v1.99.4.0 release (#4363)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
This commit is contained in:
Garry Tanandcapy-ai[bot] authored and GitHub committed 2026-11-14 14:15:59 +00:00
1 parent 3c15b7210b
commit 81fb900835
11 files changed
+19 -25

No files matched your search

+16
View File
@@ -1,5 +1,21 @@
# Changelog
## [1.99.4.0] - 2026-11-14
- **`/health` and every check that picks gbrain's health probe now look for gbrain's config under `GBRAIN_HOME`, and gstack trims `GBRAIN_HOME` exactly as gbrain does.** The health selection in `lib/checks.ts` read only `~/.gbrain/config.json`, so on a machine where `GBRAIN_HOME` names the brain (gbrain's usual unattended setup) it judged a different brain's config: it could skip the gbrain check for the brain gbrain actually uses, or run it because of an unrelated `~/.gbrain`. It now uses the same rule as gbrain's own `configDir()`: `GBRAIN_HOME` trimmed with JavaScript `trim()`, with `.gbrain` appended, else `~/.gbrain`; a `GBRAIN_HOME` gbrain would refuse (relative, or with a `..` segment) selects nothing instead of falling back to `~/.gbrain`. gstack's TypeScript gbrain helpers (`gbrainConfigDir()`, which seeds gbrain's database URL for every gbrain call gstack makes) now trim and refuse the same way, and the shell helper `bin/gstack-gbrain-home.sh` now treats a `GBRAIN_HOME` made only of Unicode spaces such as a no-break space or a byte-order mark as blank, as gbrain does, instead of refusing it. Nothing changes when `GBRAIN_HOME` is unset.
- **The gbrain source wireup and skill start now find gbrain's config under `GBRAIN_HOME`, so the wireup no longer writes to the wrong brain.** Both helpers read only `~/.gbrain/config.json` (#4325). On a machine with two brains, where `GBRAIN_HOME` names one and `~/.gbrain` holds another, a wireup run with no database URL in its environment (the `/setup-gbrain` path, `gstack-brain-restore`, the daily advance) locked the `~/.gbrain` brain's `database_url` and sent `sources add` and `sync` there, and skill start never printed its `/sync-gbrain` hint when the only config was under `GBRAIN_HOME`. Both now resolve the path the way gbrain does, through one shared helper (`bin/gstack-gbrain-home.sh`): `GBRAIN_HOME` trimmed, with `.gbrain` appended, else `~/.gbrain`; a `GBRAIN_HOME` gbrain would refuse (relative, or with a `..` segment) locks nothing instead of falling back to `~/.gbrain`. To learn whether an earlier run could have hit the wrong brain, run `gstack-gbrain-source-wireup --check-target` with the same environment your wrapper sets. It prints `TARGET: MATCH` or `TARGET: MISMATCH`, reads the two configs only, and never prints a database URL.
- **When OpenAI refuses /ship's outside voice under its cyber-security policy, Capy's /ship now retries it instead of losing the second read.** A refusal that says `cyber_policy` is its own outcome now (`cyber_policy_refused`, docs/troubleshooting.md#outside-review-cyber-policy) instead of a generic failure. On Capy, the review pass sends the brief once more to GPT-6 Astra reworded as a read-only correctness review (no attacker role-play, no exploit code or reproductions, each defect as file:line, the wrong-answer condition and the smallest fix), and if Astra refuses too, once to Gemini 4 Carbon with the brief unchanged. If every attempt is refused, the Step 11 record says the second-family read is missing, the native review still runs, and the PR body lists the gap as missing coverage, not a pass. The first brief every outside voice gets is unchanged, other failures (auth, timeout, empty replies) are handled as before, and nothing else retries. The Codex CLI and api runners report the new reason but do not retry by themselves.
- **/ship's second-model review can be rendered on its own for a PR whose diff is over 2 MB again.** Since the cyber_policy retry change (#4350), the 2 MB review-size ceiling exempts the outside voice by its role, but `gstack-review-dispatch render` built its reader without one, so `render --role outside-voice` on a large PR stopped at `REVIEW_DIFF_TOO_LARGE` instead of writing its brief with a sliced diff. The standalone render now passes the role it was asked for. The ceiling's size, which readers it exempts and the pass manifest's own stages are unchanged, and every other reader on the standalone path is still refused over the ceiling.
- **/ship's test lane now runs test files under hidden folders such as `.github/scripts/` instead of failing on them.** The lane passed each selected test to `bun test` as a bare relative path. Bun 1.4 reads a bare path as a name filter, and that search skips hidden folders, so `bun test .github/scripts/x.test.mjs` found nothing, exited 1, and the pregate reported `PREGATE_LANE_FAILED` on a green test. That blocked /ship on any gbrain PR touching `.github/scripts/`. Every relative path now reaches `bun test` as an explicit path (`./` added unless it already starts with `./`, `../` or `/`): the local lane, the small-machine remote gate command, the docs-only reader tests and /ship's base-red check all share one helper with the CI carry's per-file command, which already did this. A failing test in a hidden folder still fails the lane.
- **A data file with a config-style name, such as a results `manifest.json`, now leaves /ship's review diff when the base branch's `review_exclude_paths` names it by file name.** Files named like tool config (`manifest.json`, `package.json`, `tsconfig*.json`, files under `config/`, dotfiles) were always kept in review as config, even under a pattern written for them, so a large eval results PR whose base policy listed `docs/benchmarks/*/results*/**/manifest.json` still handed every reader its 1.42 MB manifest and stopped at `REVIEW_DIFF_TOO_LARGE`. The kept manifest also counted as code that names files, so every receipt it listed by full path stayed in review as `named-by-code`. Now a matched JSON, CSV, TSV or image file is treated as data when the pattern that matches it ends in its exact file name with no glob character. Such a file is excluded, and its list no longer keeps other files. A config file under a broader pattern (`data/**`, `**/*.json`) is still kept as `config`, config files no pattern matches are unchanged, and every other keep rule (gate paths, executables, secret lines, gate tags, `.gstack/` and `.github/`) still applies first. The policy is still read only from the base branch.
- **/ship's reviewers now state how many findings block in one machine-readable line, and gstack checks that count instead of guessing from their prose.** Every blocking reader's brief asks for `REVIEW_VERDICT: blocking=<n>` (with optional `at=<path>:<line>` locations). When the count is higher than the findings gstack parsed, collect adds a CRITICAL row for each missing blocker, so a blocker written in a form the parser misses can no longer vanish. When it is lower, nothing is lowered automatically. Rows that block only because a sentence used the word CRITICAL or P0, or because the native adversarial voice wrote FIXABLE in plain prose (FOUN-298's 'all FIXABLE, nothing blocks'), become summary candidates the lead may lower with a written reason. A missing or malformed line changes nothing. Ledgers written before this change replay exactly as before. See docs/review-policy.md#verdict-line.
## [1.99.3.0] - 2026-11-13
- **/ship's documentation audit no longer refuses Next.js and SvelteKit route files.** When the Step 14.5 docs audit listed a reviewed or updated file whose path has square brackets, such as `app/api/items/[id]/route.ts`, `[...slug]` or `[[...opt]]`, the result check read the brackets as a glob, refused the result with `DOCS_RESULT_INVALID` on both attempts, and stopped an unattended /ship at Blocked recovery even though the audit found nothing wrong (#4313). A listed path that holds `*`, `?`, `[` or `]` now counts as a file when that exact file exists from the repository root, so file-routed projects ship normally. A pattern such as `docs/*.md` or `docs/[ab].md`, or a bracketed path that names no file, is still refused as a glob.
+1 -1
View File
@@ -1 +1 @@
1.99.3.0
1.99.4.0
+1 -1
View File
@@ -1,4 +1,4 @@
# gstack digest v1.99.3.0 — regenerate/re-copy after upgrading gstack
# gstack digest v1.99.4.0 — 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,
@@ -1 +0,0 @@
- **`/health` and every check that picks gbrain's health probe now look for gbrain's config under `GBRAIN_HOME`, and gstack trims `GBRAIN_HOME` exactly as gbrain does.** The health selection in `lib/checks.ts` read only `~/.gbrain/config.json`, so on a machine where `GBRAIN_HOME` names the brain (gbrain's usual unattended setup) it judged a different brain's config: it could skip the gbrain check for the brain gbrain actually uses, or run it because of an unrelated `~/.gbrain`. It now uses the same rule as gbrain's own `configDir()`: `GBRAIN_HOME` trimmed with JavaScript `trim()`, with `.gbrain` appended, else `~/.gbrain`; a `GBRAIN_HOME` gbrain would refuse (relative, or with a `..` segment) selects nothing instead of falling back to `~/.gbrain`. gstack's TypeScript gbrain helpers (`gbrainConfigDir()`, which seeds gbrain's database URL for every gbrain call gstack makes) now trim and refuse the same way, and the shell helper `bin/gstack-gbrain-home.sh` now treats a `GBRAIN_HOME` made only of Unicode spaces such as a no-break space or a byte-order mark as blank, as gbrain does, instead of refusing it. Nothing changes when `GBRAIN_HOME` is unset.
-4
View File
@@ -1,4 +0,0 @@
---
important: data-loss
---
- **The gbrain source wireup and skill start now find gbrain's config under `GBRAIN_HOME`, so the wireup no longer writes to the wrong brain.** Both helpers read only `~/.gbrain/config.json` (#4325). On a machine with two brains, where `GBRAIN_HOME` names one and `~/.gbrain` holds another, a wireup run with no database URL in its environment (the `/setup-gbrain` path, `gstack-brain-restore`, the daily advance) locked the `~/.gbrain` brain's `database_url` and sent `sources add` and `sync` there, and skill start never printed its `/sync-gbrain` hint when the only config was under `GBRAIN_HOME`. Both now resolve the path the way gbrain does, through one shared helper (`bin/gstack-gbrain-home.sh`): `GBRAIN_HOME` trimmed, with `.gbrain` appended, else `~/.gbrain`; a `GBRAIN_HOME` gbrain would refuse (relative, or with a `..` segment) locks nothing instead of falling back to `~/.gbrain`. To learn whether an earlier run could have hit the wrong brain, run `gstack-gbrain-source-wireup --check-target` with the same environment your wrapper sets. It prints `TARGET: MATCH` or `TARGET: MISMATCH`, reads the two configs only, and never prints a database URL.
@@ -1,4 +0,0 @@
---
important: blocker
---
- **When OpenAI refuses /ship's outside voice under its cyber-security policy, Capy's /ship now retries it instead of losing the second read.** A refusal that says `cyber_policy` is its own outcome now (`cyber_policy_refused`, docs/troubleshooting.md#outside-review-cyber-policy) instead of a generic failure. On Capy, the review pass sends the brief once more to GPT-6 Astra reworded as a read-only correctness review (no attacker role-play, no exploit code or reproductions, each defect as file:line, the wrong-answer condition and the smallest fix), and if Astra refuses too, once to Gemini 4 Carbon with the brief unchanged. If every attempt is refused, the Step 11 record says the second-family read is missing, the native review still runs, and the PR body lists the gap as missing coverage, not a pass. The first brief every outside voice gets is unchanged, other failures (auth, timeout, empty replies) are handled as before, and nothing else retries. The Codex CLI and api runners report the new reason but do not retry by themselves.
@@ -1,4 +0,0 @@
---
important: regression
---
- **/ship's second-model review can be rendered on its own for a PR whose diff is over 2 MB again.** Since the cyber_policy retry change (#4350), the 2 MB review-size ceiling exempts the outside voice by its role, but `gstack-review-dispatch render` built its reader without one, so `render --role outside-voice` on a large PR stopped at `REVIEW_DIFF_TOO_LARGE` instead of writing its brief with a sliced diff. The standalone render now passes the role it was asked for. The ceiling's size, which readers it exempts and the pass manifest's own stages are unchanged, and every other reader on the standalone path is still refused over the ceiling.
@@ -1,4 +0,0 @@
---
important: blocker
---
- **/ship's test lane now runs test files under hidden folders such as `.github/scripts/` instead of failing on them.** The lane passed each selected test to `bun test` as a bare relative path. Bun 1.4 reads a bare path as a name filter, and that search skips hidden folders, so `bun test .github/scripts/x.test.mjs` found nothing, exited 1, and the pregate reported `PREGATE_LANE_FAILED` on a green test. That blocked /ship on any gbrain PR touching `.github/scripts/`. Every relative path now reaches `bun test` as an explicit path (`./` added unless it already starts with `./`, `../` or `/`): the local lane, the small-machine remote gate command, the docs-only reader tests and /ship's base-red check all share one helper with the CI carry's per-file command, which already did this. A failing test in a hidden folder still fails the lane.
@@ -1,4 +0,0 @@
---
important: blocker
---
- **A data file with a config-style name, such as a results `manifest.json`, now leaves /ship's review diff when the base branch's `review_exclude_paths` names it by file name.** Files named like tool config (`manifest.json`, `package.json`, `tsconfig*.json`, files under `config/`, dotfiles) were always kept in review as config, even under a pattern written for them, so a large eval results PR whose base policy listed `docs/benchmarks/*/results*/**/manifest.json` still handed every reader its 1.42 MB manifest and stopped at `REVIEW_DIFF_TOO_LARGE`. The kept manifest also counted as code that names files, so every receipt it listed by full path stayed in review as `named-by-code`. Now a matched JSON, CSV, TSV or image file is treated as data when the pattern that matches it ends in its exact file name with no glob character. Such a file is excluded, and its list no longer keeps other files. A config file under a broader pattern (`data/**`, `**/*.json`) is still kept as `config`, config files no pattern matches are unchanged, and every other keep rule (gate paths, executables, secret lines, gate tags, `.gstack/` and `.github/`) still applies first. The policy is still read only from the base branch.
-1
View File
@@ -1 +0,0 @@
- **/ship's reviewers now state how many findings block in one machine-readable line, and gstack checks that count instead of guessing from their prose.** Every blocking reader's brief asks for `REVIEW_VERDICT: blocking=<n>` (with optional `at=<path>:<line>` locations). When the count is higher than the findings gstack parsed, collect adds a CRITICAL row for each missing blocker, so a blocker written in a form the parser misses can no longer vanish. When it is lower, nothing is lowered automatically. Rows that block only because a sentence used the word CRITICAL or P0, or because the native adversarial voice wrote FIXABLE in plain prose (FOUN-298's 'all FIXABLE, nothing blocks'), become summary candidates the lead may lower with a written reason. A missing or malformed line changes nothing. Ledgers written before this change replay exactly as before. See docs/review-policy.md#verdict-line.
+1 -1
View File
@@ -1,6 +1,6 @@
{
"name": "gstack",
"version": "1.99.3",
"version": "1.99.4",
"description": "Garry's Stack — Claude Code skills + fast headless browser. One repo, one install, entire AI engineering workflow.",
"license": "MIT",
"type": "module",