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.

6.7 KB
Raw Blame

CI under ten minutes: queue demand and measured execution

Owner request (2026-10-13): continue the CI optimization roadmap toward a complete successful lint result under ten minutes, including runner queue waits. Not permission to skip failures, publish/deploy, reset environments, upgrade the GitHub plan, or merge without a new explicit owner instruction.

Base: master a5c6cbf17 (#2542). Session sesp2bd67rp9c owns the task branch vibe/20261013-ci-under-10m; no parallel mutable owners.

Evidence and honest acceptance

  • #2527 execution critical path reached ~9m21s but total elapsed was 15m43s. Organization Free concurrency is 20 hosted jobs. An org-wide timeline showed all 20 slots busy for 348 of one shard's 381 seconds of queue wait.
  • After #2540/#2541, successful master run 38355727304 (9efb8b95c) has 540 complete file records, 3943 launcher seconds (3745.60s before interpreter shutdown), and a 332–506s file-work imbalance across nine shards. The longest job is 551s. Rebalancing eight lanes may reduce queue demand without making execution exceed 600s; replay is evidence for a trial, not a guarantee.
  • Latest master run 38358514819 is failed, not a fast success: ARM Agent setup-python's PowerShell archive extraction crashed with an internal CLR error before any tests ran. Preserve this observation; do not weaken that gate or mix partial failed-run data into the timing snapshot.
  • Accept only complete successful exact-head runs. Report elapsed seconds, dependency critical path, ready-to-start queue gaps and total runner seconds. A single run under 600s is a measured sample, not a P95 guarantee. Saturated hosted capacity cannot be solved solely by YAML; surface capacity limits rather than repeatedly increasing shard count.

First bounded change

  1. Only test workflow lint uses latest-wins cancellation for its existing per-PR / master concurrency group. Every push trigger here is master; publishing/deployment workflows and other PRs' groups are untouched. A superseded commit may lack a completed master verdict; the newest must still pass all gates. Master pushes always use full CI, even for a docs-only before..after diff: a docs push may cancel an unvalidated earlier code push. Documentation fast-path authority is retained for PRs only. This deliberate trade-off was identified by the independent design reviewer and avoids creating a validation gap. No manual cancellation of peers' runs.
  2. Refresh per-file launcher timings from complete successful current-generation CI. Replay a separate run to check balance before trialling eight lanes rather than nine. Keep all discovered files exactly once, isolated-process watchdogs, integration exclusions and required check names.
  3. Add an observational-only impact category to the existing scope summary, reusing its complete validated Git diff. UI-local / backend-local categories are candidates for later analysis, not skip authority. Only the existing docs-only decision is exposed as a workflow output; unknown/shared/mixed changes stay full. No workflow condition consumes shadow telemetry.
  4. Run focused validation and first full PR CI; observe current-head review/CI through one durable Watch. If the target is unmet, diagnose the actual tail before another change. No inferred win from cancellation or missing jobs.

Later decisions, only after first run

  • Observe a real docs-only run when naturally available; do not fabricate a stacked PR or introduce a force-skip input.
  • Consider grouping short checks only if worth migrating required contexts; never silently delete the required checks or add placebo success wrappers.
  • Consider safe fixture overhead or batch-runner experiments only with real order-independence and hermeticity proof. The current process-isolation workaround remains needed until then; no broad local full-suite gate.
  • Keep Windows artifact/installer/engine coverage. Cache downloads rather than user state only if observed setup dominates after queue reduction.

First candidate evidence

  • Snapshot: median launcher elapsed time from complete successful master runs 38350418272 (f73106bd1, 538 files, 3843s) and 38355727304 (9efb8b95c, 540 files, 3943s); new files use their one available measured sample. The current network guard file has no sample in those runs and retains the existing structural fallback, not an invented measured duration.

  • Eight-lane median planning predicts 487–488.5s of launcher work each. Replaying the later source sample gives 476–503s; using only the older sample to plan and holding out the newer gives 466–515s. Add roughly 30–45s measured setup/cleanup overhead: eight lanes are a plausible sub-ten-minute candidate with less queue demand, not a promised service-level bound.

  • Observational labels are docs-only, ui-source-only, python-source-only, shared-or-mixed, unknown. They classify path areas, not independent behavioral boundaries: App/API/routing or migration code may still couple both sides. They appear only in stdout/Actions summary, never GITHUB_OUTPUT; the sole output remains docs_only and no new skip condition exists. Incomplete/invalid mode/event evidence yields unknown/full.

  • Independent read-only design review on contract c4c014684 found the docs-push cancellation hole above; no artifact consumer or isolation issue. All assertions, installers, branch protection and release workflows remain.

  • Focused validation: 50 docs/event/impact cases and 62 pipeline/shard cases pass in separate invocations, including a real code-B/docs-C push whose effective output must be full and a matching docs PR whose output stays true. The first combined probe exceeded its command deadline during execution and is not counted as passing; the bounded focused reruns completed normally. Ruff 0.4.9 and actionlint pass. No repository-wide local gate was run.

  • No ARM code change is needed for the observed failure: job 115166968293 failed inside setup-python extracting the x64 archive with PowerShell Internal CLR error (0x80131506) / exit 3221225477, before dependency setup and tests. It is not counted as success; future full master CI must verify Windows Agent coverage again. No retries/failure suppression were added.

Validation boundary

Existing pytest owns changed CI behavior. New cases must protect a credible failure mode (accidental gating on observational classification, incorrect path-side handling, losing required coverage or cross-PR cancellation), not pin tool versions for their own sake. Existing root/mode/rename/event validation from #2534 stays conservative. No runtime/application behavior changes.