diff --git a/CHANGELOG.md b/CHANGELOG.md index 19ce4d615..4b9ba11e7 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,95 @@ # Changelog +## [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. + +- **An unattended /autoplan run on Claude Code no longer says the publication guard is off while it is on.** `gstack-skill-start` printed `autoplan guard: not enforced by this host ... (GUARD_NOT_INSTALLED)` for every unattended session, but Claude Code runs /autoplan's phase hook whatever the session kind, so runs were blocked by a guard their own record said was off. The line now follows the host. On Claude Code with the hook installed it reads `... (GUARD_ENFORCED)`, elsewhere it keeps `GUARD_NOT_INSTALLED`, and `gstack-artifact validate` accepts either in an unattended run's record. +- **A phase report written only in reasoning now gets a denial that says so and names the fix.** When a `true autoplan-published ` message carries no visible text, the guard denies with the new `publication_not_visible` code instead of the generic `publication_missing`, and every publication denial names the exact command (for example `true autoplan-published dx`) and says reasoning does not count. A saved report file still does not count as publication. + +- **`gstack-autoplan revise` no longer refuses a plan that the first `next` accepted.** A plan with no `## Implementation plan` or `## Review record` heading is reviewed whole by `next`, but once the owner added one of those headings while editing (for example a `## Review record` noting the accepted changes), `revise` stopped with `PLAN_SECTIONS_INVALID` after a full paid review round, and its fix pointed at a `--scaffold` flag the first `next` never asked for. `revise` now adds the missing heading to its own copy of the plan and prints `SCAFFOLDED:`, the way `next --scaffold` does; the owner's file is never edited. A run whose first plan already had both headings still refuses a revision that drops one, and now says why. +- **`revise --plan` with a different file now says the run follows that file.** It prints `PLAN_INPUT: -> ` and records the move, so drift checks no longer switch to another file silently. + +- **A second-model plan review in `/autoplan` is no longer reported as unavailable when it actually ran (GST-3051).** `gstack-outside-voice run` judged autoplan's reviews by `/review`'s rule, which wants a `Recommendation: ... because ...` line autoplan never asks for, so a finished review printed `VERDICT: unavailable` (`missing_markers`) and sessions paid to re-run it or dropped it. A prompt that carries autoplan's result format is now checked against that format (its `INPUT:` line and `gstack-findings` fence; `--gate autoplan` asks for it explicitly), and a reply without them, or an empty one, is still unavailable. `gstack-outside-voice --help` also no longer prints a line that starts with `OUTSIDE_STATUS:`, which logs and models read as a real consent refusal. + +The /ship conductor trial plan now lets case reverts that conflict with later code be resolved by hand in docs, tests and the planted file, so large pull requests can be cases (GST-3003). + +- **A locked plan for a trial of /ship without its separate conductor subagent (GST-3003).** `docs/evals/ship-conductor-fold-prereg.md` fixes, before any paid run, the bars, arms (today's /ship, the folded /ship and a no-gstack review), cases, scoring and stop rules for testing whether a pull request's own thread can run /ship's phases and start its reviewers itself, at no extra cost and with nothing missed that today's /ship catches. + +- **A driven /ship now runs each phase as one command, counts its calls, and stops itself at 100 minutes.** `gstack-ship-phase run` runs a ship/DRIVER.md phase's commands in order, prints only their result lines, returns after 200 s while slow work keeps running, and on a rerun skips what already finished; on Capy, `gstack-ship-phase program` serves a checked-in program that also runs the review pass or the closing budget note in the same call. `gstack-ship-budget note` records the calls since the previous note, and `gstack-ship-receipt write --budget-record ` puts the run's total in the receipt, so `SHIP_RECEIPT:` shows `calls=` (older receipts without it still validate and carry). Past 100 minutes, the note or the phase runner prints `STOP: past 100 minutes` and writes the head, review passes and next act to `stop.json`; `gstack-ship-budget resume` restarts the clock. The driver also prices waits by length. It keeps a 200 s keep-alive under about 80 minutes, uses one wait for anything longer, and never polls every 10-15 minutes. Measured conductors made 121 to 308 calls against the driver's 9 to 12 and 10 of 12 ran past the 100-minute line while these limits were prose. + +**/ship stops asking owners to clear two kinds of bookkeeping noise: base-side merge text and a reviewer's confirmation prose.** + +### Fixed +- **A base merge that only touched base-side text no longer stops `/ship` finish for an owner's grant.** After a main merge, finish counted two base-side differences as "behavior, tests or build inputs changed since review": the shared-preamble hash id in every regenerated skill file the branch also touched (#4383, 94 files), and an expected-value list where main's new entry was spliced next to the branch's own (ghome#16). The path comparison now reads both sides with the shared-preamble ids masked (a file whose only branch change is that id compares as unchanged), and a line that is the base's new line plus exactly the tokens the branch added to the reviewed line counts as base context, so it goes to the one scoped reader instead. A resolution that drops, invents or rewrites anything on the branch's own lines still stops as a behavior change. + +- **A reviewer's confirmation prose no longer becomes a Critical finding that only an owner can clear.** The collector read prose that named a file and line while confirming code, such as "a High or Critical challenge…", as an anchored Critical finding that judgment could not lower, even when the reply's own `REVIEW_VERDICT` said `blocking=0` or it said `NO FINDINGS`, so each one cost an owner ruling and a stopped `/ship` (twice on #4431). A file:line in prose, or a list item or the native voice's FIXABLE tag in a reply that says `NO FINDINGS`, no longer anchors a finding, so under a `blocking=0` verdict such a row is a summary candidate the lead can lower with a reason. It still blocks until then, so prose that names a real blocker ("no Critical issues except the token leak in log.ts:3") is never dropped. Structured findings (JSON, a lead severity label, a label heading, a numbered finding heading, a `Severity` or `Blocks: yes` field, or the native voice's FIXABLE/INVESTIGATE tag) still collect as anchored and still disagree with the verdict, and a reply whose verdict declares a blocker is read as before. + +- **A CI run that passed only after its failed jobs were re-run is now marked wherever gstack reuses it.** When a run's `tree.json` lists `rerun_files` and `attempts` (gbrain writes these since gbrain#9690), the `CI_CARRY:` line, the `SHIP_RECEIPT:` line's `ci_carried=` field, and the receipt line `gstack-ship-queue ready` and `land` print now add `rerun= attempts=`, so a pass on a re-run reads as flaky instead of as a clean first-attempt pass. The run is still accepted. A `tree.json` without those fields prints exactly as before, and a malformed `rerun_files` or `attempts` refuses that run with `CI_CARRY_RERUN_MALFORMED`. + +- **/ship reuses CI's passing test run when merging the base changed no file, and stops instead of quietly starting hour-long test runs or reviewing an empty diff (GST-3019).** When git reads a commit's file tree as identical to an earlier commit's, a passed CI run of the earlier commit now counts for it, under the same workflow, command and freshness rules, and the output names the commit and run it used. `gstack-ship-preflight --tests-only` now stops with `SHIP_LANE_NOT_CARRIED` and the CI command to run when a mapped test suite has no passing CI run to reuse, instead of running it on this machine (`--run-uncarried-here` still runs it). `--audit --base ` now diffs against that commit, and a base that names no commit stops with `SHIP_BASE_UNKNOWN` instead of reporting zero changed files. + +- **/ship on Codex no longer stops at the test coverage audit.** The Codex version of /ship told Codex to start its audit with Claude Code's Agent tool, `subagent_type` and `run_in_background`, which Codex doesn't have, so the run stopped before it could finish (#4476). Codex skills now start every audit and reviewer subagent with Codex's own `spawn_agent` and wait on `wait_agent` for its final result: /ship's coverage, plan and docs audits, the native review voices in /review, /office-hours and the design reviews, the plan-CEO spec review, /autoplan's phases and /context-restore. Turn on `multi_agent = true` under `[features]` in `~/.codex/config.toml`; without it the dispatch fails and the step takes its failure path. The plan reviews' timed fallback reviewer needs a read-only subagent type Codex doesn't offer, so on Codex it reports the second-model review as unavailable instead of starting one. Claude Code's and every other host's instructions are unchanged. + +- **A cold first-use check before an app build reaches its owner (GST-3025).** New `gstack-first-use` freezes the first five day-one actions from the plan and the exact build, has an agent from another model family play that build on the same kind of device with a screenshot or video for each action, checks every link within 15 minutes of sending, and prints the "try this build" message only when all of it passes. The owner's score and words are kept beside the checker's predicted score in the owner's private drive. See `docs/first-use-check.md`. + +- **`skill_prefix=true` now survives updates when gbrain is installed.** Every Claude re-render (the brain-aware render in `./setup`, `gstack-config gbrain-refresh`, and the pinned-overlay render) applies your saved prefix before it goes live, so skills stay `/gstack-retro`, `/gstack-review` and so on instead of reverting to short names after each update (#4413). Thanks @fliaron for the report. +- The OpenClaw docs no longer send you to ClawHub, which held only April 2026 builds of office-hours, ceo-review and retro and no longer has gstack skills (the docs also said four skills while listing three). `docs/hosts.md` and `docs/OPENCLAW.md` now point to native registration (`./setup --host openclaw`), which already serves the current version of each, and tell you to remove the ClawHub copies if you installed them (#4421). Thanks @exGeni for the report. +- **`gstack-issue-links` no longer reads other repositories' issues as gstack's.** A parenthesis that cites several repositories, such as `(founderstar #837, harness-evals #393)`, now gives each its own cross-repo reference instead of `Refs #837` and `Refs #393` on this repository, and `owner/repo#N` naming the current repository counts as its own issue, so `Fixes garrytan/gstack#N` closes it. + +- **/ship's pre-gate finds the tests of a monorepo app that imports through its own `@/` alias.** The import graph read `paths` and `baseUrl` from the repository root's tsconfig only, so in an app like `apps/mobile` whose `tsconfig.json` maps `@/*` to `./src/*`, every touched source file failed with `PREGATE_NO_LANE` even with a test right beside it. Each file now takes its aliases from the nearest tsconfig.json or jsconfig.json above it that declares `paths` or `baseUrl`; a nested config that declares neither defers to the next one up, the root config governs the rest, and an `extends` outside the repository is still never followed. Thanks to @dtoelke for the report and the original fix (#4462). + +- **/office-hours explains why it recommends a mode.** The first question (what's your goal?) now recommends a mode by quoting your own words and saying what another mode would skip, instead of restating the mode's name. + +- **/office-hours keeps its 34 founder videos and essays out of the skill file.** The list now lives in `lib/founder-resources.ts`, and the closing runs the new `gstack-founder-resources` command, which prints the ones a user hasn't been shown yet. The resources, the dedup and the opt-out work as before; the skill's own text is about 8.6 KB smaller. + +- **Two `gstack-owner-answers add` runs for the same answer no longer collide.** When two processes recorded the same owner answer at once, about one pair in ten failed: one could read the other's row while it was still empty and report a conflict, or see the record's folder appear between two checks and refuse it as "a link that cannot be resolved". The second add now waits up to 2 seconds for the first to finish writing, and the folder check looks for the folder before resolving it. Three tests that failed intermittently under load also got fixes. The sealed-guard streaming test now treats Bun's buffer-overflow error as a failure even when git exits 0, and the two `gstack-ship-receipt` test files use the same 30-second limit as CI. + +- **A Captain can check an ask against the owner's earlier answers before sending it.** `gstack-owner-answers` keeps the owner's answers and standing yeses in one record, outside every repository: on Capy it is the owner's private Drive folder, which every thread the owner runs can read. `screen --ask-file` reads a Captain's draft ask and proposes leaving out only the lines that a verified, current standing yes in the owner's own words covers exactly. Every other line stays in, with any related answers shown beside it. That includes autoplan gate items, untyped or compound lines, and anything that always needs a fresh ask (product decisions, spend over a cap, access, publishing, thresholds, closing a contribution, destructive steps). A damaged, conflicting or unreadable record keeps every ask. Rows are one file each, written once; a row needs a second actor to verify its words before it can be taken, and the recorder can't verify its own. The file format and the two-week pilot are in docs/owner-answers.md. + +- **`gstack-post` stops reading /ship's own receipt identifiers as phone numbers, so posting a /ship PR body no longer needs a confirm.** The receipt's session run id (`--<8 hex>`, read as an E.164 number when the pid has seven digits, #4480) and all-digit 12-character commit prefixes in its `pregate` and `requires_remote` values used to stop `gstack-post pr-body` with a MEDIUM `pii.phone.e164` finding. The same went for the `SHIP_RECEIPT:` and `CI_CARRY:` lines /ship prints (`heads=`, `carried_from=`, `base_red=`, `ci_carried=:/`). Only those exact values are exempt, in a receipt block the receipt writer rendered and on lines outside code blocks that open with `SHIP_RECEIPT:`, `CI_CARRY:`, `PREGATE:` or `READY_RECEIPT:`. A phone number in prose, in any other code block or field (a receipt line pasted inside one included), or in a hand-written receipt block still needs a confirm. + +- **GitHub release notes now list every change in plain words, with a link to its pull request.** Releases v1.99.1.0 to v1.99.4.0 said only "see CHANGELOG.md", because the release cut took one headline per version and found none in entries written as bullets. Each release page now lists every change the release ships under What's new, Fixes and Changes for maintainers, one line each, with the pull request that made it, then the compare link. A release whose CHANGELOG has no entry for its version, or an entry with nothing in it, now stops with `NOTES_MISSING` instead of publishing an empty page, and notes too long for GitHub keep the changes that fit and say how many more are in the full changelog (docs/releases.md#release-notes). + +- **gstack's router no longer gives bug reports a special rule, so the model fixes a bug the way it would without gstack.** The rule said to check the cause once and write a failing test first when the cause was unclear. On 26 real gbrain bugs that plain Opus 5.5 had failed once, gstack with the rule fixed 3 and plain Opus fixed 2: +3.8 points, with a 95% interval of -11.5 to +19.2, so no clear gain. The router's description no longer lists debugging either, so a bug report doesn't load it. Planning, review, /verify, /qa and /ship are unchanged, and the bug line gstack offers for your CLAUDE.md is unchanged too. The two paid evals of the removed rule are retired with it. + +- **/ship's review plan now sorts Swift and Xcode files by risk instead of sending every iOS PR to the full review panel (GST-3041).** `gstack-diff-scope` had no rule for `.swift`, so an all-Swift diff printed `SCOPE_ERROR=unmatched` and the planner failed closed to five readers plus a coverage audit; a 55-line SwiftUI copy change got the whole panel. Swift source now reads as UI and app code (like Dart), as do storyboards, XIBs, string catalogs and asset catalogs. `Info.plist`, entitlements and the privacy manifest read as permissions (auth), and `Package.swift`, `Package.resolved`, `Podfile`(`.lock`), `project.pbxproj`, `.xcconfig` and schemes read as config, never as UI. Swift networking code (`URLSession`, `URLRequest`, `NWConnection`, Alamofire) still raises the review to full depth, the same as `fetch` in TypeScript. A file type no rule knows, such as a Metal shader, still fails closed. Reported from the GHome iOS app's PR review. + +- **/ship's review plan now sends Swift code that touches private user data to the full review panel (GST-3052).** Once `.swift` stopped failing closed, an iOS change that asked for microphone or speech permission, read HealthKit or calendar data, or wrote and deleted users' files on the device planned the light review. Swift lines that request a privacy permission (speech, microphone, HealthKit, calendars, location, camera, photos, tracking) now count as security, and lines that read HealthKit, EventKit or contacts, write, delete or protect files on the device, use an app group container, exclude files from backup, or declare how long data is kept count as user data. Plain SwiftUI views, UserDefaults settings and file reads stay light. Reported from the GHome iOS app's PR review. + +- **A round grant at /ship's review round cap now buys the fix round it names, in the same run or the next (GST-2968).** When a review stops at the cap, /ship records the stop in the PR body (`gstack-ship-receipt stop`), and a `gstack-approvals round-grant --kind fix-check` naming that stop's head, run and findings buys one scoped fix-check of the fix instead of another read-only full review. Before, the read-only review had already counted as the extra round, so the grant bought nothing, and a grant recorded between /ship runs was ignored, so the next run paid for another full review. A grant is spent by the one run that claims it (`gstack-ship-receipt claim`), and an extra-round grant recorded before review keeps its fix round, also across preflight's base merge. + +- **The secrets check stops flagging three kinds of ordinary code as a secret, and catches two misses.** TypeScript types with spaced generic arguments, Python keyword arguments that read a settings constant (`aws_secret_access_key=settings.S3_SECRET_KEY,`), and camelCase dotted field paths (`"form.section.fieldName"`) no longer stop a push. A secret built with `+` from short pieces, or written in a trailing or whole-line comment, is now reported. A secret in any of those positions still fails the check. Template keys, arrays of names and string-literal generic types still report and are documented as known limits. (#4289) + +- **A security finding's location and reason no longer go into a PR body.** When an owner ruled a security review finding a note, or a review deferred one in code the PR did not change, the PR body, its receipt block and the `Ruled note:`, `Gate disposition:` and deferred-findings lines named the file, line and reason, which on a public repository points at a live exposure. Findings whose category is `security`, `auth`, `secret` or `red-team` now print an opaque id and "security finding, details recorded privately"; the details stay in the private approval record (`gstack-approvals list`) and the local review files. `gstack-ship-receipt write` and `gstack-post` also refuse text that names a raw `path:line:security` fingerprint. Other findings print exactly as before. + +- **The token-capped /ship child on Claude Code now starts its prompt with the typed `/ship` command, so the skill actually loads.** When /ship hands its remaining steps to a fresh child (`gstack-ship-budget launch`), the child's prompt used to open with "Load gstack. … Run /ship --continue ". Claude Code expands a skill only when the prompt starts with the typed command, so the child got plain text and had to find /ship on its own, which unattended Opus 5.5 did in 0 of 60 sessions in gstack-evals#366. The prompt now opens with `/ship --continue ` (`/gstack-ship` when `skill_prefix` is on). docs/unattended.md gains the rule for any parent that starts a child agent, and the Capy skill's Claude Code line now says the same. + +- **A /ship receipt now counts the second-model review's blocking findings, and refuses to be written while one is still open.** When /ship stopped after Step 11 because the second-model review (or the native adversarial pass) raised a `Blocks: yes` finding, no review-log row recorded it, so the receipt said `blocking_found=0` and the PR looked clean to `ready` and the merge step. Step 11's own row now carries `via: ship` and its findings, `gstack-review-scope summary` counts them (they are never a pass or a round), and `gstack-ship-receipt write` refuses with `RECEIPT_BLOCKER_OPEN` while any blocker's newest action is still `open` and no owner ruling closes it. A finding below the floor still counts only as a note. + +- **A `/ship` with a token cap on Claude Code no longer stops at its first handoff.** When the budget handed steps to a fresh child, `gstack-ship-budget launch` started `claude -p` with `--allowedTools ` right before the prompt. Claude Code reads every argument after that option as one more tool name, so the child got no prompt, exited with "Input must be provided either through stdin or as a prompt argument", and the run stopped. The launch now ends the options with `--` before the prompt, which Claude Code 2.1.292 and 2.1.326 both read as the prompt. Thanks to @exGeni for the report and diagnosis (#4448). + +- **/ship's push step now always ends with one line that says whether it pushed (GST-3053).** `gstack-ship-release push` (Step 17) finishes with exactly one of `PUSH: pushed remote=origin ref= head=`, `PUSH: already_pushed remote=origin ref= head=` or `PUSH: failed reason=PUSH_REFUSED|PUSH_REJECTED|PUSH_GIT_FAILED`, so a script or eval grader can tell that /ship pushed by matching that line instead of looking for a `git push` call. Exit status is unchanged. Each reason code has a troubleshooting entry, and the format is documented in docs/unattended.md#push-line. An eval had graded 48 real /ship pushes as "no push" because it searched for the git command. The line's shape changed. The old `PUSH: ALREADY_PUSHED ` and `PUSH: pushed to origin/` lines are gone, so update anything that matched them. + +- **A PR's review-round count can no longer be overwritten by the body `/ship` posts.** When `/ship` posted a PR body whose Agent notes held their own `gstack-receipt-runs` blocks, gstack removed only the first one, so a second block placed before the receipt became the history `gstack-ship-receipt history` counted (one round where there were three). A post now drops every runs block it brings and keeps exactly one. That one is the live body's history, placed right after the posted receipt, or none when the PR had no earlier runs. A post whose notes hold no receipt yet keeps the live started stubs in place of the runs block it brings. See docs/review-policy.md#rounds-per-pull-request. + +- **One queued record with a malformed path no longer stops artifacts sync for every project.** A skill started where the project slug was empty (for example from `/`) queued `projects//timeline.jsonl`; git staged it under another spelling, the drain refused the whole batch, and every later record waited until the queue was dropped by hand. `gstack-brain-enqueue` now refuses a path with an empty, `.` or `..` component, and the drain drops such a record, syncs the rest, and reports it (#4445). +- **A failed artifacts push stays reported until a push succeeds.** A later drain with nothing new to commit used to write `idle` over `push_failed` while the commits were still unpushed, and /sync-gbrain printed "curated artifacts pushed" because `gstack-brain-sync` exits 0 on a failed push. The status now stays `push_failed`, and the brain-sync row reads the drain's status. It says "pushed" only after a push, "nothing new to push" when there was nothing to send, and `ERR artifacts not pushed (...)` otherwise (#4418). +- **The /sync-gbrain memory row no longer says OK when pages failed to import.** It read only the ingest's exit code, which stays 0 for per-page failures, so it printed `OK ... 126 failed`. The row now reads the ingest's counts: `WARN` when some pages failed, `ERR` when pages failed and none imported, with the failed count and the first cause. A transcript that cannot be parsed is listed as a note and leaves the row `OK` (#4417). + +### Tests +- A policy-floor test no longer fails once gstack's own version reaches the minor after the newest policy key; it now checks a floor above both, so the v1.101.0.0 release can pass CI. + +- CI now uploads free-suite test results to Trunk with analytics-uploader v2.1.8 and analytics-cli 0.17.0, up from v2.1.6 and 0.16.2. + +- **Packing a checkout for a Ubicloud run no longer fails when git beside it is rewriting a ref.** `ubi-runner.sh pack` asked git's upload-pack for every ref, and upload-pack advertises a ref it cannot read (a loose ref file caught mid-write or mid-delete) with a null id, so the snapshot failed with "not our ref 0000…". Pack now reads the refs once with `git for-each-ref`, which skips an unreadable ref, and fetches those exact objects. Seen twice in Free Tests on 2026-11-17 (runs 44063454800 and 44050883970); an empty loose ref file reproduces it every time. +- **The /plan-ceo-review HOLD SCOPE eval no longer fails a review because its follow-up question didn't repeat the mode name.** The check required the decision's "Project/branch/task" line to say "HOLD SCOPE", which the skill never asks for; the decision is already tied to the HOLD answer by session and order. A line naming another mode, or past or hypothetical work, still fails. + +- **Unattended /autoplan now records the owner's relayed gate reply instead of refusing it.** `autoplan/SKILL.md` told an unattended session to end at the final gate and never approve, while `autoplan/DRIVER.md` told it to record a relayed owner reply with `gstack-autoplan answer --by agent: --owner-reply`, so whether a run reached `complete` depended on which file the model followed. Both now render one rule from `scripts/resolvers/autoplan-driver.ts`: never approve on your own judgment, stop at `gate_pending` when there is no owner reply, and record the owner's own relayed words with `--owner-reply-file`, never handing the command back to the owner. `gstack-skill-start`'s unattended block names the same exception for /autoplan. + +- **Unattended /autoplan now runs the second-model review you turned on, and waits for it instead of ending the session.** With `codex_reviews: enabled` in the config, Opus 5.5 still skipped the second reviewer in 2 of 3 unattended `--full` runs, saying the owner had not agreed to send the plan off the machine; in the third it started the review in the background and ended its turn, which ends a headless session. `gstack-autoplan next` now prints `OUTSIDE_CONSENT: given by codex_reviews=enabled` beside the outside attempt, and autoplan/DRIVER.md says that an explicit `enabled` is the owner's standing consent, that the review runs in the foreground with a 9-minute limit, and that a timeout or failure is recorded with the standard failure line while the run carries on. `disabled` is still enforced by the runner, and an unset `codex_reviews` is handled as before. +- **Two /autoplan runs started in the same second no longer share a run id.** Run ids were the run directory's name plus a time stamp to the second, so two runs whose directories were both named `autoplan` got the same id and the same checkpoint folder. Each id now carries six random letters before the stamp. + ## [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. diff --git a/VERSION b/VERSION index f173507db..78df83493 100644 --- a/VERSION +++ b/VERSION @@ -1 +1 @@ -1.100.0.0 +1.101.0.0 diff --git a/agents-digest/gstack-AGENTS.md b/agents-digest/gstack-AGENTS.md index 17c59a82d..403067e4a 100644 --- a/agents-digest/gstack-AGENTS.md +++ b/agents-digest/gstack-AGENTS.md @@ -1,4 +1,4 @@ -# gstack digest v1.100.0.0 — regenerate/re-copy after upgrading gstack +# gstack digest v1.101.0.0 — 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/approvals-mirror-stays-inside-fixtures.md b/changelog.d/approvals-mirror-stays-inside-fixtures.md deleted file mode 100644 index 33f60aeba..000000000 --- a/changelog.d/approvals-mirror-stays-inside-fixtures.md +++ /dev/null @@ -1 +0,0 @@ -- **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/changelog.d/autoplan-guard-says-when-it-is-on.md b/changelog.d/autoplan-guard-says-when-it-is-on.md deleted file mode 100644 index 61c9503de..000000000 --- a/changelog.d/autoplan-guard-says-when-it-is-on.md +++ /dev/null @@ -1,2 +0,0 @@ -- **An unattended /autoplan run on Claude Code no longer says the publication guard is off while it is on.** `gstack-skill-start` printed `autoplan guard: not enforced by this host ... (GUARD_NOT_INSTALLED)` for every unattended session, but Claude Code runs /autoplan's phase hook whatever the session kind, so runs were blocked by a guard their own record said was off. The line now follows the host. On Claude Code with the hook installed it reads `... (GUARD_ENFORCED)`, elsewhere it keeps `GUARD_NOT_INSTALLED`, and `gstack-artifact validate` accepts either in an unattended run's record. -- **A phase report written only in reasoning now gets a denial that says so and names the fix.** When a `true autoplan-published ` message carries no visible text, the guard denies with the new `publication_not_visible` code instead of the generic `publication_missing`, and every publication denial names the exact command (for example `true autoplan-published dx`) and says reasoning does not count. A saved report file still does not count as publication. diff --git a/changelog.d/autoplan-revise-accepts-heading-less-plans.md b/changelog.d/autoplan-revise-accepts-heading-less-plans.md deleted file mode 100644 index 12decab39..000000000 --- a/changelog.d/autoplan-revise-accepts-heading-less-plans.md +++ /dev/null @@ -1,2 +0,0 @@ -- **`gstack-autoplan revise` no longer refuses a plan that the first `next` accepted.** A plan with no `## Implementation plan` or `## Review record` heading is reviewed whole by `next`, but once the owner added one of those headings while editing (for example a `## Review record` noting the accepted changes), `revise` stopped with `PLAN_SECTIONS_INVALID` after a full paid review round, and its fix pointed at a `--scaffold` flag the first `next` never asked for. `revise` now adds the missing heading to its own copy of the plan and prints `SCAFFOLDED:`, the way `next --scaffold` does; the owner's file is never edited. A run whose first plan already had both headings still refuses a revision that drops one, and now says why. -- **`revise --plan` with a different file now says the run follows that file.** It prints `PLAN_INPUT: -> ` and records the move, so drift checks no longer switch to another file silently. diff --git a/changelog.d/capy-autoplan-second-model-review-not-unavailable.md b/changelog.d/capy-autoplan-second-model-review-not-unavailable.md deleted file mode 100644 index 3bc3e442b..000000000 --- a/changelog.d/capy-autoplan-second-model-review-not-unavailable.md +++ /dev/null @@ -1 +0,0 @@ -- **A second-model plan review in `/autoplan` is no longer reported as unavailable when it actually ran (GST-3051).** `gstack-outside-voice run` judged autoplan's reviews by `/review`'s rule, which wants a `Recommendation: ... because ...` line autoplan never asks for, so a finished review printed `VERDICT: unavailable` (`missing_markers`) and sessions paid to re-run it or dropped it. A prompt that carries autoplan's result format is now checked against that format (its `INPUT:` line and `gstack-findings` fence; `--gate autoplan` asks for it explicitly), and a reply without them, or an empty one, is still unavailable. `gstack-outside-voice --help` also no longer prints a line that starts with `OUTSIDE_STATUS:`, which logs and models read as a real consent refusal. diff --git a/changelog.d/capy-g1-cases-amend.md b/changelog.d/capy-g1-cases-amend.md deleted file mode 100644 index 86151c378..000000000 --- a/changelog.d/capy-g1-cases-amend.md +++ /dev/null @@ -1 +0,0 @@ -The /ship conductor trial plan now lets case reverts that conflict with later code be resolved by hand in docs, tests and the planted file, so large pull requests can be cases (GST-3003). diff --git a/changelog.d/capy-plan-a-trial-of.md b/changelog.d/capy-plan-a-trial-of.md deleted file mode 100644 index 0f4dad300..000000000 --- a/changelog.d/capy-plan-a-trial-of.md +++ /dev/null @@ -1 +0,0 @@ -- **A locked plan for a trial of /ship without its separate conductor subagent (GST-3003).** `docs/evals/ship-conductor-fold-prereg.md` fixes, before any paid run, the bars, arms (today's /ship, the folded /ship and a no-gstack review), cases, scoring and stop rules for testing whether a pull request's own thread can run /ship's phases and start its reviewers itself, at no extra cost and with nothing missed that today's /ship catches. diff --git a/changelog.d/capy-run-each-driven-ship.md b/changelog.d/capy-run-each-driven-ship.md deleted file mode 100644 index 66c3a5fa4..000000000 --- a/changelog.d/capy-run-each-driven-ship.md +++ /dev/null @@ -1 +0,0 @@ -- **A driven /ship now runs each phase as one command, counts its calls, and stops itself at 100 minutes.** `gstack-ship-phase run` runs a ship/DRIVER.md phase's commands in order, prints only their result lines, returns after 200 s while slow work keeps running, and on a rerun skips what already finished; on Capy, `gstack-ship-phase program` serves a checked-in program that also runs the review pass or the closing budget note in the same call. `gstack-ship-budget note` records the calls since the previous note, and `gstack-ship-receipt write --budget-record ` puts the run's total in the receipt, so `SHIP_RECEIPT:` shows `calls=` (older receipts without it still validate and carry). Past 100 minutes, the note or the phase runner prints `STOP: past 100 minutes` and writes the head, review passes and next act to `stop.json`; `gstack-ship-budget resume` restarts the clock. The driver also prices waits by length. It keeps a 200 s keep-alive under about 80 minutes, uses one wait for anything longer, and never polls every 10-15 minutes. Measured conductors made 121 to 308 calls against the driver's 9 to 12 and 10 of 12 ran past the 100-minute line while these limits were prose. diff --git a/changelog.d/capy-ship-base-side-and-confirmation-prose.md b/changelog.d/capy-ship-base-side-and-confirmation-prose.md deleted file mode 100644 index 24dcbbb77..000000000 --- a/changelog.d/capy-ship-base-side-and-confirmation-prose.md +++ /dev/null @@ -1,9 +0,0 @@ ---- -important: blocker ---- -**/ship stops asking owners to clear two kinds of bookkeeping noise: base-side merge text and a reviewer's confirmation prose.** - -### Fixed -- **A base merge that only touched base-side text no longer stops `/ship` finish for an owner's grant.** After a main merge, finish counted two base-side differences as "behavior, tests or build inputs changed since review": the shared-preamble hash id in every regenerated skill file the branch also touched (#4383, 94 files), and an expected-value list where main's new entry was spliced next to the branch's own (ghome#16). The path comparison now reads both sides with the shared-preamble ids masked (a file whose only branch change is that id compares as unchanged), and a line that is the base's new line plus exactly the tokens the branch added to the reviewed line counts as base context, so it goes to the one scoped reader instead. A resolution that drops, invents or rewrites anything on the branch's own lines still stops as a behavior change. - -- **A reviewer's confirmation prose no longer becomes a Critical finding that only an owner can clear.** The collector read prose that named a file and line while confirming code, such as "a High or Critical challenge…", as an anchored Critical finding that judgment could not lower, even when the reply's own `REVIEW_VERDICT` said `blocking=0` or it said `NO FINDINGS`, so each one cost an owner ruling and a stopped `/ship` (twice on #4431). A file:line in prose, or a list item or the native voice's FIXABLE tag in a reply that says `NO FINDINGS`, no longer anchors a finding, so under a `blocking=0` verdict such a row is a summary candidate the lead can lower with a reason. It still blocks until then, so prose that names a real blocker ("no Critical issues except the token leak in log.ts:3") is never dropped. Structured findings (JSON, a lead severity label, a label heading, a numbered finding heading, a `Severity` or `Blocks: yes` field, or the native voice's FIXABLE/INVESTIGATE tag) still collect as anchored and still disagree with the verdict, and a reply whose verdict declares a blocker is read as before. diff --git a/changelog.d/ci-carry-marks-rerun-passes.md b/changelog.d/ci-carry-marks-rerun-passes.md deleted file mode 100644 index 0f4d5bff0..000000000 --- a/changelog.d/ci-carry-marks-rerun-passes.md +++ /dev/null @@ -1 +0,0 @@ -- **A CI run that passed only after its failed jobs were re-run is now marked wherever gstack reuses it.** When a run's `tree.json` lists `rerun_files` and `attempts` (gbrain writes these since gbrain#9690), the `CI_CARRY:` line, the `SHIP_RECEIPT:` line's `ci_carried=` field, and the receipt line `gstack-ship-queue ready` and `land` print now add `rerun= attempts=`, so a pass on a re-run reads as flaky instead of as a clean first-attempt pass. The run is still accepted. A `tree.json` without those fields prints exactly as before, and a malformed `rerun_files` or `attempts` refuses that run with `CI_CARRY_RERUN_MALFORMED`. diff --git a/changelog.d/ci-same-tree-reuse-and-preflight-stops.md b/changelog.d/ci-same-tree-reuse-and-preflight-stops.md deleted file mode 100644 index 7814a1852..000000000 --- a/changelog.d/ci-same-tree-reuse-and-preflight-stops.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship reuses CI's passing test run when merging the base changed no file, and stops instead of quietly starting hour-long test runs or reviewing an empty diff (GST-3019).** When git reads a commit's file tree as identical to an earlier commit's, a passed CI run of the earlier commit now counts for it, under the same workflow, command and freshness rules, and the output names the commit and run it used. `gstack-ship-preflight --tests-only` now stops with `SHIP_LANE_NOT_CARRIED` and the CI command to run when a mapped test suite has no passing CI run to reuse, instead of running it on this machine (`--run-uncarried-here` still runs it). `--audit --base ` now diffs against that commit, and a base that names no commit stops with `SHIP_BASE_UNKNOWN` instead of reporting zero changed files. diff --git a/changelog.d/codex-ship-starts-audits-with-spawn-agent.md b/changelog.d/codex-ship-starts-audits-with-spawn-agent.md deleted file mode 100644 index 5ca75257a..000000000 --- a/changelog.d/codex-ship-starts-audits-with-spawn-agent.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: blocker ---- -- **/ship on Codex no longer stops at the test coverage audit.** The Codex version of /ship told Codex to start its audit with Claude Code's Agent tool, `subagent_type` and `run_in_background`, which Codex doesn't have, so the run stopped before it could finish (#4476). Codex skills now start every audit and reviewer subagent with Codex's own `spawn_agent` and wait on `wait_agent` for its final result: /ship's coverage, plan and docs audits, the native review voices in /review, /office-hours and the design reviews, the plan-CEO spec review, /autoplan's phases and /context-restore. Turn on `multi_agent = true` under `[features]` in `~/.codex/config.toml`; without it the dispatch fails and the step takes its failure path. The plan reviews' timed fallback reviewer needs a read-only subagent type Codex doesn't offer, so on Codex it reports the second-model review as unavailable instead of starting one. Claude Code's and every other host's instructions are unchanged. diff --git a/changelog.d/first-use-check-before-owner-tries-a-build.md b/changelog.d/first-use-check-before-owner-tries-a-build.md deleted file mode 100644 index 38a83b8e1..000000000 --- a/changelog.d/first-use-check-before-owner-tries-a-build.md +++ /dev/null @@ -1 +0,0 @@ -- **A cold first-use check before an app build reaches its owner (GST-3025).** New `gstack-first-use` freezes the first five day-one actions from the plan and the exact build, has an agent from another model family play that build on the same kind of device with a screenshot or video for each action, checks every link within 15 minutes of sending, and prints the "try this build" message only when all of it passes. The owner's score and words are kept beside the checker's predicted score in the owner's private drive. See `docs/first-use-check.md`. diff --git a/changelog.d/fix-skill-prefix-renders-clawhub-docs-issue-refs.md b/changelog.d/fix-skill-prefix-renders-clawhub-docs-issue-refs.md deleted file mode 100644 index a8206dad0..000000000 --- a/changelog.d/fix-skill-prefix-renders-clawhub-docs-issue-refs.md +++ /dev/null @@ -1,3 +0,0 @@ -- **`skill_prefix=true` now survives updates when gbrain is installed.** Every Claude re-render (the brain-aware render in `./setup`, `gstack-config gbrain-refresh`, and the pinned-overlay render) applies your saved prefix before it goes live, so skills stay `/gstack-retro`, `/gstack-review` and so on instead of reverting to short names after each update (#4413). Thanks @fliaron for the report. -- The OpenClaw docs no longer send you to ClawHub, which held only April 2026 builds of office-hours, ceo-review and retro and no longer has gstack skills (the docs also said four skills while listing three). `docs/hosts.md` and `docs/OPENCLAW.md` now point to native registration (`./setup --host openclaw`), which already serves the current version of each, and tell you to remove the ClawHub copies if you installed them (#4421). Thanks @exGeni for the report. -- **`gstack-issue-links` no longer reads other repositories' issues as gstack's.** A parenthesis that cites several repositories, such as `(founderstar #837, harness-evals #393)`, now gives each its own cross-repo reference instead of `Refs #837` and `Refs #393` on this repository, and `owner/repo#N` naming the current repository counts as its own issue, so `Fixes garrytan/gstack#N` closes it. diff --git a/changelog.d/nested-tsconfig-path-aliases.md b/changelog.d/nested-tsconfig-path-aliases.md deleted file mode 100644 index e6dd791a3..000000000 --- a/changelog.d/nested-tsconfig-path-aliases.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship's pre-gate finds the tests of a monorepo app that imports through its own `@/` alias.** The import graph read `paths` and `baseUrl` from the repository root's tsconfig only, so in an app like `apps/mobile` whose `tsconfig.json` maps `@/*` to `./src/*`, every touched source file failed with `PREGATE_NO_LANE` even with a test right beside it. Each file now takes its aliases from the nearest tsconfig.json or jsconfig.json above it that declares `paths` or `baseUrl`; a nested config that declares neither defers to the next one up, the root config governs the rest, and an `extends` outside the repository is still never followed. Thanks to @dtoelke for the report and the original fix (#4462). diff --git a/changelog.d/office-hours-explains-mode-pick.md b/changelog.d/office-hours-explains-mode-pick.md deleted file mode 100644 index 457395c18..000000000 --- a/changelog.d/office-hours-explains-mode-pick.md +++ /dev/null @@ -1 +0,0 @@ -- **/office-hours explains why it recommends a mode.** The first question (what's your goal?) now recommends a mode by quoting your own words and saying what another mode would skip, instead of restating the mode's name. diff --git a/changelog.d/office-hours-resources-lookup.md b/changelog.d/office-hours-resources-lookup.md deleted file mode 100644 index 69a6bf2df..000000000 --- a/changelog.d/office-hours-resources-lookup.md +++ /dev/null @@ -1 +0,0 @@ -- **/office-hours keeps its 34 founder videos and essays out of the skill file.** The list now lives in `lib/founder-resources.ts`, and the closing runs the new `gstack-founder-resources` command, which prints the ones a user hasn't been shown yet. The resources, the dedup and the opt-out work as before; the skill's own text is about 8.6 KB smaller. diff --git a/changelog.d/owner-answer-adds-no-longer-collide.md b/changelog.d/owner-answer-adds-no-longer-collide.md deleted file mode 100644 index 0ee128e22..000000000 --- a/changelog.d/owner-answer-adds-no-longer-collide.md +++ /dev/null @@ -1 +0,0 @@ -- **Two `gstack-owner-answers add` runs for the same answer no longer collide.** When two processes recorded the same owner answer at once, about one pair in ten failed: one could read the other's row while it was still empty and report a conflict, or see the record's folder appear between two checks and refuse it as "a link that cannot be resolved". The second add now waits up to 2 seconds for the first to finish writing, and the folder check looks for the folder before resolving it. Three tests that failed intermittently under load also got fixes. The sealed-guard streaming test now treats Bun's buffer-overflow error as a failure even when git exits 0, and the two `gstack-ship-receipt` test files use the same 30-second limit as CI. diff --git a/changelog.d/owner-answers-record.md b/changelog.d/owner-answers-record.md deleted file mode 100644 index 64738fad4..000000000 --- a/changelog.d/owner-answers-record.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -bump: minor ---- -- **A Captain can check an ask against the owner's earlier answers before sending it.** `gstack-owner-answers` keeps the owner's answers and standing yeses in one record, outside every repository: on Capy it is the owner's private Drive folder, which every thread the owner runs can read. `screen --ask-file` reads a Captain's draft ask and proposes leaving out only the lines that a verified, current standing yes in the owner's own words covers exactly. Every other line stays in, with any related answers shown beside it. That includes autoplan gate items, untyped or compound lines, and anything that always needs a fresh ask (product decisions, spend over a cap, access, publishing, thresholds, closing a contribution, destructive steps). A damaged, conflicting or unreadable record keeps every ask. Rows are one file each, written once; a row needs a second actor to verify its words before it can be taken, and the recorder can't verify its own. The file format and the two-week pilot are in docs/owner-answers.md. diff --git a/changelog.d/post-skips-receipt-ids-as-phones.md b/changelog.d/post-skips-receipt-ids-as-phones.md deleted file mode 100644 index 347751ab0..000000000 --- a/changelog.d/post-skips-receipt-ids-as-phones.md +++ /dev/null @@ -1 +0,0 @@ -- **`gstack-post` stops reading /ship's own receipt identifiers as phone numbers, so posting a /ship PR body no longer needs a confirm.** The receipt's session run id (`--<8 hex>`, read as an E.164 number when the pid has seven digits, #4480) and all-digit 12-character commit prefixes in its `pregate` and `requires_remote` values used to stop `gstack-post pr-body` with a MEDIUM `pii.phone.e164` finding. The same went for the `SHIP_RECEIPT:` and `CI_CARRY:` lines /ship prints (`heads=`, `carried_from=`, `base_red=`, `ci_carried=:/`). Only those exact values are exempt, in a receipt block the receipt writer rendered and on lines outside code blocks that open with `SHIP_RECEIPT:`, `CI_CARRY:`, `PREGATE:` or `READY_RECEIPT:`. A phone number in prose, in any other code block or field (a receipt line pasted inside one included), or in a hand-written receipt block still needs a confirm. diff --git a/changelog.d/release-notes-list-every-change.md b/changelog.d/release-notes-list-every-change.md deleted file mode 100644 index 0f312b473..000000000 --- a/changelog.d/release-notes-list-every-change.md +++ /dev/null @@ -1 +0,0 @@ -- **GitHub release notes now list every change in plain words, with a link to its pull request.** Releases v1.99.1.0 to v1.99.4.0 said only "see CHANGELOG.md", because the release cut took one headline per version and found none in entries written as bullets. Each release page now lists every change the release ships under What's new, Fixes and Changes for maintainers, one line each, with the pull request that made it, then the compare link. A release whose CHANGELOG has no entry for its version, or an entry with nothing in it, now stops with `NOTES_MISSING` instead of publishing an empty page, and notes too long for GitHub keep the changes that fit and say how many more are in the full changelog (docs/releases.md#release-notes). diff --git a/changelog.d/remove-router-bug-fix-route.md b/changelog.d/remove-router-bug-fix-route.md deleted file mode 100644 index 148547c8c..000000000 --- a/changelog.d/remove-router-bug-fix-route.md +++ /dev/null @@ -1 +0,0 @@ -- **gstack's router no longer gives bug reports a special rule, so the model fixes a bug the way it would without gstack.** The rule said to check the cause once and write a failing test first when the cause was unclear. On 26 real gbrain bugs that plain Opus 5.5 had failed once, gstack with the rule fixed 3 and plain Opus fixed 2: +3.8 points, with a 95% interval of -11.5 to +19.2, so no clear gain. The router's description no longer lists debugging either, so a bug report doesn't load it. Planning, review, /verify, /qa and /ship are unchanged, and the bug line gstack offers for your CLAUDE.md is unchanged too. The two paid evals of the removed rule are retired with it. diff --git a/changelog.d/review-plan-sorts-swift-paths.md b/changelog.d/review-plan-sorts-swift-paths.md deleted file mode 100644 index 64a158ba7..000000000 --- a/changelog.d/review-plan-sorts-swift-paths.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship's review plan now sorts Swift and Xcode files by risk instead of sending every iOS PR to the full review panel (GST-3041).** `gstack-diff-scope` had no rule for `.swift`, so an all-Swift diff printed `SCOPE_ERROR=unmatched` and the planner failed closed to five readers plus a coverage audit; a 55-line SwiftUI copy change got the whole panel. Swift source now reads as UI and app code (like Dart), as do storyboards, XIBs, string catalogs and asset catalogs. `Info.plist`, entitlements and the privacy manifest read as permissions (auth), and `Package.swift`, `Package.resolved`, `Podfile`(`.lock`), `project.pbxproj`, `.xcconfig` and schemes read as config, never as UI. Swift networking code (`URLSession`, `URLRequest`, `NWConnection`, Alamofire) still raises the review to full depth, the same as `fetch` in TypeScript. A file type no rule knows, such as a Metal shader, still fails closed. Reported from the GHome iOS app's PR review. diff --git a/changelog.d/review-plan-swift-user-data.md b/changelog.d/review-plan-swift-user-data.md deleted file mode 100644 index 3f1491492..000000000 --- a/changelog.d/review-plan-swift-user-data.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship's review plan now sends Swift code that touches private user data to the full review panel (GST-3052).** Once `.swift` stopped failing closed, an iOS change that asked for microphone or speech permission, read HealthKit or calendar data, or wrote and deleted users' files on the device planned the light review. Swift lines that request a privacy permission (speech, microphone, HealthKit, calendars, location, camera, photos, tracking) now count as security, and lines that read HealthKit, EventKit or contacts, write, delete or protect files on the device, use an app group container, exclude files from backup, or declare how long data is kept count as user data. Plain SwiftUI views, UserDefaults settings and file reads stay light. Reported from the GHome iOS app's PR review. diff --git a/changelog.d/round-grant-buys-the-fix-round-after-a-review-cap-stop.md b/changelog.d/round-grant-buys-the-fix-round-after-a-review-cap-stop.md deleted file mode 100644 index 82d36c322..000000000 --- a/changelog.d/round-grant-buys-the-fix-round-after-a-review-cap-stop.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: blocker ---- -- **A round grant at /ship's review round cap now buys the fix round it names, in the same run or the next (GST-2968).** When a review stops at the cap, /ship records the stop in the PR body (`gstack-ship-receipt stop`), and a `gstack-approvals round-grant --kind fix-check` naming that stop's head, run and findings buys one scoped fix-check of the fix instead of another read-only full review. Before, the read-only review had already counted as the extra round, so the grant bought nothing, and a grant recorded between /ship runs was ignored, so the next run paid for another full review. A grant is spent by the one run that claims it (`gstack-ship-receipt claim`), and an extra-round grant recorded before review keeps its fix round, also across preflight's base merge. diff --git a/changelog.d/secrets-check-code-shapes.md b/changelog.d/secrets-check-code-shapes.md deleted file mode 100644 index e0645620d..000000000 --- a/changelog.d/secrets-check-code-shapes.md +++ /dev/null @@ -1 +0,0 @@ -- **The secrets check stops flagging three kinds of ordinary code as a secret, and catches two misses.** TypeScript types with spaced generic arguments, Python keyword arguments that read a settings constant (`aws_secret_access_key=settings.S3_SECRET_KEY,`), and camelCase dotted field paths (`"form.section.fieldName"`) no longer stop a push. A secret built with `+` from short pieces, or written in a trailing or whole-line comment, is now reported. A secret in any of those positions still fails the check. Template keys, arrays of names and string-literal generic types still report and are documented as known limits. (#4289) diff --git a/changelog.d/security-findings-stay-out-of-pr-bodies.md b/changelog.d/security-findings-stay-out-of-pr-bodies.md deleted file mode 100644 index 51b0b23a5..000000000 --- a/changelog.d/security-findings-stay-out-of-pr-bodies.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: security ---- -- **A security finding's location and reason no longer go into a PR body.** When an owner ruled a security review finding a note, or a review deferred one in code the PR did not change, the PR body, its receipt block and the `Ruled note:`, `Gate disposition:` and deferred-findings lines named the file, line and reason, which on a public repository points at a live exposure. Findings whose category is `security`, `auth`, `secret` or `red-team` now print an opaque id and "security finding, details recorded privately"; the details stay in the private approval record (`gstack-approvals list`) and the local review files. `gstack-ship-receipt write` and `gstack-post` also refuse text that names a raw `path:line:security` fingerprint. Other findings print exactly as before. diff --git a/changelog.d/ship-child-starts-with-typed-ship.md b/changelog.d/ship-child-starts-with-typed-ship.md deleted file mode 100644 index e832153c5..000000000 --- a/changelog.d/ship-child-starts-with-typed-ship.md +++ /dev/null @@ -1 +0,0 @@ -- **The token-capped /ship child on Claude Code now starts its prompt with the typed `/ship` command, so the skill actually loads.** When /ship hands its remaining steps to a fresh child (`gstack-ship-budget launch`), the child's prompt used to open with "Load gstack. … Run /ship --continue ". Claude Code expands a skill only when the prompt starts with the typed command, so the child got plain text and had to find /ship on its own, which unattended Opus 5.5 did in 0 of 60 sessions in gstack-evals#366. The prompt now opens with `/ship --continue ` (`/gstack-ship` when `skill_prefix` is on). docs/unattended.md gains the rule for any parent that starts a child agent, and the Capy skill's Claude Code line now says the same. diff --git a/changelog.d/ship-counts-second-model-blockers.md b/changelog.d/ship-counts-second-model-blockers.md deleted file mode 100644 index 9bd46ff81..000000000 --- a/changelog.d/ship-counts-second-model-blockers.md +++ /dev/null @@ -1 +0,0 @@ -- **A /ship receipt now counts the second-model review's blocking findings, and refuses to be written while one is still open.** When /ship stopped after Step 11 because the second-model review (or the native adversarial pass) raised a `Blocks: yes` finding, no review-log row recorded it, so the receipt said `blocking_found=0` and the PR looked clean to `ready` and the merge step. Step 11's own row now carries `via: ship` and its findings, `gstack-review-scope summary` counts them (they are never a pass or a round), and `gstack-ship-receipt write` refuses with `RECEIPT_BLOCKER_OPEN` while any blocker's newest action is still `open` and no owner ruling closes it. A finding below the floor still counts only as a note. diff --git a/changelog.d/ship-handoff-child-gets-its-prompt.md b/changelog.d/ship-handoff-child-gets-its-prompt.md deleted file mode 100644 index f2a6e3a0c..000000000 --- a/changelog.d/ship-handoff-child-gets-its-prompt.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -important: blocker ---- -- **A `/ship` with a token cap on Claude Code no longer stops at its first handoff.** When the budget handed steps to a fresh child, `gstack-ship-budget launch` started `claude -p` with `--allowedTools ` right before the prompt. Claude Code reads every argument after that option as one more tool name, so the child got no prompt, exited with "Input must be provided either through stdin or as a prompt argument", and the run stopped. The launch now ends the options with `--` before the prompt, which Claude Code 2.1.292 and 2.1.326 both read as the prompt. Thanks to @exGeni for the report and diagnosis (#4448). diff --git a/changelog.d/ship-push-prints-one-stable-line-gst-3053.md b/changelog.d/ship-push-prints-one-stable-line-gst-3053.md deleted file mode 100644 index 4a694dd9d..000000000 --- a/changelog.d/ship-push-prints-one-stable-line-gst-3053.md +++ /dev/null @@ -1 +0,0 @@ -- **/ship's push step now always ends with one line that says whether it pushed (GST-3053).** `gstack-ship-release push` (Step 17) finishes with exactly one of `PUSH: pushed remote=origin ref= head=`, `PUSH: already_pushed remote=origin ref= head=` or `PUSH: failed reason=PUSH_REFUSED|PUSH_REJECTED|PUSH_GIT_FAILED`, so a script or eval grader can tell that /ship pushed by matching that line instead of looking for a `git push` call. Exit status is unchanged. Each reason code has a troubleshooting entry, and the format is documented in docs/unattended.md#push-line. An eval had graded 48 real /ship pushes as "no push" because it searched for the git command. The line's shape changed. The old `PUSH: ALREADY_PUSHED ` and `PUSH: pushed to origin/` lines are gone, so update anything that matched them. diff --git a/changelog.d/strip-incoming-receipt-runs.md b/changelog.d/strip-incoming-receipt-runs.md deleted file mode 100644 index 05324cf8e..000000000 --- a/changelog.d/strip-incoming-receipt-runs.md +++ /dev/null @@ -1 +0,0 @@ -- **A PR's review-round count can no longer be overwritten by the body `/ship` posts.** When `/ship` posted a PR body whose Agent notes held their own `gstack-receipt-runs` blocks, gstack removed only the first one, so a second block placed before the receipt became the history `gstack-ship-receipt history` counted (one round where there were three). A post now drops every runs block it brings and keeps exactly one. That one is the live body's history, placed right after the posted receipt, or none when the PR had no earlier runs. A post whose notes hold no receipt yet keeps the live started stubs in place of the runs block it brings. See docs/review-policy.md#rounds-per-pull-request. diff --git a/changelog.d/sync-gbrain-reports-failures.md b/changelog.d/sync-gbrain-reports-failures.md deleted file mode 100644 index 7da52d1a8..000000000 --- a/changelog.d/sync-gbrain-reports-failures.md +++ /dev/null @@ -1,3 +0,0 @@ -- **One queued record with a malformed path no longer stops artifacts sync for every project.** A skill started where the project slug was empty (for example from `/`) queued `projects//timeline.jsonl`; git staged it under another spelling, the drain refused the whole batch, and every later record waited until the queue was dropped by hand. `gstack-brain-enqueue` now refuses a path with an empty, `.` or `..` component, and the drain drops such a record, syncs the rest, and reports it (#4445). -- **A failed artifacts push stays reported until a push succeeds.** A later drain with nothing new to commit used to write `idle` over `push_failed` while the commits were still unpushed, and /sync-gbrain printed "curated artifacts pushed" because `gstack-brain-sync` exits 0 on a failed push. The status now stays `push_failed`, and the brain-sync row reads the drain's status. It says "pushed" only after a push, "nothing new to push" when there was nothing to send, and `ERR artifacts not pushed (...)` otherwise (#4418). -- **The /sync-gbrain memory row no longer says OK when pages failed to import.** It read only the ingest's exit code, which stays 0 for per-page failures, so it printed `OK ... 126 failed`. The row now reads the ingest's counts: `WARN` when some pages failed, `ERR` when pages failed and none imported, with the failed count and the first cause. A transcript that cannot be parsed is listed as a note and leaves the row `OK` (#4417). diff --git a/changelog.d/test-policy-floor-next-above-version.md b/changelog.d/test-policy-floor-next-above-version.md deleted file mode 100644 index f07c2cba5..000000000 --- a/changelog.d/test-policy-floor-next-above-version.md +++ /dev/null @@ -1,2 +0,0 @@ -### Tests -- A policy-floor test no longer fails once gstack's own version reaches the minor after the newest policy key; it now checks a floor above both, so the v1.101.0.0 release can pass CI. diff --git a/changelog.d/trunk-uploader-v2.1.8.md b/changelog.d/trunk-uploader-v2.1.8.md deleted file mode 100644 index 80c543845..000000000 --- a/changelog.d/trunk-uploader-v2.1.8.md +++ /dev/null @@ -1 +0,0 @@ -- CI now uploads free-suite test results to Trunk with analytics-uploader v2.1.8 and analytics-cli 0.17.0, up from v2.1.6 and 0.16.2. diff --git a/changelog.d/ubicloud-pack-skips-refs-being-rewritten.md b/changelog.d/ubicloud-pack-skips-refs-being-rewritten.md deleted file mode 100644 index 3066cce58..000000000 --- a/changelog.d/ubicloud-pack-skips-refs-being-rewritten.md +++ /dev/null @@ -1,2 +0,0 @@ -- **Packing a checkout for a Ubicloud run no longer fails when git beside it is rewriting a ref.** `ubi-runner.sh pack` asked git's upload-pack for every ref, and upload-pack advertises a ref it cannot read (a loose ref file caught mid-write or mid-delete) with a null id, so the snapshot failed with "not our ref 0000…". Pack now reads the refs once with `git for-each-ref`, which skips an unreadable ref, and fetches those exact objects. Seen twice in Free Tests on 2026-11-17 (runs 44063454800 and 44050883970); an empty loose ref file reproduces it every time. -- **The /plan-ceo-review HOLD SCOPE eval no longer fails a review because its follow-up question didn't repeat the mode name.** The check required the decision's "Project/branch/task" line to say "HOLD SCOPE", which the skill never asks for; the decision is already tied to the HOLD answer by session and order. A line naming another mode, or past or hypothetical work, still fails. diff --git a/changelog.d/unattended-autoplan-records-relayed-owner-reply.md b/changelog.d/unattended-autoplan-records-relayed-owner-reply.md deleted file mode 100644 index f0393beb6..000000000 --- a/changelog.d/unattended-autoplan-records-relayed-owner-reply.md +++ /dev/null @@ -1 +0,0 @@ -- **Unattended /autoplan now records the owner's relayed gate reply instead of refusing it.** `autoplan/SKILL.md` told an unattended session to end at the final gate and never approve, while `autoplan/DRIVER.md` told it to record a relayed owner reply with `gstack-autoplan answer --by agent: --owner-reply`, so whether a run reached `complete` depended on which file the model followed. Both now render one rule from `scripts/resolvers/autoplan-driver.ts`: never approve on your own judgment, stop at `gate_pending` when there is no owner reply, and record the owner's own relayed words with `--owner-reply-file`, never handing the command back to the owner. `gstack-skill-start`'s unattended block names the same exception for /autoplan. diff --git a/changelog.d/unattended-autoplan-runs-its-second-voice.md b/changelog.d/unattended-autoplan-runs-its-second-voice.md deleted file mode 100644 index 23c75fae4..000000000 --- a/changelog.d/unattended-autoplan-runs-its-second-voice.md +++ /dev/null @@ -1,2 +0,0 @@ -- **Unattended /autoplan now runs the second-model review you turned on, and waits for it instead of ending the session.** With `codex_reviews: enabled` in the config, Opus 5.5 still skipped the second reviewer in 2 of 3 unattended `--full` runs, saying the owner had not agreed to send the plan off the machine; in the third it started the review in the background and ended its turn, which ends a headless session. `gstack-autoplan next` now prints `OUTSIDE_CONSENT: given by codex_reviews=enabled` beside the outside attempt, and autoplan/DRIVER.md says that an explicit `enabled` is the owner's standing consent, that the review runs in the foreground with a 9-minute limit, and that a timeout or failure is recorded with the standard failure line while the run carries on. `disabled` is still enforced by the runner, and an unset `codex_reviews` is handled as before. -- **Two /autoplan runs started in the same second no longer share a run id.** Run ids were the run directory's name plus a time stamp to the second, so two runs whose directories were both named `autoplan` got the same id and the same checkpoint folder. Each id now carries six random letters before the stamp. diff --git a/package.json b/package.json index 2bd4a94b6..22a156dcf 100644 --- a/package.json +++ b/package.json @@ -1,6 +1,6 @@ { "name": "gstack", - "version": "1.100.0", + "version": "1.101.0", "description": "Garry's Stack — Claude Code skills + fast headless browser. One repo, one install, entire AI engineering workflow.", "license": "MIT", "type": "module",