[tobert]

The functional test harness never compared Expect.ExitCode. Nine fixtures set it, and none was tested. With the check on, three fixtures showed a real bug: otel-cli exec exits 255 when its child is killed by a signal.

Behavior change for review: exec now exits 128 + signal for a signaled child, which is the shell convention: 130 for Ctrl-C, 137 for --command-timeout (SIGKILL). Before, it exited 255. Scripts can see this change. The alternative is to keep 255 and pin it in the fixtures.

What changed:

• main_test.go: checkProcess compares the exit code when the process exited on its own. A killed process reports -1, so it is skipped.
• data_for_test.go: the #360 fixture uses ExitCode: 42 in place of a CheckFunc; "protocol: bad config" with --fail expects 1. The three signal fixtures expect 130 and 137. Their old values (2 and 0) were never checked.
• otelcli/exec.go: 128 + status.Signal() when ProcessState.Sys() is a syscall.WaitStatus and Signaled().
• otelcli/exec.go: signal.Stop(signals) before close(signals). Before, a Ctrl-C during the span export panicked with "send on closed channel" (exit 2). Now SIGINT has its default action. The new fixture "exec ctrl-c during export does not panic" was red before the fix.
• CHANGELOG: two lines under 0.6.0 Fixed.

Commits are test-first: each fix follows a failing fixture.

Verification:

go vet ./... && go build && go test ./...                 # green
GOOS=windows|darwin|freebsd go build -o /dev/null .       # ok
./otel-cli exec --command-timeout 50ms -- sleep 1; echo $?   # 137
./otel-cli exec -- sh -c 'exit 3'; echo $?                   # 3

Found by the kaibo review of AGENTS.md on #49. After this merges, the AGENTS.md hazard "the harness ignores Expect.ExitCode" is no longer true; #49 drops it.

🤖 Generated with Claude Code

[tobert]

kaibo review (cast deepseek). Claude checked each applied finding against the code.

Applied:

• checkProcess skipped the exit-code check whenever TimedOut was set. The server timer sets TimedOut even when otel-cli exits on its own (main_test.go:446-456), so a real exit code could be skipped. The guard is now !CommandFailed only.
• The ExecExitCode: 1 expectation in "protocol: bad config" checked nothing. Diagnostics.ToStringMap omits exec_exit_code, and only exec writes it. Removed.
• The CHANGELOG line now says Unix. On Windows, WaitStatus.Signaled() is always false.

Confirmed correct: 128 + signal on Linux, darwin and freebsd. Diag.ExecExitCode has one writer and is set before any export-time SoftFail. --fail precedence is unchanged from #360. Background fixtures do not race on ExitCode. No fixture passes by accident (all 11 ExitCode fields traced).

Not fixed here, verified and pre-existing:

• exec.go calls close(signals) while signal.Notify is still registered. A second Ctrl-C during the export window sends on a closed channel and panics. The fix is signal.Stop(signals) before the close.
• The signal fixtures have 50-200 ms budgets. This PR adds no new timing; the budgets were already enforced through TimedOut.
• Lower priority: status always exits 0; start failures always exit 127 (shells use 126 for "not executable"); EventCount is never compared.

🤖 Generated with Claude Code