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.

9.8 KB
Raw Blame

CI Critical Path

Baseline

Read at master 9993ce41ce3669e5d05ae167564daffb6cb4cbc5. Successful lint runs on September 5, 2026:

Run Source Wall time Artifact build Install regression Slowest Python shard
33952795137 9993ce41ce 13m24s 8m00s 5m17s 9m35s
33951672367 d5bd8d1cda 11m46s 7m26s 4m13s 9m23s

The artifact build followed by install regression is the critical path. In the first run, UI validation and tests consumed approximately four minutes before the production bundle could be built. Unit dependency installation took two to three seconds; adding more dependency caches is not the first lever.

Changes and Invariants

  • UI validation runs independently from production artifact creation. Every existing theme, lint, test-typecheck, test, and production-build command remains gated. The existing unit-tests aggregate requires both Python shards and ui-checks to succeed, including fail-closed handling of skipped/cancelled jobs. Artifact consumers still require the real production artifact build.
  • Packaged upgrade tests build the two core/Memory version pairs once per module, reducing sixteen wheel builds to four. Each case copies those immutable inputs into its own wheelhouse and installs into its own virtual environment. Deleting Memory wheels in the missing-package case cannot affect another test or the shared inputs. All upgrade and optional-import assertions remain unchanged.
  • The existing six-shard planner uses a refreshed timing snapshot containing all 396 files from run 33952795137. No planner algorithm or test selection changes. Every discovered file, including newly added files without timing samples, must appear in exactly one shard; unknown files retain the structural estimate.

Replaying the measured durations with the refreshed assignment gives about 7m23s per shard instead of a 9m24s slowest test step. As a holdout, run 33951672367 gives 6m42s to 7m14s with that same assignment. These are scheduling estimates, not a promise about runner speed or queue time. Compare actual PR job timestamps after the change before claiming an end-to-end improvement.

Evidence and Boundaries

  • Workflow contract tests execute the aggregate shell gate across all dependency outcomes and verify that UI checks cannot be bypassed by the split.
  • The real packaged upgrade tests cover core-only status, paired upgrades, core-only upgrades, and refusal when Memory is missing. Their fixture isolation also has a focused mutation test. Related scenarios: MEMORY-INDEP-021 and MEMORY-INDEP-024/025.
  • Full UI checks and builds remain required; so do package distribution/sdist contracts, released migration guards, pinned Memory runtime contracts, Docker install/upgrade regressions, and Windows installation smoke tests.
  • There is one additional lightweight Node setup/npm install for the independent UI job. No new dependencies, services, path-based skips, or reduced coverage.
  • Timing snapshots remain advisory and may age. Refresh from complete successful shard logs, preserve the source run/commit, and validate against another run before changing the snapshot again. Automatic refresh is outside this change.

October 2026: Runner Lanes and Browser Packages

Baseline

Sample: the 45 latest successful lint runs up to master f225b7a0b (39 first attempts). Median elapsed was 18.7 minutes; the 90th percentile was 64.6 minutes. Job setup was small: checkout, uv, and npm restores each took seconds and every dependency cache hit.

  • ui-browser-checks was the longest job in 18 of 39 first attempts. It took a median of 14.7 minutes, running ten suites serially with one worker. Each suite also spent about 30 seconds in tsc and Vite startup.
  • The slowest of the eight unit shards finished after 11 to 14 minutes. The suite had grown to 616 files and about 4,900 launcher seconds, and 77 files were missing from the timing snapshot. Per-file CI_TEST_METRICS show that 86% of file wall time was process CPU on a single core of a four-vCPU runner. Process startup and shutdown accounted for about 13%.
  • Slow runs had two causes. In 3 of 45 runs, playwright install --with-deps downloaded 119 MB of WebKit packages from the Ubuntu mirror in 13, 22, and 44 minutes; Playwright's CDN and apt-get update stayed fast. The other cause was queue saturation. The organization has 20 hosted slots, and one run requests 33 to 39 jobs. A burst of five Dependabot PRs took 65 to 75 minutes, with up to 48 minutes of shard queueing.

Changes and Invariants

  • Each unit runner executes two planned lanes concurrently through scripts/ci_unit_test_lanes.py. Shard i of N runs virtual shards 2i and 2i+1 of 2N. The existing planner and launcher are unchanged, so every discovered file still runs exactly once, in its own interpreter, with its own watchdog. Any failing lane fails the shard. The lanes are child processes, not shell background jobs, so test processes keep Python's default SIGINT handling.
  • The timing snapshot is refreshed from three successful master runs and covers all 615 files.
  • Browser suites run in two matrix lanes that are balanced by measured suite time. Each suite still runs exactly once with the same arguments. The unit-tests gate still requires the ui-browser-checks result, which fails if either lane fails.
  • apt keeps Playwright's system packages in a directory that is cached per runner image and Playwright version. apt still resolves, verifies, and installs every package, and downloads any package that is missing.

No tests, required checks, watchdogs, or runner labels are removed. Measured results for the pull request's own runs are recorded in the PR.

October 2026: UI Check Parts and Bytecode Measurement

Baseline

Master run 40481392503 (317377ee8, after #3026 and #3029) and PR runs since #3019 show two chains of roughly the same length. ui-checks took 5.5 to 8.8 minutes, depending on the runner. It ran validate:theme (5 to 9 s), lint (49 to 81 s), and typecheck:tests (24 to 48 s), then Vitest (206 to 351 s). The longer browser lane took 7.6 to 9.4 minutes. pdf-preview has grown from 50 s to 139 s, and lane 2's suites took 50 to 60 s longer than lane 1's in each of four runs. The install chain took about 7.6 minutes. Unit shards were 4 to 6 minutes.

Changes and Invariants

  • ui-checks is a two-part matrix. The static part runs theme validation, lint and the test type-check. The tests part runs Vitest. Each check runs exactly once. The required unit-tests gate still needs ui-checks, which succeeds only when both parts succeed. Lint and Vitest no longer share a checkout, so the lint-integrity test's temporary invalid source cannot reach the linter.
  • badge-triggers moves to browser lane 1. Every suite still runs exactly once with the same arguments.
  • Peak demand: the run requests one more hosted job when it starts, 19 instead of 18. The static part usually finishes before the artifact build does, so it is not competing for runners when the build's consumers are requested.

Bytecode Compilation Was Measured and Not Adopted

These probes ran on scratch-branch runs 40481504100 and 40483710679. Each probe compared install methods on a single runner, so runner speed was not a factor.

Measurement Lazy (current) --compile-bytecode
uv pip install -e . --group dev (14,114 files) 1.5–1.8 s 4.4–7.5 s
Shard 5, two lanes 241 s 232 s
Shard 0, two lanes 310 s 312 s
Full test collection, one process (two runners) 79.8 s / 78.0 s 78.7 s / 79.8 s
memory-insight contract pytest 4.7 s 4.7 s

Each runner compiles a dependency once, on the first import. Precompiling everything costs as much as it saves, and the differences are within run noise. lark_oapi is the exception. Its first import takes about 8 s cold and about 2.7 s compiled. It is imported only by the Feishu files, which all run on one shard, and the shards are not on the critical path. Compiling repository sources adds 1.8 s. The only measurable gain came from pytest's own rewrite cache, about 1.4 s for the contract job, and compileall does not produce that cache.

October 2026: One Ubuntu Web-Fetch Job

Baseline

web-fetch-runtime and the install regression become ready together, the moment the artifact build finishes. On the critical path, the install regression waited 2.4 to 7.3 minutes for a runner in #3019's runs. Nine web-fetch legs were requested at that same moment. Five of them were ubuntu-latest legs, one per supported Python (3.10 to 3.14). Each leg took 45 to 63 seconds: 4 to 9 seconds to install and 31 to 41 seconds to test. The tests are mostly waits; locally, 40 seconds of wall time used 5.6 seconds of CPU.

Change and Invariants

  • One ubuntu-latest job verifies all five Pythons in turn. Each version uses its own setup-python interpreter, found with uv python find --system, and gets the same three uv pip install --system commands, the same wheel, and the same pytest selection and flags. Every version runs even after one fails; each failure is reported as an ::error:: naming its version, and the job fails. The arm64, macOS, and Windows legs are unchanged.
  • At the build-finish moment, the workflow requests four fewer hosted jobs. The scenario-evidence gate parses the same pytest arguments for this job. The job uploads no junit report, so its matrix size changes no verdict.
  • web-fetch-runtime is not a required check, and no document or script references its per-version job names.

Measured durations and queue times for this pull request's runs are recorded in the PR.