vanillagreencom Harness CI
Run the classifier to decide whether CI can skip product checks. Commit .kendex-generated.json with the renders after kendex refresh. The engine writes that inventory from rendered artifacts. In-place content and carrier package source remain outside it.
.agents/skills/harness-ci/scripts/harness-only \
--event pull_request --base "$BASE_SHA" --head "$HEAD_SHA"
Flags and exit codes: harness-only --help. Consumer setup: README.md. Workflow shapes to copy: references/wiring.md.
Use --mode docs for the docs-only path set that harness-only --help defines. It prints docs_only=true|false.
scripts/change-class answers the wider question every gate, workflow and lane asks: what kind of change is this diff. It prints one of render, trivial, micro, small and standard, takes the same event and endpoint flags, and hands every read of a diff range to harness-only or to orch's branch measurement rather than deriving one, the revision the base end of the range resolves to included. render is proved by re-rendering: kendex verify has to report files checked and none failed, and each changed path has to be one that run itself names, which today is the Claude instruction shim it listed and nothing else. Neither the inventory nor the install record is taken as proof of its own provenance, because the branch can rewrite both, so every other changed path answers standard with cause=render-path-unowned: a re-rendered skill, agent, hook or command, and kendex's own .kendex-lock.json and .kendex-generated.json with them, so every refresh diff answers standard. That stands until KEN-1673 has kendex verify print the rendered positions of each row it prints for the classifier to own paths from. A changed configuration or instruction source refuses every narrow class, not only render; the paths trivial, micro and small all refuse are orch's narrow-change.conf, whose ceilings are the last two classes' alone and which micro.md § Escape condition 3 states in prose. Flags and settings: change-class --help.
Required-context aggregators call scripts/aggregate-needs. Pass the full toJSON(needs) object, the classifier job name, its verdict, and each job that the verdict may skip. The helper rejects a failed classifier, a failed or cancelled job, and a skipped job outside that explicit set.
This package never edits a workflow
Nothing here writes .github/. Wire the one step yourself, once, from references/wiring.md.
The rules to hold when wiring it
Classify inside a job, never in on.<event>.paths. A path filter stops the workflow from starting, the required context is never created, and a merge queue waits forever on a check nothing will report.
Keep the required-context job unconditional. Gate the expensive lanes with a job-level if: off a changes job's output, or a step-level if: inside an aggregate, and let the aggregate that carries the required name run on every event.
A job-level if: needs a status function. Without one it keeps the implicit success() and skips the lane whenever the classifying job failed, which stands the expensive lanes down on exactly the diffs nothing classified. An aggregate accepts a skipped lane only after checking that the classifier ran and cleared the diff.
A lane reading a path family beside the verdict needs more than the status function. A dead classifying job publishes no outputs, so the family term reads empty and skips the lane on its own. Lift it behind needs.changes.result != 'success', the two-gate shape in references/wiring.md.
A step that installs a tool for an unconditional lane stays unconditional. A harness-only if: on the install, while the lane that runs the tool runs on every event, fails that lane on a harness-only diff. The tool commit-guards needs is commit-guards CHECKS.md § py-names.
Reading a verdict
stdout is the selected verdict line alone; changed paths and reasons go to stderr; exit 2 is a wiring error that prints no verdict.
Fail-closed
Every unprovable case answers false, which runs every lane (DEVELOPMENT.md § Invariants). --no-renames is fixed. change-class answers standard on the same terms.
A class is never read from an author-writable field. Not a label, not a branch name, not a pull request title, and no flag carries one: the author of the diff being judged writes all of them, so trusting one fails open on exactly the diffs that most want to pass. A caller that acts on a verdict without review runs the DEFAULT BRANCH's copy of the script against the pull request's tree, because the branch can change the script too.
Nothing out of the judged tree runs, and nothing in it is read as configuration. change-class fixes its measurement settings from its own process environment rather than loading the tree's kendex.settings.toml, and it refuses the render class outright in a checkout whose git directory carries a kendex arming record, because kendex verify would run that package's declared checker out of the tree under judgement.
The judged checkout is read, never written, and it must be a commit. A checkout with anything uncommitted is refused the render class: the proof would otherwise attest to whatever a step before the classifier left on disk. What the check establishes is that the tree matches the checkout's own HEAD, which the wiring shape makes the pull request's merge ref rather than --head; the ways those two commits differ all cost a package its row, never grant one. Nothing in the wiring writes there.