diff --git a/CHANGELOG.md b/CHANGELOG.md index debc49fc6..19ce4d615 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,80 @@ # Changelog +## [1.100.0.0] - 2026-11-17 + +- **/autoplan's approval page no longer hides a reviewer's product objection behind "keep your direction".** A plan can now quote its owner under `### Owner's words`, and only those quotes count as the plan's direction. When a reviewer recommends changing anything else the plan says (its scope, build order, direction, cost or packaging), at any severity and even from one reviewer, the approval page shows it as a lettered choice that recommends the reviewer's change, and the cost line says what taking it actually does. A plan with no Owner's words section keeps today's default and gets one question asking the owner to confirm the direction is theirs. A challenge raised in an earlier review round stays on the page until the owner answers it or a later round marks it fixed. App plans with a `### Day one` table (the owner's first actions, how often, the feature, the way in from the first screen, the first build that has it, and how it's checked) show it first on the approval page, turn each guessed row into a question for the owner, and letter every cut that drops a day-one row or a quoted owner line; an app plan without the table gets one item asking for it. Declining a reviewer's change on a High or Critical finding keeps the plan as written and carries the finding into the build as a blocking task. Quoted owner words, including ones a later revision removed, are redacted from public exports. + +- **Every skill that loads the shared preamble, on every host, now says to ask a question once and to check a file before handing it over.** The Ground Rules section in the shared preamble gains two lines: ask each question once, after checking what you can yourself, and until it is answered list it as one waiting line in later updates instead of asking again; and open every file you hand over and remove placeholders first (TODO, ``, lorem ipsum, broken links and images). The ask-once rule used to live only in the Capy wrapper's "how we work" page, which now points at the shared rule. + +- **The Captain manuals now name the `Failed or skipped:` and `Also changed:` lines that READY and lane reports carry.** A controller asks for a missing one in a follow-up and never refuses or stales a READY over it, the same as a missing `Unverified:` (GST-1455). + +- **gstack's shell locks hold again on Ubuntu 26.04, where `mkdir` reports success when another process won the race (GST-2247).** Ubuntu 26.04, and GitHub's `ubuntu-latest` from 2026-11-09, ships Rust coreutils (uutils) 0.8.0 as `mkdir`. It exits 0 even when the directory already existed because a racing process just made it, so every `mkdir "$LOCK"` lock in gstack could admit two holders: two updates moving one checkout at once, or two `gstack-config set` calls where one dropped the other's key (a test with eight parallel writers under a uutils-style `mkdir` lost six of the eight keys). Every lock now goes through one helper, `gstack_lock_take` in `bin/gstack-lock.sh`: `mkdir` only makes the directory, and the holder is whoever creates the lock's pid or owner file with the shell's noclobber redirect, which bash, zsh and dash open `O_CREAT|O_EXCL`, so the kernel admits exactly one. No `flock` (macOS has none). This covers the session-start updater, `gstack-upgrade`, the Capy session pin switch, `gstack-ff-update` (inlined, pinned to the helper by a test, because it must run standalone), setup's dependency-migration and Chromium-install locks, `gstack-browser-ensure`, `gstack-config`, the install registry, the Claude render lock, the developer profile and its takeover lock, `gstack-memorable`, the settings.json hook, the brain-sync drain lock, the Codex model-probe lock and its breaker, the lease directory lock, freeze's state lock and /setup-gbrain's concurrent-run lock. The Claude Code wrapper's copied runtime (`CLAUDE_CODE_RUNTIME_FILES`) and `bin/gstack-render-claude.sh` now carry or source the helper themselves, so a copied runtime and every script that sources the render library still take the lock. Each caller keeps its own waiting, stale-lock recovery, timeout and release. If the owner file is created but can't be fully written (disk full, a quota or a file-size limit), the helper removes it and the directory again and returns 3, so a failed write never leaves a lock that nobody holds. A lock directory that already exists is still never taken, even with no owner file in it, so a lock left by a killed holder waits for the caller's existing stale rule, as before. +- `test/gstack-lock-race.test.ts` runs the old pattern and the helper against a PATH shim that behaves like uutils 0.8.0 `mkdir`, so it runs on any runner image: the old pattern lets two of four racers win, the helper admits exactly one in each of 60 races and never overlaps across take, hold and release loops, and parallel `gstack-config` writers keep every key. `test/mkdir-lock-guard.test.ts` fails when shipped shell, a template, a workflow, a Markdown bash fence or TypeScript that embeds shell decides anything on a non-`-p` `mkdir`'s exit status; it found 26 such sites before this change, 12 more than the first report listed. The /setup-gbrain lock test now expects `mkdir:` on stderr only when the parent cannot be created, because the helper never runs `mkdir` on a path that already exists; no other assertion, threshold or timeout changed. + +- **Windows Setup E2E now runs when `bin/gstack-lock.sh` changes.** `./setup` sources the lock helper on every install, but the workflow's path filter and its `windows-scope` list left it out, so a pull request that changed only the helper skipped the Windows install check. Both lists now name it. + +- **Claude Code's skill list now says when to use each gstack skill, so Opus loads /review and /plan-eng-review when a request fits.** gstack cut every Claude Code description to its first sentence, which says what a skill is but usually not when to load it. In a test with Opus 5.5 in Claude Code, "Review my changes before I land them" loaded /review 0 of 3 times (its line read "Use when you review this PR's diff."), and "plan this feature" loaded no gstack skill in 9 tries. Almost every Claude Code description now keeps a short when-to-use clause after its first sentence. The clause is the skill's own trigger phrases ("Use for: review my changes, look over my diff, review this pr."), or its "Use when" sentence when it has no trigger list; /design-consultation and /eval-plan keep only their first sentence. /review's line no longer assumes a PR exists (local changes and a branch count), /plan-eng-review says it plans a feature before you build it, and the `gstack` router's line says what it does instead of "Router for the gstack skill suite." The skill list Claude Code loads every session grows by about 840 tokens. The clause is added only to Claude Code's trimmed list, so `--catalog-mode=full` and the other hosts don't get it. + +- **A pull request's merge head can no longer keep its `/ship` receipt without a full review.** Three gaps let a receipt through on less than the full review R51 asks for. A head whose own added lines match the reviewed ones but whose removed lines changed with the base (the base took `SAVE_VERSION := 83` under a branch reviewed as `82` to `84`) used to carry after one scoped reader; it's now a product change, so carry refuses and `/ship` reviews the head in full. A receipt an older gstack carried through a scoped re-check no longer carries on through a later base merge or stamp, and `gstack-ship-queue` won't bind it (`QUEUE_RECEIPT_STALE`), until a fresh full review replaces it. And a `/ship` body post keeps the PR body's `gstack-receipt-runs` block, so `gstack-ship-receipt history` counts every run instead of only the last. The read-only review's `PR_ROUND_CAP` stop no longer offers a round grant it can't honor, and the `stamp-only-carry` no-stop class is gone: a refused carry already runs Steps 2-16 without a stop, so a receipt that still lists that class is invalid and its PR runs `/ship` again. See docs/review-policy.md#rounds-per-pull-request. + +- **`/ship` now counts review rounds per pull request, and the head that merges always gets a full review.** gstack's two-round review cap (R51) was counted inside each `/ship` run, so a PR shipped five times could be reviewed ten times. `/ship` Step 1 now reads the PR body's receipt history (`gstack-ship-receipt history`), prints `PR_REVIEW_HISTORY:` and `PR_REVIEW_ROUTE:`, and passes the earlier rounds to `gstack-review-plan` (`--prior-rounds`, and `--prior-full` for the full rounds the cap counts). A new head whose own lines changed gets a fresh full review, the PR's last before merge; a PR already at two full rounds gets one read-only full review with no fix round, whose blocker goes to the lane's Captain. A scoped re-check of only what changed is now advice only: `carry` prints `RECHECK: advisory` and refuses `--recheck`, because in the replay a diff-only re-check caught none of 8 real faults that full rounds had caught (docs/evals/pr-round-count-results.md). A run that stops before its receipt still counts (`gstack-ship-receipt stub`), and the lane or Captain can buy one more round with `gstack-approvals round-grant --kind extra-round`; gstack cannot grant itself one. See docs/review-policy.md#rounds-per-pull-request. + +- **/ship's documentation audit no longer accepts a passing result when an earlier line from the same audit said "blocked".** A blocked result line followed by an unfenced "current" JSON line for the same audit as the very last line was accepted as current, and a last "current" line whose only fault was its format could be repaired back to current. Now a last line that drops an earlier blocked status or blocker is never accepted. It gets the free one-way re-ask, whose reply must stay blocked and keep those blockers. This corrects the previous entry's claim. A later JSON line, fenced or not, and a re-ask reply can no longer turn a blocked audit into a passing one. A blocker written only in prose after a passing line still depends on the re-ask counting it, which the parser can't check (GST-2349). + +- **/ship's documentation audit no longer spends an attempt when the audit child writes text after its result.** Step 14.5 reads the child's last nonempty line. A child that ended with its JSON result and then restated the section in prose (founderstar #837 and harness-evals #393, both "docs current, no edits") had its result treated as unreadable content, which skipped the free format re-ask. Each rejection spent one of the two attempts, and after both, /ship stopped for an owner risk grant. Text after a JSON line that carries the audit id is now a format slip, provided every such line is free of content problems: `FORMAT_REPAIR:` re-asks once for the whole result line without spending an attempt. The re-ask only works one way. If any such line said blocked or listed a blocker, the reply must stay blocked and keep those blockers, so a later correction or a fenced "current" example can never turn a blocked audit into a passing one. A content problem on such a line (a missing status, "current" with edits) still spends an attempt. The re-ask and the retry note now tell the child to put the section inside the JSON as a string and write nothing after the line (GST-2349). + +- **The README lists the two new guides, and the install doc counts its hooks right.** The README's docs table links the guides for landing a fix in someone else's repository and for running a human playtest. `docs/install.md` no longer says setup registers "up to three hook families" before naming a fourth, and it now mentions the two hooks that stay off until you turn them on, skill routing and skill contracts. The manual uninstall steps in `docs/uninstall.md` now name those two hooks too. + +- **docs/releases.md now says the stable release cut treats a failed commit status as red and a pending one as still going, as it has since #4257.** The docs' definition of a green commit only mentioned check runs. + +- **/ship's `finish --docs-requested` now accepts a docs fix when the audit named the file with a line number.** An audit that wrote `changelog.d/engine-parity-decisions.md:8: wording` or `...md:8:12 wording` in its blockers or decisions was refused with "is not named in the audit result's blockers or decisions", even though `...md: line 8` worked. The path may now be followed by `:` or `::`. A different file whose name starts with the same path still doesn't match. + +- Document the `.gstack/approval-classes.json` schema, `gstack-approval-classes` CLI commands, and owner approval citation format in `docs/ship-policy.md#approval-classes`. + +- `/eval-plan` now runs four result checks instead of leaving them to judgment. `prereg --check` requires a `## Decisions` section (what each possible result changes, at least two different actions) and a `## Scoring` section (a judge from another model family than every arm, checked against at least 40 hand labels at kappa 0.6 or better, and paired results with a 95% interval and the smallest detectable difference stated before the run). The new `judge --check --labels` scores the judge against hand labels, and `results --check` holds back a judged result without passing labels, a zero from a counter never shown to find a planted case, and a failures column with no answer-only column beside it. Preregistrations frozen before 2026-11-15 keep their rules. + +- **The /review coverage-audit eval no longer fails a run that really read the source file when zsh stops a command line at an `echo ====` separator.** zsh treats an unquoted word that starts with `=` as a command lookup, so `cat -n src/billing.ts; echo ====; cat -n test/billing.test.ts` prints the source, then `(eval):1: === not found`, and runs nothing after it. The read check now credits the files printed before that error when the error is the last line of the output and everything before it matches the usual read rules; files after the separator still count only when they are read again (gate census 43575953190). + +Sixteen in-repo documentation links reach their section again. `docs/skills.md` linked 12 skills to `#anchors` with no section (13 links); those rows now open each skill's own `SKILL.md`. `docs/review-policy.md`'s three links to `#targeted-only-read-after-the-loop` lost their target when the heading gained `(GST-1598)`; the heading now carries that anchor. Contributed by @Davmunrey. + +- **/autoplan's approval page closes three gaps left by the last change.** When both reviewers challenge the plan and one calls it a build-order change while the other calls it a scope change, the page now treats it as a scope change, so taking it sends the plan back for review instead of settling it without one. A plan with no Owner's words section now gets the "confirm the direction is yours" question when both reviewers challenge it and the page recommends keeping the plan, as it already did for a challenge from one reviewer. And a quote is checked only against this repository's approvals record, so another repository's record for a branch with the same name can no longer verify it. + +- **gstack agents now open every status report with what failed, what they didn't verify and what else they changed.** The shared closing rule used to ask for "a few short lines". Now a report stays under 250 words and opens with three lines, `Failed or skipped:`, `Unverified:` and `Also changed:`, each naming every item or saying none. Lists longer than 10 get the top 10 and a count of the rest. Capy crews follow the same shape (factory.md point 10, the lane report and READY templates). On a 40-task dev split with Claude Opus 5.5, this wording raised the share of key facts disclosed in a report's first 250 words from 0.442 to 0.520 against the old closer, and in Capy lane reports from 0.552 to 0.604. Those are dev numbers, and the sealed held-out read is still to come (GST-1455). + +- **Two new short guides.** `docs/howto-land-a-fix-in-someone-elses-repo.md` covers reading the other project's rules, searching for the same fix first, running every guard their CI runs before you open, keeping the change small, and answering each review point in a new commit with the proof rerun. `docs/howto-run-human-playtests.md` covers asking about moments and where people would stop, and why a back-to-back pick between two builds measures the order they were played in, not the builds. + +- **A skill you wrote whose contract still lists `check_after_last_edit` keeps the rest of its contract.** This corrects the entry in this release that says such a contract is now invalid and none of its items are checked. The optional Stop hook now skips only that item and keeps checking the skill's other items. Until you delete the item, the hook prints a `SKILL_CONTRACT_SKIPPED` line naming the file and the item. A contract whose only items use the removed kind is still invalid, so the hook checks nothing for that skill. + +- **The ship policy guide now explains how to merge any pull request with `gstack-ship-queue land`, even in a repository with one maintainer.** It names the two policy keys land needs before it will run, lists what land already checks, gives a pinned `gh pr merge` fallback for repositories that can't use land yet, and covers four common traps. CONTRIBUTING.md links to it. + +- **/ship's test step runs a large unit suite on Linux.** A lane with thousands of unit files used to stop with `PREGATE: incomplete` and `E2BIG: argument list too long` before any test ran, because every file went into one command argument and Linux caps a single argument at 128 KiB. gstack now splits a big file list into runner processes of about 96 KB each, run one after another and counted as one run, so the result is the same as before: any failing chunk fails the lane. An `E2BIG` launch failure is now reported as its own cause (`arg-limit`) instead of telling you to install a missing tool. (#4424) + +- **Every skill that loads the shared preamble, on every host, now says never to print a whole environment, git config or process table.** The rule used to load only in the Capy wrapper, and its list missed `git config --list`, which can print a proxy's login header, process or container listings that show arguments or environment (`ps aux`, `ps -ef`, `ps -o args`, `docker top ... args`, `docker inspect`), where run tokens ride, and `set -x` or an `export $(...)` that expands to a bare `export`. A new Ground Rules section in the shared preamble lists these commands and the safe forms, which are to read one value by name (`git config --get `), check a variable is set without printing it, and list processes by name or pid. It also says to wait on CI with one `gstack-pr-wait` call instead of a poll loop or `gh pr checks --watch`, which spends the shared GitHub GraphQL quota. The section is pinned, so no model loadout drops it. + +- **Pull requests from /ship now open with at most three plain lines, and every merge lands as one commit titled with the PR title.** /ship writes `## ` sections and `gstack-post` folds each into a collapsed section when it posts, so a reader sees what changed, why, and what they have to do, with the rest one click away and the Agent notes section last. `gstack-post fold --body-file F` shows the folded body without posting it. In a repository with `.gstack/plain-titles.json`, `gstack-post` and the Plain title check refuse a body that shows more than three lines (or 600 characters) before its first fold, or uses a listed in-house word above the Agent notes section (`PLAIN_BODY_REFUSED`); gstack now ships a default word list (outside voice, pregate, receipt, carry, stamp, train, lane, soak and more), which a policy extends or replaces with `replace_default_terms`, and `body_since` keeps PRs opened before that date to warnings. Every `gstack-ship-queue land` form now passes the subject ` (#n)` and a body of the PR's opening lines plus the branch's `Co-authored-by:` and other trailers, so GitHub no longer writes its commit list; `--dry-run` prints it and `--body-file` replaces the opening. Release PRs list one plain line per change with its PR. Two new policy settings: `history: no-rewrite` (squash landings, never a rebased or force-pushed branch) and `land.merge_commit_paths` (keep a merge commit for PRs that change those paths, such as eval results). See docs/writing-prs.md. + +- **`gstack-pr-wait`, `gstack-ci-gate` and the release cut no longer call a commit green while a check on it failed.** They read CI from GitHub's rollup or its default check-runs list, which hold only each check suite's latest attempt, so when a rerun attempt created no jobs the failed run before it disappeared, and a workflow run that failed before creating any job never appeared at all. That is the false green the merge queue had on gstack-evals#303 (#4235). All three now also read CI at the commit through the merge queue's own helper, which reads every attempt and every workflow run, and keep the worse answer per check. So `gstack-pr-wait --until green` blocks with `check failed: checks` and `check failed: sealed (run <id>, no jobs)`, the land gate says `VERDICT FAIL`, and the release cut skips the commit. The release cut now also counts a failed commit status as red, as the merge queue does. A rerun that passed still reads green, and an unreadable read is an error, never a pass (GST-2252). + +- **`gstack-runaway` now says what a thread's cold and empty wakes cost, not just how many there were (GST-2770).** Each wake it reads is marked cold when it started more than 5 minutes after the thread's last activity, so the prompt cache had lapsed, and is priced in warm model calls, with a cold call at 19.4 (measured on Opus 5.5: $2.42 against $0.125 per million counted tokens; a call after a tool that ran past 5 minutes is cold too). The RUNAWAY line adds `wakes`, `cold_wakes`, `empty_wakes`, `cold_extra_spend` and `empty_spend`, the empty-wake warning gives the empty wakes' share of spend, and `scan --wakes` prints one WAKE line per wake. A message sent to another thread now counts as action even when its state is "working", and more than 5 minutes idle inside a turn (not counting a tool's own run time) starts a new wake. On four coordinating threads over 4.7 days the meter's cold and empty shares came within 1.3 points of a hand count; before this, sends were counted as empty and the empty share read 4 to 12 points high. /runaway-stopper's fetch program now keeps each message's target and each call's run time. + +- **The optional Stop hook no longer checks for a test run after the last edit, and /verify asks for it in plain words instead.** With /verify loaded, Opus 5.5 and GPT-6.1 Sol already ran a test after their last edit in 96% and 89% of runs, and the hook's check added between -0.5 and +2 points, inside the noise on both models. The hook's other checks (report labels, specialist dispatches, the adversarial pass) are unchanged. If you wrote your own skill whose contract lists `check_after_last_edit`, that contract is now invalid. The hook then checks none of that skill's items and prints one line naming the file until you delete the item. /verify now says to run the full check after your last edit. It also no longer says it never edits CHANGELOG, because a changelog entry the repo expects with a change is part of the change; /ship still folds branch entries into the release entry. + +- **/ship's test step says when it ran out of time, and you can set its budget.** `gstack-ship-preflight` used to stop gstack-pregate after a fixed hour and print `PREGATE: unreadable (exit 143)`, which hid the cause. The whole step's budget is now `--step-timeout <s>` (default 3600 s), and a stop at that limit prints `PREGATE: time limit hit` with the way out: push and let CI's full run carry the tests, or give the step a larger budget. +- **The remote test command no longer puts every file on one command line.** On a small machine with a remote `test_backend`, the command pregate prints for gstack's Ubicloud runner used to carry every selected test file in one argument, which Linux refuses past 128 KiB (`E2BIG`). The files now go to the VM in a setup script, so the command stays short however many files are selected. docs/troubleshooting.md has a new `arg-limit` entry for this kind of failure. + +- **The stable release no longer waits on the nightly Windows tests (GST-2815).** The nightly Windows free suite never blocks a pull request, but the daily stable cut still treated its red or unfinished results as a reason to skip a commit, so stable stopped moving while one Windows shard kept timing out. The cut now sets that workflow's check runs aside, matched by workflow file, the same way it ignores its own job. Every other check, the required `free-tests` check and the release-gate evidence still block exactly as before. + +- **Ubicloud fan-outs can no longer overshoot the vCPU limit, and owners no longer delete each other's VMs.** `ubi-runner.sh up` now reads the project's vCPU usage and books its new VM under one lock per machine, so shards started together see each other's bookings instead of all passing on the same reading. A new `ubi-runner.sh room [-s SIZE]` command prints how many more VMs fit now, so a fan-out can size each batch to the room left. When usage can't be read and `UBI_QUOTA_VCPUS` is set, the create is refused (`UBI_QUOTA_UNREADABLE`) instead of going ahead. VM names now carry the whole owner id (up to 31 characters; longer is refused with `UBI_OWNER_TOO_LONG`) instead of its first 12, so two owners sharing a prefix no longer reap each other's expired VMs (#3748); a VM an older runner copy named under the old cut form is still reaped by the long owner on the machine whose VM record lists it, and `recover` accepts the longer owner ids. Merge checks and canaries set no `UBI_QUOTA_VCPUS` and behave as before. + +**hono is pinned to 4.13.16, which fixes five advisories the weekly OSV scan found in 4.13.13 (gstack#4406).** The 2026-11-16 scan (run 43825448246) flagged GHSA-3gg2-335m-rxc6 (CVSS 3.4, the `css` helper) and GHSA-46wx-v44x-9hh2, GHSA-4h7h-rv7c-7m2v, GHSA-4m76-xfj5-7975 and GHSA-h7mg-cq63-vhpq (CVSS 5.3 each: `csrf()`, `validator('form')`, `mount()` and RegExpRouter wildcard middleware). hono reaches gstack only through the dev and test toolchain (`@anthropic-ai/claude-agent-sdk` pulls in `@modelcontextprotocol/sdk`), gstack's own code never imports it or starts a hono server, and none of the affected helpers is used, so no advisory was reachable. Contributor checkouts get the patched version on their next `bun install`. + +- **`package.json` overrides**: `hono` moves from 4.13.13 to 4.13.16, the release that fixes all five. `bun.lock` holds the one copy. +- `test/dependency-security.test.ts`: the lockfile floor check now covers `hono >= 4.13.16`. It fails on main. + +- **The nightly Windows test run finishes again.** Since 2026-11-06 one of its six shards ran past the 15-minute job limit almost every night, because 95 newer test files had no recorded Windows duration and the planner packed them blind, so one shard got 26 minutes of work while another got 12. The Windows durations are refreshed to a recording that covers all 902 files, and the run now uses eight shards of about 12 minutes each, so the suite fits with room to grow. No test was dropped and the time limit is unchanged. + +- **A short write-up checklist in ETHOS.md.** Explain the Number now ends with three checks before a post or write-up goes out: its words match the page it links, every number, "green" and "ready" names its source and is read at the current head, and a fresh model reads it cold once. The READY template in `captain/templates/READY.md` points to the second check. Anything carried over from an earlier head goes under `Unverified:` until it is re-read. + ## [1.99.5.0] - 2026-11-15 - **CONTRIBUTING.md now says plainly how outside reports and PRs land.** It explains that AI agents working for the maintainer triage, rebuild and land outside patches with the author credited, which is why commits read as the maintainer's. It lists the three paths a PR can take (a docs-only PR for people can merge as it is after its links, commands and install steps are checked; a plainly correct fix of about 40 lines is taken even when small; everything else is rebuilt on main when it has proof, a real consequence and matters to most users), the 48-hour first-reply goal, and the narrowed open-PR limit, which never applies to anyone with a PR taken in the last 90 days. The stale wave process, which merged outside PRs on a collector branch, is gone. `crew/docs/captain.md`, the Captain's CREW template and `/captain` match. diff --git a/VERSION b/VERSION index 7a9079c44..f173507db 100644 --- a/VERSION +++ b/VERSION @@ -1 +1 @@ -1.99.5.0 +1.100.0.0 diff --git a/agents-digest/gstack-AGENTS.md b/agents-digest/gstack-AGENTS.md index 370b4f309..17c59a82d 100644 --- a/agents-digest/gstack-AGENTS.md +++ b/agents-digest/gstack-AGENTS.md @@ -1,4 +1,4 @@ -# gstack digest v1.99.5.0 — regenerate/re-copy after upgrading gstack +# gstack digest v1.100.0.0 — 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/approval-page-keeps-reviewer-product-challenges-open.md b/changelog.d/approval-page-keeps-reviewer-product-challenges-open.md deleted file mode 100644 index 7486b88ad..000000000 --- a/changelog.d/approval-page-keeps-reviewer-product-challenges-open.md +++ /dev/null @@ -1 +0,0 @@ -- **/autoplan's approval page no longer hides a reviewer's product objection behind "keep your direction".** A plan can now quote its owner under `### Owner's words`, and only those quotes count as the plan's direction. When a reviewer recommends changing anything else the plan says (its scope, build order, direction, cost or packaging), at any severity and even from one reviewer, the approval page shows it as a lettered choice that recommends the reviewer's change, and the cost line says what taking it actually does. A plan with no Owner's words section keeps today's default and gets one question asking the owner to confirm the direction is theirs. A challenge raised in an earlier review round stays on the page until the owner answers it or a later round marks it fixed. App plans with a `### Day one` table (the owner's first actions, how often, the feature, the way in from the first screen, the first build that has it, and how it's checked) show it first on the approval page, turn each guessed row into a question for the owner, and letter every cut that drops a day-one row or a quoted owner line; an app plan without the table gets one item asking for it. Declining a reviewer's change on a High or Critical finding keeps the plan as written and carries the finding into the build as a blocking task. Quoted owner words, including ones a later revision removed, are redacted from public exports. diff --git a/changelog.d/ask-once-and-check-handoff-files.md b/changelog.d/ask-once-and-check-handoff-files.md deleted file mode 100644 index dc76b9535..000000000 --- a/changelog.d/ask-once-and-check-handoff-files.md +++ /dev/null @@ -1 +0,0 @@ -- **Every skill that loads the shared preamble, on every host, now says to ask a question once and to check a file before handing it over.** The Ground Rules section in the shared preamble gains two lines: ask each question once, after checking what you can yourself, and until it is answered list it as one waiting line in later updates instead of asking again; and open every file you hand over and remove placeholders first (TODO, `<fill in>`, lorem ipsum, broken links and images). The ask-once rule used to live only in the Capy wrapper's "how we work" page, which now points at the shared rule. diff --git a/changelog.d/captain-docs-report-lines.md b/changelog.d/captain-docs-report-lines.md deleted file mode 100644 index a580b24c2..000000000 --- a/changelog.d/captain-docs-report-lines.md +++ /dev/null @@ -1 +0,0 @@ -- **The Captain manuals now name the `Failed or skipped:` and `Also changed:` lines that READY and lane reports carry.** A controller asks for a missing one in a follow-up and never refuses or stales a READY over it, the same as a missing `Unverified:` (GST-1455). diff --git a/changelog.d/capy-shell-locks-hold-when-mkdir-lies-gst-2247.md b/changelog.d/capy-shell-locks-hold-when-mkdir-lies-gst-2247.md deleted file mode 100644 index baa8e6560..000000000 --- a/changelog.d/capy-shell-locks-hold-when-mkdir-lies-gst-2247.md +++ /dev/null @@ -1,5 +0,0 @@ ---- -important: blocker ---- -- **gstack's shell locks hold again on Ubuntu 26.04, where `mkdir` reports success when another process won the race (GST-2247).** Ubuntu 26.04, and GitHub's `ubuntu-latest` from 2026-11-09, ships Rust coreutils (uutils) 0.8.0 as `mkdir`. It exits 0 even when the directory already existed because a racing process just made it, so every `mkdir "$LOCK"` lock in gstack could admit two holders: two updates moving one checkout at once, or two `gstack-config set` calls where one dropped the other's key (a test with eight parallel writers under a uutils-style `mkdir` lost six of the eight keys). Every lock now goes through one helper, `gstack_lock_take` in `bin/gstack-lock.sh`: `mkdir` only makes the directory, and the holder is whoever creates the lock's pid or owner file with the shell's noclobber redirect, which bash, zsh and dash open `O_CREAT|O_EXCL`, so the kernel admits exactly one. No `flock` (macOS has none). This covers the session-start updater, `gstack-upgrade`, the Capy session pin switch, `gstack-ff-update` (inlined, pinned to the helper by a test, because it must run standalone), setup's dependency-migration and Chromium-install locks, `gstack-browser-ensure`, `gstack-config`, the install registry, the Claude render lock, the developer profile and its takeover lock, `gstack-memorable`, the settings.json hook, the brain-sync drain lock, the Codex model-probe lock and its breaker, the lease directory lock, freeze's state lock and /setup-gbrain's concurrent-run lock. The Claude Code wrapper's copied runtime (`CLAUDE_CODE_RUNTIME_FILES`) and `bin/gstack-render-claude.sh` now carry or source the helper themselves, so a copied runtime and every script that sources the render library still take the lock. Each caller keeps its own waiting, stale-lock recovery, timeout and release. If the owner file is created but can't be fully written (disk full, a quota or a file-size limit), the helper removes it and the directory again and returns 3, so a failed write never leaves a lock that nobody holds. A lock directory that already exists is still never taken, even with no owner file in it, so a lock left by a killed holder waits for the caller's existing stale rule, as before. -- `test/gstack-lock-race.test.ts` runs the old pattern and the helper against a PATH shim that behaves like uutils 0.8.0 `mkdir`, so it runs on any runner image: the old pattern lets two of four racers win, the helper admits exactly one in each of 60 races and never overlaps across take, hold and release loops, and parallel `gstack-config` writers keep every key. `test/mkdir-lock-guard.test.ts` fails when shipped shell, a template, a workflow, a Markdown bash fence or TypeScript that embeds shell decides anything on a non-`-p` `mkdir`'s exit status; it found 26 such sites before this change, 12 more than the first report listed. The /setup-gbrain lock test now expects `mkdir:` on stderr only when the parent cannot be created, because the helper never runs `mkdir` on a path that already exists; no other assertion, threshold or timeout changed. diff --git a/changelog.d/capy-windows-setup-e2e-watches-gstack-lock.md b/changelog.d/capy-windows-setup-e2e-watches-gstack-lock.md deleted file mode 100644 index 4ef32983d..000000000 --- a/changelog.d/capy-windows-setup-e2e-watches-gstack-lock.md +++ /dev/null @@ -1 +0,0 @@ -- **Windows Setup E2E now runs when `bin/gstack-lock.sh` changes.** `./setup` sources the lock helper on every install, but the workflow's path filter and its `windows-scope` list left it out, so a pull request that changed only the helper skipped the Windows install check. Both lists now name it. diff --git a/changelog.d/claude-skill-descriptions-say-when-to-use.md b/changelog.d/claude-skill-descriptions-say-when-to-use.md deleted file mode 100644 index fdef5b07e..000000000 --- a/changelog.d/claude-skill-descriptions-say-when-to-use.md +++ /dev/null @@ -1 +0,0 @@ -- **Claude Code's skill list now says when to use each gstack skill, so Opus loads /review and /plan-eng-review when a request fits.** gstack cut every Claude Code description to its first sentence, which says what a skill is but usually not when to load it. In a test with Opus 5.5 in Claude Code, "Review my changes before I land them" loaded /review 0 of 3 times (its line read "Use when you review this PR's diff."), and "plan this feature" loaded no gstack skill in 9 tries. Almost every Claude Code description now keeps a short when-to-use clause after its first sentence. The clause is the skill's own trigger phrases ("Use for: review my changes, look over my diff, review this pr."), or its "Use when" sentence when it has no trigger list; /design-consultation and /eval-plan keep only their first sentence. /review's line no longer assumes a PR exists (local changes and a branch count), /plan-eng-review says it plans a feature before you build it, and the `gstack` router's line says what it does instead of "Router for the gstack skill suite." The skill list Claude Code loads every session grows by about 840 tokens. The clause is added only to Claude Code's trimmed list, so `--catalog-mode=full` and the other hosts don't get it. diff --git a/changelog.d/close-last-recheck-carry-paths.md b/changelog.d/close-last-recheck-carry-paths.md deleted file mode 100644 index ad0af674a..000000000 --- a/changelog.d/close-last-recheck-carry-paths.md +++ /dev/null @@ -1 +0,0 @@ -- **A pull request's merge head can no longer keep its `/ship` receipt without a full review.** Three gaps let a receipt through on less than the full review R51 asks for. A head whose own added lines match the reviewed ones but whose removed lines changed with the base (the base took `SAVE_VERSION := 83` under a branch reviewed as `82` to `84`) used to carry after one scoped reader; it's now a product change, so carry refuses and `/ship` reviews the head in full. A receipt an older gstack carried through a scoped re-check no longer carries on through a later base merge or stamp, and `gstack-ship-queue` won't bind it (`QUEUE_RECEIPT_STALE`), until a fresh full review replaces it. And a `/ship` body post keeps the PR body's `gstack-receipt-runs` block, so `gstack-ship-receipt history` counts every run instead of only the last. The read-only review's `PR_ROUND_CAP` stop no longer offers a round grant it can't honor, and the `stamp-only-carry` no-stop class is gone: a refused carry already runs Steps 2-16 without a stop, so a receipt that still lists that class is invalid and its PR runs `/ship` again. See docs/review-policy.md#rounds-per-pull-request. diff --git a/changelog.d/count-review-rounds-per-pr.md b/changelog.d/count-review-rounds-per-pr.md deleted file mode 100644 index 5fb3fa1c8..000000000 --- a/changelog.d/count-review-rounds-per-pr.md +++ /dev/null @@ -1 +0,0 @@ -- **`/ship` now counts review rounds per pull request, and the head that merges always gets a full review.** gstack's two-round review cap (R51) was counted inside each `/ship` run, so a PR shipped five times could be reviewed ten times. `/ship` Step 1 now reads the PR body's receipt history (`gstack-ship-receipt history`), prints `PR_REVIEW_HISTORY:` and `PR_REVIEW_ROUTE:`, and passes the earlier rounds to `gstack-review-plan` (`--prior-rounds`, and `--prior-full` for the full rounds the cap counts). A new head whose own lines changed gets a fresh full review, the PR's last before merge; a PR already at two full rounds gets one read-only full review with no fix round, whose blocker goes to the lane's Captain. A scoped re-check of only what changed is now advice only: `carry` prints `RECHECK: advisory` and refuses `--recheck`, because in the replay a diff-only re-check caught none of 8 real faults that full rounds had caught (docs/evals/pr-round-count-results.md). A run that stops before its receipt still counts (`gstack-ship-receipt stub`), and the lane or Captain can buy one more round with `gstack-approvals round-grant --kind extra-round`; gstack cannot grant itself one. See docs/review-policy.md#rounds-per-pull-request. diff --git a/changelog.d/docs-audit-blocked-line-carries-gst-2349.md b/changelog.d/docs-audit-blocked-line-carries-gst-2349.md deleted file mode 100644 index b5190db44..000000000 --- a/changelog.d/docs-audit-blocked-line-carries-gst-2349.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: blocker ---- -- **/ship's documentation audit no longer accepts a passing result when an earlier line from the same audit said "blocked".** A blocked result line followed by an unfenced "current" JSON line for the same audit as the very last line was accepted as current, and a last "current" line whose only fault was its format could be repaired back to current. Now a last line that drops an earlier blocked status or blocker is never accepted. It gets the free one-way re-ask, whose reply must stay blocked and keep those blockers. This corrects the previous entry's claim. A later JSON line, fenced or not, and a re-ask reply can no longer turn a blocked audit into a passing one. A blocker written only in prose after a passing line still depends on the re-ask counting it, which the parser can't check (GST-2349). diff --git a/changelog.d/docs-audit-ignores-prose-after-result-gst-2349.md b/changelog.d/docs-audit-ignores-prose-after-result-gst-2349.md deleted file mode 100644 index e5ec338b0..000000000 --- a/changelog.d/docs-audit-ignores-prose-after-result-gst-2349.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: blocker ---- -- **/ship's documentation audit no longer spends an attempt when the audit child writes text after its result.** Step 14.5 reads the child's last nonempty line. A child that ended with its JSON result and then restated the section in prose (founderstar #837 and harness-evals #393, both "docs current, no edits") had its result treated as unreadable content, which skipped the free format re-ask. Each rejection spent one of the two attempts, and after both, /ship stopped for an owner risk grant. Text after a JSON line that carries the audit id is now a format slip, provided every such line is free of content problems: `FORMAT_REPAIR:` re-asks once for the whole result line without spending an attempt. The re-ask only works one way. If any such line said blocked or listed a blocker, the reply must stay blocked and keep those blockers, so a later correction or a fenced "current" example can never turn a blocked audit into a passing one. A content problem on such a line (a missing status, "current" with edits) still spends an attempt. The re-ask and the retry note now tell the child to put the section inside the JSON as a string and write nothing after the line (GST-2349). diff --git a/changelog.d/docs-index-and-install-hooks.md b/changelog.d/docs-index-and-install-hooks.md deleted file mode 100644 index 94daa40d6..000000000 --- a/changelog.d/docs-index-and-install-hooks.md +++ /dev/null @@ -1 +0,0 @@ -- **The README lists the two new guides, and the install doc counts its hooks right.** The README's docs table links the guides for landing a fix in someone else's repository and for running a human playtest. `docs/install.md` no longer says setup registers "up to three hook families" before naming a fourth, and it now mentions the two hooks that stay off until you turn them on, skill routing and skill contracts. The manual uninstall steps in `docs/uninstall.md` now name those two hooks too. diff --git a/changelog.d/docs-release-cut-green-counts-commit-statuses.md b/changelog.d/docs-release-cut-green-counts-commit-statuses.md deleted file mode 100644 index 847426185..000000000 --- a/changelog.d/docs-release-cut-green-counts-commit-statuses.md +++ /dev/null @@ -1 +0,0 @@ -- **docs/releases.md now says the stable release cut treats a failed commit status as red and a pending one as still going, as it has since #4257.** The docs' definition of a green commit only mentioned check runs. diff --git a/changelog.d/docs-requested-accepts-path-line-blockers.md b/changelog.d/docs-requested-accepts-path-line-blockers.md deleted file mode 100644 index cdee4b9d8..000000000 --- a/changelog.d/docs-requested-accepts-path-line-blockers.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship's `finish --docs-requested` now accepts a docs fix when the audit named the file with a line number.** An audit that wrote `changelog.d/engine-parity-decisions.md:8: wording` or `...md:8:12 wording` in its blockers or decisions was refused with "is not named in the audit result's blockers or decisions", even though `...md: line 8` worked. The path may now be followed by `:<line>` or `:<line>:<column>`. A different file whose name starts with the same path still doesn't match. diff --git a/changelog.d/docs-ship-policy-approval-classes-4320.md b/changelog.d/docs-ship-policy-approval-classes-4320.md deleted file mode 100644 index c6e94573c..000000000 --- a/changelog.d/docs-ship-policy-approval-classes-4320.md +++ /dev/null @@ -1 +0,0 @@ -- Document the `.gstack/approval-classes.json` schema, `gstack-approval-classes` CLI commands, and owner approval citation format in `docs/ship-policy.md#approval-classes`. diff --git a/changelog.d/eval-plan-result-checks.md b/changelog.d/eval-plan-result-checks.md deleted file mode 100644 index 8613ccc0a..000000000 --- a/changelog.d/eval-plan-result-checks.md +++ /dev/null @@ -1 +0,0 @@ -- `/eval-plan` now runs four result checks instead of leaving them to judgment. `prereg --check` requires a `## Decisions` section (what each possible result changes, at least two different actions) and a `## Scoring` section (a judge from another model family than every arm, checked against at least 40 hand labels at kappa 0.6 or better, and paired results with a 95% interval and the smallest detectable difference stated before the run). The new `judge --check --labels` scores the judge against hand labels, and `results --check` holds back a judged result without passing labels, a zero from a counter never shown to find a planted case, and a failures column with no answer-only column beside it. Preregistrations frozen before 2026-11-15 keep their rules. diff --git a/changelog.d/fix-coverage-audit-zsh-separator.md b/changelog.d/fix-coverage-audit-zsh-separator.md deleted file mode 100644 index d01c3736a..000000000 --- a/changelog.d/fix-coverage-audit-zsh-separator.md +++ /dev/null @@ -1 +0,0 @@ -- **The /review coverage-audit eval no longer fails a run that really read the source file when zsh stops a command line at an `echo ====` separator.** zsh treats an unquoted word that starts with `=` as a command lookup, so `cat -n src/billing.ts; echo ====; cat -n test/billing.test.ts` prints the source, then `(eval):1: === not found`, and runs nothing after it. The read check now credits the files printed before that error when the error is the last line of the output and everything before it matches the usual read rules; files after the separator still count only when they are read again (gate census 43575953190). diff --git a/changelog.d/fix-docs-missing-anchors.md b/changelog.d/fix-docs-missing-anchors.md deleted file mode 100644 index 27823e38e..000000000 --- a/changelog.d/fix-docs-missing-anchors.md +++ /dev/null @@ -1 +0,0 @@ -Sixteen in-repo documentation links reach their section again. `docs/skills.md` linked 12 skills to `#anchors` with no section (13 links); those rows now open each skill's own `SKILL.md`. `docs/review-policy.md`'s three links to `#targeted-only-read-after-the-loop` lost their target when the heading gained `(GST-1598)`; the heading now carries that anchor. Contributed by @Davmunrey. diff --git a/changelog.d/fix-gate-gaps-after-day-one-gate.md b/changelog.d/fix-gate-gaps-after-day-one-gate.md deleted file mode 100644 index 3a1d3ebe2..000000000 --- a/changelog.d/fix-gate-gaps-after-day-one-gate.md +++ /dev/null @@ -1 +0,0 @@ -- **/autoplan's approval page closes three gaps left by the last change.** When both reviewers challenge the plan and one calls it a build-order change while the other calls it a scope change, the page now treats it as a scope change, so taking it sends the plan back for review instead of settling it without one. A plan with no Owner's words section now gets the "confirm the direction is yours" question when both reviewers challenge it and the page recommends keeping the plan, as it already did for a challenge from one reviewer. And a quote is checked only against this repository's approvals record, so another repository's record for a branch with the same name can no longer verify it. diff --git a/changelog.d/gstack-reports-lead-with-failures.md b/changelog.d/gstack-reports-lead-with-failures.md deleted file mode 100644 index dd0dda6ec..000000000 --- a/changelog.d/gstack-reports-lead-with-failures.md +++ /dev/null @@ -1 +0,0 @@ -- **gstack agents now open every status report with what failed, what they didn't verify and what else they changed.** The shared closing rule used to ask for "a few short lines". Now a report stays under 250 words and opens with three lines, `Failed or skipped:`, `Unverified:` and `Also changed:`, each naming every item or saying none. Lists longer than 10 get the top 10 and a count of the rest. Capy crews follow the same shape (factory.md point 10, the lane report and READY templates). On a 40-task dev split with Claude Opus 5.5, this wording raised the share of key facts disclosed in a report's first 250 words from 0.442 to 0.520 against the old closer, and in Capy lane reports from 0.552 to 0.604. Those are dev numbers, and the sealed held-out read is still to come (GST-1455). diff --git a/changelog.d/guides-for-outside-fixes-and-playtests.md b/changelog.d/guides-for-outside-fixes-and-playtests.md deleted file mode 100644 index 6baed792a..000000000 --- a/changelog.d/guides-for-outside-fixes-and-playtests.md +++ /dev/null @@ -1 +0,0 @@ -- **Two new short guides.** `docs/howto-land-a-fix-in-someone-elses-repo.md` covers reading the other project's rules, searching for the same fix first, running every guard their CI runs before you open, keeping the change small, and answering each review point in a new commit with the proof rerun. `docs/howto-run-human-playtests.md` covers asking about moments and where people would stop, and why a back-to-back pick between two builds measures the order they were played in, not the builds. diff --git a/changelog.d/keep-custom-contracts-with-removed-item.md b/changelog.d/keep-custom-contracts-with-removed-item.md deleted file mode 100644 index a8789f001..000000000 --- a/changelog.d/keep-custom-contracts-with-removed-item.md +++ /dev/null @@ -1 +0,0 @@ -- **A skill you wrote whose contract still lists `check_after_last_edit` keeps the rest of its contract.** This corrects the entry in this release that says such a contract is now invalid and none of its items are checked. The optional Stop hook now skips only that item and keeps checking the skill's other items. Until you delete the item, the hook prints a `SKILL_CONTRACT_SKIPPED` line naming the file and the item. A contract whose only items use the removed kind is still invalid, so the hook checks nothing for that skill. diff --git a/changelog.d/land-normal-merge.md b/changelog.d/land-normal-merge.md deleted file mode 100644 index b76774f3a..000000000 --- a/changelog.d/land-normal-merge.md +++ /dev/null @@ -1 +0,0 @@ -- **The ship policy guide now explains how to merge any pull request with `gstack-ship-queue land`, even in a repository with one maintainer.** It names the two policy keys land needs before it will run, lists what land already checks, gives a pinned `gh pr merge` fallback for repositories that can't use land yet, and covers four common traps. CONTRIBUTING.md links to it. diff --git a/changelog.d/lanes-large-file-lists-on-linux.md b/changelog.d/lanes-large-file-lists-on-linux.md deleted file mode 100644 index 68f8401de..000000000 --- a/changelog.d/lanes-large-file-lists-on-linux.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship's test step runs a large unit suite on Linux.** A lane with thousands of unit files used to stop with `PREGATE: incomplete` and `E2BIG: argument list too long` before any test ran, because every file went into one command argument and Linux caps a single argument at 128 KiB. gstack now splits a big file list into runner processes of about 96 KB each, run one after another and counted as one run, so the result is the same as before: any failing chunk fails the lane. An `E2BIG` launch failure is now reported as its own cause (`arg-limit`) instead of telling you to install a missing tool. (#4424) diff --git a/changelog.d/never-print-secrets-shared-rule.md b/changelog.d/never-print-secrets-shared-rule.md deleted file mode 100644 index 0288caa48..000000000 --- a/changelog.d/never-print-secrets-shared-rule.md +++ /dev/null @@ -1 +0,0 @@ -- **Every skill that loads the shared preamble, on every host, now says never to print a whole environment, git config or process table.** The rule used to load only in the Capy wrapper, and its list missed `git config --list`, which can print a proxy's login header, process or container listings that show arguments or environment (`ps aux`, `ps -ef`, `ps -o args`, `docker top ... args`, `docker inspect`), where run tokens ride, and `set -x` or an `export $(...)` that expands to a bare `export`. A new Ground Rules section in the shared preamble lists these commands and the safe forms, which are to read one value by name (`git config --get <name>`), check a variable is set without printing it, and list processes by name or pid. It also says to wait on CI with one `gstack-pr-wait` call instead of a poll loop or `gh pr checks --watch`, which spends the shared GitHub GraphQL quota. The section is pinned, so no model loadout drops it. diff --git a/changelog.d/plain-pr-bodies.md b/changelog.d/plain-pr-bodies.md deleted file mode 100644 index a73331f9f..000000000 --- a/changelog.d/plain-pr-bodies.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -bump: minor ---- -- **Pull requests from /ship now open with at most three plain lines, and every merge lands as one commit titled with the PR title.** /ship writes `## ` sections and `gstack-post` folds each into a collapsed section when it posts, so a reader sees what changed, why, and what they have to do, with the rest one click away and the Agent notes section last. `gstack-post fold --body-file F` shows the folded body without posting it. In a repository with `.gstack/plain-titles.json`, `gstack-post` and the Plain title check refuse a body that shows more than three lines (or 600 characters) before its first fold, or uses a listed in-house word above the Agent notes section (`PLAIN_BODY_REFUSED`); gstack now ships a default word list (outside voice, pregate, receipt, carry, stamp, train, lane, soak and more), which a policy extends or replaces with `replace_default_terms`, and `body_since` keeps PRs opened before that date to warnings. Every `gstack-ship-queue land` form now passes the subject `<title> (#n)` and a body of the PR's opening lines plus the branch's `Co-authored-by:` and other trailers, so GitHub no longer writes its commit list; `--dry-run` prints it and `--body-file` replaces the opening. Release PRs list one plain line per change with its PR. Two new policy settings: `history: no-rewrite` (squash landings, never a rebased or force-pushed branch) and `land.merge_commit_paths` (keep a merge commit for PRs that change those paths, such as eval results). See docs/writing-prs.md. diff --git a/changelog.d/pr-wait-and-ci-gate-read-every-check-attempt.md b/changelog.d/pr-wait-and-ci-gate-read-every-check-attempt.md deleted file mode 100644 index 1cdff6fea..000000000 --- a/changelog.d/pr-wait-and-ci-gate-read-every-check-attempt.md +++ /dev/null @@ -1 +0,0 @@ -- **`gstack-pr-wait`, `gstack-ci-gate` and the release cut no longer call a commit green while a check on it failed.** They read CI from GitHub's rollup or its default check-runs list, which hold only each check suite's latest attempt, so when a rerun attempt created no jobs the failed run before it disappeared, and a workflow run that failed before creating any job never appeared at all. That is the false green the merge queue had on gstack-evals#303 (#4235). All three now also read CI at the commit through the merge queue's own helper, which reads every attempt and every workflow run, and keep the worse answer per check. So `gstack-pr-wait --until green` blocks with `check failed: checks` and `check failed: sealed (run <id>, no jobs)`, the land gate says `VERDICT FAIL`, and the release cut skips the commit. The release cut now also counts a failed commit status as red, as the merge queue does. A rerun that passed still reads green, and an unreadable read is an error, never a pass (GST-2252). diff --git a/changelog.d/price-runaway-wakes.md b/changelog.d/price-runaway-wakes.md deleted file mode 100644 index d1df30a85..000000000 --- a/changelog.d/price-runaway-wakes.md +++ /dev/null @@ -1 +0,0 @@ -- **`gstack-runaway` now says what a thread's cold and empty wakes cost, not just how many there were (GST-2770).** Each wake it reads is marked cold when it started more than 5 minutes after the thread's last activity, so the prompt cache had lapsed, and is priced in warm model calls, with a cold call at 19.4 (measured on Opus 5.5: $2.42 against $0.125 per million counted tokens; a call after a tool that ran past 5 minutes is cold too). The RUNAWAY line adds `wakes`, `cold_wakes`, `empty_wakes`, `cold_extra_spend` and `empty_spend`, the empty-wake warning gives the empty wakes' share of spend, and `scan --wakes` prints one WAKE line per wake. A message sent to another thread now counts as action even when its state is "working", and more than 5 minutes idle inside a turn (not counting a tool's own run time) starts a new wake. On four coordinating threads over 4.7 days the meter's cold and empty shares came within 1.3 points of a hand count; before this, sends were counted as empty and the empty share read 4 to 12 points high. /runaway-stopper's fetch program now keeps each message's target and each call's run time. diff --git a/changelog.d/remove-test-after-edit-stop-hook-check.md b/changelog.d/remove-test-after-edit-stop-hook-check.md deleted file mode 100644 index 8fa40aaa7..000000000 --- a/changelog.d/remove-test-after-edit-stop-hook-check.md +++ /dev/null @@ -1 +0,0 @@ -- **The optional Stop hook no longer checks for a test run after the last edit, and /verify asks for it in plain words instead.** With /verify loaded, Opus 5.5 and GPT-6.1 Sol already ran a test after their last edit in 96% and 89% of runs, and the hook's check added between -0.5 and +2 points, inside the noise on both models. The hook's other checks (report labels, specialist dispatches, the adversarial pass) are unchanged. If you wrote your own skill whose contract lists `check_after_last_edit`, that contract is now invalid. The hook then checks none of that skill's items and prints one line naming the file until you delete the item. /verify now says to run the full check after your last edit. It also no longer says it never edits CHANGELOG, because a changelog entry the repo expects with a change is part of the change; /ship still folds branch entries into the release entry. diff --git a/changelog.d/ship-step-time-limit-and-remote-file-lists.md b/changelog.d/ship-step-time-limit-and-remote-file-lists.md deleted file mode 100644 index 243013fcb..000000000 --- a/changelog.d/ship-step-time-limit-and-remote-file-lists.md +++ /dev/null @@ -1,2 +0,0 @@ -- **/ship's test step says when it ran out of time, and you can set its budget.** `gstack-ship-preflight` used to stop gstack-pregate after a fixed hour and print `PREGATE: unreadable (exit 143)`, which hid the cause. The whole step's budget is now `--step-timeout <s>` (default 3600 s), and a stop at that limit prints `PREGATE: time limit hit` with the way out: push and let CI's full run carry the tests, or give the step a larger budget. -- **The remote test command no longer puts every file on one command line.** On a small machine with a remote `test_backend`, the command pregate prints for gstack's Ubicloud runner used to carry every selected test file in one argument, which Linux refuses past 128 KiB (`E2BIG`). The files now go to the VM in a setup script, so the command stays short however many files are selected. docs/troubleshooting.md has a new `arg-limit` entry for this kind of failure. diff --git a/changelog.d/stable-cut-ignores-windows-nightly.md b/changelog.d/stable-cut-ignores-windows-nightly.md deleted file mode 100644 index e932e4ee0..000000000 --- a/changelog.d/stable-cut-ignores-windows-nightly.md +++ /dev/null @@ -1 +0,0 @@ -- **The stable release no longer waits on the nightly Windows tests (GST-2815).** The nightly Windows free suite never blocks a pull request, but the daily stable cut still treated its red or unfinished results as a reason to skip a commit, so stable stopped moving while one Windows shard kept timing out. The cut now sets that workflow's check runs aside, matched by workflow file, the same way it ignores its own job. Every other check, the required `free-tests` check and the release-gate evidence still block exactly as before. diff --git a/changelog.d/ubicloud-quota-booking-full-owner.md b/changelog.d/ubicloud-quota-booking-full-owner.md deleted file mode 100644 index c38a11286..000000000 --- a/changelog.d/ubicloud-quota-booking-full-owner.md +++ /dev/null @@ -1 +0,0 @@ -- **Ubicloud fan-outs can no longer overshoot the vCPU limit, and owners no longer delete each other's VMs.** `ubi-runner.sh up` now reads the project's vCPU usage and books its new VM under one lock per machine, so shards started together see each other's bookings instead of all passing on the same reading. A new `ubi-runner.sh room [-s SIZE]` command prints how many more VMs fit now, so a fan-out can size each batch to the room left. When usage can't be read and `UBI_QUOTA_VCPUS` is set, the create is refused (`UBI_QUOTA_UNREADABLE`) instead of going ahead. VM names now carry the whole owner id (up to 31 characters; longer is refused with `UBI_OWNER_TOO_LONG`) instead of its first 12, so two owners sharing a prefix no longer reap each other's expired VMs (#3748); a VM an older runner copy named under the old cut form is still reaped by the long owner on the machine whose VM record lists it, and `recover` accepts the longer owner ids. Merge checks and canaries set no `UBI_QUOTA_VCPUS` and behave as before. diff --git a/changelog.d/update-hono-security-advisories.md b/changelog.d/update-hono-security-advisories.md deleted file mode 100644 index a360fb109..000000000 --- a/changelog.d/update-hono-security-advisories.md +++ /dev/null @@ -1,7 +0,0 @@ ---- -important: security ---- -**hono is pinned to 4.13.16, which fixes five advisories the weekly OSV scan found in 4.13.13 (gstack#4406).** The 2026-11-16 scan (run 43825448246) flagged GHSA-3gg2-335m-rxc6 (CVSS 3.4, the `css` helper) and GHSA-46wx-v44x-9hh2, GHSA-4h7h-rv7c-7m2v, GHSA-4m76-xfj5-7975 and GHSA-h7mg-cq63-vhpq (CVSS 5.3 each: `csrf()`, `validator('form')`, `mount()` and RegExpRouter wildcard middleware). hono reaches gstack only through the dev and test toolchain (`@anthropic-ai/claude-agent-sdk` pulls in `@modelcontextprotocol/sdk`), gstack's own code never imports it or starts a hono server, and none of the affected helpers is used, so no advisory was reachable. Contributor checkouts get the patched version on their next `bun install`. - -- **`package.json` overrides**: `hono` moves from 4.13.13 to 4.13.16, the release that fixes all five. `bun.lock` holds the one copy. -- `test/dependency-security.test.ts`: the lockfile floor check now covers `hono >= 4.13.16`. It fails on main. diff --git a/changelog.d/windows-nightly-fits-its-time-limit.md b/changelog.d/windows-nightly-fits-its-time-limit.md deleted file mode 100644 index d54095383..000000000 --- a/changelog.d/windows-nightly-fits-its-time-limit.md +++ /dev/null @@ -1 +0,0 @@ -- **The nightly Windows test run finishes again.** Since 2026-11-06 one of its six shards ran past the 15-minute job limit almost every night, because 95 newer test files had no recorded Windows duration and the planner packed them blind, so one shard got 26 minutes of work while another got 12. The Windows durations are refreshed to a recording that covers all 902 files, and the run now uses eight shards of about 12 minutes each, so the suite fits with room to grow. No test was dropped and the time limit is unchanged. diff --git a/changelog.d/write-up-checklist.md b/changelog.d/write-up-checklist.md deleted file mode 100644 index 230ebf5b0..000000000 --- a/changelog.d/write-up-checklist.md +++ /dev/null @@ -1 +0,0 @@ -- **A short write-up checklist in ETHOS.md.** Explain the Number now ends with three checks before a post or write-up goes out: its words match the page it links, every number, "green" and "ready" names its source and is read at the current head, and a fresh model reads it cold once. The READY template in `captain/templates/READY.md` points to the second check. Anything carried over from an earlier head goes under `Unverified:` until it is re-read. diff --git a/package.json b/package.json index 306cd7dda..2bd4a94b6 100644 --- a/package.json +++ b/package.json @@ -1,6 +1,6 @@ { "name": "gstack", - "version": "1.99.5", + "version": "1.100.0", "description": "Garry's Stack — Claude Code skills + fast headless browser. One repo, one install, entire AI engineering workflow.", "license": "MIT", "type": "module",