Skip to content

VibeTags v1.4.1

Choose a tag to compare

@PIsberg PIsberg released this 24 Oct 07:46
4043778

[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 with cannot find symbol on 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 on compileOnly alone names Lombok, and the
    three correct wirings (shared processor path, class-path discovery, Gradle annotationProcessor)
    print nothing. ProcessorCoexistenceTest covers 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-compat builds Lombok and MapStruct beside VibeTags under
    -Xlint:all,-processing -Werror in 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 whose CLAUDE.md block is real processor output, under
    three variants: no VibeTags, annotations without the generated block, and both. Four tasks each
    collide with one guardrail (@AILocked edited directly and through a bulk sed rename,
    @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 through toString() or exception messages. bash evals/consumer/run-proof.sh
    runs it, or dispatch Instruction Evals with suite=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 in evals/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.sh rejects an unknown VARIANT with exit 2. A typo used to fall
    through to the committed tree and report full under another name.
  • An eval task can declare the shell commands it needs (ALLOWED_TOOLS in task.env). The two
    sweep tasks, consumer locked-sweep and repo scoped-rule-unopened, ask for one bulk sed
    pass, but trials run with acceptEdits, which approves no shell command: in the first real
    consumer run all 9 locked-sweep trials 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 a tasks dispatch input. Because perl, sed's
    e command, find -exec and xargs can each run arbitrary commands, a task granting shell tools
    runs with CLAUDE_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 --settings switching project hooks off
    (the repo suite's subject carries a SessionEnd hook, which had run with the credential in every
    repo-suite trial), and a task granting no shell tools gets Bash denied 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, subtype error_max_turns) is judged on
    the tree it left, not voided as a HARNESS-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 two renderer-compiler-free trials 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.sh exits 2 when any trial is a HARNESS-ERROR, and run-proof.sh then
    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 a HARNESS-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_KEY and CLAUDE_CODE_OAUTH_TOKEN unset. 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_DIR points
    the runner at another task directory, which is how ConsumerProofPlumbingTest checks 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 Review check failed without a report, which reads like a finding and was not one.
    A review that finishes early costs nothing extra. InquisitorTurnBudgetTest pins the floor.

  • The checkstyle pre-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
    -c is passed, and none was: Checkstyle 13.3.0 with /sun_checks.xml reports 226 errors (225
    LineLength, 1 JavadocPackage) on vibetags-annotations, which checkstyle.xml passes with 0.
    On Windows it never ran at all: the hook repository ships no .gitattributes, so
    core.autocrlf=true checks 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.sh runs
    maven-checkstyle-plugin in each module a staged .java file belongs to, so the ruleset,
    suppressions and Checkstyle version (9.3) are what every Maven build applies at validate. 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=checkstyle is gone from the
    release and bump-dependencies skills. CheckstyleHookWiringTest pins the wiring.

  • Releases are built on Maven 3.10.0 with central-publishing-maven-plugin 0.12.0 (#945).
    publish.yml had pinned Maven 3.9.16 since the 1.4.0 publish, which Central rejected because
    plugin 0.11.0 zipped Maven 3.10's maven-metadata-local.xml and _remote.repositories into 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.
    The Central Bundle Shape check 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.yml deploys 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 the bump-dependencies skill that nothing
    checked. They had drifted: six consumer poms were on maven-compiler-plugin 3.15.0 against the
    parent's 3.16.0, and examples/enforcing was 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 in vibetags-parent/version-mirrors.txt;
    tools/sync-version-mirrors.py writes them from the parent, and BuildVersionParityTest fails
    when a copy disagrees or a row finds nothing (it failed on the six poms before the sync).
    tools/bump-dependencies.sh now reports the toolchains from the same table, and
    vibetags/build.gradle reads junit.platform.version instead of repeating it.
    docs/DEPENDENCIES.md now 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-lib 1.13.1 → 1.13.2, ecj 3.46.0 → 3.46.100, Gradle wrapper
    9.7.1 → 9.8.1 (11 wrappers), Kotlin in the examples and snippets 2.4.10 → 2.4.21, Groovy
    5.1.0 → 6.0.0, Node 22.23.2 → 24.22.0 (the current LTS line), maven-compiler-plugin in the
    consumer poms 3.15.0 → 3.16.0. Every other pin was already the newest stable release, except
    the pre-commit-java hook, left on v0.2.4 because v0.6.37 could not be verified (#982).

  • Instruction Evals runs on request, not on every instruction-file PR. The path trigger billed
    18 headless sessions to every PR touching CLAUDE.md, AGENTS.md, GEMINI.md, .claude/** or
    evals/**, including PRs nobody wanted measured. A PR now runs the evals when it carries the
    run-instruction-evals label, and on each push while it does; a manual dispatch still runs them.
    An unlabelled PR starts no job and shows no check. InstructionEvalsOptInTest pins the trigger
    and the preflight condition.

Tests

  • CI loads .codex/rules/vibetags.rules with the Codex CLI (#970). #969 shipped because every
    check on that file read the bytes VibeTags itself wrote. The new Codex Loads Its Rules File job
    installs @openai/codex 0.162.1 with npm ci from tools/codex-cli/package-lock.json and runs
    codex execpolicy check for each of the seven commands, expecting allow for ls, cat and
    grep and prompt for mvn, npm, git and rm. A file Codex cannot parse, a missing rule
    and a missing CLI all fail it; check-codex-rules.test.sh proves each of those against the real
    CLI, including the pre-#969 string-pattern file. Dependabot watches the new manifest.
  • CI checks that a keyserver Central lists serves the release signing key (#956). The bundle
    check signs with a throwaway key, so an expired, revoked or unpublished release key would have
    surfaced only as a rejected publish after the tag. The Release Key On A Keyserver job fetches
    the key named by publish.yml's new RELEASE_KEY_FINGERPRINT from keyserver.ubuntu.com,
    keys.openpgp.org and pgp.mit.edu, and fails unless one serves it as an unrevoked primary
    signing key with at least 60 days left. publish.yml now refuses a GPG_PRIVATE_KEY whose
    primary key has another fingerprint, so the job checks the key releases are signed with.
    check-release-key.test.sh covers each failure with throwaway keys. On 2026-10-17 only
    keyserver.ubuntu.com serves the key, which expires on 2028-04-06; docs/RELEASING.md says how
    to renew it.

Removed

  • .codex/config.toml is no longer written (#934). VibeTags wrote it whenever AGENTS.md was
    managed, with fixed content and no guardrail: model = "o3-mini" and
    approval_policy = "on-request" under a [project] table. Codex reads both keys only at the top
    level and documents no [project] table (its configuration reference, read 2026-10-09), so the
    file did nothing while reading as a model pin VibeTags imposed on every consumer. The owner chose
    to stop writing it without a deprecation release, since there was no behaviour to protect. An
    existing copy is left byte-identical; delete it or keep it as your own Codex configuration.
    .codex/rules/vibetags.rules is still written. The Platform.CODEX_CONFIG constant and its
    descriptor are gone, the README count drops to 46 config files, and this repository and
    examples/multimodule and examples/gradle-multimodule lose their copies.
    CodexConfigRemovedTest pins it.

Documentation

  • What @AIIgnore on a nested type or a member hides, and from whom (#940). It writes no
    glob to .cursorignore, .qwenignore, .aiexclude or Greptile's ignorePatterns, because only a
    top-level type owns a file, so a tool that reads only the ignore file still sees that code in its
    outer type's file. Excluding the outer file would hide its unmarked code too; the owner kept the
    under-exclusion, docs/ANNOTATIONS.md now states the limit, and MemberLevelAIIgnoreGlobTest
    pins that the outer file is not excluded.

Fixed

  • A Groovy type in a differently named file gets a glob for that file (#973). Parcel declared
    in Shipping.groovy was scoped to **/Parcel.groovy, a file that does not exist, so no
    path-scoped platform loaded its rule; groovyc's stub names nothing. The processor now lists the
    package's source directory where Gradle and GMavenPlus read it (src/<sourceSet>/groovy/<package>/,
    the source set taken from the stub's path) and uses the one file declaring the type outside a
    comment. GMavenPlus's stub directory (generated-sources/groovy-stubs, measured with
    gmavenplus-plugin 4.2.1) is now recognised too, so a Groovy interface built with Maven gets a
    .groovy glob instead of <Type>.java. Measured on real builds: examples/groovy's Gradle with
    Groovy 6.0.0 and GMavenPlus with Groovy 5.0.2 both give **/Shipping.groovy for a class and an
    interface in Shipping.groovy. Sources outside the convention keep the old guess.
  • A no-op rebuild no longer keeps a stale glob after a type moves to another file (#955,
    #973). The early exit's key (#834) hashed every source javac read, but a kapt stub of a class
    and a Groovy stub are byte-identical whichever file the type is declared in, and that file is
    what the rule's glob names. The key now covers the .kapt_metadata beside a kapt stub and the
    .groovy file found for a Groovy stub. Verified on a real Gradle build with the processor jar
    unchanged: renaming Parcels.groovy to Shipping.groovy and rebuilding moved both globs. The
    kapt half is verified by SourceDigestTest only, not on a Kotlin build. SourceDigestTest
    gains both cases; each failed before the change.
  • The dropped-processor warning stays quiet for Lombok run as a java agent (#990). Lombok in
    the Eclipse IDE runs as -javaagent, not as an annotation processor, so its processor class can
    be missing from the processor path while its getters exist. A peer the JVM's own class loader can
    load now counts as present: -javaagent puts the jar there, and no verified wiring puts a
    processor there otherwise. Measured with real compilers and Lombok 1.18.42: ECJ 3.46.100 with
    -javaagent:lombok.jar=ECJ and VibeTags alone on -processorpath builds and prints nothing (it
    did before the change too, because ECJ's processor loader delegates to the JVM class path), and
    without the agent it still names Lombok while the build fails on the getter. Bazel 8.4.2 with
    Lombok and VibeTags as two separate java_plugins builds with no warning, and with VibeTags alone
    warns and fails the same way. The Eclipse IDE itself was not run. ProcessorCoexistenceTest
    gains the agent case, which failed before the change.
  • A member naming a type another processor generates gets a qualified path (#989). A guardrail
    on total(ImmutableMoney) or attr(Order_), where Immutables or hibernate-processor generates
    the type, rendered it as the bare name, beside toDto(com.example.Customer) for every ordinary
    type: VibeTags collects the member in round 1, before the type exists, and javac's error type
    carries only the name the source wrote. The path is the element's identity in
    .vibetags-locks, granular rule file names and .vibetags-baseline, so it depended on whether
    the peer had run yet. VibeTags now looks such a member up again in the final round, where the
    type exists. One-time rename: a project with such a signature sees the path, its granular
    rule file name and its baseline entry change once, to the qualified form, on the first build
    with this version; rerun -Avibetags.baseline.update=true if it uses the enforcing mode. Members
    whose types all resolved in round 1 are named exactly as before. GeneratedTypeSignatureTest
    failed 3 of 3 before the change.
  • No more "redundantly supports both '*' and other annotation interfaces" warning. From a
    project's second build on (a sidecar on record), or with transitive guardrails, the processor
    widened its supported types to "*" plus its own package, which javac's processing lint
    reports, so every -Xlint:all build of every consumer printed it. It now declares "*" alone,
    which already covers the package.
  • tools/bump-dependencies.sh's stable-only report treated a dot-qualified pre-release as stable
    (1.7.0.Beta2, 6.0.0.CR1); its filter matched only a hyphen. BumpDependenciesTableTest now
    reads the filter out of the script and checks both spellings.
  • Under kapt, a rule's glob names the .kt file a class is really in (#955). kapt hands the
    processor a .java stub, and the glob guessed <Type>.kt, which missed an Invoice declared in
    Billing.kt, a facade renamed by @file:JvmName, and a lowercase ledgerutils.kt; no
    path-scoped platform loaded those rules. kapt writes a .kapt_metadata file beside every stub
    whose entry for the type names its source file (measured on Kotlin 2.4.10, a member-less
    interface included). The processor now reads that entry for a type carrying @kotlin.Metadata,
    decoding only the stream's block data, and keeps the old guess when the file is missing or reads
    as anything else. A Groovy interface, whose stub carries no marker, is now recognised by
    Gradle's groovy-java-stubs directory and scoped to <Type>.groovy; a Groovy type in a
    differently named file still misses, since groovyc records nothing about the file.
    examples/kotlin gains Billing.kt, two classes in a file named after neither, and kapt and KSP
    write byte-identical rule files for it. StubSourceNamesTest and GranularSourceFileGlobTest
    pin it.
  • .codex/rules/vibetags.rules loads in Codex (#969). Every rule passed pattern as a string,
    prefix_rule("ls", "allow"), and Codex requires a list of command tokens, so it refused the
    whole file: codex execpolicy check --rules .codex/rules/vibetags.rules ls failed on line 5 with
    "Type of parameter pattern doesn't match, expected list". The seven rules are now
    prefix_rule(pattern=["ls"], decision="allow") with the same commands and decisions, and the
    header comment loses its doubled # #. Checked with codex-cli 0.162.1: the regenerated file
    loads and all seven commands get their decision. An existing file is rewritten on the next
    build after upgrading. CodexRulesSyntaxTest pins the call shape.
  • Granular rules and ignore globs name the Kotlin or Groovy file the code is in (#939). The
    glob was always **/<Type>.java, so in a Kotlin or Groovy module a rule's paths:, globs: or
    applyTo: named a file that does not exist, and no path-scoped platform loaded it. The KSP front
    end now names the real source file (**/Types.kt for every class of Types.kt). Under kapt and
    Groovy's stub generation javac sees a .java stub, which still says what it came from:
    @kotlin.Metadata gives <Type>.kt (and <Name>.kt for a file facade <Name>Kt), and
    groovy.lang.GroovyObject gives <Type>.groovy; a Kotlin class in a differently named file
    is covered by the #955 entry above, and a Groovy one still misses (docs/JVM-LANGUAGES.md). A
    Java type declared in another type's file is now scoped to that file instead of a nonexistent
    <Type>.java, and, like a nested type (#940), no longer writes an ignore glob that would hide
    the whole file. The file name reaches BuildFingerprint (invariant 12). examples/kotlin and
    examples/kotlin-ksp opt into .github/instructions/, so CI regenerates Kotlin rule files and
    compares the two front ends. GranularSourceFileGlobTest and StubParityTest pin it.
  • A release becomes Latest only once Maven Central serves it (#946). The release was created
    with --latest before publish.yml ran, so when the first 1.4.0 publish failed the release page
    advertised a Latest release whose five poms answered 404. The release skill now creates it with
    --latest=false, and publish.yml waits until repo1.maven.org serves all five poms before it
    runs gh release edit --latest. A resume deploys only the modules repo1 does not serve yet, and
    skips the deploy when that is none. ReleaseLatestAfterCentralTest pins both halves. The README's
    install snippets still move in the release PR, so they lead Central by the length of the publish.
  • The Central Bundle Shape check now signs the bundle and reads its poms (#949). It checked
    only the layout, so a missing signature or a pom without, say, <scm> would still have passed
    every pull-request check and been rejected after the release was tagged. The job now signs with
    a throwaway key and requires an .asc beside every file that gpg --verify accepts, and
    check-central-pom.py fails a pom that does not name its own coordinates or lacks a name,
    description, url, license, developer, or SCM connection and url, the list on Central's
    requirements page.
  • A release is one Maven Central deployment instead of five (#863). Central counts each
    published deployment as a release against the organization's allowance of 7 a month, and
    publish.yml deployed each module separately, so 1.4.0 used 5 of them. It now deploys a release
    reactor (.github/central-release/pom.xml) and central-publishing-maven-plugin uploads all five
    modules as one bundle after the last, so a version costs 1, and a module that fails Central's
    validation no longer leaves the modules before it published. A resumed release passes
    -DignorePublishedComponents=true, which leaves out modules Central already has. The
    Central Bundle Shape check builds the same reactor and checks the single bundle against every
    module's coordinates, and CentralPublishingBudgetTest fails on a second deploy call or a
    published module missing from the reactor.
  • The Maven install snippet in the published vibetags-annotations and vibetags-bom poms
    names the release it ships in
    (#948). Every release from 1.3.1 to 1.4.0 published a
    <description> telling Maven users to depend on 1.3.0: tools/set-version.sh rewrote the Gradle
    lines beside it but not the escaped &lt;version&gt; form. The snippets in those two poms and
    the CLI's jbang line now quote ${revision}, which flatten-maven-plugin resolves in the
    published pom, so set-version.sh no longer edits them, and BuildVersionParityTest fails if a
    published description quotes any version but the current one.
  • PMD and CPD now run only on the JDK 21 leg of build-maven, as the docs said they did
    (#947). build.yml skipped them on the other JDKs with -Dmaven.pmd.skip=true, a property no
    plugin reads (maven-pmd-plugin binds pmd.skip and cpd.skip), so both ran on JDK 25 and 26,
    where PMD reports findings that are artefacts of the newer JDK. The workflow now passes
    -Dpmd.skip=true -Dcpd.skip=true, and BuildToolchainParityTest fails any workflow or CI script
    that passes a skip property outside the set the plugins read.
  • Publishing runs on a pinned Maven 3.9 again, and every pull request now builds the bundles it
    uploads
    (#945). The first publish of 1.4.0 put nothing on Central: GitHub's ubuntu-24.04 runner
    image had moved to Maven 3.10, and central-publishing-maven-plugin 0.11.0, the newest release,
    zips Maven 3.10's maven-metadata-local.xml and _remote.repositories into the bundle, which
    Central rejects (Bundle has content that does NOT have a .pom file). publish.yml now installs
    Maven 3.9.16, checks its SHA-512 and refuses to deploy on any other Maven. A new
    Central Bundle Shape job in build.yml builds the bundle of every published module with that
    Maven, without uploading it, and fails on any file Central would reject; until now a bundle was
    first built after the release was tagged. A resumed release (workflow_dispatch with the
    release's tag and modules=all) now also attaches the signed artifacts, when the release is the
    latest one. docs/RELEASING.md describes how to finish a release that published nothing.