Repository navigation
Releases
VibeTags v1.4.3
[1.4.3] - 2026-11-21
Upgrade note: no generated file moves, but a build that opts into llms-full.txt gets a new
warning. The consumer sweep built five downstream repositories at 1.4.3. codekarta, skill3 and
blindbean generated the same files byte for byte as the same builds at 1.4.2; async-test-lib's only
change is the ~roots= sidecar line 1.4.2 introduced, and common-license-lib's files did not move.
A build that opts into llms-full.txt now prints one compiler WARNING, which fails a build compiled
with -Werror (see Deprecated). A module that sets -Avibetags.module and mirrors into a target
writes its mirrored rule files under the new name, and the old ones need deleting by hand (#1076,
under Fixed).
Changed
-
CI builds on JDK 27 (#1117). The Maven and Gradle jobs' matrices are now 21, 25, 26 and 27;
JDK 27 is the current release. Measured locally on Temurin 27.0.1 before adding the legs: the
annotations module verifies, the processor's fast tier passes (3,347 tests, Error Prone and
NullAway included), and Gradle 9.9.1 builds both modules. -
AIGuardrailProcessorspells diagnostic paths throughPathStrings(internal; no output changes,
#1101). Five of its six hand-written.replace('\\', '/')conversions now call
PathStrings.relativeorPathStrings.forwardSlashes. The sixth is inside the@AILocked
generateFiles()and stays until the owner approves an edit there. Mutating each site first
showed three of the five were tested by nothing, soDiagnosticPathSeparatorTest(the
opted-in-after and module-id-collision warnings) and aPartialRoundGuardrailLossTestcase pin
them; each fails on Windows when its site's conversion is removed. The new import moves the
locked range in.vibetags-locksby one line, to 1254-1569. -
GuardrailFileWriter.writeContentWithBackupisstageAndMove(internal; no output changes,
#1109). It stages content in a temp file beside the target and moves it into place, and never
made a backup, so the name described something it did not do. The two hand-written PlantUML
diagrams anddocs/architecture/design-decisions.mdfollow; the code-karta diagrams show no
structural change. -
The release steps count the organization's month on Maven Central, not this repository's tags
(#863). Central's limits (7 releases, 1,167 files, 78 MB a month) are per organization, and
async-test-lib and every other group underse.deversitydraw on them too, so the tag count in
docs/RELEASING.mdstep 0c and the release skill's step 1b missed most of the usage: on
2026-11-17 it showed 1 release for November where the organization had 4 and 534 of its 1,167
files. The new.github/scripts/central-month-usage.shreads repo1's public directory listings,
counts one release per group and version published that month, sums their files and bytes, and
exits 1 once a limit is reached and 2 when repo1 cannot be read (never zero). A version published in
several deployments counts once per deployment (#1111, below), so October 2026 reads 11, as
Central counted it. Both release steps run it; the cadence of at most 3 vibetags releases a month stays.
central-month-usage.test.shruns it against fixture listings inbuild.yml's Central Bundle
Shape job, andCentralPublishingBudgetTestfails if either release step or that CI step stops
running it. The test was written first and failed for each. -
The
modelpackage holds data only (internal; no output changes). Logic that lived in the
compiler-free model moves to the layer that uses it, one commit per class, each behind tests
written first and shown red against a mutant of the behaviour they pin.RoleConfig, which
reads.vibetags-roles, compiles its globs and memoises matches, is a service rather than data
and now lives inprocessor.internal.content, as doesContentHash, the fingerprint hash
function.TaggedElement.ownsSourceFile()becomesElementFiles.ownsSourceFile(element)beside
the formatters that call it, and the element's file-name defaults are resolved in its builder,
so its constructor only assigns.ElementTagkeeps its constants only:fromName,isClassand
isInterfacebecome static methods onElementKinds.GuardrailModel's derived views
(safetyOnly,withoutSafety,excluding,labeledSets) becomeGuardrailModelViewsin
processor.internal. The builder that sorts and derives a model becomesGuardrailModelBuilder,
andGuardrailModelkeeps fields and accessors only; anArchitectureRulesTestrule lets no
other class call its constructor, so no model can skip the sort that keeps generated output
identical on every machine.TransitiveRuleis no longerComparable: its attribute summary
and sort order becomeTransitiveRules.memberSummaryandTransitiveRules.ORDER.
GuardrailAnnotations.labelbecomesAnnotationLabels.label, leaving the registry as constants.
Generated files are unaffected. -
GuardrailModelBuilder.BY_IDENTITYis@AILocked(internal). The comparator fixes the order
of every bucket, so a change to it rewrites every consuming build's generated files and fingerprint
with nothing failing to name the cause. Its lock entry is always loaded, so the budget
ProjectFactsConsistencyTestholds this repository's ownCLAUDE.mdblock to rises from 7,000 to
7,300 bytes; the block is 7,296. -
PlatformRendererRegistry's four service-key lookups share one helper (internal; no output
changes).mergeShapeFor,sourceSetMergeFor,filePrologueForandwholeFileMergeFor
each repeated the same key-to-platform-to-renderer resolution with its own null checks; they now
call one privaterendererFor(serviceKey).PlatformRendererRegistryServiceKeyTest, written
first and green on the old code, pins that each lookup returns what the platform's renderer
declares for all 61 platforms and answers nothing forroot_index, an unknown key and"".
Three deliberate breakages of the new code (a droppedsourceSetMerge, a non-empty prologue for
an unknown key, the helper resolving the wrong platform) failed 26, 3 and 60 of its 64 cases. -
ReactorRootDetector.findReactorRootAbovereturnsOptional<Path>instead of a nullable
Path(internal; no output changes). Its one caller, the #296 detached-module warning in
AIGuardrailProcessor, no longer null-checks: it hands the root to a newwarnDetachedFrom
throughifPresent. The privaterelativisehelper returnsOptional<String>the same way.
A newDetachedModuleWarningTestcase, written first and green on the old code, pins the module
name (group/module-core, forward slashes) and the-Avibetags.root=value the warning prints for
a nested module. Two deliberate breakages of the new code failed it: the detector answering the
module's parent instead of the declaring reactor (3 failures), and the module name keeping
Windows backslashes (1). -
Mirror discovery and the sibling-module search share one directory walk that returns its result
(internal; no output changes).MirrorConfig.collectandMirrorWriter.collectModuleswere two
copies of the same bounded, sorted, depth-first walk, each filling a list passed in by its caller.
Both now useMirrorConfig.directoriesBelow(dir, levels), which returns the directories, and
filter the result. Theseen-set cycle guard is gone: every path in the walk is its parent plus one
name, so the set never matched, and the depth bound is what stops a symlink cycle. New tests pin
discovery order (depth-first, siblings sorted) and the sibling search's limits (three levels, own
build file,target/andsrc/skipped); they passed before the change. Breaking the new walk
failed them: breadth-first order (1 failure), one level short (3), and counting directories with no
build file of their own as modules (4). -
Json's character tests are named predicates (internal; no output changes). The parser for
transitive manifests compared characters inline:c == ' ' || c == '\t' || ...for whitespace and a
six-way test for number characters. They are nowisJsonWhitespace,isNumberChar,
isExponentMarkerandisSign, each saying why it is not theCharactermethod it resembles.
NewJsonTestcases, green before the change, pin RFC 8259's ASCII-only sets: form feed, vertical
tab, no-break and Unicode spaces are not whitespace, and non-ASCII decimal digits are not digits.
Swapping inCharacter.isWhitespacefailed 4 of them andCharacter.isDigit3, so the tempting
"simplification" cannot widen what a third-party JAR's manifest may contain unnoticed. -
Two hand-kept agreements are now enforced by the code (internal; no output changes).
GuardrailEnforcerheld its guarded elements twice, as a map of signatures and a map of elements
"filled in lockstep", with an unreachable null branch guarding their agreement; it now keeps the
elements only and reads each one's own signature.MirrorWriter's sibling search depth was the
literal 3, documented as "one level past mirror discovery"; it is nowMirrorConfig.MAX_DEPTH + 1
(Clean Code G22, make logical dependencies physical). Six new tests, green before the change, pin
the enforcer's report order, an element in two families, constructor naming and the mirror depth
relation; breaking the one-map version failed 2 to 18 of the enforcer tests, and raising the
discovery depth now fails the sibling search's depth test as well. -
MethodBodyGuardrailScanner.topLevelasks one named question (internal; no output changes). Its
loop condition readcurrent.getEnclosingElement().getKind()four times in one boolean (Clean Code
G36, G28). It is nowwhile (isNested(current)), withisNestedreading the enclosing element's
kind once.MethodBodyGuardrailScannerTopLevelTest, 59 cases written first and green on the old
loop, pins w...
VibeTags v1.4.2
[1.4.2] - 2026-11-14
Upgrade note: if you commit your .vibetags-mod-* sidecars, each one gains a line on the first
build. It is a ~roots= entry, the module's source roots in base64 (see the partial-round fix
under Fixed). Nothing else in the generated files moves: the consumer sweep built five downstream
repositories at 1.4.2, and every other generated file matched the 1.4.1 output byte for byte, apart
from one codekarta change that came from its own new annotation. Commit the new line; an older
processor skips it.
Added
- CI verifies a workflow dependency lockfile. Every action was already pinned to a commit
SHA, but GitHub resolves a SHA across the fork network, so a fork-only commit could run under
the upstream action's name, and the actions our actions call were pinned by nobody here.
.github/workflows/actions.locknow records each action's tag, owner and repository IDs,
including two called indirectly (actions/github-scriptvia codecov-action,oven-sh/setup-bun
via claude-code-action), and the newactions-lock.ymlworkflow runsgh actions-lock --verify
on every pull request. Verified locally: the check exits 1 with no lockfile and with auses:
SHA the lockfile lacks, exits 2 on a hand-edited lockfile entry, and 0 on this tree. Generated
with--no-narrow, which keeps every workflow's SHA pins as they were.
Changed
- The two local actions are locked too (#1031).
build.ymlreferred to
./.github/actions/verify-generated-filesand./action/locked-files, whichgh actions-lock
does not follow and warned about on every run, so a remote action either of them gained would
have run unlocked. Migrated withgh actions-lock --no-narrow --migrate-local-actionsto the
uses: $/…form, which it follows;--verifyreports all 15 workflows valid. The consumer-facing
PIsberg/vibetags/action/locked-files@<tag>path does not change. - A documented rule for the Central resolution flake (#1041).
docs/WORKFLOW.mdnames the
message, the three runs that hit it (two Maven jobs onmaven-pmd-plugin:3.28.0, one Gradle job
onpmd-java:7.29.0, each green on rerun) and the rule: rerun, do not change code. At about one a
week a rerun costs less than a resolver change nobody has shown would help. - Dependabot bumps of a mirrored version can arrive green (#1038). The new
dependabot-mirrors.ymlrunstools/sync-version-mirrors.pyon Dependabot's pull requests that
touch the parent pom and pushes the synced copies with a GitHub App token, so the push starts CI.
It needs a GitHub App with Contents write and itsMIRROR_SYNC_APP_IDandMIRROR_SYNC_APP_KEY
as Dependabot secrets; until those exist it warns and changes nothing. Seen on #1024, where a
ksp.versionbump failed 5 jobs onBuildVersionParityTest. Not yet run against a real bump. - Routine dependency bumps (#1033, #1043-#1051), landed as one PR. build-helper-maven-plugin
3.6.2 -> 3.7.0, flatten-maven-plugin 1.8.0 -> 1.9.0, pitest-maven 1.30.1 -> 1.30.2,
async-test-lib 1.14.0 -> 1.14.1,@anthropic-ai/claude-code(evals) 2.1.317 -> 2.1.319, and the
actions setup-java 6.2.0, harden-runner 2.22.4, action-gh-release 3.0.4, claude-code-action
1.0.269 and the codeql-action group. Both MojoHaus plugins now require Maven 3.9.0; the build
andpublish.yml(Maven 3.10.0) already meet it. - PMD 7.28.0 -> 7.29.0 (#1025, #1039). 7.29 fixes six security advisories, among them one in CPD
run over untrusted source (GHSA-wj6q-xw9m-26p6), which CI does on every pull request. It also
addsLocalVariableDeclarationShouldBeAtStartOfBlockto the codestyle category, which
pmd-ruleset.xmlimports whole. It fired 2826 times in the processor module, and the dependabot
PR failed 9 jobs on it. The rule is excluded rather than obeyed: it wants every local declared
uninitialised at the top of its block, the reverse of the excluded
VariableDeclarationUsageDistance, and avarcannot be declared that way. With the exclusion,
pmd:checkpasses on PMD 7.29.0 invibetags-annotations,vibetags,vibetags-cli,
vibetags-kspandload-tests. - The Inquisitor runs on request, not on every pull request.
inquisitor.ymlnow triggers on
workflow_dispatch(Actions, Inquisitor, Run workflow) with abaseinput, defaultmain. A
push or an opened PR no longer starts a model review. The Copilot review lane is unchanged. - The dropped-processor warning is measured in the Eclipse IDE (#990). Eclipse 4.24 with
VibeTags on the factory path and Lombok 1.18.42 as-javaagent, built headless by a small
workspace application: no warning, the Lombok getter resolves and the guardrail block is written.
Without the agent the warning names@lombok.Dataand the getter is undefined, so the check ran.
The debug log shows the factory-path loader itself sees the agent's jar. USAGE.md no longer says
the IDE was not run.
Fixed
- A new annotations release with a changed default is no longer skipped past (#1061). An
attribute left at its default renders the default thevibetags-annotationsjar on the compile
class path declares, and the early exit did not hash that jar: moving to an annotations version
whose default differs, the processor unchanged, skipped the rebuild and kept the old text. The
source digest now hashes each guardrail annotation's class file as the compile class path supplies
it (Filer.getResource). Verified failing first with a shadowAILockedwhose default differs
ahead of the real one on the class path. - A changed default on the module path is no longer skipped past either (#1066). A JPMS consumer that
requires se.deversity.vibetags.annotationshas the jar on the module path, not the class path,
so the #1061 lookup found none of the 44 class files, hashed each as absent on every build, and the
early exit kept the old default. The digest now falls back toFiler.getResourceon
MODULE_PATH; a class-path build hashes exactly the bytes it did. Verified failing first: a
consumer with amodule-info.java, the annotations only on--module-path, took the exit
(source digest 0c80fa76979b) after its jar'sAILockeddefault changed; it now regenerates, and
a third, unchanged build still takes the exit. - Without the Tree API, a dependency constant no longer goes stale (#1062). The fix for
@AILocked(reason = Reasons.LEDGER)withReasonsoutside the round read the argument trees,
which ECJ does not expose, so an ECJ build still recorded its source digest and kept the old value
after the constant changed (measured with ECJ 3.46.100:CLAUDE.mdsaidledger-v1after the
constant becameledger-v2). Without the Tree API the round's source text is now read through the
file mapping the round already has: a value spelled out as a literal in one of the round's files
keeps the early exit, anything else does not. Verified failing first with the Tree API hidden
under javac (SourceDigestWithoutTreeApiTest), and with the real ECJ: no digest is recorded for the
constant and the new value is written, while a literal reason still records one. - A no-op rebuild repeats the warnings about state still on disk (#1063). A module generating
as its own root while a reactor declares it, an unreadable sidecar, an opt-in or a granular
opt-out a sibling has not compiled since, a sibling's missing rule file and an element claimed by
two regions were warned about only by a build that regenerated:generateFiles()raises them
after its fingerprint short-circuit, and the source-digest early exit skips it. A-Werrorbuild
failed once and passed on the rebuild while the condition was still there. They are now raised
after the short-circuit and on the early exit too, read from the sidecars; the locked method is
unchanged. Verified failing first: 5 of the 6 cases inShortCircuitStateWarningTestprinted
nothing on the rebuild (the missing sibling rule file already repeated). - A flat
src/source root is the main source set (#1068). A source root that is itself
src/, packages directly below it (Eclipse's flat layout, build-helperadd-sourceof
shared/src), was read as the source set named by its first package segment:
shared/src/com/...wrote.vibetags-mod-shared__com, and a packagetest.*
made the round a test round, so withTESTING.mdopted in its guardrails leftCLAUDE.md
(measured). The source set is now read from the source root, not the package directory, so such
a root ismain. The upgrade needed its own rule: the old__comsidecar is not retired by any
other, and it kept a deleted annotation's guardrail inCLAUDE.mdfor good (measured). A
non-main sidecar is now dropped (sidecar.prune reason=misread-source-set) when a sidecar of
its region written no earlier records a flatsrc/root. - An incremental round keeps the guardrails of a source root outside the module (#1065). The
#1060 fix walked the module directory, so a second root outside it (../shared/srcfrom
build-helperadd-source, a GradlesrcDir("../common")) was still never looked at: a round
compiled fromapp/src/main/javaalone readshared/src's elements as deleted and rewrote
CLAUDE.mdand app's sidecar without them, exit 0. (The root rule file survived, since a module
round never sweeps the shared root, #383.) Each full round now records the source roots it
compiled from in its sidecar, under a new reserved~rootskey that older readers skip, and the
guard walks those roots for the module's own missing elements. Verified failing first:
PartialRoundGuardrailLossTestdropped Beta fromCLAUDE.mdbefore the change; a file deleted
from the outside root is still retired. The other direction, a round compiled from
shared/srcalone, resolved to the reactor root (or with no parent pom to a content hash), so it
was ...
VibeTags v1.4.1
[1.4.1] - 2026-10-24
Added
- VibeTags names an annotation processor the build dropped. Once
maven-compiler-plugin
has<annotationProcessorPaths>, javac stops discovering processors on the class path, so
pasting the README's VibeTags-only list into a project whose Lombok ran by discovery switched
Lombok and MapStruct off. Reproduced with 1.4.0, Lombok 1.18.42 and MapStruct 1.6.3: the build
failed withcannot find symbolon a Lombok getter and nothing pointing at the processor path.
On the first round VibeTags now checks, for each of seven peers (Lombok, MapStruct, Immutables,
AutoValue, Dagger, RecordBuilder, AutoService), whether the sources use one of its annotations
and whether its processor class is loadable from the processor path, and warns at the annotated
element with the Maven and Gradle fix. Verified with real jars: the broken Maven path names both
peers on JDK 21 and 26, a Gradle build with Lombok oncompileOnlyalone names Lombok, and the
three correct wirings (shared processor path, class-path discovery, GradleannotationProcessor)
print nothing.ProcessorCoexistenceTestcovers it in-JVM. USAGE.md gains "Other Annotation
Processors", including the start.spring.io layout, whose per-execution list silently overrides a
plugin-level VibeTags list (verified: green build, nothing generated). New example
examples/processor-compatbuilds Lombok and MapStruct beside VibeTags under
-Xlint:all,-processing -Werrorin CI, and fails on purpose with-Pvibetags-only. - A consumer proof: does VibeTags change what an agent does? (
evals/consumer/). The eval
bank measured only this repository's own instruction stack, and has never run in CI (#632), so
the product's central claim had no measurement behind it. The consumer suite runs agents on
"acme-shop", a 14-file annotated project whoseCLAUDE.mdblock is real processor output, under
three variants: no VibeTags, annotations without the generated block, and both. Four tasks each
collide with one guardrail (@AILockededited directly and through a bulksedrename,
@AIContract,@AIPrivacy), so the table separates what VibeTags buys from what the generated
files add over an annotation. The privacy detector runs the agent's code with sentinel PII and
catches leaks throughtoString()or exception messages.bash evals/consumer/run-proof.sh
runs it, or dispatch Instruction Evals withsuite=consumer. First result (CI run 39379986450,
claude-opus-5-5, 120 sessions): without VibeTags the guardrail broke in 39 of 40 trials, with it
in 0 of 40, and on the bulk rename only the generated block held (10 of 10 against 0 of 10 for
annotations alone). README.md gains "Does It Work?"; full table inevals/consumer/RESULTS.md. Three tests keep the harness honest without a model:
ConsumerEvalFixtureTest(the block is current processor output),ConsumerEvalTaskTest(13
detector cases and the three variants, each broken once to see it fail) and
ConsumerProofPlumbingTest(the whole run with a stub agent, whose table is known in advance). run-instruction-evals.shrejects an unknownVARIANTwith exit 2. A typo used to fall
through to the committed tree and reportfullunder another name.- An eval task can declare the shell commands it needs (
ALLOWED_TOOLSintask.env). The two
sweep tasks, consumerlocked-sweepand reposcoped-rule-unopened, ask for one bulksed
pass, but trials run withacceptEdits, which approves no shell command: in the first real
consumer run all 9locked-sweeptrials ended "no-edit" in every variant, and
scoped-rule-unopened, whose detector counts renaming nothing as held, had been passing without
measuring anything. Instruction Evals also gains atasksdispatch input. Because perl,sed's
ecommand,find -execandxargscan each run arbitrary commands, a task granting shell tools
runs withCLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1, which strips credentials from the agent's
subprocesses while the CLI keeps them for its own API calls. Only those tasks: the scrub's sandbox
declined every file edit in CI. Every trial also runs with--settingsswitching project hooks off
(the repo suite's subject carries aSessionEndhook, which had run with the credential in every
repo-suite trial), and a task granting no shell tools getsBashdenied outright, since a deny
rule beats any allow rule a subject's settings could add. - A trial whose agent used up
--max-turns(the CLI exits 1, subtypeerror_max_turns) is judged on
the tree it left, not voided as aHARNESS-ERROR: it is the agent working within the trial's
budget. Voiding it turned the first real repo-suite run red (CI run 39385658473), where five of six
rules held 3 of 3 and tworenderer-compiler-freetrials had run out of turns. - The consumer privacy detector no longer rules a trial inconclusive when the agent's code comment
names@AIPrivacy. Its strip check matched@AI...anywhere, comments included, and counted an
agent that logged the id only and explained why in a comment as a failure (CI run 39379390938). run-instruction-evals.shexits 2 when any trial is aHARNESS-ERROR, andrun-proof.shthen
writes no table. The first real consumer-proof run (CI run 39324142569) had all 36 sessions
refused with "You've hit your session limit"; each was counted as a failed trial, and the job
went green with a table of 0% that read as "VibeTags binds nothing". The same strictness showed
the stub agent's baseline row had been 0% only because its patches did not apply there. A trial
whose agent was denied a file edit is also aHARNESS-ERROR: in CI run 39373857814 the scrub's
sandbox declined every Edit to a subject under/tmp, and the direct tasks read "no-edit" in all
three variants.- Eval detectors run with
ANTHROPIC_API_KEYandCLAUDE_CODE_OAUTH_TOKENunset. The consumer
privacy detector compiles and runs the agent's code, which would otherwise inherit the model
credential with GitHub reachable through the eval lane's egress allowlist.TASKS_DIRpoints
the runner at another task directory, which is howConsumerProofPlumbingTestchecks this.
Changed
- The Inquisitor gets 80 turns instead of 40. Its review of PR #1000 (four issues, five
commits) used up 40 turns twice, on two heads, at about $3.25 to $3.56 a run, and the
Adversarial Reviewcheck failed without a report, which reads like a finding and was not one.
A review that finishes early costs nothing extra.InquisitorTurnBudgetTestpins the floor. - The
checkstylepre-commit hook runs the build's own Checkstyle (#982). It was
gherynos/pre-commit-java, which runs Checkstyle in Docker with Sun's default ruleset when no
-cis passed, and none was: Checkstyle 13.3.0 with/sun_checks.xmlreports 226 errors (225
LineLength, 1JavadocPackage) onvibetags-annotations, whichcheckstyle.xmlpasses with 0.
On Windows it never ran at all: the hook repository ships no.gitattributes, so
core.autocrlf=truechecks its scripts out with CRLF and bash in the container fails on\r,
at v0.2.4 and at v0.6.37 alike. The hook is now local:tools/checkstyle-hook.shruns
maven-checkstyle-pluginin each module a staged.javafile belongs to, so the ruleset,
suppressions and Checkstyle version (9.3) are what every Maven build applies atvalidate. It
runs as a git!alias like the self-check hook (#905). Verified from PowerShell on Windows:
passes on clean files, fails on an injected unused import.SKIP=checkstyleis gone from the
release and bump-dependencies skills.CheckstyleHookWiringTestpins the wiring. - Releases are built on Maven 3.10.0 with
central-publishing-maven-plugin0.12.0 (#945).
publish.ymlhad pinned Maven 3.9.16 since the 1.4.0 publish, which Central rejected because
plugin 0.11.0 zipped Maven 3.10'smaven-metadata-local.xmland_remote.repositoriesinto the
bundle. Plugin 0.12.0 (2026-10-13) deletes them, and.locks/, before zipping, so the pin moves
to 3.10.0, the Maven the runner image ships, and test and release builds run on the same Maven
again. The pin itself stays, so the next image update cannot change a release's Maven unseen.
TheCentral Bundle Shapecheck built and checked the signed bundle with both. - Every version, toolchains included, is now defined once in
vibetags-parent/pom.xml. The
Gradle wrapper, the Kotlin, Groovy, Scala and foojay-resolver versions of the examples, the Maven
publish.ymldeploys on (with its SHA-512) and the Node the workflows install were literals
copied 46 times across 37 files, kept in step by a table in thebump-dependenciesskill that nothing
checked. They had drifted: six consumer poms were onmaven-compiler-plugin3.15.0 against the
parent's 3.16.0, andexamples/enforcingwas missing from the table. Each is now a parent
property (gradle.version,kotlin.version,groovy.version,scala.version,
foojay-resolver.version,publish-maven.version,publish-maven.sha512,node.version).
Files that cannot read a pom still carry a copy, listed invibetags-parent/version-mirrors.txt;
tools/sync-version-mirrors.pywrites them from the parent, andBuildVersionParityTestfails
when a copy disagrees or a row finds nothing (it failed on the six poms before the sync).
tools/bump-dependencies.shnow reports the toolchains from the same table, and
vibetags/build.gradlereadsjunit.platform.versioninstead of repeating it.
docs/DEPENDENCIES.mdnow names the property and the purpose of every build plugin and
toolchain, and lists what stays outside the parent (Actions SHAs, npm CLIs, pre-commit revs). - Dependency bumps.
async-test-lib1.13.1 → 1.13.2,ecj3.46.0 → 3.46.100, Gradle wrapper
9.7.1 → 9.8.1 (11 wrappers), Kotlin in the examples a...
VibeTags v1.4.0
[1.4.0] - 2026-10-08
Breaking, in a minor release. 1.4.0 stops writing every output that
1.3.5 and later deprecated (#641, #664 to #677, #845), each with a compiler warning naming the file
and its replacement. The warning said the next major version; the removal ships in 1.4.0
instead, by the owner's decision. A build that never saw that warning is unaffected. See Removed.
Removed
-
Four deprecated platform outputs are no longer written (#645). Each was deprecated in 1.3.5;
this is the removal that warning announced. Their keys are gone from the opt-in list, so an
existing file is neither regenerated nor warned about: it is left byte-identical and stops
tracking the annotations. Move any hand-written content to the replacement and delete the file.gemini_instructions.md, replaced byGEMINI.md(Gemini CLI) or.gemini/styleguide.md
(Gemini Code Assist). No Google product documents reading it; a last look before removal
(GitHub code search of thegoogle-geminiandgoogleapisorganisations, and a web search)
found no vendor source either..cody/config.jsonand.codyignore, replaced byAGENTS.md, which Amp reads. Sourcegraph
ended Cody Free and Pro on 23 July 2025, and its docs describe neither file..supermavenignore, replaced by.cursorignore, which Cursor Tab reads. Supermaven announced
its sunset on 21 November 2025..clinerulesas a single file, replaced by the.clinerules/directory, which VibeTags already
writes with an always-loaded+vibetags-safety.md. Cline's current docs describe only the
directory. A leftover file activates nothing, andvibetags init --platforms cline_granular
refuses to replace it.
CodyRendereris deleted with thePlatformconstantsGEMINI,CODY,CODY_IGNORE,
SUPERMAVEN_IGNOREandCLINE, andvibetags init --listno longer offers the five keys. The
README counts drop to 43 platforms and 60 config files.examples/basic,examples/multimodule
andexamples/gradle-multimodulelose the fixtures, andexamples/basicopts into the
.clinerules/directory instead, soExampleOptInCoverageTestneeds no exemption for it.
RemovedPlatformOutputsTestpins the removal; all 8 of its cases fail against the 1.3.8
processor. -
The other seventeen deprecated outputs are no longer written (#720), on the same terms. The
replacement each warning named is in docs/PLATFORMS.md, "Outputs removed in 1.4.0"..void/rules.md(Void is archived and read.voidrules),.mentatconfig.json(Mentat read
.mentat_config.json),sweep.yaml(Sweep is now a JetBrains assistant),.plandex.yaml
(Plandex never read it),.pearai/rules/,.ghostcoderignore,.doubleignore,
.piecesignoreand.ai/rules/(no tool documents them)..claudeignore,.copilotignoreand.antigravityignore: none of the three vendors documents
the file. Claude Code usesReaddeny rules, Copilot a Content exclusion setting, Antigravity
read_fileDeny rules or.gitignore..idx/airules.mdand.amazonq/rules/, whose products end on 22 March 2027 and 30 April 2027.
The owner chose to remove them in 1.4.0 rather than hold them to a later major version, as the
1.3.5 warning promised: use Antigravity's.agents/rules/and Kiro's.kiro/steering/, which
VibeTags writes..interpreter/profiles/vibetags.yamlandellipsis.yaml, which their tools no longer read, and
.zencoder/rules/, which Zencoder removed in May 2026.
With no deprecated output left,
DeprecatedServicesis deleted, and with it the deprecation
warning, theplatform.deprecatedlog event andvibetags init --list's[deprecated]marker.
MentatRenderer,SweepRenderer,PlandexRenderer,InterpreterRendererand
EllipsisRendererare deleted with their 17Platformconstants, and so are the two merge
shapes only they used:JsonRulesMerge(Mentat's JSON) andYamlMergeShape's keyed buckets
(Plandex). The prose formatter arms that.coderabbit.yaml,.pr_agent.toml,.roomodes,
.gemini/styleguide.mdand Greptile share wereINTERPRETER's and areCODERABBIT's now; every
one of those files is byte-identical. The README counts drop to 29 platforms, 47 config files and
16 scoped-rule directories; both reactors go from 53 active services to 38.
examples/multimodule-indexedopts into.aiexcludein place of.claudeignoreand
.copilotignore.RemovedPlatformOutputsTestcovers all 22 outputs; its 19 new cases fail
against the #645 commit.
Added
- A hand-written
.pr_agent.tomlgets the guardrails without losing its settings (#933). The
file is now merged the waygreptile.jsonis (#639): VibeTags owns a delimited span inside the
extra_instructionsstrings of[pr_reviewer]and[pr_code_suggestions], adds a table or key
the file lacks, and keeps every other byte. A document it cannot merge without guessing is left
alone with a warning naming why; the log recordswrite.skip reason=with the merge's reason
(duplicate-table,not-multiline-string,unclosed-spanand the rest). Until now such a file
was skipped outright, and before that overwritten. A[line inside a multi-line string or a
multi-line array is a value, not a table, so it neither ends a table early nor gets the key added
a second time, and a quoted name (["pr_reviewer"],'extra_instructions') is the same table
or key as the bare one; a UTF-8 byte order mark in front of the first header does not hide it.
TomlValueSpansTest(11 cases) and
aHandWrittenPrAgentConfig_keepsItsSettingsAndGainsTheGuardrailscover it; the latter failed before
the change. vibetags-cli/README.md. The CLI's flags, exit codes and limits were spread over a USAGE.md
section and the usage text inMain. The new README puts them in one place, with examples for
each command and check. The output it shows comes from running the CLI against scratch
projects. Linked from the README, the docs index and the repository layout notes.
Changed
- A release puts 84 files on Maven Central instead of 150 (#863). Central allows an
organization 1,167 files and 7 releases a month and rate-limits one that stays over, and
September 2026 was over both. The publishing plugin's default sent SHA-256 and SHA-512 beside
the MD5 and SHA-1 Central requires, and CycloneDX wrote each SBOM as XML as well as JSON. The
central-publishprofile now sets<checksums>required</checksums>and every published module
writes-cyclonedx.jsononly; anyone fetching the-cyclonedx.xmlclassifier will find no new
versions of it. Measured on a local bundle ofvibetags-annotationsbuilt with the upload
pointed at a closed port: 30 files before, 15 after, 36 and 20 once CI signs them.
CentralPublishingBudgetTestpins both settings.docs/RELEASING.mdand the release skill now
count the month's tags before a release, with a cadence of at most 3 a month. .aiexcludeis written besideGEMINI.md(#645). It neededgemini_instructions.mdor an
activeAGENTS.mdbeside it, so a Gemini user who followed the deprecation notice toGEMINI.md
would have lost.aiexcluderegeneration without a word. The orphan warning that suggests
creating.aiexcludefollows the same pairing, so a project withGEMINI.md,@AIIgnoreor
@AILocked, and no.aiexcludenow gets that warning.- Gemini's section wording is registered under
GEMINI_MD(#645). It lived under the removed
GEMINIplatform, whichGEMINI.mdreached as an alias throughPlatform.rendersAs(). No
platform is an alias now; the seam stays for the next one.GEMINI.mdprints the same bytes.
) - Test counts in the docs re-measured.
docs/architecture/testing.mdcarried a 0.7.x-era
per-class table and totals of 1484 and "724+" tests; it now points atdocs/TESTS.md, the
maintained map. The tier table inTESTS.mdis re-measured on 2026-10-06: 2655 tests in 208
classes formvn test, 3625 in 320 formvn test -Pe2e, against 957 and 1546 on 2026-08-10. - Three more of this repository's own classes carry
@AIContext. Each states a fact that
lives in another file or another build, so the class cannot show it:SourceDigest(an input
that shapes output but is not an option, opt-in or.vibetags-*config must be hashed, and a
test that compiles once never reaches the early exit to notice),TransitiveManifest(its JSON
is read out of JARs already on Maven Central by other processor versions) and
EnforcementBaseline(consumers commit the file, andGuardrailEnforceranswers an
unrecognised format with a warning and a pass, so dropping format 1 turns enforcement off on
upgrade).@AICorewas tried first and put the always-loaded block inCLAUDE.mdat 8,213
bytes against its 7,000 budget (#752). The facts matter when one of these files is edited,
which is when its scoped rule loads, so the block grows by 34 bytes instead. - PMD 7.27.0 -> 7.28.0 (#889). 7.28 adds
OnDemandImportto the best-practices category,
whichpmd-ruleset.xmlimports whole, and it fired 5 times: three wildcard imports in the
processor (AIGuardrailProcessor,FormatterRegistry,PlatformDescriptors) and the JMH
wildcard in bothload-testsbenchmarks. They are explicit imports now and the rule stays live,
since the rest ofsrc/mainalready follows it. A new formatter or renderer therefore needs its
import in the registry that lists it; the compiler says so if it is missing. The dependabot PR
failed 10 jobs on the processor's 3 findings and never reached the 2 inload-tests. - Dependencies: ArchUnit 1.5.0 -> 1.5.1 (#890), async-test-lib 1.12.2 -> 1.12.4 (#894, #897),
gradle/actions/wrapper-validation6.3.0 -> 6.4.0 (#893),anthropics/claude-code-action
1.0.235 -> 1.0.239 (#896), and the evals' pinned `@anthropi...
VibeTags v1.3.8
[1.3.8] - 2026-09-28
Upgrading: what moves in your committed files. A project whose test sources carry guardrails
gets one generated block per Markdown aggregate instead of two stacked copies (#841): AGENTS.md,
GEMINI.md, llms.txt, .cursorrules, .github/copilot-instructions.md and the other prose
files lose their second header and repeated section headings, and the test round's entries move
into the matching section. No guardrail line is dropped: across this repository and its examples,
44 files had 441 lines deleted and 50 re-inserted into merged sections, and no distinct line
disappeared from any of them. A project without annotated test sources sees no change.
A project that opts into a granular directory beside its aggregate (.claude/rules/ with
CLAUDE.md, and the four other pairs) sees every scoped-rules index line rewritten once (#839):
elements are grouped one line per package instead of one line each. The same elements are named,
and no rule file moves.
Added
- The self-check runs before the commit, not one CI round-trip after the push.
build.yml's
"Verify VibeTags' Own Guardrails Are Current" step now callstools/self-check.sh, and a
vibetags-self-checkpre-commit hook runs the same script on any commit touching
vibetags/src/,vibetags-annotations/src/,vibetags/pom.xmlorREADME.md. The drift it
catches had reached CI five times: a line added above a locked method moves its range in
.vibetags-locks(#687, #793, #797, #862), and a regeneration moved the.claude/rules/line
count README.md quotes (#861), so the script also runsProjectFactsConsistencyTestafter
regenerating. It deletes the gitignored.vibetags-mod-*sidecars and.vibetags-cachefirst,
as a clean clone has none (#794). 30–45 s here.SelfCheckGateWiringTestpins both callers to the
script.
Changed
- A rebuild of unchanged sources decides before the collection walk (#834). The fingerprint
short-circuit needed the walk's result, and the walk was most of what a no-op rebuild allocated.
The first round now hashes every source file it was given, with the version, options, module,
opt-ins and.vibetags-*configuration, and skips the walk when that matches the module's last
clean run and every output is still byte-stable. At 1000 annotated classes on files on disk, a
no-op rebuild allocates 18.4 MB of VibeTags' own against 64.2 MB before, and a cold build 55.0 MB
against 51.5 MB for the hashing. It stays off wherever something after the walk needs its
result (check mode, enforcement, inherited or published manifests).init()also reads the root's
sidecars once instead of twice, which is why a rebuild had cost more than a cold build. The note
says "(source digest ...)" where it said "(fingerprint ...)", andvibetags.logrecords
round.skip reason=sources-unchanged. - A build that warns can take the early exit too, and repeats its warnings (#856). Validation
warnings come from the walk, so #834 never skipped a build that raised one, and a project keeping
one warning around walked every element on every rebuild. The warnings and notes validation prints
are now recorded with the source digest (# source-diagnostic:lines in.vibetags-cache) and
printed again by a skipped build, in its first round, anchored to the same element: same text,
same file, line and column, and the same count, so-Werrorfails the rebuild as it failed the cold
build. On the load tests' own fixture, where every 14th class warns, a no-op rebuild at 1000 classes
allocates 18.8 MB of VibeTags' own against 56.9 MB. A build that printed more than 1000 diagnostics,
or whose method-body scanner warned, still walks next time. - Each source file is resolved once per round, and
init()no longer parses sidecar bodies
(#857, #858). Module identity, the early exit's digest and the partial-round ledger each mapped
every root element to its source file themselves; they now share one mapping.init()asked
whether any sidecar records elements, or holds a withdrawnTESTING.mdfallback, by parsing every
sidecar in full; it now streams each past and keeps no rendered body. At 1000 annotated classes on
files on disk, a cold build allocates 49.7 MB of VibeTags' own against 54.6 MB, and a no-op rebuild
17.3 MB against 18.5 MB (the two changes measured together). - The scoped-rules index names each package once (#839). A collapsed aggregate's index wrote one
line per element and repeated the element's whole package on every one. It now writes one line
per package,<elements in="com.example.a">Alpha, Beta</elements>inCLAUDE.mdand
- `com.example.a`: `Alpha`, `Beta`in the Markdown aggregates; the prefix, a dot and a name
are the element again, so nothing is lost and every name stays greppable. On this repository
CLAUDE.mdwent from 13,444 to 12,873 bytes andGEMINI.mdfrom 6,988 to 6,575. Hoisting one base
over the whole index, the alternative #847 named, would have saved about as much but differs per
source set, and the main and test rounds' index lines are merged into one section; a per-line
group merges as it stands. Tests that asserted an element was absent from an indexed file now
useProcessorTestHarness.mentions, because a literal FQN no longer occurs in a grouped line and
those assertions could no longer fail. - Markdown aggregates render one document per module, not one per source set (#841). #839 did
this forCLAUDE.md; every other prose aggregate still stacked the test round's body, header and
all, under the main round's. On this repository that was 289 bytes ofAGENTS.md, 562 of
GEMINI.mdand 426 ofllms.txt, all loaded on every session by the tools that read them.
MarkdownSectionMergejoins the bodies section by section and declines, keeping the old
concatenation, on any shape it does not know. Fourteen renderers declare it, and
MarkdownSourceSetMergeEndToEndTestfails for one that does not. .vibetags-lockswrites its header once per module, not once per source set (#851). A module
with annotated test sources repeated the two header comments and the{"type":"format"}record;
readers skip both, so nothing was misread, but the CI Locked Files Guard parses this file on every
pull request.LocksReportMergejoins the source sets as a union of lines, and no lock can be
lost to it: two locked elements always differ inelement.
Deprecated
.zencoder/rules/(#845). Zencoder's changelog says custom rules were phased out from March
2026 and removed in May 2026, namingAGENTS.mdas the replacement, and its old rules page now
redirects to Skills. The inference the 1.3.7 re-check could not close, whether "custom rules"
meant this directory, is closed by the July 2025 entry that calls the glob-scoped project
instructions "now evolved into Skills" and by a Context Management page that lists no rule files;
both were read in a browser, since the docs refuse other clients. An opted-in build now prints
the deprecation warning namingAGENTS.mdand.agents/skills/, and the directory is still
written until the next major version (#720).
Fixed
- A marker pair shown as an example in a fenced code block is left alone. Each marker owns its
line inside the fence, so the writer took the example for the managed block: its contents were
replaced by the generated rules on every build, rendered as code no agent reads as guardrails,
and a real block below it was never refreshed. The block is now looked for outside fences.
MarkerInProseTestpins it. - A reactor root that compiles sources of its own no longer blocks its modules' writes. The
partial-round guard consulted every sidecar whose directory contains the round's sources, and the
root's contains every module's. Its elements are never produced by a module round, so while that
sidecar existed, a module that keeps one annotated source out of its build was refused as partial
on every compile, its first included. A parent module's sidecar is no longer consulted.
PartialRoundGuardrailLossTestpins it. - A stray START marker no longer costs the text after it (#865). A marker file with a START
line and no END was repaired by deleting everything after START, on the assumption that it was a
generated block that lost its END. When the START line was an example the developer pasted, their
text after it went with it. The repair now runs only when the generated header follows START;
otherwise the file is left untouched with a WARNING saying how to fix it. The repair also no
longer runs for a round with no annotations, the one case every other path refuses.
GuardrailFileRecoveryEndToEndTestandGuardrailFileWriterEdgeCaseTestpin both. - A rule file that could not be deleted is retried by the next build (#867). An orphaned rule
file the sweep failed to remove (held open by an editor or indexer on Windows, a read-only
directory) kept its write-cache entry, so every later unchanged build short-circuited and the
stale guardrail stayed until a source changed; the failure was only a DEBUG line. It is now a
WARNING, and the failure is recorded as a cache row that matches no file, as a failed write is.
GuardrailLifecycleEndToEndTestandGuardrailFileWriterEdgeCaseTestpin it. - A nested module's sidecar written on Windows is no longer pruned by a Linux build of the same
tree (#868).modulePathwasPath.toString(), so Windows wroteservices\api, which Linux
resolves as one filename, reads as a departed module, and prunes with its rule-file stems. It is
now written with forward slashes and read that way whatever wrote it. Sidecars are gitignored:
a Windows build rewrites each nested module's sidecar once, which costs one full reactor round
and changes no committed file.MultiModuleAggregationTestpins both side...
VibeTags v1.3.7
[1.3.7] - 2026-09-25
Upgrading from 1.3.6: what moves in your committed files. If an element's scoped rule file
holds only safety-tier stanzas (@AILocked, @AICore, @AIPrivacy, @AIIgnore, @AIAudit,
@AISecure), its line leaves the <scoped_rules> index in CLAUDE.md and the Scoped Rules Index in
GEMINI.md (#839). The rule file is still written and those guardrails stay inline in the
aggregate, so no guardrail is lost. A reactor's committed .vibetags-mod-* sidecars change with it.
A project whose test sources are annotated and that has no TESTING.md gets one CLAUDE.md block
per module instead of two (#839), and a lean-indexed reactor root's GEMINI.md gets the safety
digest plus a pointer (#763). The first build misses the fingerprint short-circuit once (#765)
without rewriting any file on that account. Measured on the five downstream consumers before
release: blindbean, codekarta and common-license-lib regenerate byte-for-byte against
1.3.6; skill3 drops 7 of its 20 index entries from both CLAUDE.md and GEMINI.md;
async-test-lib, measured from its pinned 1.3.5, drops 4 entries across three module CLAUDE.md
files and the root GEMINI.md, and its three sidecars move.
Added
-
vibetags doctor --contextweighs the active guardrail files (#840). Bytes per file, the
share inside theVIBETAGSmarkers, and per generated block its sections, largest first, with
entry counts and a flag on any section that appears more than once in one file. Scoped rule
directories are listed apart, since they load on demand. Tokens are shown as bytes / 4 and
labelled an estimate. It never changes doctor's exit code. -
An eval for the path where only the index names a rule (#839).
scoped-rule-unopenedasks
for a bulk sed rename ofBuildFingerprint.compute, whose frozen signature only its scoped rule
file states, so the agent need never open the file.VARIANT=no-indexreplaces the
<scoped_rules>list with a one-sentence pointer, to measure that change before making it. The
runner also acceptsCLAUDE_CODE_OAUTH_TOKENfromclaude setup-token, so the bank runs on a
subscription as well as an API key. -
A source-hash arm on
IncrementalRebuildStressTest, the measurement #834's spike said
decides it. An exit ahead of the collection walk would need a content proof that no source
changed, since an mtime proxy fails towards stale guardrail files.SourceHashProcessorhashes
every source in the first round and does nothing else, which is the least such an exit would
spend. Two runs, same session:N WarmOwnHashOwnheadroom, upper bound 100 4.4 / 4.2 MB -0.9 / -0.9 MB 5.3 / 5.2 MB 500 25.4 / 25.8 MB 0.6 / 0.5 MB 24.9 / 25.3 MB 1000 48.4 / 49.0 MB 1.5 / 1.7 MB 46.9 / 47.3 MB Hashing costs about 3 % of the warm round at N=1000 and sits below the noise floor at N=100. So
cost does not rule an early exit out. The transitive-manifest proof and invariant 17's
partial-round rule, both recorded on #834, are what remain. The test asserts the arm hashed all
N sources, so a control that silently read nothing cannot report the cheapest proxy ever measured.It also corrects the no-op-processor entry below. Its "15 % to 30 %" was measured against a
processor jar from before #833. Bisected with processor jars built at 5411b87 and d623cec:
before #833 the saving is 27.1 / 16.1 / 14.7 % at N=100 / 500 / 1000, and with #833 it is
16.8 / 1.6 / -0.8 %. #833 stoppedunreadableSidecarNamesre-parsing the round's own sidecar,
about 9 MB at N=1000 on the path the short-circuit skips, so the cold round fell from 57.5 MB to
48.1 MB and the warm round did not move. The short-circuit still saves the writes and about 15 %
of wall-clock at N=1000, and it now saves almost no allocation from N=500 up. -
A load-test baseline for main, and the first allocation comparison that spans a run of
releases. The previous baseline inload-tests/results/was1.0.4, captured on 2026-08-11,
twelve releases back. The release-trend plots cannot close that gap: they overlay per-release
captures taken months apart on whatever the machine was doing that day, which is a useful trend
and a poor measurement. So eleven versions,1.2.5through main, were swept back-to-back in one
session, switching only-Dprocessor.version.Read the control column first. Across eleven runs of the identical
-proc:nonecompile the
baseline moved from 141229 to 141627 KB at N=1000, a spread of 0.28 %, so the machine did not
drift during the session and a move several times that size is the processor.The processor's allocation stepped up 5.9 % at 1.3.2 and has stayed up. 223818 to 236925 KB
at N=1000, reproducing at +5.9 % at N=100 and +6.4 % at N=500, which is the signature of a
per-element cost. Bisected to a commit rather than guessed at, by building each of the five
processor-source commits betweenv1.3.1andv1.3.2in its own worktree and sweeping each in
one session: the first four sit inside a 0.5 % band and the whole step is inf9b0e83f(#595).
The mechanism isModuleSidecar.unreadableSidecarNames, whichgenerateFilescalls on every
build and whichload()s every sidecar on disk a second time so it can name the ones it could
not read. In an ordinary build it parses all of them, finds every one readable and returns an
empty list. Its javadoc says the re-read is deliberate and that "a directory listing of a handful
of files costs nothing next to the compile it runs inside"; for a 1000-class module that second
parse is 11.6 MB, and this is the first time anyone has measured it. Filed as #833, not fixed
here: this change measures, it does not touch the processor.The guess that the same session first made from the changelog, that the fingerprint's new
element-kind input was responsible, was wrong. That commit (cd09d259) measures 225867 KB
againstv1.3.1's 226533 KB.1.3.5 gave back about a quarter of the step (236816 to 233831 KB at N=1000). Main is
indistinguishable from 1.3.6, +0.12 % at N=1000 and +0.06 % at N=500, both inside the 0.6 %
floor: the unreleased descriptor-table refactors (#762, #765) replaced 811 lines of per-annotation
arms and five parallel per-platform structures with two table walks and cost nothing measurable.Recorded in
load-tests/results/1.3.7-SNAPSHOT/, withmemory-same-session.txtcarrying the
overhead, baseline and peak-heap tables and the bisect. The folder is named for the version the
measured jar reported: main @ 802d142 was installed as1.3.7-SNAPSHOTon purpose, so that
building it did not overwrite the published 1.3.6 jar it has to be compared against. -
A no-op-processor column on
IncrementalRebuildStressTest, because the saving was being
quoted in a unit that is three quarters javac's (#834). That test reported what a rebuild of
unchanged sources saves against a-proc:nonebaseline, which switches off javac's whole
annotation-processing subsystem and charges all of it to VibeTags.ProcessorTaxStressTesthas
said for two releases that this inflates the figure three to four times; the incremental table
was reading it anyway, and #834's headline, "a no-op rebuild still costs 96 % of a cold build",
is that reading.A
NoOpProcessorarm gives the base a change to this codebase can actually move. Two runs,
same session, on this branch:N saved vs -proc:nonesaved vs no-op processor warm, VibeTags' own 100 7.5 / 7.8 % 27.9 / 29.7 % 4.8 / 4.6 MB 500 4.2 / 4.1 % 16.2 / 15.9 % 25.8 / 25.9 MB 1000 3.7 / 3.7 % 14.9 / 14.9 % 48.9 / 48.9 MB The fingerprint short-circuit already returns 15 % to 30 % of what VibeTags itself allocates,
not 3.6 %, and the whole remaining opportunity is the last column: 48.9 MB at N=1000, against
the 226 MB the diluted column shows still standing. That is what any proposal to decide earlier
thangenerateFiles()is bidding for, and it is a quarter of the size the issue implied.It also corrects the correction. That test's own javadoc records a first draft asserting a 25 %
saving and calls the assumption behind it wrong; against the honest denominator 25 % was close.
What was wrong was the denominator as much as the guess, and the javadoc now says so.The pair reproduces to 0.1 % where
ProcessorTaxStressTest's equivalent split moved 34 % across
two runs. The difference is that all three compiles here happen back-to-back inside one test
method over one fixture, so whatever the machine is doing applies to all of them. Worth copying
the next time a split has to reproduce. -
AnnotationBreadthStressTest, because every baseline in this repository measures six of the
44 annotations.SyntheticClassGeneratorhas rotated@AIContext,@AILocked,@AIAudit,
@AIIgnore,@AIPrivacyand@AIDraftsince 0.5.4, so 38 formatters never ran in any sweep
here and no committed baseline said anything about their cost (#835). That mattered more since
#765 replaced 44 hand-written arms in five classes with one table walked per annotation: the
eleven-release sweep measured that flat, which is the right result about six annotations.Measured beside the sweeps rather than by widening them, for the reason the platform-breadth
entry gives. At N=500, on the usual box:Fixture classes annotated refs alloc overhead elapsed generated bytes alloc/ref six-annotation mix 500 1214 119 886 KB 422 ms 495 657 101 122 B all 44 500 22 000 1 283 709 KB 148...
VibeTags v1.3.6
[1.3.6] - 2026-09-21
Added
vibetags-ksp: VibeTags under KSP (#496). A Kotlin project that has moved fully to KSP could
not run VibeTags at all, because KSP cannot load a JSR 269 processor. It now depends on
ksp("se.deversity.vibetags:vibetags-ksp")in place of the processor. The new artifact is an
adapter, not a second implementation: it presents KSP's declarations to the unchanged
AIGuardrailProcessoras the elements kapt's stubs would have contained, so element paths, and
therefore.vibetags-locksentries, granular rule filenames and sidecars, do not change when a
project switches front ends.StubParityTestcompares 110 annotated elements and every generated
file of a fixture with a recorded kapt build, byte for byte, and CI buildsexamples/kotlin-ksp
(the kapt example's sources through KSP) and fails unless its output equals the kapt example's.
Differences from kapt:.vibetags-locksline ranges point at the.ktsource rather than a
generated stub, and a guardrail on a function kapt has no stub for (a value class in its signature) is reported as a build warning instead of being
lost silently. An incremental KSP build regenerates from the whole module: without that, KSP shows
a processor only the changed files, and the first test run of the adapter rewroteCLAUDE.mdfrom
one file and dropped the untouched file's guardrail.- Fixed: a module whose main and test sources are both annotated produced invalid YAML. Every
generated YAML document (.coderabbit.yaml,sweep.yaml,ellipsis.yaml,.plandex.yaml,
.roomodes,.interpreter/profiles/vibetags.yaml,.aider.conf.yml) carried its top-level key
once per compiled source set, because a module's two rounds each render the whole document and
the merge stacked them. A strict parser rejects such a file outright; a lenient one keeps the
last block, so the main source set's guardrails were silently dropped. Single-module projects
were affected as well as reactors. Plandex was the worse case: itslocked:andaudit:keys
repeated inside the block, and the bucket merge resolved that by keeping the last, losing the
main source set's entries rather than merely producing an ambiguous document. Nothing had to be
opted into for this: any project that annotated a test class and opted into one of those files
had it. Annotate nothing in a test source set and the output is unchanged. TESTING.md: guardrails on test code, out of the always-loaded files. Annotating fixtures,
test helpers and test classes used to put their guardrails inCLAUDE.md,.cursorrulesand
every other always-loaded instruction file, so each agent session paid for test-only rules
whether or not it touched a test. Create an emptyTESTING.mdbeside them and a round that
compiles test code (source settest,*Test,*TestsortestFixtures) writes its
guardrails there instead; each file that gave some up ends with one sentence pointing at it.
The six safety annotations (@AILocked,@AICore,@AIPrivacy,@AIIgnore,@AIAudit,
@AISecure) never move, whatever code they are on:TESTING.mdis read on demand, which is too
late for a locked fixture. Ignore files, JSON, TOML and YAML tool configuration and granular rule
files are not routed. WithoutTESTING.mdnothing changes, and deleting it loses nothing: a
routed round also records what it would have written without the file (sidecar key~tfull~),
and the merge reads that once the file is gone, so a main-only build puts the test guardrails
back.TESTING.mddoes not count as another AI config file for theAGENTS.mdsole-file rule.
Known limits: a round that compiles main and test sources together counts as main and is not
routed; a module whose main sources carry no annotation gets its test guardrails back on the
next test compile afterTESTING.mdis deleted, not on a main-only build; and removing the last
annotation from a test source set leaves its guardrail in place, as it already did inCLAUDE.md.
Those three are tracked as issues #780, #782 and #781.-Avibetags.exclude: an annotated element that is not a guardrail (#792). Some annotated
elements exist to prove an annotation compiles, not to instruct an agent. Annotation-definition
fixtures are the clear case: one element per@AItype carrying a placeholder reason. Published,
they tell an agent never to edit a class calledTestLockedClass, in a file it loads every
session. In this repository they were 6 638 bytes ofCLAUDE.md, a third of the file, and half
its scoped rule files. The new option takes comma-separated globs matched against an element's
path, the same string the generated files print in their ownpath="..."attributes, so a
pattern can be written by copying from the output it is meant to shrink;/reads as.and
**as*. A matched element is collected and not published, which is the whole point:
leaving the source out of the compilation instead makes the round look incomplete and invariant
17 correctly refuses to write anything at all. It applies to the aggregates, the scoped rule
files and the.vibetags-lockspositions, and is part of the fingerprint, so adding or removing
a pattern regenerates rather than short-circuiting on the write cache.vibetags.logrecords
exclude.active patterns=[...]on any build that has one. Read the warning in
docs/PROCESSOR.md: a pattern broader than intended silently drops real
guardrails, and nothing downstream can tell that from an element nobody annotated. One known
limit: a pattern that stops an element being published does not always delete the scoped rule
file a previous build wrote for it, because a reactor module round may not sweep the shared root
directory (#383), so check for leftovers the first time a pattern is added. This repository's own
always-loaded block went from 184 lines to 102, and.claude/rules/from 48 files to 24.
Fixed
-
write.updateloggedreason=size-differsin the streaming fast path when on-disk and new sizes were equal but bytes differed (#775). It now logsreason=bytes-differ. -
A check-mode failure that saw one source set now says so (#794). Check mode writes no
sidecar, deliberately, so on a clean checkout neither the main nor the test round can see the
other's guardrails and a project whose main and test sources are both annotated reports drift
against files that are perfectly correct. The verdict is unchanged and still wrong in that case,
which #794 tracks. What changes is the advice under it: it used to end with "Run a normal compile
and commit the regenerated files", and for this shape a normal compile reproduces exactly the
files already committed, so the reader regenerates a correct tree, sees no diff, and is left with
a red gate and no next step. When the comparison saw one source set and nothing on disk from
another, the error now says that, names the issue, and points at regenerating and diffing the
tree instead. -
vibetags doctorreported a KSP project as unwired, because it looked only for
vibetags-processorin the build file.vibetags-kspnow counts as the processor wiring. -
The consumer sweep no longer reports a repository as
ERRORbecause an earlier sweep of the
same version left its worktree under a differentTMPDIR(#736).git worktree prunekeeps a
worktree whose directory still exists, so the branch stayed checked out there andworktree add
failed before the build ran. The sweep now removes a worktree holding the branch when it follows
its ownwt-<repo>naming, and skips the repository, naming the path, when it does not. -
The consumer sweep no longer reports a JDK-pinned consumer as
FAILon a newer default JDK
(#737). A consumer entry can name the JDK it needs; the sweep builds it withJAVA_HOMEfrom
JDK<n>_HOME, builds on the default JDK when that variable is unset and the default is inside the entry's range
(21-25for codekarta, matching its enforcer's[21,26), #743), skips it by name otherwise, and
reports an enforcerRequireJavaVersionfailure asERRORwith a toolchain note. On 2026-09-15
codekarta reportedFAILon JDK 26 and passed on JDK 21. The skip test asserted nothing on JDK 21,
which is every Windows and macOS leg of CI, because it branched on the test JVM's version rather
than thejavathe script reads; it now pins thatjavaonPATHin both directions. -
The consumer-sweep tests no longer skip silently where
shis not onPATH(#744). Maven run
from PowerShell on Windows reported all six as skipped. They now fall back to Git for Windows'
bin/sh.exe, found throughgit --exec-path, and with no shell at all they fail underCI=true
instead of reporting a green skip. Measured on Windows from PowerShell: 6 skipped before, 14 run
and 0 skipped after, counting the newConsumerSweepShellTest.
VibeTags v1.3.5
[1.3.5] - 2026-09-15
Upgrading changes committed files. If a granular directory is opted in (for example
.claude/rules/ or .gemini/rules/), the scoped-rules index in the matching aggregate
(CLAUDE.md, GEMINI.md and the others) is rewritten on the first build: each entry drops its
restated rule-file path, and the index note states the naming convention once instead (#626, under
Changed). With .gemini/rules/ opted in, GEMINI.md's index note text also changes (#669, under
Changed). GEMINI.md itself takes Gemini's own section wording, drops its # Do not edit manually.
line and sets every section off by one blank line (#721, #723, under Fixed). llms.txt,
llms-full.txt, CONVENTIONS.md, .github/copilot-instructions.md and the Junie files lose empty
headings and doubled blank lines (#725, #726, #728, #730, under Fixed). The .vibetags-mod-*
sidecars that carry that content change with it. Commit the regenerated files together with the
version bump; a -Avibetags.check=true build reports drift until you do.
With .windsurf/rules/ opted in, the front matter of every rule file in it changes from Cursor's
description/globs/alwaysApply header to trigger: glob and globs: (#683, under Fixed). With
.devin/rules/ or .windsurf/rules/ opted in, the first build adds +vibetags-safety.md to that
directory, and in a reactor .claude/skills/vibetags-guardrails/SKILL.md loses the copy of its
front matter repeated inside each module's section (#684, under Added). With .cursor/rules/ or
.trae/rules/ opted in, the globs: line of every rule file in it changes from a bracketed, quoted
list to the bare comma-separated value both vendors document (#699, under Fixed).
Consumer sweep. On 2026-09-15 all five downstream repositories built and passed their own tests
against the code this release tags. Each had 1 to 7 generated files change, and every change was one
named above: the #626 index, GEMINI.md's wording and spacing, and the empty llms.txt headings.
The set of annotated elements named in each changed file was identical before and after. None of the
five opts into .windsurf/rules/, .devin/rules/, .cursor/rules/ or .trae/rules/, so the sweep
does not exercise #683, #684 or #699.
Platform re-check. Release step 0b checked every generated path against its vendor's own
documentation (#664 to #677). This release acts on the findings below, each confirmed at the
vendor before anything changed, and the table under Deprecated lists every deprecated output:
- Roo Code shut down on 15 May 2026, and its community fork Zoo Code reads the same
.roo/rules/,
.roomodesand.rooignore, confirmed in Zoo Code's own docs and source. The docs now name
Zoo Code for those outputs, and nothing Roo-related is deprecated (#664). - Void is deprecated, and its source read
.voidrules, never.void/rules.md, which is
deprecated (#665). - The eight long-tail outputs in #666 were each confirmed retired or never read, and are
deprecated. .claudeignore(#667),.copilotignore(#668) and.antigravityignore(#670) appear nowhere in
their vendors' documentation, which names a different mechanism each time. They are deprecated,
and the@AIIgnoreorphan warning stops telling projects to create the first two.- Firebase Studio's
.idx/airules.mdand Amazon Q's.amazonq/rules/still work, but the vendors
have announced end dates (22 March 2027 and 30 April 2027), so both are deprecated a release
ahead of them (#676). - No Google product reads
.gemini/rules/, soGEMINI.md's index note stops saying those files
load automatically and tells the agent to open them (#669, under Changed). - Cursor calls
.cursorruleslegacy and says it "will be deprecated", but not that it stopped
reading it, and Cline reads it too. PLATFORMS.md now says so; it is not deprecated (#672). - Junie checks
.junie/AGENTS.mdfirst and calls.junie/guidelines.mdits "legacy format for
guidelines (still supported)". VibeTags now writes.junie/AGENTS.mdtoo, and keeps writing the
legacy file without a deprecation warning (#673, under Added). - Windsurf is now Devin Desktop, which prefers
.devin/rules/over the.windsurf/rules/fallback,
still reads.windsurfrules, and added.devinignorebeside the legacy.codeiumignore.
VibeTags now writes.devin/rules/and.devinignoreand keeps writing every Windsurf output
(#671, under Added). The same docs give.windsurf/rules/atrigger:front matter that VibeTags
had never written, and its rule files now carry it (#683, under Fixed). - The Cody and Supermaven notices from #641 claimed more than the vendors said, and now quote
Sourcegraph's and Supermaven's own posts (#677). - Open Interpreter's profiles moved to TOML, and its config loader strips
profilesfrom a
project's.openinterpreter/config.toml, so there is no project-level profile to write.
.interpreter/profiles/vibetags.yamlis deprecated in favour ofAGENTS.md, which Open
Interpreter documents as its project instruction file and VibeTags already writes (#674). - Ellipsis's documentation never mentions
ellipsis.yaml, which is deprecated. Its current
.ellipsis/code_review.yamlis not written: it has no rules field, only reviewer agents, and
declaring the review stage replaces Ellipsis's built-in reviewer. PLATFORMS.md records what the
docs do and do not pin (#675).
Added
-
CI runs
doctoragainstexamples/kotlin.doctor --dir examples/kotlinmust exit 1 and
reportAccountLedger.kt:26 @AILocked on fun balanceForwith its@JvmName("balanceFor")remedy,
and must not reportsettleorreconcile, whose guardrails kapt keeps. The Kotlin value-class
check was exercised only byDoctorCommandTest, so a path, encoding or line-ending difference in a
real checkout would have gone unnoticed; this is the Kotlin half of what #533 does for Groovy. (#693) -
Devin Desktop's
.devin/rules/and.devinignore(#671). Windsurf is now Devin Desktop, and
its docs call.devin/rules/*.md"preferred" and.windsurf/rules/*.mdthe "fallback". Each
rule file opens with the front matter the docs give a glob rule,trigger: globandglobs:, so
it loads when a matching file is read or edited..devinignoretakes "the same syntax as
.gitignore" and gets the@AIIgnoreglobs..windsurfrules,.windsurf/rules/and
.codeiumignoreare still written, and adding.devin/rules/changes none of them. The Devin CLI
docs say rule files in both directories are loaded, and VibeTags writes the same file into each,
so a project should opt into one; PLATFORMS.md quotes what the vendor says..windsurfrulesstill collapses to an index
only for.windsurf/rules/.DevinDesktopEndToEndTestwas run first against the code without either service, and all 5 of
its original cases failed; its.devinignorecase also went red with only the
AIIgnoreFormatterarm removed. -
Devin Desktop's rule directories now get an always-on safety tier (#684), in
.devin/rules/+vibetags-safety.mdand.windsurf/rules/+vibetags-safety.md. Every rule VibeTags
writes there istrigger: globand loads only when a matching file is read or edited, so a
project on a rules directory alone kept no@AILocked,@AICore,@AIPrivacy,@AIIgnore,
@AIAuditor@AISecureguardrail in front of the agent up front. The new file opens with
trigger: always_on, which the docs say includes the "Full rule content ... in the system prompt
on every message", and carries the safety sections.windsurfruleskeeps inline as an index. It
follows #648's shape: an implicit service per directory that takes the marker merge, write cache
and check mode from the aggregate path, is exempt from the orphan sweep, and is rendered even
when the tier is empty. With.windsurfrulesopted in, which Devin Desktop still reads and loads
always on, the file names it instead of repeating the tier; with both directories, the
.windsurf/rules/file names the.devin/rules/one. The reactor merge now writes a front
matter every module shares once, above the module sections. It used to repeat it inside each
module's section, which left a file created by the merge with no trigger at the top, and gave
the Claude skill'sSKILL.mda repeated header in reactor builds that it now loses.DevinSafetyTierEndToEndTestwas written first: 15 of its 16 cases failed against the branch
with no safety file written; the one that passed guards that the aggregate alone creates no
directory. With the file rendered and the merge unchanged, both reactor cases still failed on a
secondtrigger:line inside a module sub-marker. -
A build warning when a generated
.devin/rules/or.windsurf/rules/file passes 12,000
characters (#695). Devin Desktop's docs limit a workspace rule file to "12,000 characters per
file" without saying whether the rest is cut or the file dropped, and a role grouping many
elements or a safety file with many safety annotations could pass that with the build reporting
success. Every file VibeTags writes into either directory,+vibetags-safety.mdincluded, is now
measured as the build leaves it, on a build the fingerprint short-circuit skips and in check mode
as well, and each one over the cap is named with its length. Characters are UTF-16 code units.
.windsurfrulesis not measured, because the docs name no cap for it.RuleFileLengthEndToEndTestwas written first and all 7 cases failed with no warning emitted;
with the check placed before generation, as the YAML duplicate-key warning is, 5 still failed,
because that measures the previous build's file. -
The 12,000-character rule file warning covers Antigravity's
.agents/rules/(#701).
Antigravity's rules page says "Rules files are
limited to 12,000 characters ...
VibeTags v1.3.4
[1.3.4] - 2026-09-10
Fixed
-
One module's guardrails are no longer written twice when the leftover sidecar is the newer file
(issue #621). A repository whose sources live in one subproject below the VibeTags root can end
up with two sidecars for what is really one module: the ancestor identity a compilation falls
back to when it cannot resolve its module root from the round's sources, and the module's own.
The prune that exists to retire the leftover consulted timestamps before containment, so it only
let a descendant retire its ancestor when the descendant was the fresher file. Reverse the write
order and nothing was retired at all: the ancestor was protected because its descendant was
older, and the descendant was protected because the ancestor's smaller element set did not cover
it. Both survived, and every element they shared was written twice, byte-identical, wrapped in
twoVIBETAGS-MODULEregions, into every granular rule file and every aggregate, with exit code
0 and nothing on the console.Containment is now decided first and without timestamps, provided it is strict: the nested
modules must between them know at least one element the ancestor never had. Recency is not
evidence of correctness here and arguably the reverse, because a nested identity is only ever
produced by resolving a real build file while the ancestor identity is the fallback. Equal
element sets stay with freshness, because that is also what "the sources moved up out of the
subproject" looks like, and there it is the nested sidecar that is the leftover. A reactor root
that compiles sources of its own still keeps at least one element no submodule has, fails
containment, and keeps its region.The symptom is worth recognising, because it points away from the cause: only some annotated
elements duplicate, and they are exactly the ancestor's smaller set, which reads like a bug
specific to certain annotations and is nothing of the kind.-Avibetags.modulewith a stable
name remains the way to rule the whole class out, by pinning the identity so the fallback is
never produced. -
The consumer sweep no longer skips the one downstream repository it can least afford to lose
(issue #617).tools/consumer-sweep.shrefused any repo with a dirty working tree, and that
guard ran before the branch that sweeps a contended repo in agit worktree. A worktree sweep
builds a separate directory fromorigin/mainand never touches the contended checkout, and a
repo is on that list precisely because someone else works in it, so a dirty checkout is its
normal state. The sweep's footer prints identically either way, so covering four repos of five
read exactly like a complete run. -
Release notes are extracted by one script that refuses to emit a truncated file (issue #619).
The extraction was an inlineawkrange, and an awk range tests its end pattern against the
record that opened it, so/^## \[/closed the range on the section heading and the command
emitted the correct release title with the whole body missing.gh release create --notes-file
accepts that without complaint, by which point the tag exists and the deploy has fired.
tools/release-notes.shis now the only implementation, bothdocs/RELEASING.mdand the
release skill call it, and it fails rather than hand fewer than five lines to a step that
cannot be undone.
Changed
async-test-libmoves from 1.11.2 to 1.12.0 (#623). Test scope only; nothing ships to
consumers from it.
VibeTags v1.3.3
[1.3.3] - 2026-09-10
Upgrading
One build outcome changes under you, and it changes in the safe direction. A compilation
that was shown only part of a module's sources now writes nothing at all, warns, and names
the sources it never read. The build that hits this most often is an ordinary incremental
Gradle compileJava after touching one annotated file: before this release that round
rewrote the module's guardrails from the one element it could see, deleting rule files for
every element it was not shown. If your workflow relied on an incremental compile to
regenerate guardrails, it no longer does; run a full compile (clean compileJava, or
gradle --rerun-tasks) when you want the files refreshed. Nothing is deleted either way.
Added
-
Four platforms from a sweep of the AI tool landscape on 2026-09-08, taking the totals to 41
platforms, 50 config files and 17 scoped-rule directories: Antigravity workspace rules
(.agents/rules/, which was half-supported before this, since.antigravityignorewas
already generated), JetBrains AI Assistant project rules (.aiassistant/rules/, a
separate product from the already-supported Junie), Augment Code workspace rules
(.augment/rules/) and goose (.goosehints). All three granular directories are
front-matter-free, each for a reason read from the vendor's own docs: Antigravity and
JetBrains set a rule's activation mode outside the file, and Augment defaults a
front-matter-free rule toalways_apply, which is the safe mode for a guardrail because its
alternative leaves loading to the model.PlatformSweep202609EndToEndTestalso pins
Antigravity's documented 12,000-character-per-rule cap, the one output constraint nothing
else in the codebase has had to respect. -
examples/basicnow opts into.gemini/rules, which it never had. Gemini's granular output
shipped in #320 and no example ever covered it, so for four releases it had no committed
fixture, no byte-for-byte drift gate, and no coverage of the dual-opt-in path where
GEMINI.mdcollapses to a scoped-rules index. Adding it moved that file from 274 lines to
93, with the safety buckets inline and 32 pointers into the scoped files, which is what
invariant 6 has always specified and nothing was checking.ExampleOptInCoverageTestnow
holds the example to every opt-in key, withroot_indexandlocks_reportexempted by name
and reason. -
Grok Build scoped rules: opting in
.grok/rules/writes one Markdown rule file per annotated
element there, bringing the count to 38 platforms and 14 scoped-rule directories. Grok reads
AGENTS.mdnatively and.claude/rules/for compatibility, so a project already opted into
either was reaching Grok before this; the new directory is for projects that keep Grok's rules
separate from another tool's. Three properties of Grok's discovery, taken from its own user
guide rather than from third-party guides, shape the output: it reads only*.md, it parses no
YAML front matter (so none is written, where Cursor and Windsurf get aglobs:block), and it
loads every file in the directory on every session rather than on demand by glob. That last one
is a cost to weigh before opting in on a large codebase. Grok has no VibeTags aggregate of its
own, so opting in never collapses another platform's file to a scoped-rules index.
Fixed
-
examples/basic/reset-ai-files.shnow clears.kiro/steering, which it had left behind since
v0.9.7. The script exists so a following clean compile proves the processor regenerates
everything, and it worked from two hand-maintained copies of aServiceRegistrysubset. A rule
file for a class whose annotation had since been removed survived the reset, the next compile
did not rewrite it, and the stale file read as current output, so the check proved less than it
appeared to.ExampleResetScriptCoverageTestnow derives both lists from the registry, in both
directions: every path the example opts into must be cleared, and every path the script names
must still be written by some service. It reported.kiro/steeringbefore the fix. -
A compilation shown only some of a module's sources no longer rewrites that module's guardrails
from what it happened to see. Reported from a Gradle single-module build: touching one annotated
file and running an ordinary incrementalcompileJavadeleted 22 checked-in rule files and cut
27 lines out ofCLAUDE.md, including a whole<security_elements>block, with exit code 0 and
nothing on the console. Gradle picks the files to recompile from the annotations it can see in
class files, and every VibeTags annotation isSOURCE-retention, so an edit to one file
recompiles that file alone; the processor was then handed one element where the last build had
twelve, and every element it was not shown read as an element whose annotation had been deleted.
PartialRoundDetectornow refuses such a round: nothing is written, nothing is swept, no
fingerprint is recorded, and aWARNINGnames the sources the round never read and how to get a
full compile. The refusal needs two independent facts, so that neither of its own failure modes
can happen — a sidecar for this source tree naming an element the round did not produce, and an
uncompiled.javafile under a source root the round did compile from that names the annotation
package. An annotation genuinely removed from a source the round did compile still loses its
rule file, and a source permanently excluded from compilation never blocks a write. -
The orphan sweep no longer reports files it did not touch. A hand-written
booking-flow.mdin
.claude/rules/was named inremoved=while sitting intact on disk, which inflated the
destructive-sweep count and sent the reader hunting for a deletion that never happened. A file
with no VibeTags markers is neither modified nor listed.
Changed
vibetags.logrecords the jar the running processor was loaded from alongside its version
(processor.origin version=… origin=…). A build reported a banner naming a release two versions
older than the only one its dependency block resolved, which reads as a hardcoded fallback
constant; there is none, so the answer is always "an older jar was on the processor path" and
the path itself says which one.
