsimmons for the report and the fix (#4396). diff --git a/changelog.d/flaky-fix-2026-11-19.md b/changelog.d/flaky-fix-2026-11-19.md deleted file mode 100644 index 2620208bc..000000000 --- a/changelog.d/flaky-fix-2026-11-19.md +++ /dev/null @@ -1 +0,0 @@ -- **Free-suite CI no longer fails before any test runs when Ubuntu's package mirror is slow.** Playwright's system packages and the gate tools are cached per runner image and Playwright version, so a run on a known image installs them with dpkg instead of downloading about 38 MB from archive.ubuntu.com, and apt waits for a timed-out attempt's lock instead of failing the retry in seconds. On 2026-11-19, 27 free-suite and quick-tier jobs failed this way. Two flaky free tests are fixed too: `first-use` now lists several user drives in sorted order on every filesystem, and the gstack-lease lock test waits for the owner line instead of reading a file that was created but not yet written. Two eval checkers stop refusing correct reads. The read-only command grammar accepts `\"` inside double quotes, as bash and zsh do, and the coverage audit follows a `cd` into a subfolder followed by `cd ..` in the next call. diff --git a/changelog.d/gbrain-read-check-per-repo-sync.md b/changelog.d/gbrain-read-check-per-repo-sync.md deleted file mode 100644 index 725c7fdf8..000000000 --- a/changelog.d/gbrain-read-check-per-repo-sync.md +++ /dev/null @@ -1 +0,0 @@ -- **Syncing a second repo no longer makes the first one ask for /sync-gbrain again.** The GBrain read check looked only at the most recent code sync on the machine, so once you synced another repo, every earlier repo reported itself as unsynced at session start. It now falls back to the per-repo sync record `/sync-gbrain` already keeps. A repo that never synced still asks, and so does one whose record is missing, malformed or points at another folder, or whose own sync is the most recent on the machine and failed. Reported by @SatriaSuriaAriffin (#4499). diff --git a/changelog.d/helper-hooks-on-by-default.md b/changelog.d/helper-hooks-on-by-default.md deleted file mode 100644 index a58563f69..000000000 --- a/changelog.d/helper-hooks-on-by-default.md +++ /dev/null @@ -1,8 +0,0 @@ ---- -bump: minor ---- -- **gstack's helper hooks are now on by default in Claude Code, OpenClaw and Hermes, and the prompt-routing hook fires on real bug reports.** `./setup` now registers the prompt-routing hook (one line naming the gstack skill a prompt fits), the skill-contract hook (holds /review once if it stops before its specialist reviewers ran), and a new destructive-command hook that runs /careful's check before every Bash and PowerShell command, so `rm -rf`, `DROP TABLE`, force-push and the like ask before they run even when nobody loaded /careful (in `claude -p` and other sessions nobody is watching, it prints a warning instead of asking, so nothing waits on an answer). New OpenClaw and Hermes registrations get the start hook (a one-line note on each turn telling the agent to run a skill's start command). Why: on Opus 5.5, unattended, a hook line naming a skill got it loaded in 58 of 60 sessions, where a skill relying on its description loaded in 0 of 60. The routing hook used to skip any prompt over 8,000 characters and miss bug reports that never say "bug"; it now scores a long prompt's first 4,000 and last 1,000 characters, reads words like error, exception, traceback, crash and fails as a bug, and keeps paths such as `site-packages` whole. On 60 real bug reports it now fires on 40 instead of 22, and on none of 60 plain prompts that are not bug reports, reviews, QA or checks. - - **Upgrading:** the first `./setup` or `/gstack-upgrade` after this release turns these hooks on once and prints what it turned on. An off you chose earlier (a `./setup --no-...` flag, `gstack-config set off` or an environment variable) is kept. An OpenClaw or Hermes start hook that is off stays off on upgrade, and the setup output prints one line saying how to turn it on: `gstack-harness register openclaw --start-hook` (or `hermes`). - - **Turning them off:** `gstack-config set skill_routing off`, `gstack-config set skill_contracts off` or `gstack-config set destructive_guard off` (each takes effect on the next prompt or command), or `./setup --no-skill-routing`, `--no-skill-contracts`, `--no-destructive-guard`. For OpenClaw or Hermes: `gstack-harness register openclaw --no-start-hook` (or `hermes`), or `gstack-config set harness_start_hook off` and `./setup`. Details: docs/skill-routing-and-contracts.md and docs/hosts.md#start-hook. diff --git a/changelog.d/pregate-clear-accepts-regenerated-release-outputs.md b/changelog.d/pregate-clear-accepts-regenerated-release-outputs.md deleted file mode 100644 index 2035e32ac..000000000 --- a/changelog.d/pregate-clear-accepts-regenerated-release-outputs.md +++ /dev/null @@ -1 +0,0 @@ -- **`gstack-pregate clear --receipt` accepts a remote gate run from before a merged release when the release only regenerated a declared release output.** A PR that merged a release commit was refused because a generated file the policy lists in `release_outputs` was checked against the `stamp_files` list first, and `clear` never reran the release tool the way the tree receipt does. A `release_outputs` path is now judged by the release-output rule, and with `--allow-repo-commands` `clear` reruns `release_tool` on the head and accepts an output it regenerates byte for byte. Without the flag the receipt is refused as before, and the refusal now says to pass it. A hand edit to the generated file, or any path outside `release_outputs`, is still refused. diff --git a/changelog.d/report-review-budget-stop-not-login.md b/changelog.d/report-review-budget-stop-not-login.md deleted file mode 100644 index afcc70a30..000000000 --- a/changelog.d/report-review-budget-stop-not-login.md +++ /dev/null @@ -1,4 +0,0 @@ -**The second-model review now keeps the reviewer's own result, and reports a used-up budget as a budget stop.** Two fixes to the Codex review that /autoplan, /review, /ship and the other skills run: - -- When the reviewer wrote its review to the result file it was asked for, Codex then replaced that file with its short closing message ("Done, I wrote the review"), so the review was thrown away and reported as missing its markers. Codex's closing message now goes to a separate `.last-message` file, and gstack keeps whichever of the two carries the review. -- When the provider refuses the call with HTTP 402 Payment Required (usually a metering proxy that hit its allowance, or an account out of credits or quota), gstack used to record only "exit 1". It now reads the cause from Codex's JSON error events and reports "the provider refused the review with HTTP 402 (payment required)", and says to check the proxy's allowance, or the account's quota and credits. It has its own reason code, `budget_exhausted`, in the run record and in docs/troubleshooting.md, and the skills' "authentication failed, run `codex login`" rule no longer fires on it. A 401 or 403 is still a login problem, and its record now keeps Codex's status line instead of "exit 1". diff --git a/changelog.d/review-collector-and-verify-timeout-fixes.md b/changelog.d/review-collector-and-verify-timeout-fixes.md deleted file mode 100644 index 722a94b9b..000000000 --- a/changelog.d/review-collector-and-verify-timeout-fixes.md +++ /dev/null @@ -1,3 +0,0 @@ -- **The pre-commit /verify check now fails a check that times out, and its checks share one time budget (GST-3201).** A check that ran out of time used to be listed as "not run", so the commit still got `VERIFY PASS`. Now it is `VERIFY FAIL: TIMEOUT`. `gstack-checks quick --timeout` (540 seconds by default) is now one budget for the whole pass instead of a fresh allowance per check: each check gets the time that is left, and one that can't start in time is a TIMEOUT. The test stage starts with the time that is left, but its runner can still run past it, so a very slow test run can still outlast Claude Code's 10-minute command limit. -- **/ship's review collector no longer reads a reviewer's own scope list or list of passed checks as findings (GST-3201).** A bullet such as "a HIGH finding still refused" in the list of what the tests cover, or "`missing_verdict` can't occur alongside P0 or P1 findings" in the list of checks that held, collected as a HIGH or P0 blocker even though the reviewer declared `REVIEW_VERDICT: blocking=0`. When the reviewer declares no blockers, a plain bullet with no file:line, no Severity or Blocks field and no severity other than one naming findings as a kind keeps its label but is marked for the lead to lower with a written reason, so a bullet that really reports an open blocker still holds the merge until someone judges it. -- **Merge-queue stack tests fail when git cannot list a PR's added test files (GST-3201).** That failure used to read as "no new test files" and pass the slot. diff --git a/changelog.d/review-content-patterns-for-published-data.md b/changelog.d/review-content-patterns-for-published-data.md deleted file mode 100644 index a09052ea4..000000000 --- a/changelog.d/review-content-patterns-for-published-data.md +++ /dev/null @@ -1,6 +0,0 @@ ---- -important: blocker ---- -- **/ship can now review results PRs that publish raw session data.** A repo's ship policy can list folders under a new key, `review_content_patterns`, next to `review_exclude_paths`. A file both keys match counts as data whatever its name, so /ship's reviewers leave it out, the same way they already leave out JSON, CSV, logs and images. Source files in that folder (`.ts`, `.py`, `.sh` and the like, in any letter case) still get reviewed. Before this, files such as `.patch` diffs, a `MANIFEST.sha256` and licence texts with no extension always counted as code, so a folder of them could never be left out, and a PR like gstack-evals #352 (13.8 MB, of which 13.6 MB is published session data) could not get under /ship's 2 MB review limit. Nothing is guessed: only a pattern on the base branch counts, a content pattern on its own excludes nothing, and the secrets scan and the sealed-output guard still read every excluded file. Details: docs/ship-policy.md#review-content-patterns. - - **Also fixed:** when a `review_exclude_paths` pattern kept too many files for the reviewers' read command (more than 24,000 bytes of file names), /ship quietly dropped the whole pattern and sent the folder to the reviewers anyway. It now stops before review with `REVIEW_EXCLUDE_KEEP_TOO_LARGE`, naming the pattern, how many files it keeps and why, and the fix (docs/troubleshooting.md#review-exclude-keep-too-large). diff --git a/changelog.d/review-readers-start-in-another-checkout.md b/changelog.d/review-readers-start-in-another-checkout.md deleted file mode 100644 index 0403e99fc..000000000 --- a/changelog.d/review-readers-start-in-another-checkout.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship's review pass works when the checkout you ship isn't the project's default checkout.** On Capy, review readers start in the thread's working directory, and /ship used to stop with `STOP: readers on the shared machine start in , not in ` whenever that was a different checkout, for example a project whose default checkout is another repository. Every reader's brief already opens by telling it to `cd` into the shipped checkout and report that checkout's tree on its first line, and a reply naming any other tree or path is still refused, so the pass now dispatches. It still stops before anything starts if readers can't reach the shipped checkout at all, or if they start in another worktree or clone of the same history, where a reader that skipped the `cd` would quietly review the wrong diff. (GST-3145) diff --git a/changelog.d/round-2-review-runs-after-a-gate-fix.md b/changelog.d/round-2-review-runs-after-a-gate-fix.md deleted file mode 100644 index 3ee74289b..000000000 --- a/changelog.d/round-2-review-runs-after-a-gate-fix.md +++ /dev/null @@ -1 +0,0 @@ -- **A second review round no longer reports done when nothing was reviewed.** After the first review pass, a fix that touched a gate-class path (CI, auth, merge checks) was planned as "the step applied no edits" unless another fix flag came with it, so /ship's round 2 printed done and no reviewer read the fix. That fix now gets its full review pass on the changed range. When the plan is told nothing was fixed but the branch has new commits since the reviewed head, it now refuses with `review_required_no_readers` instead of printing done. `gstack-review-dispatch` also stops reading its own flags out of the review-plan flags after `--`, so a plan's `--pass 1` can no longer stand in for the pass id. diff --git a/changelog.d/routing-hook-stops-sending-bug-reports-to-qa.md b/changelog.d/routing-hook-stops-sending-bug-reports-to-qa.md deleted file mode 100644 index 8cc23541b..000000000 --- a/changelog.d/routing-hook-stops-sending-bug-reports-to-qa.md +++ /dev/null @@ -1 +0,0 @@ -- **The prompt-routing hook now suggests skills you install yourself, and no longer sends ordinary bug reports to /qa (GST-3123).** Any installed skill that lists `triggers:` in its own SKILL.md can now be named by the hook next to gstack's /review, /qa and /verify, so a bug-fix skill with a "fix this bug" trigger is suggested for bug reports. One line names at most two different skills, gstack's own skill wins a tie, and a skill whose name or trigger text holds anything but letters, digits, spaces and simple punctuation is skipped, so its text can't break out of the quoted trigger or add markup to the line. gstack's own generated skill copies count as gstack's even without an install marker, so they still never route. Separately, /qa no longer takes bug reports on everyday words ("cannot find" plus "fails", or "call site" and "tests" far apart), while "please test the site" and "find bugs on my site" still route to it. `gstack-skill-route --explain` and the routing log show where each suggested skill came from. diff --git a/changelog.d/script-runner-and-merge-queue-refuse-cut-off-output.md b/changelog.d/script-runner-and-merge-queue-refuse-cut-off-output.md deleted file mode 100644 index 4a13297b9..000000000 --- a/changelog.d/script-runner-and-merge-queue-refuse-cut-off-output.md +++ /dev/null @@ -1 +0,0 @@ -- **gstack's script runner and merge-queue checks now also refuse cut-off output.** The helper that runs gstack's own scripts and `bash -c` commands, and the merge queue's branch-ownership, stacked-test, release-version and merge-group record checks, now fail with `SPAWN_OUTPUT_INCOMPLETE` when Bun cuts a command's output off at its read limit, as the pre-push hook and the pre-merge scans already do. The earlier entry about cut-off output said every other gstack command that reads large git or gh output was covered; it was not, because these went straight to Node's `spawnSync`. A new test now stops any gstack command from calling Node's `spawnSync` directly. The 68 files that still do so mostly run short commands or already check for this error. The rest (the co-author trailer helper's commit-log read, the CI health report and the release-stamp precondition) never decide a security check, a ship record or a merge check. The test lists them, and the list can only shrink. The browser's process-list read is still not covered. diff --git a/changelog.d/second-model-review-final-failure-decides.md b/changelog.d/second-model-review-final-failure-decides.md deleted file mode 100644 index 327baa854..000000000 --- a/changelog.d/second-model-review-final-failure-decides.md +++ /dev/null @@ -1 +0,0 @@ -- **A failed second-model review reports the cause that actually stopped it, and the runner never deletes its own prompt.** When Codex logged a 402 on an early attempt and then failed for good with a 401, gstack called it a used-up budget; the final failure now decides. A request id or a duration that happens to contain "402" no longer reads as a budget stop. Claude Code exiting 0 with `subtype: error_during_execution` and a 402 is now a budget stop instead of a generic provider error, and a Claude Code usage limit is reported as `budget_exhausted` instead of with Codex's usage-limit advice. `gstack-outside-voice run` refuses a `--prompt` that is one of the files the run writes for `--out` (such as `.last-message`) instead of deleting it. When the reviewer's own result file has no final verdict, or its first line names another plan snapshot, Codex's complete closing message is used instead. diff --git a/changelog.d/ship-tooling-small-fixes.md b/changelog.d/ship-tooling-small-fixes.md deleted file mode 100644 index 79bfc8113..000000000 --- a/changelog.d/ship-tooling-small-fixes.md +++ /dev/null @@ -1,3 +0,0 @@ -- **`gstack-ship-queue ready` now fills in the trial merges and the second-model trail read instead of printing lines to fill.** It lists the open PRs that change a file this PR changes and prints each one's `git merge-tree` result, and with `--trail ` it names the reader from the trail's newest `trail-read` row, under a train policy too; a gate, security, money, auth or data PR with no such row prints `MISSING: cross-family trail read`. -- **Five small ship-tool fixes.** `Plan: exempt:` accepts bug-fix wording such as "bug fix with failing-first tests", and ready uses the Plan line `/ship` finish resolved when the PR body has none. `gstack-approvals round-grant` refuses only gstack's own names, so "gstack-evals Captain" can grant a round. The Capy PR hand-off follows `GH_REPO` the way gh's own write does and refuses a PR URL that names another repository, and a pull request opened from a fork names its head as `owner:branch`. A targeted review read on a path with non-ASCII characters now carries across a later docs-only change. -- **A ruling the owner gives while a `/ship` run is open can be relayed.** `gstack-approvals rule --owner-ref :` records the owner's words with the message they came from and marks the ruling relayed in the record, the receipt and the PR body; without it the refusal now says `APPROVAL_RULING_RUN_OPEN`. `rule --not-security` lets the owner say a red-team finding is not a security finding, so its PR text shows the finding instead of "security finding, details recorded privately"; it is refused while a `/ship` run is open. diff --git a/changelog.d/stall-check-counts-own-session-workers.md b/changelog.d/stall-check-counts-own-session-workers.md deleted file mode 100644 index ac9395b69..000000000 --- a/changelog.d/stall-check-counts-own-session-workers.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship no longer kills a working test run as stalled when the test command starts its workers in their own session.** The stall check read CPU time from the test command's process group only, so a quiet suite run through a launcher that starts its workers in a new session (pnpm 11's `pnpm exec` running Vitest) looked idle and was killed as `PREGATE_STALLED`. It now counts every process descended from the command, and a stall or timeout kill reaches those workers too. A truly hung run is still killed at the same stall timeout. Reported by @dtoelke in #4477. diff --git a/changelog.d/state-matrix-check-accepts-reader-header.md b/changelog.d/state-matrix-check-accepts-reader-header.md deleted file mode 100644 index 59cc99966..000000000 --- a/changelog.d/state-matrix-check-accepts-reader-header.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: blocker ---- -- **/ship no longer stops a review because the state-machine reader started its reply the way its instructions say.** Every /ship reader is told to open its reply with a `REVIEWED_TREE: ` line and, for readers that can block, a `REVIEW_VERDICT: blocking=` line under it. The state-machine check read those two lines as unexplained text and reported the review as incomplete (`STATE_MATRIX_CHECK: incomplete gaps=2`), which stopped /ship before the adversarial review step on any change that touches a proposal, hold or approval flow. The check now skips exactly those two lines at the top of the reply, using the same parsers /ship's collector uses, so the two can't drift apart. Any other line that isn't a matrix record or a finding, including a misplaced or malformed header line, is still reported as a gap. diff --git a/changelog.d/verify-quick-check-before-commit.md b/changelog.d/verify-quick-check-before-commit.md deleted file mode 100644 index 2090aca55..000000000 --- a/changelog.d/verify-quick-check-before-commit.md +++ /dev/null @@ -1 +0,0 @@ -- **When Claude Code runs /verify right before a commit, it now gets a quick check instead of the full fix loop (GST-3150).** Claude Code 2.1.286 and later tell the model to run any skill named `verify` before every commit that isn't docs-only or tests-only, so gstack's /verify was starting its whole build-and-fix workflow, first-run notices included, on ordinary commits. Now that call runs one command, `gstack-checks quick`: the repo's existing checks (typecheck, lint) and the tests /ship would pick for the change, with no code edits, no new tests and no questions, ending in one line: `VERIFY PASS`, `VERIFY FAIL`, or `VERIFY skipped: `. A repo with no `origin` remote, check commands you haven't approved yet (type /verify once in that repo to review and approve them), or nothing to run is skipped and the commit goes ahead; a failure makes the model stop and ask you before committing. Typing /verify yourself still runs the full skill. To turn the pre-commit check off, run `gstack-config set verify_before_commit off`; to stop Claude Code from asking for it at all, install with `./setup --prefix`, which names the skill /gstack-verify. diff --git a/package.json b/package.json index 22a156dcf..4e73fcbcd 100644 --- a/package.json +++ b/package.json @@ -1,6 +1,6 @@ { "name": "gstack", - "version": "1.101.0", + "version": "1.102.0", "description": "Garry's Stack — Claude Code skills + fast headless browser. One repo, one install, entire AI engineering workflow.", "license": "MIT", "type": "module",