Repository navigation
assurance: probe binaries, the trend tier and per-commit measure history #81
Description
Measures that need a process-wide allocator, a fresh process for RSS or wall-clock timing run as tools/probe binaries. A scheduled workflow runs every tier on the default branch and attaches each run report under refs/notes/measures (assurance.measure.history).
Ledger entries: audit-open-bounded-heap (retained heap across 1k and 100k entries) and audit-query-digest-keyed (unkeyed dictionary matches), each with its clause, and the trend entries.
Activity
assurance: measure history reads Dispatcher execution ids, and the no…
assurance: measure history reads Dispatcher execution ids, and the no…
assurance: measure history reads Dispatcher execution ids, and the no…
debuggingfuture commentedon Oct 16, 2026
#563 settles the note credential and remote attestation (
assurance.measure.note-credential,assurance.measure.remote-attest) and pinsassurance.measure.run-report:measure-reportreads FlareDispatch execution ids as run ids. The nightlycontextful-measuresrun already attaches reports underrefs/notes/measures. The remaining step is fractalboxdev/flare-dispatch#199, which builds the note line withcontextful-ci measure-reportso nightly notes carry trend annotations; this issue closes once it deploys.debuggingfuture commentedon Oct 18, 2026
The 2026-10-18 nightly run is the first to post a
flare-dispatch/contextful-measurescheck run. It was admitted at 02:17 UTC and is still queued an hour later: every sandbox slot is held, with no run queued ahead of it, and its admission deadline is 08:17 UTC. The newestrefs/notes/measuresnote is still the 2026-10-15 run.That pattern fits the missing notes: a nightly run that never gets a slot writes no note. Several check executions on merged pull requests are still
in_progresshours after dispatch, which points to stalled executions holding the pool rather than to a failing measures step.Next: read this run's terminal state. If it times out in admission, the fix belongs on the dispatcher side, either by releasing slots held by stalled executions or by reserving capacity for scheduled runs, rather than in this repository.
debuggingfuture commentedon Oct 18, 2026
At 03:10 UTC on 2026-10-18 the pool is busy with gate checks, not with measures.
check:*children dispatched between 02:30 and 03:00 UTC are running for pull requests that have already merged (#608, #612, #613, #614, #618), next to children stillin_progresssince 2026-10-16 and 2026-10-17 (#605, #607). Merged pull requests keep consuming sandbox capacity, and the scheduled measures run waits behind them for its whole admission window.Two dispatcher-side fixes would address this: cancel a pull request's gate executions when it merges or closes, and give scheduled runs admission ahead of pull-request checks.
debuggingfuture commentedon Oct 18, 2026
Tonight's run did get a sandbox slot. It then ended at 03:59 UTC with Failing step:
measure—StepFailed:ExecFailed: exec failed (exit -1): Connection closed: this Durable Object instance is no longer active. The container was recycled mid-measure, and themeasurestep ran withretries: 0, so no note was written. This is the failure that removes notes. The queueing described above delays a run but does not by itself drop a note.The fix is fractalboxdev/flare-dispatch#204, merged as 759aaa5a. The
measurestep now rebuilds a recycled checkout inside its retry and retriesExecFailedandCheckoutFailedtwice, and the report steps read the workspace it measured. This issue closes once a nightly note withannotationslands.