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 render-candidate only to gate the engine installation and mirror refresh. It prints render_candidate=true|false, permits new head-owned paths, and grants no CI skip. change-class must prove render before CI skips product checks.

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 --json has to report files checked and none failed. The proof passes --at-record beside --base, so each package that follows its source renders at the commit the install record names, held to its source's history and to no older than the base's record, and a render the catalog has moved past since its push is still weighed and prints one render-stale: line per source commit it trails; and each changed path has to be owned by a position a passing record of that run prints — a file, a tree, or keys in a shared file whose rest kendex reports unchanged since the range's base. Neither the inventory nor the install record is taken as proof of its own provenance, because the branch can rewrite both; each is owned from the record kendex verify prints for it, which passes only where the file is as kendex writes it. Every other surviving path answers standard with cause=render-path-unowned, and a shared file kendex cannot vouch for answers cause=render-path-partial. Both ownership refusals carry measured=true: missing ownership evidence, including foreign=unknown, selects standard with full checks. A required classifier read failure or a failed verifier result stays unmeasured and stops consumer refresh. Deletion requirements and refusals: change-class --help. A changed kendex configuration file 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. That list names the files a package's risk sits in, never the package whole. In the size path, a names-only inventory change that unlists only paths the diff deletes leaves the path set the list and the subsystem rule read, under the conditions in change-class --help. An agent instruction file, an instruction line of that list (AGENTS.md and SKILL.md), that earns trivial or micro answers small, cause=instruction-file, the lowest class the default review policy gives a bot round; item-tier holds a Location naming one to small from the same lines. Every verdict also prints a queue-only: queue_only=true|false line on stderr, off the queue lines of the same list, the repository's own HARNESS_CI_QUEUE_PATHS, the lanes its default branch's .github/ci-lanes.conf marks :queue, which the change-class action defers off a pull request to the merge group, and the jobs its HARNESS_CI_QUEUE_SELECTOR command names for the merge group; change-class --help states when it reads true. github.sh pr-merge reads that line before it takes the admin route. Flags and settings: change-class --help.

Required-context aggregators call scripts/aggregate-needs. Pass the full toJSON(needs) object, the classifier job name, and each job a verdict may skip: --skippable beside the one --waiver, --lane JOB=LANE, whose skip the lane's own verdict in the classifier job's outputs authorizes, or both. The helper rejects a failed classifier, a failed or cancelled job, and a skipped job no verdict stood down. A workflow with several lanes declares the paths each reads in its default branch's .github/ci-lanes.conf, and the change-class action answers one verdict per lane: references/wiring.md § Per-lane verdicts.

This package never edits a workflow

Nothing here writes .github/. Wire the one step yourself, once, from references/wiring.md. A repository under the organization standard reports the aggregate CI context: it copies templates/ci.yml to .github/workflows/ci.yml, or names its own workflow's aggregate job CI, per references/wiring.md § The CI context.

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 tools commit-guards needs are in commit-guards CHECKS.md § py-names and § secrets.

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 takes its measurement settings from its own process environment, else from the [env] tables of kendex.settings.toml and .kendex/settings.toml, the second winning, as the commit --base names holds them (the pull request's base tip or a merge group's base, never the merge base), parsed and never sourced; the judged tree's settings and the private env file never reach it. It runs kendex verify in a private checkout with a git directory of its own, which carries no kendex arming record, so no package's declared checker runs out of the tree under judgement.

The judged checkout is read, never written, and never weighed. The render proof weighs the commit --head resolves to, checked out privately, whatever the judged checkout's own HEAD is and whatever uncommitted content it holds.