Skip to content

armgit / avibe

Public Imported from GitHub
0

Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.

7.6 KB
Raw Blame

Documentation-only CI fast path

Current scheduling policy: the fast path applies to PRs. Latest-wins master cancellation requires every replacement master run to validate the full tree, including code from an older cancelled run; even a docs-only push uses full CI. See ci-under-10m.md for that follow-up contract. The original diff-classifier and failure-closed gate design below remain unchanged.

Contract

Owner request: avoid a full lint run for changes that are proven to affect only narrative documentation, reducing contention for the organization's 20 hosted-runner slots. Base: 9b63a3009 (#2527). No branch-protection changes, production behavior, deployment, regression update, or merge authorization.

  • Always trigger the workflow; do not use workflow-level path filters, changed file API pagination, or the push event's possibly truncated commits array.
  • A single lightweight changes job classifies the complete Git tree diff. PR: compare the event base to the checked-out synthetic merge tree, verifying its two parents match the event base/head. Push: compare before to after, verifying the checked-out SHA and rejecting new/deleted/forced pushes. Checkout depth two suffices for PR merge parents; fetch the validated prior SHA on a multi-commit push only if missing. Unknown/missing/unreadable evidence selects full CI; failure to execute classification leaves the final gates red.
  • Disable rename detection so both old and new paths count. Require ordinary non-executable Git blobs (100644 / absent); symlinks, executable files, submodules, type changes, malformed/empty diffs, mixed code/docs and unsupported events select full CI. No user input can request a skip.
  • Initial allowlist: root AGENTS.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, VISION.md, VISION_ZH.md; .github/PULL_REQUEST_TEMPLATE.md; narrative .md files beneath docs/, excluding the entire docs/plans/ subtree, contract directories (contracts / *-contracts) and docs/INSTALL_FOR_AI{,_ZH}.md. Tests dynamically scan all plans, not just a finite list of known contract files. README.md and README_ZH.md remain full (distribution metadata/install instructions). Runtime Markdown in core/prompts, skills, UI fixtures and test assets is never inferred to be documentation by extension alone. New executable docs must be excluded when their consumer is introduced.
  • changes emits an explicit lowercase docs_only=true|false and a human reason. A docs-only classification also runs the existing AGENTS/review-rule contract in tests/test_repository_guidance.py, extracted unchanged from the application-dependent Harness tests. A failed lightweight contract fails changes; its output alone cannot authorize gate success. All expensive root jobs require a successful classification with docs_only=false. Artifact consumers remain dependent on the artifact job.
  • Preserve all nine required check names. The three existing aggregate jobs remain always(), also require changes, and accept skipped producers only for an explicit successful docs-only classification. Full-CI paths still reject failure, cancellation, skip, or missing results. Leaf jobs skipped by an intentional job-level condition are completed/skipped GitHub checks, not absent/pending workflow checks. No fake check-runs or extra write permissions.
  • Documentation fast path uses four small hosted jobs (classifier + existing three aggregate checks), no application dependency installs or test matrices. Only Python and pinned pytest are prepared for the lightweight guidance contract, with no app conftest/plugins or cache writes. It checks scope and existing guidance requirements, not prose style. Code PR adds only the classifier's small startup dependency, not full Git history downloads.

Validation plan

Use existing pytest (no new test framework) with temporary real Git repositories and event fixtures: complete multi-commit diffs, changes beyond 300 files, renames across code/docs, non-ASCII names, mixed changes, merge parent/SHA mismatches, forced/new pushes, absent revisions and ambiguous events. Verify real workflow aggregate shell behavior for docs/full/error and each producer result, and changed job selection/required contexts. Keep existing CI boundary coverage rather than adding assertions of tool spelling.

First PR run must use full CI because workflow/classifier/tests change. Exercise the classifier against a real historical docs-only PR's merge and event SHAs with the new script, and execute the actual aggregate shell steps locally for both paths. A real docs-only GitHub scheduling result remains a post-merge acceptance observation: do not add a bypass input, create a stacked PR or alter another Agent's PR just to manufacture it. No unrelated PRs are modified or cancelled. Exact-head Codex, all expected CI results and zero open review threads remain close-out gates.

Implementation and evidence

  • scripts/ci_docs_only.py is standard-library-only. Classification does not install Avibe, execute documentation, read user data or call a changed-file API. Errors choose full; missing/invalid outputs never authorize gate success. Docs-only guidance validation installs pytest, not Avibe or its dependencies.
  • Workflow-level triggers and all nine required names are unchanged. Expected docs result is: changes and three aggregate jobs succeed; expensive leaf checks explicitly skip. These skipped leaf statuses are intentional completed checks, not a claim that their underlying tests ran.
  • Initial focused tests: 89 passed. After the review-boundary correction: 94 passed (classifier, pipeline and lightweight guidance); the remaining Harness guidance tests pass separately (9 passed). Ruff/actionlint pass. The 31 assertions moved to the lightweight owner are AST-identical to the previous function. A real Git docs-only mutation that removes a required AGENTS rule still classifies docs-only, but the actual standalone guidance test fails; aggregate failure cases prove a failed changes cannot pass.
  • Real docs-only PR #2533 replayed read-only with a depth-two checkout: base 9b63a3009bef1c81d95c6d866f909179fb7c9cc0, head a66890ba25aa47b63314f1d2274fa6a605e17c80, synthetic merge ea74a8cc110de06f0d1faef864ecda75463472fc. It changes only AGENTS.md; the new classifier emits docs_only=true. No changes, comments, cancellations or pushes were made to that PR.
  • First full-CI run, exact-head Codex review, and a future real docs-only scheduling observation remain separate verification layers. No time saving is claimed measured until the corresponding GitHub run exists.

Review boundary correction

Codex reviewed 6b664a178 and found one root-cause class: Markdown consumed by existing tests was incorrectly allowlisted. Independent read-only review identified the same cases. test_model_hub_config.py also reads implementation and routing plans; test_show_page_legacy_ui_removed.py globs all top-level plans; test_harness_skill_guidance.py reads root AGENTS.md. The fix is not a larger finite exception list: all plans remain full CI, and AGENTS retains its existing assertions in a dependency-light owner. No asserted contract was removed or weakened. This is the first findings-bearing reviewed head; no review-loop circuit breaker has fired. Full CI for that head passed in 11m16s, and its scope job explicitly reported full due to the workflow change.

Ownership

Session sesp2bd67rp9c owns implementation, independent review decisions and the single delivery PR. A read-only design reviewer may challenge the skip boundary; no parallel mutable lane is needed for this tightly coupled workflow change.