Interestingly i had been building a terminal in rust and libghostty(with Linux and windows supported too) with built-in agent that understands terminal, too.
And the motivation was warp is doing a little bit more than a terminal.
I had been building con for a few weeks, it’s built in rust(GPUI) and libghostty, plus rig, the rust agent framework.
It’s built for the old-school terminal users, who, still do lots of work within terminal(not tui only), yet, with ai harness to help with your workflow anytime, and take over on our own , also any time.
I spent some efforts on:
- the harness, how can we abstract the state, structure that’s best for harness to control, leverage and manipulate the native panes, SSHed and/or TMUXed panes, or TUIs when needed, so that we can actually count on the agent to do things, like you create your own release workflow as a skill, effectively and smoothly. The way to do so was in a benchmark driven, self-evolve approach that’s inspired by the auto-research and Ralph loop way, with, both benchmark infra, cases and judges etc AND the harness env design in the evolve loop, and each loop are ChatGPT 5.4 120min runs, and interestingly I made it in a shape that’s acceptable as 0.1 beta in a few days.
- ensure it’s an elegant terminal, I am still working on this, but this process takes heavy me in the loop when we are using rust on ux improvements and performance tuning :)
- supports windows, well, I cannot stop doing it, although there are some many missing pieces out there and till today, windows version is runnable, still needs perf tune and refactorings, and in the meantime I am still working on the Linux desktop version!
And now, I decided to share it first before any landing page or blogs/tweets on hacker news!
I was reading an article earlier today, and it brought me back to a question I’ve heard over and over again in real data/infra teams:
Do we just accept vendor lock-in because it’s convenient,
or do we take the pain and build an open, multi-engine metadata stack?
For context (not my product, just what triggered the thought):
https://medium.com/p/35cc5b15b24e
I’m not trying to argue Gravitino vs. UC here — I’m more interested in the architectural mindset behind these two approaches.
On the vendor-integrated side, the upsides are obvious:
smoother UX
one place for lineage/policies
fewer moving parts
But so are the downsides:
cost keeps creeping up
you end up tied to one engine/format
migrations basically don’t happen in real life
And on the open/composable side:
Spark/Trino/Flink/Ray all first-class
Iceberg/Hudi/Delta can actually coexist
Metadata isn’t tied to compute
But again:
inconsistent metadata models everywhere
no unified governance layer
someone eventually owns a pile of glue code forever
So I’m curious: what actually works in practice?
If your company had to make this choice:
Did you go all-in on a vendor, or build something open?
Did the decision age well after a year or two?
Has anyone actually avoided metadata sprawl without getting locked in?
Where do lineage, ACLs, policies, and the “source of truth” actually live in your setup?
Really interested in what folks think, especially if you're juggling multiple engines, table formats, and clouds.
I was reading an article earlier today, and it brought me back to a question I’ve heard over and over again in real data/infra teams:
Do we just accept vendor lock-in because it’s convenient,
or do we take the pain and build an open, multi-engine metadata stack?
For context (not my product, not promotional — just what triggered the thought):
https://medium.com/p/35cc5b15b24e
I’m not trying to argue Gravitino vs. UC here — I’m more interested in the architectural mindset behind these two approaches.
On the vendor-integrated side, the upsides are obvious:
smoother UX
one place for lineage/policies
fewer moving parts
But so are the downsides:
cost keeps creeping up
you end up tied to one engine/format
migrations basically don’t happen in real life
And on the open/composable side:
Spark/Trino/Flink/Ray all first-class
Iceberg/Hudi/Delta can actually coexist
Metadata isn’t tied to compute
But again:
inconsistent metadata models everywhere
no unified governance layer
someone eventually owns a pile of glue code forever
So I’m curious: what actually works in practice?
If your company had to make this choice:
Did you go all-in on a vendor, or build something open?
Did the decision age well after a year or two?
Has anyone actually avoided metadata sprawl without getting locked in?
Where do lineage, ACLs, policies, and the “source of truth” actually live in your setup?
Really interested in what folks think, especially if you're juggling multiple engines, table formats, and clouds.
I think the ultimate version in such use case would be carefully wire-up the baremetal one with ad-hoc in-mem-disk or tempdir :), this could be a future backend of py-pglite(planned in v2).
For now, it's more accessible for me to hack it in hours and it works.
yeah, w/o py-pglite attempt this should be the only approach, the pglite ideally could make it more flexibly/lightweight in unittest cases, but as you mentioned it's never felt slow, it should be fine to working on it.
And actually, more e2e cases I think it's way better to not use the lite backend.
the non-container solutions would do more like the lifecycle mgmt/isolated env prep/tear-down with elegantly designed abstractions. While I think similar abstractions could be done on top of containers.
Maybe we ideally could have unified abstractions on both container-based, wasm evantually to boost dx yet with different expectation of speed vs compatibility.
aha, SF is pure managed service, thus 100% local/or even lite-local of SF seems not feasible, maybe clickhouse/databend could be consider to enable such flexibility of lightweight testing?
And the motivation was warp is doing a little bit more than a terminal.
https://con.nowledge.co
Glad to see now warp is open-sourced