Repository navigation
|
Thanks for ECC. We forked its The questionWhen one machine runs many agents at once (several parallel lanes, each a chain of nested sessions, each lane running its own repo's test suite), the host itself becomes the bottleneck: CPU, RAM, and the test processes each lane starts. Do you have, or plan, a version of the rules or a skill that covers this? In particular:
If this already lives somewhere we missed, a pointer would be all we need. What we found in ECC todayEvidence class: the files at HEAD
What we measuredEvidence class: measured on one machine, summarised from our own logs.
What we changed on our sideThis is shared as data and as one possible answer, not as a proposal for you to adopt:
Next stepIf this is outside ECC's scope, that is a fine answer too. If it is in scope, we are happy to share more detail on any of the above, or to test a draft of a rule or skill on this machine and report back. |
Replies: 2 comments 5 replies
|
I am not an ECC maintainer, so this is one practitioner's answer. I looked for the same thing in I run several Claude Code sessions at once on an 8-core Mac, each with its own worktree and test What we measured when we looked:
A test that ran a CLI six times, each run a dozen git spawns and two vitest children, took 30 s What we changed, in the order it paid off:
On your two questions:
On Track B of the roadmap: as I read it, that executor is about containing one agent's processes Happy to share the test-file diffs behind the numbers if useful. |
|
I am not an ECC maintainer, so this is one practitioner's answer. I looked for the same thing in
rules/andskills/and found what you found: parallelism is encouraged, the host is nevermentioned.
I run several Claude Code sessions at once on an 8-core Mac, each with its own worktree and test
suite. Load averages between 90 and 235 are normal there. What broke first was not the agents, it
was the tests: suites started failing by timeout, with no assertion failing.
What we measured when we looked:
gitprocess