https://nf-co.re logo
Join Slack
Powered by
m
must use codex more.
ai thumbsdown 1
By this I meant it produced some funky combined grep commands to find the intersection of topics implemented but not with a typical eval 😛 I am playing some some automation/agent frameworks for topic version migration and have a few open PRs from that atm. Not sure how useful/automated it will be but I will be considerate and not bomb the repo with too many requests at once (will keep to <5 and won't automate the actual PRs) Does anyone else know if anyone's working on mass migration of topic versions? I figure automation getting even 30% of the more simple modules completed would be a big win. Things get tricky when there are lots of subworkflow and module dependencies involved
🙏 1
b
actually I suspect codex and claude code should be very good at this I just did a similar thing with the experimental record type syntax. I refactored an nf-core pipeline to use record types, then I changed the record syntax, then I wrote a spec for claude on how to update from one to the other it worked quite well! I'm sure you could do a similar thing for topics with a sufficiently detailed spec the only thing was, I had to tell it to do one file at a time... otherwise it tried to read all modules at once and blew up the context window...
m
Yeah nice, thanks for sharing I have had it make some good progress but it is easy to get in the weeds and try to meet lots of different edge cases which adds too much complexity and things start to break
I have been experimenting with cron jobs to trigger an agent to spawn every x minutes and process another module
Such that it doesn't have that same context explosion issue
b
I have heard that claude takes more liberties with your instructions, whereas codex only does exactly what you say, so it may be that with codex you have to be more verbose (and with claude maybe I got lucky with its guessing 😅 )
m
I'm actually playing with openclaw and Claude sonnet on a sandbox device
On my work computer I use codex
The models definitely behave a bit differently 😛
Trying to build an openclaw skill for it but I imagine that handing this set of scripts and markdown files from that skill to codex or Claude would have similar results
m
I don't think mass conversion will be that helpful because they all need to get reviewed and nobody wants to see a flood of open PRs ready to be reviewed (and I keep seeing errors/missed things in these individual PRs, so reviews are needed. but there is a hackathon project concerning this https://nf-co.re/hackathon-projects/hackathon-march-2026/modules (which is also a great way for beginners to join the hackathon (a classical "LLM taking away experience from juniors"-situation I would say))
m
I hear you, to be honest I think there are plenty of modules to go around and the automation framework is hardly sophisticated enough to get through the 1000 plus modules that remain to be migrated Conscious of the hackathon topic and not interfering with it
And agree that there's no point flooding the modules repo with 1000 PRs. More thinking some small batch automated framework feeding in modules with a manual (human) curation layer pre PR review
2 Views
Open in Slack
PreviousNext
Powered by