diff --git a/.gstack/pregate.json b/.gstack/pregate.json index 74a99f1ad..6be73a295 100644 --- a/.gstack/pregate.json +++ b/.gstack/pregate.json @@ -1,6 +1,6 @@ { "$schema": "https://github.com/garrytan/gstack/docs/pregate.md#pregate-v1", - "runner": "bun test {files}", + "runner": "bun test --timeout=30000 {files}", "remote_gate": "bash scripts/ubicloud/pregate-remote.sh", "dependencies": { ".github/workflows/**": ["test/pregate-selection.test.ts", "test/free-tests-workflow-wiring.test.ts"], diff --git a/BROWSER.md b/BROWSER.md index 13aa70909..69be0410d 100644 --- a/BROWSER.md +++ b/BROWSER.md @@ -99,7 +99,7 @@ through CDP before `goto`, as the cookbook does), and no `file://` navigation. | `/scrape` | Prototype the extraction with `aside repl`, hand back the table, list, or prices as structured data. Read-only. | | `/devex-review` | Walk the real onboarding flow and time it, carrying the cookbook inline. | | `/ship`, `/spec`, `/land-and-deploy`, `/setup-deploy`, `/office-hours` | Third-party web actions (vendor dashboards, API keys, webhooks) offered as an Aside drive across your real sessions, with the one-question consent gate for anything mutating. | -| `/plan-ceo-review`, `/plan-eng-review`, `/plan-devex-review`, `/design-consultation`, `/review`, `/investigate`, `/cso`, `/office-hours` | Web research runs through `aside exec` in your real browser (`{{ASIDE_RESEARCH}}`), one read-only request per question, answers treated as untrusted content. No Aside → the same queries go to the host's WebSearch tool; no WebSearch either → "Search unavailable" once and the skill proceeds on in-distribution knowledge. | +| `/plan-ceo-review`, `/plan-eng-review`, `/plan-devex-review`, `/design-consultation`, `/review`, `/cso`, `/office-hours` | Web research runs through `aside exec` in your real browser (`{{ASIDE_RESEARCH}}`), one read-only request per question, answers treated as untrusted content. No Aside → the same queries go to the host's WebSearch tool; no WebSearch either → "Search unavailable" once and the skill proceeds on in-distribution knowledge. | ### Local-HTML rendering @@ -619,7 +619,7 @@ Storage stays intact by default. With explicit approval on a Chromium target, `- | Command | Description | |---------|-------------| -| `tabs` | List open tabs | +| `tabs` | List open tabs, with each owned tab's owner (a pair-agent client or `session:`) | | `tab ` | Switch to tab | | `newtab [url] [--json]` | Open new tab; `--json` returns `{tabId, url}` for programmatic use | | `closetab [id]` | Close tab | @@ -634,6 +634,27 @@ in a long script, `status`, `tabs`, `tab `, `newtab`, `closetab`, not responding. A command that needs the stuck tab times out after 30s and names it, so you can switch away or close it without `--force-restart`. +**Sessions keep parallel agents on their own tabs (#1961).** A call from a +known agent harness (Claude Code, Codex and others) gets a session named after +that agent process; its first command opens the session's own tab and later +commands stay there without moving the shared active tab. Name one with +`--session ` or `BROWSE_SESSION=`; `BROWSE_SESSION=` turns it off. +`--tab-id` and `BROWSE_TAB` win over a session and no longer move the active +tab either. Sessions isolate tabs, not cookies or storage, and only the local +root token can name one. + +### Phone emulation + touch + +This is Chromium emulating a phone on one tab, not Safari; a real device still goes through `/ios-qa`. + +| Command | Description | +|---------|-------------| +| `device [ [--landscape] \| off \| --list]` | Emulate a phone on the current tab only (size, pixel ratio, touch, `pointer: coarse`, user agent) from Playwright's devices table; other tabs stay desktop. `off` restores; `viewport` and `responsive` refuse while a device is active | +| `tap <@ref\|sel\|x y> [--hold ]` | Touch and release at an element's center or at x y. Needs `device` first | +| `swipe <@ref\|sel\|x y> [--distance ] [--ms ]` | One-finger swipe (default 300px over 300ms). Needs `device` first | +| `pinch <@ref\|sel\|x y> ` | Two-finger pinch; scale above 1 zooms in, below 1 zooms out. Needs `device` first | +| `touch [down\|move \| up \| cancel]` | Hold, move and lift numbered fingers for multi-finger flows; no action lists held fingers. Needs `device` first | + ### Extraction | Command | Description | @@ -1515,6 +1536,19 @@ capacity each), flushed to disk asynchronously via `Bun.write()`: - Network: `.gstack/browse-network.log` - Dialog: `.gstack/browse-dialog.log` +The console, network, dialog and startup-error logs start empty at each daemon +start and can grow while one daemon runs; `browse-daemon.log` and the +`browse-audit.jsonl` command audit build up across restarts, and codex scratch +files collect in `.gstack/tmp/`. +Run `gstack-clean` in the repository to see what it would reclaim, then +`gstack-clean --apply` to do it: it cuts these logs to their last 256 KB +once they pass 1 MB, removes `tmp/codex-*` files older than 7 days, and +trims the audit log only when you pass `--audit-days N`. It never touches +`browse.json`, `session-state.json`, `chromium-profile/`, `browser-skills/` +or `browse-states/`, and it skips the browse files while the daemon runs. +Deleting the whole `.gstack/` folder instead also deletes your cookies and +browser skills. + The `console`, `network`, and `dialog` commands read from the in-memory buffers (not disk) so capture is real-time even when disk is slow. @@ -1732,6 +1766,9 @@ the global `~/.gstack/browser-skills/foo/` only inside project-a. | `BROWSE_HEADLESS_SKIP` | 0 | Skip Chromium launch entirely (test harness only) | | `BROWSE_TUNNEL` | 0 | Activate the dual-listener tunnel architecture (requires `NGROK_AUTHTOKEN`) | | `BROWSE_TUNNEL_LOCAL_ONLY` | 0 | Test-only — bind both listeners locally without ngrok | +| `BROWSE_SESSION` | unset (agent harness calls get `agent-`) | CLI side: this caller's session name; empty or `off` turns sessions off | +| `BROWSE_SESSION_IDLE_MINUTES` | 120 | Daemon side: close (headless) or release (headed) a session's tabs after this many idle minutes; 0 keeps them | +| `BROWSE_MAX_SESSIONS` | 16 | Daemon side: most sessions at once; the least recently used idle one is closed to make room | | `CHROMIUM_PROFILE` | unset | Explicit headed Chromium profile directory (used by gbrowser's gbd per-workspace); honored by headed launch and profile-lock cleanup, not used by headless sessions | | `GSTACK_DISABLE_GPU` | unset | On macOS, headless Chromium starts with a GPU-taming flag set (it stops runaway GPU-process spin) that also disables WebGL: `getContext('webgl')`/`'webgl2'` return `null`, so WebGL pages (three.js, Mapbox GL, sigma.js) show their fallback. Set `GSTACK_DISABLE_GPU=off` to enable WebGL (`off` is the only recognized value). | | `GSTACK_BROWSE_MAX_HTML_BYTES` | 52428800 (50MB) | `load-html` size cap | @@ -1776,7 +1813,11 @@ browse/ │ ├── cdp-inspector.ts # $B inspect — persistent CDP session per page │ ├── activity.ts # ActivityEntry, CircularBuffer, SSE subscribers, privacy filtering │ ├── buffers.ts # Console/network/dialog circular buffers (O(1) ring) -│ ├── tab-session.ts # Per-tab session state (load-html replay, ref map scope) +│ ├── tab-session.ts # Per-tab session state (load-html replay, ref map scope, device profile, held fingers) +│ ├── session-name.ts # CLI: which session a call belongs to (--session, BROWSE_SESSION, agent harness pid) +│ ├── session-tabs.ts # Daemon: per-session and per-tab-id request scopes, idle sweep, session cap +│ ├── device-emulation.ts # $B device: per-tab phone emulation over the tab's own CDP session +│ ├── touch-commands.ts # $B tap/swipe/pinch/touch via Input.dispatchTouchEvent │ ├── token-registry.ts # Mint/validate/revoke for root + setup keys + scoped tokens │ ├── sse-session-cookie.ts # 30-min HttpOnly cookie for /activity/stream + /inspector/events │ ├── pty-session-cookie.ts # Separate scope: live Claude PTY auth diff --git a/CHANGELOG.md b/CHANGELOG.md index 209155077..a1e1d10ae 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,168 @@ # Changelog +## [1.93.8.0] - 2026-11-02 + +**/ship now runs Step 11's voices before any gate-class stop, and a recorded owner grant runs its fix-check.** On a gate-class change, a blocker fixed in review round 2 used to stop the run for the owner before Step 11 ran. If the owner then chose "ship with open notes", gate code shipped with no adversarial, outside or targeted review (C6-36 after-opus). `gstack-review-plan` now answers `NEXT: voices` for any gate-class stop while Step 11 has not run. /ship runs Step 11 as it does on `done`, asks the plan again with `--adversarial-ran`, and the stop then stands. The round cap still limits fix rounds; it no longer skips the required voices. Second, the owner's grant of one scoped fix-check after round 2 had no way in: the plan took no grant input, and the pass program printed `PASS: none next=stop`. Now `gstack-review-plan --grant ` checks that the branch's `gstack-approvals` record holds that id as an approved grant naming `review_loop_grant` and this run's stop head (exit 3 for any other row, a grant recorded for an earlier /ship on the branch included), and plans the one fix-check in place of the stop after a round-2 full pass. The pass program then dispatches that fix-check like any other. A fix-check never earns another, so the grant cannot repeat. After `NEXT: voices`, a blocking finding the voices queue goes back to the plan as `--blocking-open` instead of starting a new review pass. + +### Tests +- `test/review-depth.test.ts`: a gate-class round-2 blocker fix plans `voices` (adversarial, outside and targeted) before every stop and stops once they ran. A grant plans one fix-check after round 2, after the voices on gate class, and never a second. The transitions table and rule 3 in `docs/review-policy.md` say both. All four failed at main 1022f4f. +- `test/review-plan-cli.test.ts`: `--grant` with a recorded grant prints `NEXT: fix-check`; an unrelated approval, an unrecorded id, a `review_loop_exception` acceptance row or a grant for an earlier head exits 3; after a fix-check the grant never plans another. Failed at main 1022f4f (`NEXT: stop`). +- `test/review-depth.test.ts`: an invariant over every transitions-table state (gate class or not, Step 11 run or not, grant or not) and a dispatch-superset no-loss check against main's plans. +- `test/review-pass.test.ts`: the pass program runs the granted fix-check reader instead of `PASS: none next=stop`. Failed at main 1022f4f. + +### Docs +- `docs/review-policy.md`: rule 3 and two new rows in the transitions table. `ship/DRIVER.md`, `ship/sections/review-army.md` and `ship/sections/adversarial.md` route `voices` and `--grant`. + +## [1.93.7.0] - 2026-11-02 + +**A version stamp the branch added at its fork point's version now counts as stale after the base moves (GST-963).** On gbrain #8425, a branch added a row that named the fork point's version. After main moved past it, `gstack-restamp` skipped the row ("no tentative mention") and `gstack-ship-release finish` routed the line as a behavior change, so the stamp kept the old version and the docs stage flagged a product edit. + +### Fixed +- **`lib/restamp.ts`:** a stamp_patterns match on a branch-added line that names the fork point's version is rewritten to the new version, like a mention of a newer tentative version. Lines the branch did not add, and versions below the fork point, are left alone. +- **`lib/ship-bin-common.ts`:** finish routes such a line as metadata (a stamp file), not behavior. +- **One stamp bound for the writer and the stamp checks** (`lib/stamp-only.ts` `stampOnlyLine` with the branch's side): the tree receipt behind `gstack-ship-queue stamp`, `verify` and the gate-ahead replay, the merge train's stamp and verify checks (read against the slot's predecessor), and finish's route now accept a moved version on a line the branch added at or above the fork-point version. Before, the tree receipt and the carry still floored every line at the gated head's VERSION, so the stamp commit the writer now makes would have read `TREE: changed` and stopped the train with `QUEUE_STAMP_NOT_STAMP_ONLY`. The same move on a base line is still refused, unless that base line's text duplicates a line the branch added (GST-979). The tree receipt and the train checks also accept the move when the branch's own VERSION already equals the new version, as it does in a train slot that kept its claim. Two gaps remain, tracked in GST-980: the /ship receipt carry can still refuse the move (`RECEIPT_CARRY_PRODUCT_CHANGED`), and when the branch keeps its VERSION claim, carry and finish still route the move as behavior. + +### Tests +- `test/restamp-mentions.test.ts`: a branch-added row at the fork version after the base moved is restamped and its file is routed as metadata. On main, the stage reports `skipped` and ROUTE counts the file as behavior. +- `test/ship-rereview-triggers.test.ts`: the GST-895 fork-version case now routes as metadata. +- `test/restamp-mentions.test.ts`: `gstack-tree-receipt` reads the GST-963 stamp commit as `same-modulo-stamps` (it read `TREE: changed` before the shared bound), refuses the same move on a base row, and still accepts the GST-895 lost-slot stamp. `test/gstack-ship-receipt-docs.test.ts`: the carry takes the fork-version bound only on the branch's own line. +- `test/ship-queue-train-stamp.test.ts`: a pin claiming 1.2.4.0 with a row left at the base's 1.2.3.0 stamps, and the row moves to 1.2.4.0. Without the train checks' bound, `stampChain` refused it with `QUEUE_TRAIN_STAMP_MISMATCH`. On main it stamped, but the row stayed at 1.2.3.0. + +## [1.93.6.0] - 2026-11-02 + +**/plan-tune now learns from answers to questions that aren't in its registry, when the skill says which preference the question is about (#3442).** Before, an answer to an ad-hoc question id such as `plan-ceo-review-add-caching` moved no dimension, because only registered ids carry a signal. Most logged answers are on ad-hoc ids, so inferred values stayed near 0.5. + +### Added +- **Signal hints at ask time:** a skill can write `` in the question and `"signal_key":""` in the `gstack-question-log` row, choosing the key from the fixed list that `gstack-question-preference --signal-keys` prints. `--derive` uses the hint only when the id has no registry entry, the key is a known signal, and the answer matches one of that signal's options, by the exact key or the key's own words. Hinted answers fall under the same per-day cap as registered ones. +- **`hinted_events`** in the profile, and `(hinted)` on `--trace` rows, so you can see which answers came from hints. + +### Changed +- One shared marker parser (`scripts/qid-marker.ts`) now serves the log hook, the preference hook and the Codex session import. A hint never changes auto-decide or one-way-door handling. An unknown hint, or a hint on a registered id, is dropped with `QLOG_SIGNAL_KEY_DROPPED`, and the row is still logged. + +### Tests +- New tests that fail on main check that derive moves a dimension for a hinted ad-hoc answer and not for an unhinted one. Unknown keys, choices outside the signal, and registered ids ignore the hint. Hinted answers share the per-day cap. The hooks and the Codex import carry the hint into the log. + +## [1.93.5.0] - 2026-11-02 + +**The weekly OSV scan is clean again: qs and @hono/node-server are pinned past two Medium advisories (#3843).** The 2026-11-02 scan (run 41406416003) flagged qs 6.16.0 (GHSA-6pq8-jhm2-2vcw, CVSS 6.3, a `qs.stringify` comma-format crash) and @hono/node-server 2.1.1 (GHSA-rmxm-3fg6-px4f, CVSS 5.3, a `serveStatic` double decode that skips middleware). Both reach gstack only through the dev and test toolchain (`@anthropic-ai/claude-agent-sdk` pulls in `@modelcontextprotocol/sdk`), and nothing setup, the build or an installed skill runs loads either one. Installs set up since v1.91.231.0 don't contain them at all. Older installs and contributor checkouts that still carry the dev toolchain get the patched versions on their next `./setup`. + +### Fixed +- **`package.json` overrides**: `qs` is pinned to 6.16.2 and `@hono/node-server` to 2.1.6, the newest patch releases inside the ranges express, body-parser and the MCP SDK ask for. `bun.lock` now holds one copy of each, and the separate `body-parser/qs` entry is gone. + +### Tests +- `test/dependency-security.test.ts`: the lockfile floor check now covers `qs >= 6.16.1` and `@hono/node-server >= 2.1.3`. Both cases failed on main. + +## [1.93.4.0] - 2026-11-02 + +**/ship's finding collector reads severity from a finding's own Severity field, so a severity word in the finding's text no longer raises it.** + +### Fixed + +- When a reader's finding carries its own `Severity:` field, the collector uses that field. A word like "P0" in the finding's text (a plan phase name, for example) no longer makes it a blocker the reader marked `Blocks: no`. A label leading the line (`CRITICAL:`, `[P0]`, `- High:`, a label heading) still sets the floor, `Blocks: yes` still makes a finding CRITICAL, and the native voice's FIXABLE still collects as CRITICAL. A field that names more than one label ("High/Critical") counts as no field, and a lowercase `severity:` keeps its label. Lines with no Severity field collect as before. (GST-976) + +## [1.93.3.0] - 2026-11-02 + +**The queue now refuses a head whose required targeted review is missing or failed (GST-966).** On gstack-evals#140, the receipt showed `REVIEW_TARGETED: required=auth receipt=legacy verdict=missing` and `gstack-ship-queue ready` still exited 0, so a skipped required reader looked the same as a clean one. + +### Fixed +- **`lib/ship-receipt-depth.ts`:** `ready`, `train build` and `land` refuse `QUEUE_RECEIPT_TARGETED_MISSING` when the head needs a targeted class (money, gate, auth or secret) that no done targeted pass in the receipt covers: no pass recorded, a failed or unrun pass, or a pass on other classes. The line says `verdict=failed` when the recorded pass failed. A head that needs no targeted class passes as before. `train build` and `land` print the `REVIEW_TARGETED:` line with the refusal. + +### Tests +- `test/gstack-ship-queue.test.ts`: `ready` on an auth head with #140's receipt shape (missing) and with a failed pass refuses with the new code; a done pass and a head that needs no class pass. On main the missing case exits 0. + +## [1.93.2.0] - 2026-11-02 + +**Native cookie import now waits for its browser to actually exit before it reports cleanup, and three more flaky tests from the 2026-11-01 flaky pass are repaired.** Windows drops a Job's process count to zero while terminated Edge processes are still shutting down and still hold the profile's files, so cleanup that trusted the count returned early and the next profile reset failed with a sharing violation. + +### Fixed +- **`browse/src/cookie-import-native-job.ts`, `browse/src/cookie-import-native-worker.ts`, `browse/src/cookie-import-native.ts`**: the supervisor holds a handle to every process it sees in the import's Job and counts a process as active until that handle signals exit, not only until the Job's counter drops. The cleanup budget is now 20 s (was 5 s), because terminated Edge took up to 17.8 s to exit on a loaded Windows runner and 2 to 57 ms idle; an idle import returns as soon as its processes are gone. The parent's kill timer allows 2 s beyond that budget. Cleanup that still cannot confirm the exit reports `native_cleanup_failed`, as before. The real-Edge tests take their kill timers and time limits from the same constant. +- **`.gstack/pregate.json`**: the local pregate lane runs `bun test --timeout=30000 {files}`, the 30 s per-test timeout CI and the remote pregate already use, so a case that takes 5 s on a busy machine no longer fails only locally. The runner string is part of the pregate trust grant, so each machine records trust again on its first run after this change (the crew's remote pregate already passes `--allow-repo-commands`). +- **`test/gstack-ship-receipt.test.ts`**: the full `write` case, about 62 git calls, gets a 15 s timeout instead of bun's 5 s default. +- **`test/helpers/coverage-audit-evidence.ts`**: the plan-eng coverage audit detector credits complete reads placed by a literal `wc -c` byte count over in-cwd files, as it already did for `wc -l` (train gate 41209768188). + +### Changed +- `test/helpers/docsync-contract.ts` names the missing `documentation_section` when a document-release completion fails the contract, so the next paid red says which field was absent. + +## [1.93.1.0] - 2026-11-02 + +- **`gstack-pregate clear --run-url` accepts a reusable-workflow lane's caller run.** A lane whose workflow file runs only through `workflow_call` (gbrain's `native-locks.yml`, run as a job of `test.yml`) was refused `workflow_mismatch` for every run since v1.92.62.0, because GitHub reports the caller's file as the run's workflow. A run of a workflow in the branch's own `.github/workflows/` with a job that says `uses: ./.github/workflows/` now clears the lane, a job link inside that run too, as long as the lane's own job succeeded under the calling job (` / `, else `caller_job_skipped`), with callers read at pregate.json's recorded tree; the red and unfinished-run refusals still apply. Fix lines only suggest `gh workflow run` for a file with a `workflow_dispatch` trigger; otherwise they name the callers. An unrelated workflow still refuses. `test/pregate-clear-runs.test.ts` failed before the fix (GST-971). + +## [1.93.0.0] - 2026-11-02 + +**`gstack-clean` reclaims the throwaway files that pile up in a repository's `.gstack/` folder, and changes nothing unless you pass `--apply`** (#2627). Until now nothing removed old codex scratch files or trimmed the browse logs, so the folder only grew (the reporter measured 8.3 MB in six weeks), and the only way out was deleting the whole folder, which also deletes your cookies, Chromium profile and browser skills. + +### Added +- **`bin/gstack-clean`** (rules in `lib/repo-state-clean.ts`): cleans `/.gstack` only. It removes top-level `tmp/codex-*` files older than 7 days, cuts `browse-console.log`, `browse-network.log`, `browse-dialog.log`, `browse-daemon.log` and `browse-startup-error.log` to their last 256 KB once they pass 1 MB (in place, starting after a line break, so a running writer keeps appending), and with `--audit-days N` drops `browse-audit.jsonl` entries older than N days while keeping every line without a readable timestamp. Everything else, including `browse.json`, `session-state.json`, `chromium-profile/`, `browser-skills/`, `browse-states/` and the other `tmp/` entries, is never touched. +- **Safety:** it refuses (exit 3) outside git, when `.gstack` is a symlink, when `BROWSE_STATE_FILE` names another folder, and when the folder is, holds or sits inside the gstack state root. It skips symlinks, logs with more than one hard link, files it cannot read and files changed in the last hour, skips browse files while the browse daemon runs and `tmp/` while codex runs, and skips a group when it cannot tell. Right before each change it checks that the folder is still the one it verified and that the file still has the size, time and inode it measured, and it stops that change if either moved. `--json` prints the same result as one versioned object. +- Result codes `GSTACK_CLEAN_USAGE`, `GSTACK_CLEAN_TARGET_REFUSED`, `GSTACK_CLEAN_PARTIAL` and `GSTACK_CLEAN_INTERNAL`, each with a docs/troubleshooting.md entry; BROWSER.md names the command where it lists the log files. + +### Tests +- `test/gstack-clean.test.ts`: the dry run changes no byte or mtime, `--apply` removes exactly the allowlist, protected files survive at any age, every refusal touches nothing, links out of the folder, hard-linked and unreadable logs and a symlinked `tmp/` are left alone, a log with no line break in its tail and a malformed audit file are left alone, an unlistable `tmp/` skips only its group, `--json` reaches a pipe whole with hundreds of actions and a reader that closes early is not an error, live and unknown liveness skip their group, the cut keeps the inode and a line boundary, audit trimming keeps the right lines byte for byte, exit codes 0, 1, 2 and 3, and the `--json` shape. +- `test/gstack-clean-producers.test.ts`: every tracked producer of a `.gstack/tmp/` entry is classified as scratch or kept, and any new `codex-*` name must be added to the scratch list on purpose. + +**Capy sessions on Claude Opus 5.5 can read a smaller set of gstack skills (GSTA-161).** Before, every model read the full shared preamble on every skill. Now Opus 5.5 gets the `lite` loadout: the same skills without eleven coaching sections of the preamble, about 10.7 KB less on land-and-deploy. Every other model, and any model gstack is not told about, still gets the full skills, and the checkout's committed SKILL.md files stay full. + +### Added +- **`lite` loadout and a model map** in `modules/loadouts.json`. `lite` drops the voice, writing-style, completeness, confusion, evidence, context-health, question-tuning, search-before-building, repo-mode, brain-health and context-recovery pieces and keeps the upgrade check and brain sync. `models` maps only `claude-opus-5-5` (and its dated snapshots) to it; a point release or any other suffix gets full. +- **Explicit model detection** (`lib/model-detect.ts`): an explicit loadout, then `--model`, then `GSTACK_MODEL`, else full. An unmapped, invalid or unreadable model id serves full and never errors (`LOADOUT_INVALID_MODEL` names an id that is not a plain model id). +- **`gstack-capy-install --model `** (passed through `scripts/install.sh`), or `GSTACK_MODEL=` in its environment, which the Capy skill uses because an older installer ignores it: prints `LOADOUT` and `SKILL_ROOT` lines and, for lite, writes a second launch file that adds `GSTACK_LOADOUT_SKILL_ROOT`. The install verdict and its key do not change. +- **`gstack-modules install` and `resolve`**: `install` publishes the lite tree under `/loadouts/-/` as a complete skill tree (rendered files plus every tracked runtime asset), keyed by the selection, model, full root, host and source tree, reused only after its files re-hash against the lock, and never replaced in place. A failure serves full with `LOADOUT_FALLBACK`. `render --bind-model --full-root ` renders the same start lines for eval arms. +- **`./setup --loadout `** for Claude Code: saved as `claude_loadout` and rendered through setup's own Claude render into the untracked render dir; `--loadout full` returns to the committed files. +- **The start line carries the model.** A lite skill's start command names its model and the full install root, so `gstack-skill-start` prints `LOADOUT` and `LOADOUT_RULE` lines, plus `LOADOUT_MISMATCH` when `GSTACK_MODEL` names another model, telling that session to read the full SKILL.md instead. + +### Tests +- `test/model-detect.test.ts`, `test/module-loadouts-install.test.ts`, `test/loadout-setup.test.ts`, and new cases in `test/module-loadouts.test.ts` and `test/gstack-capy-install.test.ts` cover the mapping and its near misses, hostile model ids, the start-line rules, two models on one machine, runtime assets from `SKILL_ROOT`, identity changes, concurrent publishers, and `./setup --loadout`. + +**Shadow mode for the CEO-only decider: in a driver run that turns it on, every approval gate gets one hidden model prediction per question, recorded before the owner answers (GST-537).** Before the CEO-only mode is allowed to decide for the owner, we need to know how often it would pick what he picks. The only decider before this was a fixed rule that always keeps the plan's direction, which agrees with the reviewers on every past item, so it measured nothing. Now a model (Opus 5.5 by default, about $0.07 per gate on your own API key) predicts each answer, and taste-replay round 2 in gstack-evals compares the predictions with the owner's real answers. Nothing routes to the owner differently. + +### Added +- **`gstack-config set autoplan_shadow on`** (off by default; `GSTACK_AUTOPLAN_SHADOW=1|0` overrides it for one run) turns shadow mode on at a run's first `next`, stored in `run.json` so a resumed run cannot flip it. `autoplan_shadow_model` picks the model (default `claude-opus-5-5`). It stays off with `SHADOW: off ()` when there is no API key for the model, the model has no price, the run is `--no-outside` or CEO-only. +- **Every driver command that prints a gate** (`next`, `submit`, `close`, `resume`, `revise`, `run`, `check`) runs the shadow first and prints `SHADOW: recorded model= usd=` or `SHADOW: failed `, then the gate, then the checkpoint. The prediction goes to the private `/shadow/ceo-only.r.jsonl`: action, pick, confidence, cites and reason per question, beside the keep-direction rule's pick, the frozen question, the prompt hash and the cost, plus a 10% audit mark for a later live switch. The run lock is free during the call; another writer pays nothing while a call is in flight, `answer` waits for it, and after a crash `answer` records that gate first, marked retrospective. A shadow failure never blocks a gate or an answer. +- **Shadow calls count against `--spend-cap`** under their own ledger ids labeled `shadow`, show as `shadow=` on `GATE_SPEND`, and never change the per-voice price on `GATE_REC_COST`. A call that was sent and then lost to a crash settles `unknown`, never released as undispatched. +- **`gstack-autoplan shadow --out | --runs [--json]`**, a read-only export for gstack-evals: one row per item of each answered gate, its shadow status (recorded, malformed, failed, missing, retrospective, off) and the owner's answer with how he gave it. An unanswered or interrogated gate refuses `SHADOW_BEFORE_ANSWER`. +- **Answered decision rows record how the owner answered**, parsed from his own words: `owner_mode` (`rec`, `rec_except`, `tokens`, `as_is`, `unparsed`), `owner_named`, the `rec_hash` answered against and `answered_role`. +- `gstack-artifact validate ` checks each shadow record against its journaled sha256. + +### Changed +- The shared model-call transport can write a caller's own egress receipt, returns the provider's stop reason (a reply cut at the output limit is retried at 8,000 tokens, not treated as malformed), and sends the output limit to OpenAI models as `max_completion_tokens` when a caller sets one. Existing callers send the same bytes as before. + +- **Parallel agents no longer take over each other's browser tab.** When several agents drive gstack's own browser at once, each agent process now gets its own tab on its first command, and its `goto`, `snapshot`, `click` and chain commands stay on that tab (#1961). Name a session yourself with `--session ` or `BROWSE_SESSION=` (subagents that share one agent process need this); `BROWSE_SESSION=` turns it off. Sessions isolate tabs, not logins, so cookies and storage stay shared. A session idle for 120 minutes is closed (`BROWSE_SESSION_IDLE_MINUTES`, 0 keeps it), at most 16 run at once (`BROWSE_MAX_SESSIONS`), and only the local root token can name one. `tabs` now shows each tab's owner. +- **`BROWSE_TAB` and `--tab-id` no longer move the active tab for everyone else.** A command pinned to a tab used to switch the daemon's active tab while it ran, so another caller's command during that window landed on the pinned tab, and two overlapping pinned commands could leave the daemon on the wrong tab. Each request now runs on its own tab without touching the shared one. +- **QA can now test phone pages with real touch input.** `$B device "iPhone 15"` makes the current tab act like that phone (size, pixel ratio, touch support, mobile user agent and `pointer: coarse`) without resizing any other tab; `$B device off` restores it and `$B device --list` names every profile. `$B tap`, `$B swipe`, `$B pinch` and `$B touch down|move|up ` send real touch events, including several fingers at once (hold a joystick while tapping fire) (#3058). This is Chromium emulating a phone, not Safari; /qa's report names the phones it emulated and what still needs a real device through /ios-qa. When the Aside browser drives, its cookbook has a touch recipe, checked on Aside 1.26. +- New result codes, each with a troubleshooting entry: `BROWSE_SESSION_INVALID`, `BROWSE_SESSION_ROOT_ONLY`, `BROWSE_SESSION_TAB_RESET`, `BROWSE_SESSION_LIMIT`, `BROWSE_DEVICE_UNKNOWN`, `BROWSE_DEVICE_CDP_FAILED`, `BROWSE_TOUCH_NO_DEVICE`. + +- **`/investigate` drops its five-phase step procedure** (GST-938). Harness Evals watched Opus 5.5 run `/investigate` on 46 bug-fix sessions: the procedure loaded every time and the model followed none of its steps (0 of 30 with Claude Code's real system prompt), while a gstack session cost $0.46 against $0.22 without gstack (HARN-14, gn-hard-follow-2026-11). The skill keeps the parts that still do a job. Those are the Iron Law, the arguments table, a short "How to work" with the three lessons that cost real time before (read call sites before calling two paths duplicates, refresh the git index before timing, test a shared premise before a third fix), the 3-strike and more-than-5-files asks, the Scope Lock with its freeze hooks, the recipes and `--backlog`, the prior-learnings search and the DEBUG REPORT. The generated Claude skill shrinks from 29.8 KB to 20.2 KB. Gone: the Phase 1-5 walkthrough, the pattern table, the web-research block and the hypothesis-keyed learnings re-pull. The research block's rule stays as a guard, so a web search names the error class and library, never raw error text, hostnames, paths or secrets. BROWSER.md and docs/skills.md now describe the short skill. +- Behavior change: the model is no longer told to work through Phases 1-5 in order, and `/investigate` no longer carries the Aside web-research block. Hook-enforced versions of the two steps that could change results (reproduce first, verify after) are planned as an opt-in Claude Code candidate that ships only if a preregistered eval shows it helps. + +- **Opening a tab for one browser session no longer undoes another caller's tab switch.** If a different caller switched the active tab while a session's new tab was opening, the daemon used to switch back; it now keeps the newer choice. +- **`gstack-artifact validate` catches a deleted shadow record.** A revision the journal says was recorded, but whose record file is gone, now fails as `ARTIFACT_STALE` instead of passing. +- **A shadow that fails while saving its record is journaled and never stops the run.** The failure is written as `shadow_failed` under the run lock and its lease is dropped; if the lock stays taken, the lease expires and the next command settles that revision. + +## [1.92.83.0] - 2026-11-02 + +**The sealed guard no longer rescans commits the destination already holds, so a stacked branch retargeted after its parent squash-landed stops being refused for the parent's history.** gstack-evals#22 listed its parent #6's seven commits after #6 landed as one squash. One of them had added and later deleted a held-out file, so the guard refused a push whose every byte GitHub already had as `refs/pull/6/head`, and the queue sat blocked for about 50 minutes until a waiver landed. + +### Changed +- **`lib/sealed-guard.ts`** (`evaluateSealed`, shared by the pre-push hook and `gstack-pregate sealed`): a listed commit is skipped when it is an ancestor of the destination's own `refs/pull//head` for a squash `(#n)` on the base since the merge-base (at most 50 refs, a 20-second budget, an egress receipt per remote call, fetched refs deleted afterwards), or when its own `git patch-id --verbatim` and diff equal a base commit's (off above 2,000 base or 400 listed commits; a landed commit still gets the `sealed.label_join` check, which reads the whole file). Nothing is matched against a combined diff, and any failed lookup scans everything with one line saying why. docs/pregate.md#sealed-skip. +- **`gstack-pregate sealed`**: the `SEALED:` line gains `skipped=` (and `commits=` counts the commits scanned), stderr names each skipped group, and `--no-skip` scans every listed commit. + +### Tests +- `test/sealed-skip.test.ts`: the #22 shape is clear with seven commits skipped as `refs/pull/6/head`, the hook and the preflight agree, `--no-skip` refuses, no fetched ref is left behind and the egress ledger records the lookup. The same parent never published refuses although its squash equals the combined diff. An unreachable remote scans everything and says so. A cherry-pick that landed on the base is skipped and named. A whitespace near-miss (`/home/alice/` against `/home/ alice/`) is scanned and refuses, and so is a landed commit that differs only in bytes that aren't valid UTF-8. + +## [1.92.82.0] - 2026-11-02 + +**`gstack-review-dispatch verify` now catches a merged findings file edited after collect.** Verify used to compare only output hashes and line counts. So a `merged.jsonl` row whose label was lowered, or whose text was replaced with another line's summary, passed with `VERIFY: clean` once the judge agreed with the edited row. Verify now re-parses each reader's unchanged output and checks every merged row against the finding at its index, label and raw text byte for byte. A mismatch prints `merged_line_differs` on the OUTPUT and JUDGE rows, and verify is dirty. A missing output file or an output event with no file recorded (even for a reply with no findings), a judgment run on a copy of the merged file, merged rows that collect never wrote, and an output whose finding count no longer matches its recorded ids are dirty too. A reply refused for naming the wrong checkout is still allowed to hold findings collect never merged, but only when verify re-checks its tree line and finds it wrong. + +### Known +- Deleting a whole output event together with its merged rows still verifies clean, because the ledger keeps no anchor for it. It is listed for the structured-findings plan. + +### Tests +- `test/review-dispatch.test.ts`: a forged label, a summary copied onto a blocker's id, a deleted output file, a judgment on a forged copy and an extra merged row all verify dirty. Each verified clean at main c830c9b. +- `test/review-dispatch.test.ts`: an output event edited to look like a wrong-tree refusal (no ids, `review: false`, `tree_ok: false`) verifies dirty, even after a re-sent brief is answered, and so does one edited to drop its file and ids. A real wrong-tree reply stays `NOT_REVIEWED` until a re-sent brief gets a review, then verifies clean. + +### Docs +- `docs/review-policy.md` (verify) describes the per-row check. + ## [1.92.81.0] - 2026-11-02 **`gstack-ship-release finish` now accepts the owner's edit after a fix-check instead of asking for another review round (GST-965).** On gbrain #8425, the fix-check found a blocker, the owner's pre-agreed fallback was applied and "then ship" was granted. Finish still classed the edit as behavior and stopped with `insert 5–11.5 → 12–14 → 16`, a third round the R51 cap forbids, so /ship could not get through Step 16 without running its stages by hand. diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 634968ea6..f394d2e3f 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -206,6 +206,7 @@ the new defaults. | Validate periodic cases plus the weekly gate census on a branch | `gh workflow run evals-periodic.yml --ref -f include_gate_census=true` | CI runners + API spend | Pushed branch, `gh` with workflow rights | | Run the opt-in ML, gitleaks and Swift checks | `gh workflow run platform-qualification.yml --ref ` | Free CI runners | Pushed branch, `gh` with workflow rights | | Render one module loadout as an eval arm ([loadouts](docs/loadouts.md)) | `bin/gstack-modules render --host claude --modules full,-section:/ --out-dir ` (`--loadout full` for the control) | Free, about a second per host | Bun 1.4.2 | +| See which loadout a model gets, or publish it as Capy would ([loadouts](docs/loadouts.md#how-gstack-picks-a-loadout)) | `bin/gstack-modules resolve --model `; `bin/gstack-modules install --model --full-root ~/.claude/skills --store ` | Free, under a second | Bun 1.4.2 | | Look at past local eval runs | `bun run eval:list` / `eval:compare` / `eval:summary` | Free | Local eval history | | See a case's pass rate across recent weekly runs | `bun run eval:pass-rates --case ` | Free | `gh` with repo read access | | Measure one red case alone against the ship bar (MEETS 9/10) | `bun run scripts/ship-measure.ts measure --case --round baseline` ([bar](docs/TESTING_INTERNALS.md#ship-measure)) | API spend for 10 trials (12 for behavior); asks above $2/trial, stops at $25/case | Same as paid coverage | @@ -431,7 +432,7 @@ eval files, and misses the strict classifier. No API keys needed. - **Catalog budget** (`test/catalog-budget.test.ts`) — Caps the aggregate discovery surface: the sum of every skill's frontmatter `name` + `description` (what every host loads at discovery, every session) must stay under 1,423 token-equivalents, with a 260-byte per-skill cap. Counting goes through the shared census in `test/helpers/skill-census.ts` (physical files vs authored skills vs registry entries — three deliberately different counts). Adding a skill? The failure message carries the re-measure + ratchet protocol. - **Context-budget ratchet** (`test/context-budget-ratchet.test.ts`) — CI ceilings on the two token ledgers the catalog budget doesn't cover: the always-on full-frontmatter aggregate and each skill's per-invocation eager tokens (SKILL.md + forced-read references), graded against `test/fixtures/context-budget.json` via `lib/context-bill.ts`. New skills fail until they have a ceiling; ceilings for removed skills must be pruned. Legitimate growth or a landed reduction: re-run `bun test/helpers/capture-context-budget.ts` and commit the refreshed fixture in the same commit, so the change is a visible decision in the diff. - **Skill-to-bin drift ratchet** (`test/skill-bin-drift.test.ts`) — `scripts/skill-bin-drift.ts` statically reads every gstack bin call in skill templates, carved sections, references and the Claude-host renders (it never executes a bin) and fails with `DRIFT_UNKNOWN_BIN` or `DRIFT_UNKNOWN_FLAG` at file:line when a skill names a bin missing from `bin/` or a flag that bin's parser does not accept. Known misses live in `test/fixtures/skill-bin-drift-allowlist.json`, which only shrinks: a new finding fails, and so does an entry that no longer reproduces. Check one skill with `bun run scripts/skill-bin-drift.ts --skill `; bins with permissive parsers or no usage text are reported, never failed. -- **Dependency security regressions** (`test/dependency-security.test.ts`) — Run `bun test test/dependency-security.test.ts` to check the resolved `sharp` and `adm-zip` version floors, load Sharp, verify ordinary ZIP extraction, and reject extraction through destination-file and destination-directory symlinks. The symlink cases skip Windows. These checks complement the OSV scan; they do not change its existing exceptions. +- **Dependency security regressions** (`test/dependency-security.test.ts`) — Run `bun test test/dependency-security.test.ts` to check the resolved `sharp`, `adm-zip`, `qs` and `@hono/node-server` version floors, load Sharp, verify ordinary ZIP extraction, and reject extraction through destination-file and destination-directory symlinks. The symlink cases skip Windows. These checks complement the OSV scan; they do not change its existing exceptions. ### Tier 2: E2E via `claude -p` diff --git a/TODOS.md b/TODOS.md index 215b3b0aa..9cb7eb0e1 100644 --- a/TODOS.md +++ b/TODOS.md @@ -2,6 +2,13 @@ ## NEXT PRIORITY +### P3: gstack-clean follow-ups (filed 2026-10-31, GSTA-226 E2, #2627) + +- **Codex templates delete their own scratch.** The `/codex` review, challenge and consult blocks leave `codex-err-*`, `codex-resp-*`, `codex-out-*` and `codex-prompt-*` mktemp files under `TMP_ROOT` (which falls back to `.gstack/tmp`) and the `codex-prompt.*` and `codex-focus.*` free-text files in `.gstack/tmp`; `gstack-clean` removes them after 7 days. Delete each file at the end of its block instead. Skill prose for every user, so it needs its own review. **Effort:** S. **Priority:** P3. +- **Codex scratch lands at the git toplevel.** `bin/gstack-paths` falls back to the relative `.gstack/tmp` when `TMPDIR` and `TMP` are unset, so a /codex run from a subfolder leaves scratch in `/.gstack/tmp`, where `gstack-clean` does not look. Make the fallback `$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.gstack/tmp`. **Effort:** S. **Priority:** P3. +- **`gstack-clean` reclaims `browse-daemon.log.1`.** The browse CLI renames `browse-daemon.log` to `.1` past 10 MB and nothing ever removes the copy. Add it to the browse group (remove when older than an hour and the daemon is not running). **Effort:** S. **Priority:** P3. +- **Browse hints `gstack-clean` when `.gstack/tmp` is large.** Print one line at daemon start when `.gstack/tmp` or the browse logs pass a size threshold, naming `gstack-clean`. **Effort:** S. **Priority:** P3. + ### P2/P3: Gemini design provider follow-ups (filed 2026-10-31, from the #990 build) - **Run the live Gemini check once a key exists** — every Gemini request and response shape in `design/src/provider.ts` follows Google's docs and stubbed fixtures; nobody has sent one to the live API. `GEMINI_API_KEY=... bun run design/scripts/live-gemini-check.ts` (about $0.04) settles the image body, the thinking-image pick and the OpenAI-compatible vision endpoint. A JPEG answer is refused today. **Effort:** S. **Priority:** P2. @@ -78,7 +85,7 @@ The 158 older open issues were re-checked against main 5cb5e1c (v1.91.83.0). Of Each item was decided in the Wave 9 plan and parked with its reason. Credit the reporter or contributor when it is picked up. - **Fold `triggers:` into skill descriptions (#2118)** — `scripts/gen-skill-docs.ts` folds only `voice-triggers:`, and the trimmed description keeps one sentence, so natural-language triggers never reach the listing. Parked because it conflicts with the catalog-trim design and would regenerate every skill. Mechanism when picked up: a `when_to_use:` frontmatter key for the Claude host, built from `triggers:`, deduped, and kept inside the catalog byte budget (Claude Code's 1,536-character combined listing limit); hosts that allowlist name and description get a bounded "Use when…" sentence or nothing. **Effort:** M. **Priority:** P2. -- **Plan-tune `{skill}-{slug}` ids never reach a signal (#3442, split from #2623)** — the label-to-key capture (#3394), `contributing_events` and the per-day bound (#3397) closed the rest of #2623. Remaining: `scripts/resolvers/question-tuning.ts` lets skills use `{skill}-{slug}` ids, which have no registry entry and so no `signal_key`; those answers never move a dimension. Acceptance: `{skill}-{slug}` ids get a hinted fallback to a registered signal key. Needs a design call on the key source. Credit @v639dragoon. **Effort:** M. **Priority:** P2. +- **Plan-tune qid marker follow-ups (#3442 review notes, filed 2026-11-02)**: `QID_MARKER_RE` in `scripts/qid-marker.ts` drops the whole marker when the `;signal=` tail is malformed (an underscore, a space, an empty key). Match the id first and drop only a bad hint. `question-preference-hook.ts` still looks up the raw marker id, while the log hook lowercases it. Use `parseQidMarker(...).id` in both hooks. Add hinted-marker cases to `test/question-preference-hook.test.ts`. **Effort:** S. **Priority:** P3. - **Stateless, instruction-only host mode (#3061, @m-check1B)** — hosted multi-tenant runtimes have no install, state root or helper binaries, but skill bodies still read and write `$GSTACK_STATE_ROOT`, call `bin/gstack-*` and log telemetry. Deferred to the new-hosts theme below. Requirements kept on record: tenant-owned state (a `capabilities.state: 'none'` host renders host-provided hooks instead of shell steps), named artifact hooks (`ARTIFACT_SAVE`/`ARTIFACT_FIND` per kind, `WEB_RESEARCH`, `SECRET_SCAN`, `USER_PREF`) mapped the way the gbrain resolvers already are, a generator check that fails a stateless render containing `~/.gstack`, `$GSTACK_`, `bin/gstack-*` or `--event-type`, and the `claude` host's output byte-identical. The contributor offered an office-hours overlay plus the generator check as a first PR. **Effort:** L. **Priority:** P3. - **Legacy unprefixed Cursor skill folders (#3084)** — 1.31 Cursor installs created `~/.cursor/skills//SKILL.md` links; 1.91 links `gstack-*` folders and `_prune_stale_generated` in `setup` only looks at `gstack-*` names, so Cursor can load two skills with the same name. `setup` already removes old unprefixed Claude entries; Cursor needs the same owned-link cleanup (remove only links that point into a gstack checkout). Split from #3084's data-loss fix, which Wave 9 Lane A0 ships. **Effort:** S. **Priority:** P3. - **`/retro global` says what a "session" is (#1315, third problem)** — `bin/gstack-global-discover.ts` counts transcript files and rollouts, so cron- or agent-driven repos show hundreds of "sessions" with no interactive work. Annotate the retro output that sessions are tool invocations or files, not distinct human working sessions. Problems 1 and 2 are fixed (v1.91.66.0 and Wave 9 Lane F). **Effort:** S. **Priority:** P3. @@ -95,7 +102,7 @@ The 73 open feature requests (75 with #3061 and #3122, which the themes below al - **Ship, review and safety (9 issues)** — `/review` shadowing the native command (#422), a silent-deploy-failures checklist section (#2089), Conductor TODO proposals and a /ship block on unresolved ones (#2090), a light-ship mode (#2188), auditing an issue before filing it (#2642), maintainer-gated CI versus failures in landing-report (#2766), Flutter pubspec versions with build metadata (#2833), a deterministic fact-consistency check for document skills (#2887), a dual-engine /careful guard (#2888). **Priority:** P3. - **Planning reviews (8 issues)** — marketplace-aware skill recommendations (#419), founder-signal reasoning in the YC recommendation (#513), grounded verification of paths, math and flags in spec reviews (#973), a substrate-neutral existing-capability check (#1685), `/autoplan --tier` (#1955), autoplan workspace safeguards (#1959), splitting one-way architecture forks from two-way defects (#2490), an evidence ledger for "is demand real?" (#2350). **Priority:** P3. - **State root and project identity (6 issues)** — repo-local portable context artifacts (#1161), design-doc archive and prune (#1536), umbrella directories over several repos (#1923), reclaiming per-repo `.gstack` state (#2627), nested repos declaring themselves projects (#2649), references and disposition on decision events (#2678). **Priority:** P3. -- **Browser and mobile QA (6 issues)** — a remote browser via `BROWSER_WS_ENDPOINT` (#974), per-session tab isolation (#1961), an opt-in User-Agent suffix and default headers (#2459), skipping the Aside probe on Linux and Windows (#2864), activity-ring copies on caught-up reads (#2973), touch input and mobile Safari (#3058). **Priority:** P3. +- **Browser and mobile QA (5 issues)** — a remote browser via `BROWSER_WS_ENDPOINT` (#974), an opt-in User-Agent suffix and default headers (#2459), skipping the Aside probe on Linux and Windows (#2864), activity-ring copies on caught-up reads (#2973), and mobile Safari, the half of #3058 that Chromium touch emulation (GSTA-227) did not cover. Per-session tabs (#1961) shipped in GSTA-227. **Priority:** P3. - **Prompt cost and skill shape (4 issues)** — a response-header template (#1376), caching preamble output within a session (#1572), the Fable 5 prompting-guide audit (#1994), per-skill benefit in `gstack-context-bill` (#2889). **Priority:** P3. - **Customization and localization (3 issues)** — locale-aware skills (#481), a tutorial on new customization dimensions (#489), a `SKILL.local.md` overlay (#1759). **Priority:** P3. - **Retro (2 issues)** — a team-scoped multi-repo retro (#402), ADR and issue awareness in the narrative (#2284). **Priority:** P3. diff --git a/VERSION b/VERSION index 9a1e28b7f..d4a1cffc9 100644 --- a/VERSION +++ b/VERSION @@ -1 +1 @@ -1.92.81.0 +1.93.8.0 diff --git a/agents-digest/gstack-AGENTS.md b/agents-digest/gstack-AGENTS.md index 45e52d15a..5b748bf34 100644 --- a/agents-digest/gstack-AGENTS.md +++ b/agents-digest/gstack-AGENTS.md @@ -1,4 +1,4 @@ -# gstack digest v1.92.81.0 — regenerate/re-copy after upgrading gstack +# gstack digest v1.93.8.0 — 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/autoplan/SKILL.md b/autoplan/SKILL.md index 9dcdf7751..5413fe13a 100644 --- a/autoplan/SKILL.md +++ b/autoplan/SKILL.md @@ -67,9 +67,9 @@ If `SKILL_PREFIX` is `true` use `/gstack-*` names; paths stay `~/.claude/skills/ ## Question Tuning (skip entirely if `QUESTION_TUNING: false`) -Before each decision brief (AskUserQuestion or Conductor/fallback prose), choose `question_id` from `~/.claude/skills/gstack/scripts/question-registry.ts` or `{skill}-{slug}`, then run `~/.claude/skills/gstack/bin/gstack-question-preference --check ""`; for an unregistered id, write the question summary to `.gstack/tmp/qt.txt` (file-write tool) and append `--summary-file .gstack/tmp/qt.txt` (one-way keyword check). `AUTO_DECIDE` means choose the recommended option and say "Auto-decided [summary] → [option] (your preference). Change with /plan-tune." `ASK_NORMALLY` means ask. +Before each decision brief (AskUserQuestion or Conductor/fallback prose), choose `question_id` from `~/.claude/skills/gstack/scripts/question-registry.ts` or `{skill}-{slug}`, then run `~/.claude/skills/gstack/bin/gstack-question-preference --check ""`; for an unregistered id, write the question summary to `.gstack/tmp/qt.txt` (file-write tool) and append `--summary-file .gstack/tmp/qt.txt` (one-way keyword check); its `--signal-keys` prints an optional signal-key hint. `AUTO_DECIDE` means choose the recommended option and say "Auto-decided [summary] → [option] (your preference). Change with /plan-tune." `ASK_NORMALLY` means ask. -**Embed the question_id as a marker in every asked brief**, ad hoc IDs included, with one ID for check, marker and log. Include `` once in the question text itself, not only a command or log. On prose paths, use the explicit reply line. Without the marker, the PreToolUse hook treats AskUserQuestion as observed-only and never auto-decides. +**Embed the question_id as a marker in every asked brief**, ad hoc IDs included, with one ID for check, marker and log. Include `` once in the question text itself, not only a command or log. On prose paths, use the explicit reply line. Without the marker the hook never auto-decides. **Embed the option recommendation via the `(recommended)` label suffix** on exactly one option per AUQ. The PreToolUse hook parses it first, falls back to "Recommendation: X" prose, and refuses when ambiguous (two labels = refuse). diff --git a/autoplan/shadow-decider.md b/autoplan/shadow-decider.md new file mode 100644 index 000000000..aaafdfcbc --- /dev/null +++ b/autoplan/shadow-decider.md @@ -0,0 +1,76 @@ + +# Predict the owner's answer at an autoplan approval gate + +You are predicting how one specific person, the owner of this plan, will answer +the questions at the final approval gate of an autoplan review. You are not +deciding for him and nothing you write is shown to him. Your prediction is +recorded before he answers and later compared with what he actually chose. + +Be a forecaster, not an advocate. A good prediction is one that turns out to +match his answer. Do not argue for what you think is best for the plan. + +## What you have + +- The plan's Implementation plan section, as reviewed. +- Each gate question: its id, its class, its title, why it was asked, its + options (each with its text and whether choosing it changes the plan), the + option the reviewers recommended and their stated basis, and the review + findings it came from, with severity and a one-line summary. + +The reviewers' recommendation is one input among several. It is often right +and sometimes wrong. Owners override recommendations when a finding is weak, +when the recommended change adds scope they did not ask for, or when the +reviewers missed something about their goals. Weigh it; do not copy it. + +## How this owner tends to decide + +These are the six principles the owner's review process applies. Use them to +judge which option he is likely to pick, not as rules you must follow. + +1. Choose completeness: ship the whole thing; prefer the approach that covers + more edge cases. +2. Boil lakes: fix everything in the blast radius (files the plan modifies and + their direct importers) when it is small. +3. Pragmatic: when two options fix the same thing, pick the cleaner one, fast. +4. DRY: reject what duplicates existing functionality; reuse what exists. +5. Explicit over clever: a short obvious fix beats a large abstraction. +6. Bias toward action: merging beats more review cycles; flag concerns rather + than block. + +## For each question, predict + +- action: what the owner would do with it. + - `decide`: he answers it now by picking one option. + - `send_back`: he sends the plan back for more work before answering. + - `hold`: he leaves it open or asks questions first. +- pick: the option key he would choose (one of that question's option keys), + required for every action. For `send_back` or `hold`, give the option he + would most likely settle on afterwards. +- confidence: a number from 0 to 1, the probability that he picks exactly that + option. +- cites: the finding ids from that question's findings that most shaped your + prediction (an empty list when none did). Cite only ids listed for that + question. +- reason: one or two sentences, at most 400 characters. + +## The plan + +{{PLAN}} + +## The questions + +{{QUESTIONS}} + +## Output contract + +Reply with exactly one fenced `json` block holding one JSON array and nothing +else. The array has exactly one object per question id listed below, in any +order, with no duplicates and no other ids: + +{{IDS}} + +Each object has exactly these fields: + +```json +{"id": "", "action": "decide | send_back | hold", "pick": "