Please do not open a public issue for a security problem. Email sn@cbox.dk, or use GitHub's private vulnerability reporting on this repository. Best effort, no promised response time.

By design, the minimum that answers "where did the time go" and "what was it doing when it died":

• Function names, class names, declaring file paths and declaration lines, as part of profiling.
• Durations and counts of a fixed set of native operations, by type.
• Engine counters (GC runs, sample counts, buffer overflows).
• Fixed-width breadcrumbs whose labels are operation and unit names only.

It does not capture function arguments, request bodies, query parameters, SQL text or bindings, environment variables, connection strings, cookies, headers, or any other application value — not by default and not behind a flag. There is no code path that reads them.

Crash records carry the same bounded data plus a signal number and si_code, a pid, a timestamp, the trace/span ids the caller supplied, and — for diagnosing crashes — three raw memory addresses: the faulting address, the program counter, and where this extension is mapped. Those are machine addresses, not application data, but they do describe the process's memory layout, so treat a crash record as something to keep inside your own infrastructure. Breadcrumb labels are truncated to 48 bytes at write time and are never free-form application strings.

Application-level redaction is cboxdk/laravel-telemetry's job, and it remains so. This extension has nothing to redact.

Crash records are written to cbox_telemetry.crash.dir (default /tmp/cbox-telemetry). The directory is created 0700 and the file 0600; it is opened O_NOFOLLOW, and a pre-existing directory is refused unless it is a real directory owned by the current user. If any of that fails the recorder disables itself and reports why in cbox_telemetry_status() — it never falls back to somewhere less safe.

The extension installs handlers for SIGSEGV, SIGABRT, SIGBUS and SIGILL. They chain to whatever was installed before and let the process die exactly as it would have, core dump included. cbox_telemetry.crash.enabled=0 disables them entirely.