4 ms·
diffwatch is kinda general purpoure, besides the agent work you could watch different processes doing stuff in your homedir, for example
by deemkeen 11mo ago
diffwatch is kinda general purpoure, besides the agent work you could watch different processes doing stuff in your homedir, for example
- westurner 11mo agoCool tool! Is the inotify directory/file watch count the limit? I can't seem to remember the name of the pre-containers tool that creates a virtual build root and traps all the file syscalls. It's not strace. Easier to trace everything an AI runs by running the agent in a container with limited access to specific filesystem volumes. eBPF is the fastest way to instrument in Linux AFAIU: Traceleft: https://github.com/ShiftLeftSecurity/traceleft https://github.com/ShiftLeftSecurity/traceleft Tracee: https://github.com/aquasecurity/tracee https://github.com/aquasecurity/tracee Falco docs > Supported events: https://falco.org/docs/reference/rules/supported-events/ https://falco.org/docs/reference/rules/supported-events/ Tetragon: https://github.com/cilium/tetragon https://github.com/cilium/tetragon strace could have a --diff-fs-syscall-files option: strace -p PID -f -F -e trace=file -s 65536
- deemkeen 11mo agoit uses the os independant fsnotify lib, it surely has its limits. eBPF is great, but linux only, yeah
- westurner 11mo agoOn MacOS: sudo dtrace -n 'vfs::*:entry { printf("%-16s %-6d %s", execname, pid, probefunc); }' sudo dtrace -n 'vfs:lookup:entry { printf("%-16s %-6d %s", execname, pid, copyinstr(arg1)); }' TIL Dtrace is included in recent builds of Windows 11 and Server 2025: https://learn.microsoft.com/en-us/windows-hardware/drivers/devtest/dtrace https://learn.microsoft.com/en-us/windows-hardware/drivers/d... ; # Must be run as Administrator dtrace -n "syscall::NtCreateFile:entry, syscall::NtReadFile:entry, syscall::NtWriteFile:entry { printf(\"%s (%d) - %s\", execname, pid, probefunc); }" It's possible to trace file system calls in Windows with procmon.exe by saving a .pmc config file and then loading it from the CLI: procmon.exe # uncheck everything except "Show File System Activity" # Filter > Drop Filtered Events # File > Export Configuration... # Must be run as Administrator procmon.exe /AcceptEula /Quiet /Minimized /LoadConfig C:\Tools\fs-only.pmc /BackingFile C:\Logs\FileSystemTrace.pml It's also possible to trace lower level file system calls in Windows with logman.exe but it's necessary to parse the traces that it generates. Then with just bpftrace on Linux: sudo bpftrace -e 'tracepoint:syscalls:sys_enter_openat { printf("%-6d %-16s %s\n", pid, comm, str(args.filename)); }' sudo bpftrace -e 'kprobe:vfs_read, kprobe:vfs_write, kprobe:vfs_open { printf("%-16s %-6d %s\n", comm, pid, probefunc); }' ... According to 2.5pro on the cli strs strace, dtrace, and bpftrace could have a --diff-fs-syscall-files option.
- deemkeen 11mo agogreat insights, i'll read up on it and see if it can be useful, thx
- westurner 11mo agonp. there's a diagram, "Linux bcc/BPF tracing tools" [-1] in the bcc readme [0] that's also in [1] which explains ebpf and bcc and bpftrace. filetop, dirtop, and vfsstat use bpf to trace the VFS layer. [4] [-1] "Linux bcc/BPF tracing tools" https://www.brendangregg.com/BPF/bcc_tracing_tools_early2019.png https://www.brendangregg.com/BPF/bcc_tracing_tools_early2019... [0] iovisor/bcc: https://github.com/iovisor/bcc https://github.com/iovisor/bcc [1] "Linux Extended BPF (eBPF) Tracing Tools", Dtrace book: https://www.brendangregg.com/ebpf.html https://www.brendangregg.com/ebpf.html If running an AI agent in a container --- with devcontainers and e.g. vscode, Good container policy prevents granting a container the CAP_SYS_ADMIN capability; the least-privileges thing to do is to grant limited capabilities to the container like CAP_BPF and (CAP_PERFMON, CAP_NET_RAW, CAP_SYS_PTRACE) [,3]. [3] https://medium.com/@techdevguides/using-bpftrace-with-limited-capabilities-inside-docker-a-capability-focused-guide-851695a13197 https://medium.com/@techdevguides/using-bpftrace-with-limite... [4] bpfcc-tools manpages: https://manpages.debian.org/unstable/bpfcc-tools/index.html https://manpages.debian.org/unstable/bpfcc-tools/index.html though ripgrep wins, vscode fails at monitoring large workspaces due to inotify limits too; so some way to parse fs events from bcc and libdtrace with python would be great prompt 1: Create a python project named idk dbpftrace with a pyproject.toml and a README and sphinx /docs, with bcc and python-dtrace as dependencies to, then in dbpftrace/, parse pid and descendents' fs syscall events from bcc (ebpf) or python-dtrace (dtrace), depending on which os we're running Edit: Prompt 1B: Create a Go package named dbpftrace with a README and docs, parse pid and descendents' fs syscall events from bpftrace or dtrace stdout, depending on which os we're running
- westurner 11mo agoPrompt 1C: Create a Go package named dbpftrace with a README and docs, then create a cli utility named dbpftrace to: parse pid and descendents' fs syscall events (like bpftrace) using libbpfgo and godtrace Use either (cilium/ebpf or libbpfgo or gobpf) or (godtrace or (CGO or FFI) bindings to libdtrace) depending on which OS, by default cilium/ebpf: https://github.com/cilium/ebpf https://github.com/cilium/ebpf aquasecurity/libbpfgo https://github.com/aquasecurity/libbpfgo https://github.com/aquasecurity/libbpfgo iovisor/gobpf w/ bcc: https://github.com/iovisor/gobpf https://github.com/iovisor/gobpf chzyer/godtrace: https://github.com/chzyer/godtrace https://github.com/chzyer/godtrace oracle/dtrace-utils/tree/devel/libdtrace: https://github.com/oracle/dtrace-utils/tree/devel/libdtrace https://github.com/oracle/dtrace-utils/tree/devel/libdtrace From https://news.ycombinator.com/item?id=45755142 https://news.ycombinator.com/item?id=45755142 re eBPF for WAF: > awesome-ebpf > Kernel docs, examples, Go libraries: https://github.com/zoidyzoidzoid/awesome-ebpf#go-libraries https://github.com/zoidyzoidzoid/awesome-ebpf#go-libraries : >> Go libraries: >> cilium/ebpf - Pure-Go library to read, modify and load eBPF programs and attach them to various hooks in the Linux kernel. >> libbpfgo - eBPF library for Go, powered by libbpf. >> gobpf - Go bindings for BCC for creating eBPF programs