ADR 0130: The CI nobody was reading¶

Status¶

Accepted

Context¶

CI on main had never passed. Not one run since the first, on 2026-09-11 — through a hundred commits, two CI jobs added to it, and a beta release cut and published with a CI badge in the README rendering red the whole time.

Nobody looked, including while adding jobs to it. Every "the tests pass" in that period was a local ctest run, which was true and is not the same claim.

Decision¶

Six failures, found by reading the logs in order and fixing what each one actually said.

The test that segfaulted on every Ubuntu run¶

memmem() is a GNU extension. Without _GNU_SOURCE it is not declared, so C implicitly typed it as returning int, the returned pointer came back truncated to 32 bits, and the memcpy through it landed on nothing. It had been doing that since ADR 0110's commit.

Replaced with a five-line local search rather than reaching for _GNU_SOURCE, because this file also has to build where memmem does not exist. For the same reason it no longer uses dirent.h, unistd.h or utime.h; cleanup asks ase_recovery_list() what is in the directory, which is portable and exercises the API besides.

The configure that failed before compiling anything¶

set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS ...) ran unconditionally, but that cache entry is only created under if(NOT CMAKE_CONFIGURATION_TYPES) and Visual Studio is a multi-config generator. Every Windows run died there.

It also took the macOS and Linux core jobs with it through fail-fast, so those had never reported either. fail-fast is off: one platform's failure hiding two others' results is the opposite of what the matrix is for.

The core that had never compiled where it claimed to¶

With configure fixed, Windows reached the compiler for the first time and stopped on <strings.h>. The list was longer than one header: strcasecmp (now ase_strcasecmp, _stricmp on Windows), usleep (now Sleep), and <strings.h> in two more places.

The failure that only ever happened on a runner¶

ase_lsp_client_tests aborted under ASan on CHECK(client != NULL), and reproduced on nothing here — not 22.04, not 24.04, not with sanitizers, not under artificial load.

A diagnostic printing errno answered it in one run: Broken pipe.

A command that does not exist execs, fails, and _exit()s. If it loses the race to the parent's initialize write, the pipe has no reader and the write takes EPIPE; if it wins, the bytes sit in the pipe buffer and the write succeeds. ase_lsp_client_start() tore the client down on one side of that race and returned it on the other — so a misconfigured lsp_command behaved differently depending on machine load.

write_framed() already sets alive = false, so the client was correctly dead either way and start() was throwing it away. It is returned now, and poll() reports the death, which is what ADR 0093 says happens when the handshake is not awaited.

The first version of that diagnostic printed to stdout and never reached the log, because CHECK calls abort() and an abort discards whatever is sitting in stdout's buffer. stderr, flushed.

The three that still crash¶

Windows now builds the core and runs its suites. Three segfault:

suite	dies at
`ase_config_tests`	inside `ase_config_create_default()`
`ase_theme_tests`	reading a static table it has just been handed a valid pointer into
`ase_plugin_host_tests`	immediately after `register_command` returns

That much is known because a crash prints nothing, so RUN() now names each case on stderr before running it and the last name in the log is the one that died. Struct packing, stdbool layout and the obvious ABI mismatches were checked and ruled out.

The job keeps running and does not gate. Deleting it would throw away the only signal anyone has about Windows; gating on it would block work on the platform this editor actually ships for. README.md and the install guide now say plainly that Windows is unverified — the previous wording, "the core builds and runs on Linux, macOS and Windows in CI", was untrue and was part of why nobody looked.

Consequences¶

RUN() is worth keeping regardless of Windows: any suite that crashes now says which case it was in, on every platform.

The LSP fix is a real behaviour change on Linux too, and it is the kind that only a busy machine would ever have shown. It is the one defect here that a green CI would have caught years of local runs could not.

What this cost: CI was added, extended twice, and never read. A red badge on the front page of the repository is a signal that decays to zero the moment it is normal, and it had been normal since the first run.