feat(supervisor): record the command line of each process - #5254

Summary

Process Activity named each process only by its executable, so /usr/bin/bash or /usr/bin/python3 didn't say what it was asked to run. With the separate, off-by-default ocsf_process_cmdline_enabled setting on, Launch records carry process.cmd_line.

Last of a stack of four for #4268 (based on #5253):

1. feat(supervisor): identify the process behind each network event #5251 Identify the process behind each network event
2. feat(supervisor): record launch and exit of sandbox sessions #5252 Record launch and exit of sandbox sessions
3. feat(supervisor): record each exec in the workload and its exit #5253 Record each exec in the workload and its exit
4. feat(supervisor): record the command line of each process #5254 This PR: record the command line of each process

Related Issue

Closes #4268

Changes

• The runtime records argv with each launch: the spawn's own program and args for the entrypoint and sessions, and the argv each exec passes, read from the caller's memory. It keeps at most 256 args and 4 KiB per process, and a shortened list ends with ….
• ReadProcessEvents.command_lines is an optional bool, applied alongside exec recording. The supervisor withholds command lines unless the current setting asks for them, which covers takeover and turning the setting off.
• cmd_line quotes arguments the way a shell would read them, and ANSI-C quotes control and bidi characters. The shorthand [cmd:…] escapes control characters and ], so a workload can't forge log lines.
• New setting ocsf_process_cmdline_enabled; docs in ocsf-json-export.mdx ("Record command lines") and logging.mdx.

Before: no process.cmd_line on Launch records.

After (585/585 Launches carry cmd_line):

{"launch_type":"Exec","pid":81,"name":"/usr/bin/dash","cmd_line":"sh -c 'echo \"in-sandbox: sh pid=$$\"; /bin/true; /bin/sleep 0.01; echo done'"}
{"launch_type":"Exec","pid":86,"name":"/usr/bin/sleep","cmd_line":"/bin/sleep 0.01"}
{"launch_type":"Exec","pid":91,"name":"/usr/bin/bash","cmd_line":"bash -c 'exec 3<>/dev/tcp/example.com/443 && echo connected'"}

Log-injection attempt, sh -c '/bin/echo "$(printf "x\nOCSF NET:OPEN [INFO] ALLOWED forged -> evil.example:443")"':

OCSF PROC:LAUNCH [INFO] /usr/bin/echo(98) [cmd:/bin/echo $'x\nOCSF NET:OPEN [INFO\] ALLOWED forged -> evil.example:443']

No forged NET line appears.

Known limitations:

• Arguments can hold secrets, and with the setting on they reach every OCSF destination: shorthand logs, openshell logs, and JSON export. There is no redaction, which is why the setting is opt-in and separate from exec recording.
• An exec's argv is read before the kernel runs it, so another thread can change it first (audit only).
• An older runtime ignores command_lines without a warning.
• Terminate records don't carry cmd_line. The network actor's cmd_line is still the policy-matched paths.
• argv is read before the rate-limit check, so the read cost is paid even for suppressed execs.

Release Notes

Set ocsf_process_cmdline_enabled to include each process's command line in OCSF Process Activity Launch events. Command lines can contain secrets and appear in every OCSF destination.

Testing

• Lockfile checks, rust:format:check, and every CI clippy invocation (--locked, including perf-harness) pass on macOS and the aarch64 Linux target for this commit
• Unit tests added/updated (argv limits and truncation, quoting, shorthand escaping, withholding); sandbox (305), sandbox-backend (111), and supervisor (300) suites pass in an Ubuntu 24.04 guest (kernel 6.8); openshell-ocsf (180) passes on macOS
• Manual end-to-end: gateway + Docker sandbox (kernel 7.0), before/after OCSF output above

Checklist

• Follows Conventional Commits
• Commits are signed off (DCO)
• Architecture docs updated (if applicable)