Repository navigation
Releases
v1.6.0
Lessons from running the kit's rules in a multi-repo setup: db-guard read only the first SQL field of
a call, the audit read an unreachable registry as a clean report, a promotion had no way to skip a
scan the pull request before it already ran, and checks crashed on a path another session deleted
from a shared checkout. The guards then gained what running them from a workspace root asked for:
unlocks and env helpers that work across repos, fewer false refusals, and an OpenAPI contract or a
first migration that is guarded wherever it lands.
Reusable workflows and actions
Added
- The quality gate takes
promotion, one exact<head>:<base>pair such asdev:prod. On that pair
alone the Skill Security Scan stands down, uv is not installed for it, and the summary names what
stood down. The pair keys on both ends, and the workflows pass no head for a fork, so a skip can
never follow the base alone. Empty, the default, changes nothing. - The iOS static job runs
terms.mjs.
Fixed
- strip-ai's back-merge re-instates only the strip set. It restored every path the merge deleted, so a
revert on the production branch half-applied to the development branch, and it kept the development
copy of every conflict, so the stop for a conflict outside the stripped paths could never fire. Any
such deletion or conflict now stops the run with the development branch as it was.
agent-core 1.4.0
Added
- Workspace mode reaches a session at the workspace root. With
AGENT_WORKSPACE_ROOTset, that
session can run each repo'sscripts/env/helpers,set.shanswers to the unlock of the repo that
owns it and its refusal names that repo, and any repo'sdbunlock opens db-guard there. A session
inside a repo keeps to its own unlock, and without the setting nothing changes. generatedPathstakes**/name, a file or folder at any depth. A folder-widereplace_in_files
finds those files through git, so an ignored folder such asnode_modulesis never walked.scripts/check/terms.mjs, the words a repo has decided not to use about itself. The vocabulary is
the repo's own, inscripts/check/terms.config.json, and starts empty, so the check passes and says
so until you list a word. An allow entry names the exact phrase, its one file and a reason; an entry
whose phrase left its file fails. It reads the index, not the worktree. A repo set up before this
version adds the linenode scripts/check/terms.mjs(kindall) toscripts/check/gates.list.
Changed
- Breaking: the
generatedPathsdefault guards the OpenAPI contract at any depth
(**/openapi.json,**/openapi.yaml,**/openapi.yml), not only at the repo root. A repo that
writes a spec below its root by hand and relies on the defaults sets its own list without it. - Only
show.sh,set.sh,envfile.pyand thescripts/env/folder itself are locked. A project's
own tool beside them is ordinary code again:show.shandset.shstartenvfile.pywith
python3 -I, which imports nothing from that folder. /agent-core:promoteresolves each thread it answered, a declined one included, as
/agent-core:resolve-pr-reviewalready did:pr-ready.shcounts unresolved threads, so an answered
thread left open held the promotion.
Fixed
- db-guard judges every SQL field of a call (
sql,query,statement, each ofqueries). It read
only the first, so a read insqllet a write inquerythrough. - db-guard's default
dbWriteGuard.toolPatterncoversexplain_queryandanalyze_query_indexes,
which run the SQL they are handed. A repo that sets its own pattern keeps it. - db-guard treats a semicolon, even a trailing one, as a write: it is where a second statement would
start. The refusal says to send one statement without it. - safety-check no longer reads a
$that expands nothing as an unresolved word, so a pattern that
only looks for a.envname (grep -E "(^|/)\.env(\.|$)") passes. zsh's$=x,$~x,$^x,$+x
and$#xstill count. - A glob is judged by what it can match. A
[with no closing]is a plain character, so jq's
.[].nameno longer counts as a.env*read;.[e]nv,.e*and.env.[s]tagingstill do. - With neither python3 nor jq, the plain-text rules refuse
rm -ron every default protected path.
They knew six of them, soapp,components,content,tests,.agent,.agentsand the
top-level instruction files could be deleted.
agent-fe-nextjs 1.5.0
Added
gates.listrunsterms.mjs.
Changed
- Breaking: generated-guard follows the new
generatedPathsdefault of agent-core 1.4.0: a repo
with no list of its own that writes a spec below its root by hand sets one. - generated-guard takes
**/nameingeneratedPaths.lib.shfollows agent-core 1.4.0.
Fixed
scripts/check/audit.tsrefuses a run with no report.bun auditwrites nothing and exits 1 on an
unreachable registry, which read as no advisories.audit.tsholds anACCEPTEDlist: an advisory stays accepted only while npm has no patched
release, and an unreadable registry fails it.check-comment-style.tsskips a tracked path gone from the worktree instead of crashing.
agent-docs-nextra 1.4.0
Added
gates.listrunsterms.mjs.
Changed
- generated-guard as in agent-fe-nextjs 1.5.0. The list setup seeds is unchanged.
lib.shfollows
agent-core 1.4.0.
Fixed
- The
audit.tsandcheck-comment-style.tsfixes in agent-fe-nextjs 1.5.0.
agent-ai-fastapi 1.5.0
Added
gates.listrunsterms.mjs.
Changed
- Breaking: migration-guard judges a path whether or not its folder exists yet, so the first
generated migration comes from its tool too. A repo that keeps something else under a default
migrationsDirsfolder sets its own list, and"migrationsDirs": []empties it. lib.shfollows agent-core 1.4.0.
agent-be-hono 2.2.0
Added
gates.listrunsterms.mjs.
Changed
- Breaking: migration-guard judges a path whether or not its folder exists yet, as in
agent-ai-fastapi 1.5.0. lib.shfollows agent-core 1.4.0.
agent-be-axum 0.3.0
Added
gates.listrunsterms.mjs.
Changed
- migration-guard judges a path whether or not its folder exists yet. For
appendOnlyMigrationsDirsthat means a file a protected branch has stays refused on a branch that
lost the folder. - generated-guard as in agent-fe-nextjs 1.5.0. The list setup seeds is unchanged.
lib.shfollows
agent-core 1.4.0.
Fixed
rust-source.mjsskips a tracked path gone from the worktree instead of crashing.
agent-be-go 0.3.0
Added
gates.listrunsterms.mjs.
Changed
- migration-guard as in agent-be-axum 0.3.0, generated-guard as in agent-fe-nextjs 1.5.0. The list
setup seeds is unchanged.lib.shfollows agent-core 1.4.0.
agent-ios-swiftui 1.2.0
Added
gates.listrunsterms.mjs.
Changed
- generated-guard as in agent-fe-nextjs 1.5.0. The list setup seeds is unchanged.
lib.shfollows
agent-core 1.4.0.
Fixed
i18n-catalog.mjsskips a tracked catalog gone from the worktree instead of failing.
agent-fe-nextjs-static 1.2.0
Added
gates.listrunsterms.mjs.
v1.5.3
The deploy caller is pinned to the v1.5.2 release commit, so setup and sync install the fresh-runner
retry.
agent-deploy 1.2.1
Fixed
- The deploy caller is pinned to the v1.5.2 release commit: when the webhook fails from one runner,
Deploy (fresh runner)retries from a new one.
v1.5.2
Reusable workflows and actions
Fixed
- deploy-webhook retries from a fresh runner. Twice, one GitHub runner timed out on all four webhook
attempts to a VPS that logged no dropped packet, firewall block or load, while a rerun on another
runner was accepted at once: the runner's address did not reach the host. The deploy job now
reports the webhook's outcome instead of failing, and a second job,Deploy (fresh runner), runs
the same step on a new runner when it failed. The docs dispatch goes out from whichever job
deployed.
v1.5.1
The strip-ai caller is pinned to the v1.5.0 release commit and passes the optional push-key secret.
agent-deploy 1.2.0
Added
- The strip-ai caller passes
push-key: ${{ secrets.STRIP_AI_PUSH_KEY }}. Store a write deploy key's
private key asSTRIP_AI_PUSH_KEYand the strip and the back-merge push over SSH, so a ruleset
that requires the quality gate onprodanddevcan let deploy keys bypass it. Without the
secret nothing changes.
Fixed
- The strip-ai caller is pinned to the v1.5.0 release commit.
v1.5.0
Two blockers met while syncing an app's repos to 1.4.1. safety-check refused agent-core's own
agent-sync plan and check in every repo already set up, since the engine reads its templates from
~/.claude/plugins/, a guard area; every plugin's /sync failed at its first step. And a repository
ruleset could not require the quality gate on prod and dev without stopping strip-ai's pushes,
because GitHub Actions cannot be a bypass actor there; deploy keys can.
Reusable workflows and actions
Added
- strip-ai takes an optional
push-keysecret, the private key of a deploy key with write access.
With it, the strip commit and the back-merge are pushed over SSH, trusting only the host keys
GitHub publishes in its meta API, so a ruleset that requires status checks on both branches and
lets deploy keys bypass it accepts them. The key lives in a private file for the one job and is
removed afterwards. Without it nothing changes.
agent-core 1.2.3
Fixed
- safety-check lets the setup engine read:
agent-sync checkandplan,agent-setup questions
andplan, run from a plugin'sbin/(under.claude/plugins/, or the running plugin's own), may
name the plugin's templates.applyandownstill refuse, since they rewrite the lock and the
guard settings. A./agent-syncof the project's, one found through aPATHset or exported on
the line, or a function or alias of that name, is not the engine. Twelve new probe rows cover both
sides; the four that must pass fail on 1.2.2's library. - setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-deploy 1.1.5
Fixed
- setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-be-hono 2.0.2
Fixed
- The hook library (
scripts/lib.sh) is agent-core 1.2.3's: safety-check lets the setup engine's
read-only modes name the plugin's templates, so this plugin's/syncdrafts again in a repo that
is already set up. - setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-be-go 0.1.3
Fixed
- The hook library (
scripts/lib.sh) is agent-core 1.2.3's: safety-check lets the setup engine's
read-only modes name the plugin's templates, so this plugin's/syncdrafts again in a repo that
is already set up. - setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-be-axum 0.1.3
Fixed
- The hook library (
scripts/lib.sh) is agent-core 1.2.3's: safety-check lets the setup engine's
read-only modes name the plugin's templates, so this plugin's/syncdrafts again in a repo that
is already set up. - setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-ios-swiftui 1.0.2
Fixed
- The hook library (
scripts/lib.sh) is agent-core 1.2.3's: safety-check lets the setup engine's
read-only modes name the plugin's templates, so this plugin's/syncdrafts again in a repo that
is already set up. - setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-ai-fastapi 1.3.2
Fixed
- The hook library (
scripts/lib.sh) is agent-core 1.2.3's: safety-check lets the setup engine's
read-only modes name the plugin's templates, so this plugin's/syncdrafts again in a repo that
is already set up. - setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-docs-nextra 1.2.2
Fixed
- The hook library (
scripts/lib.sh) is agent-core 1.2.3's: safety-check lets the setup engine's
read-only modes name the plugin's templates, so this plugin's/syncdrafts again in a repo that
is already set up. - setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-fe-nextjs 1.3.2
Fixed
- The hook library (
scripts/lib.sh) is agent-core 1.2.3's: safety-check lets the setup engine's
read-only modes name the plugin's templates, so this plugin's/syncdrafts again in a repo that
is already set up. - setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-fe-nextjs-static 1.1.4
Fixed
- setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
agent-fe-threejs 1.0.4
Fixed
- setup and sync say what to do when safety-check refuses
applyorownin a repo that is
already set up: hand the user the exact line to run with!.
v1.4.1
The ios-swiftui quality-gate caller is pinned to the v1.4.0 release commit, so setup and sync
install the gate that builds first and tests last.
agent-ios-swiftui 1.0.1
Fixed
- The quality-gate caller is pinned to the v1.4.0 release commit: the macOS job runs
build-for-testing, the String Catalog check, Periphery and the client drift check before
test-without-building, so a failing or timed-out test no longer hides them.
v1.4.0
agent-ios-swiftui leaves 0.x: an app built with it reached production, and the lessons of that run
are in this release. The iOS gate builds first and tests last, so a failing or timed-out test no
longer hides what the build-dependent checks found.
Reusable workflows and actions
Changed
- The iOS gate's macOS job runs
xcodebuild build-for-testing(zero warnings, read from its own
result bundle), then the String Catalog check against the build, Periphery and the API client
drift check, and only thenxcodebuild test-without-building. Before, a failing test skipped
those checks, and a timeout cancelled the job before they ran: one app carried a stale catalog key
and four unused properties for days behind a gate that never reached them.
agent-ios-swiftui 1.0.0
Changed
- Stable: the plugin is no longer marked experimental. A rename, or a removed rule, check or
setting, is now a major release. scripts/ios/app.shtakesIOS_SIMULATOR_ID, and otherwise picks an iPhone of the newest runtime
that is not booted. A booted simulator is usually the one a developer is using: the unsigned test
build replaced their signed app, whose Keychain then refused every write
(errSecMissingEntitlement, -34018).- The I4 finding for an extracted key no code uses says what usually causes it: a key written
inside a ternary or a tuple, which the build does not extract. Mark it manual. - The quality-gate caller and
.claude/rules/ios/testing.mdexplain the CI scheme (<App>CI): the
same app and UI test target with only the quick UI tests, once the accessibility audits outgrow
the hosted macOS runner. The audits run on a developer's simulator. AGENTS.md Rule 24's row says
so.
Added
- Anti-patterns:
full-accessibility-audit-outlasts-hosted-runner.md(90 minutes on about 120
screens, contrast sampled mid-fade),periphery-reads-a-stale-index.md(an incremental build
keeps old index records; never cacheDerivedDatain CI) and
string-literal-in-ternary-is-not-extracted.md.
agent-core 1.2.2
Fixed
scripts/check/ai-config.shno longer reports a rule AGENTS.md defines as undefined. It piped
printfintogrep -qunderpipefail:grep -qexits at its first match,printfdies of
SIGPIPE, and the pipeline read as "not found", now and then, on a busy runner.
scripts/sync/workflows.shhad the same pipe. Both read a here-string now, and a test refuses the
pattern in every template script that setspipefail.
Changed
/agent-core:helpno longer marks agent-ios-swiftui experimental.
v1.3.5
The ios-swiftui quality-gate caller is pinned to the v1.3.4 release commit, so setup and sync
install the simulator boot fix.
agent-ios-swiftui 0.1.4
Fixed
- The quality-gate caller is pinned to the v1.3.4 release commit: the macOS job boots the simulator
and waits for it beforexcodebuild test, instead of failing on a cold simulator that cannot
launch the UI test runner.
v1.3.4
Reusable workflows and actions
Fixed
- The iOS gate boots the simulator and waits for it (
xcrun simctl bootstatus <udid> -b) before
xcodebuild test. Left to xcodebuild, a cold simulator on a busy macOS runner sometimes failed
to launch the UI test runner ("Failed to launch", "operation never finished bootstrapping") after
ten minutes, a runner fault that read as a failing test. One app hit it in 4 of its first 20 runs.
A simulator that does not boot now fails the gate with that message, before the build.
v1.3.3
The callers whose reusable workflows 1.3.2 fixed are pinned to its release commit, so setup and
sync install those fixes. The Workflows Lint docs add a prerequisite: an Actions policy that allows
only selected actions must allow docker://rhysd/actionlint:* and zizmorcore/zizmor-action@*, or
every run ends in Startup failure.
agent-ios-swiftui 0.1.3
Fixed
- The quality-gate caller is pinned to the v1.3.2 release commit: the static job no longer runs
swiftlint-static, and the macOS job has gitleaks forsecrets.sh.
agent-deploy 1.1.4
Fixed
- The strip-ai caller is pinned to the v1.3.2 release commit, whose back-merge no longer stops on
index.lockwhen the development branch edited a stripped file since the last strip.