Repository navigation
Closed
Description
Open items after 2026.11.19.1 (#966), continuing the items of #953 that the round did not take up (design record .agents/docs/2026-11-19-ecosystem-rule-convergence-design.md §1 states its scope). Each is stated with what was measured and what was not.
Intermittent CI failures
e2e/openkal/images1033 (an ok image with astd::thread, run by okrun) timed out three times in CI with status 137. A timeout now asks the program to end before it kills it (mtest 0.10.0), and okrun names the openkal call each task of the image waits in (openkal-image 0.7.0): the next occurrence leaves its state. No occurrence since.e2e/openkal/sets986: oneopenkal: fault: an access to memory that is not there (SIGSEGV) at 0x2ewhile packing the carrier formacos-aarch64, in four local runs. Not reproduced since.e2e/toolchain/custom875 expects one reason for the fast path to decline; under a shared home a parallel test rewrites the runtime description first, and the case's expectation is narrower than its subject.windows / bare Windows: no Visual Studio:mingw/182starts from a fresh home and downloads its toolchains; across the round's runs it took 48 s, 185 s, 238 s, 298 s, and more than its 300 s budget once (run 44592183424). The variance is the source xlings downloads from with the mirror left onauto: in that runxim:llvm@23.1.3(312 MB) came through ghfast.top in 155 s; in run 44572146728, on the same runner image, it came from github.com in 6.7 s. Which source a GitHub runner is given, and whether a source is chosen by its throughput, is xlings' to answer; the case's budget is the default one.e2e/openkal/setsande2e/openkal/treesshare one Linux unit shard with each other's parts: sets (1/2), sets (2/2) and trees (2/2) ran beside one another in each of the round's last four runs. They took 1134, 1367 and 1615 s (run 44553014658) and 1224, 1459 and 1720 s (44531373098) of their 1800 and 2400 s, and exceeded them in 44572146728 and 44592183424; theretrees/992's polyglot pack reached its 600 s while compiling its seventh carrier. Run one after another in one process, the trees cases take up to 236 s each, 992 163 s (44592183424, attempt 2). The work did not change: from a cold dependency cache the sets program ran in 375 s under fe72481, the last commit before per-program resolution, and in 372 s under 7b3f707, the round's last change of code, with 3392 and 3424 s of user time and the same 602 MB cached; a polyglot pack took 94.5 and 93.7 s. What the shard plan places together, and what a part may take beside others, is open.- A selection of a suite's cases runs under the run's limit, not the suite's.
mcpp test --failed(a CI rerun) and--case 'sets/*'select a suite's cases without the suite, and the time the suite declares (timeout{minutes{30}}) is counted only when the suite's own entry is selected: the rerun of 44592183424 ran ten openkal cases as one process under--timeout 1200and ended it during the tenth, each of the nine before it having passed; locally,--case 'sets/*'ended the program at the default 300 s.
Found in the release's verification
- The skeleton's mtest is 0.9.0.
mcpp newwritesmcpplibs.mtest = "0.9.0"(kMtestRelease, src/build/test_protocol.cppm), while mcpp's own tests and modules moved to 0.10.0 in this release, whose timeout asks a program to end before it kills it. Both speak protocol 1 and a new project builds and tests (clean room of 2026.11.19.1 through xlings);.github/tools/check_version_pins.shdoes not hold the two equal.
Not decided
- Upstream feedback for the external validation project built with mcpp and openkal on its fork: every check passes there; nothing was sent upstream.
Settled in the round, for reference
The Termux app sandbox (a build without units says so, compiler-run/no-unit), the refused admission's condition, the package instances (SPEC-004 §9 item 17), zlib in the Windows LLVM payload's lld (xim-pkgindex#1196), and the arch leg of ci-fresh-install (pacman -Syu).
Metadata
Metadata
Assignees
Labels
No labels
Activity
speak-agent commentedon Nov 20, 2026
A further measurement for the CI budgets of the openkal programs: in run 44646887291 (#978, a workflow-only change on main at 2026.11.19.1),
e2e/openkal/imagesended at the run's limit of 1200 s, six of its cases having passed, withimages/1033running (status 137). The suite states no time of its own, so it runs under the run's--timeout; it took 949.55 s for its eight cases in run 44553014658.mcpp test's limit ends the program's unit by a kill (src/run/run.cpp,end_unitandkill), without the request that makes okrun write the openkal call each task waits in: an occurrence ended by the program's limit, rather than by the 10 min a case's command may take, leaves no record of where the image waited. Whether this one waited or was only slow is therefore not known; the rerun of the failed shard follows.docs: the test and CI design for mtest, mcpp's test system and CI (mc…
feat: a test and a build share the machine's budget of slots and a me…
2026.11.21.1: the test and CI design (mtest 0.11), the home's unit ca…
speak-agent commentedon Nov 22, 2026
Released in mcpp 2026.11.21.1 (#984), by the test-architecture design of 2026-11-20: one budget of slots for the machine (
MTEST_BUDGET, SPEC-017 O11) bounds the compilers that nested tests start; a case selected with--caseor--failedruns within the timeout its listed suite states; the shared home is a run fixture and a case that changes a home gets its own, sotoolchain/custom875 no longer meets a parallel case's rewrite; the skeleton's mtest is mcpp's own release (0.11.2) andcheck_version_pins.shholds the two equal. The download time ofmingw/182depends on the source xlings selects; its suite states 15 minutes for a cold cache. The release's full CI run (44930053323) passed on every host. After the release, #998 gives a command of the kit the time its case has left (instead of a ten-minute default below it), judges the surface ledger in thefulltier only, and counts the Linux unit shards from the longest shard of main's last run, so that the openkal programs do not share one shard past their stated times.