Looks like a good start - running through some tests, a few notes:

• cpu usage isn't actually tracked
• mem_free isn't accurate
• I'd suggest gathering timestamps, both to verify the sampling interval (in case anything stalls the tracker), and to independently provide per-PID resource net consumption. - realizing this is what 'ts' is

cpu usage isn't actually tracked

This is now fixed. CPU usage is defined as the percentage of use over the interval; eg. our default calculates the usage over the last two seconds

I also added some new CPU features:

• cores parameter to be passed in which is used in the new cpu_avg summary stat
  • If the program uses less cores than the system total, this helps track the efficiency of the program itself
• sys_cpu_per_core retrieves the CPU usage of every core
  • Summary stat sys_cpu to get the average of these

mem_free isn't accurate

This has been replaced with mem_avail. Previously, mem_free just reported the amount of memory that was completely unused and didn't account for memory that was actually available on the system. We care about availability, not unused

Some other changes:

• The first row written to the jsonl contains information that doesn't change, such as cores, mem_unit, mem_total, polling_interval, etc -- may be useful to power users
• The memory units is now configurable. Default is GB but could be changed to anything else, if that's ever desired
• ts renamed to timestamp for clarity

Working great for me now, thanks for the improvements!