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.
CI Wait Attribution
Outcome and Order
The owner approved diagnosing long-tail waits, then removing proven repeated
test preparation, then reviewing the remaining installation chain. Start from
master 3619b19d9. This first PR is observational, not completion of that order.
Seven first-attempt green runs each covered all 405 selected files. Workflow wall time ranged from 6m04s to 9m00s. On the identical final PR/merge source trees, config-save tests took 18.72s/46.35s while process CPU fell from 16.51s to 11.20s. Show and media tests on the same shard also grew in elapsed time. Existing CPU, context-switch and byte counters cannot attribute that gap to scheduler delay, disk latency, database locks, child processes, or application waits.
Scope and Measurement Contract
Only the existing per-file metrics helper, its consuming pipeline tests and this plan change. Retain the original selection, one interpreter per file, six shards, 300-second visible watchdog, 20-minute job bound, all 17 checks and their fail-closed aggregates. No new dependency, sampling thread, runtime hook, privileged command, kernel setting or application change is needed.
Take two bounded, read-only snapshots: at metrics initialization (after the launcher imports pytest) and just after pytest returns, before interpreter shutdown. Emit their interval separately from the existing launcher wall and process-lifetime counters. The existing shell Finished line remains authoritative.
- Thread CPU and Linux schedstat deltas belong only to the launcher thread, not all application threads or children. A changed thread identity invalidates thread deltas. Runqueue time means runnable but waiting to receive a CPU.
- Require schedstats enabled at both endpoints before exposing scheduler deltas. Disabled, unavailable, malformed or decreasing counters produce null, not a fabricated zero. No attempt is made to enable kernel accounting.
- CPU, I/O and memory PSI deltas describe the whole host during that interval, not this process. They are contextual evidence, not per-test I/O/lock latency. System-wide CPU full pressure is undefined and deliberately omitted.
- Do not subtract overlapping or differently scoped counters to label a residual as disk time. Neither absent counters nor host pressure alone proves a cause. These observations cannot retroactively explain prior runs.
The kernel documents schedstat counters and PSI scope and units.
Verification and Delivery
Contracts cover counter parsing, availability, disabled accounting, changes at the interval boundaries, reset counters, thread identity, nanosecond/microsecond units, and host/thread separation. Existing real subprocess contracts continue to verify collection, execution, teardown, child CPU and failed process exits.
Require a new head-bound bot pass, complete review/thread inventory, every exact-head lint run and all 17 checks before guarded integration. Inspect all 405 file exits and metrics, plus actual distribution/installation suites, then the exact merged-source CI. Record the observed wait evidence before choosing the next bounded preparation optimization. Do not rebalance from variance or replace real migration, locking, legacy-import or package-install coverage.
Local validation: 60 pipeline/metrics cases, 6 shard contracts and 12 private database fixture contracts passed in separate processes. Ruff 0.4.9, dependency consistency (84 packages) and whitespace checks passed. The Linux readers have synthetic parser/interval coverage on macOS; actual hosted counter availability and full-suite noninterference still require this head's CI. Review inventory starts at zero findings-bearing heads. No previous PASS applies to this PR.