Peter Edmonds

@thdxr the #1 most annoying thing with oc right now is stale lsp diagnostics. this happens basically on every edit: > agent adds import > lsp: "Error [78:10] Cannot find name 'foobar'." > agent goes crazy looking through dts files trying to debug is this a known issue?

@tim_smart what’s “lalph”?

Assigning tasks to different models seems to be working quite well. Codex for the serious coding work, opus for things like frontend tasks or more "creative" work.

@mitsuhiko probably...durable sandboxes? opencode is great but doesn't have a solution for non-fs persistence (yet? @thdxr) which makes it a non-starter for building widely used agents on top of it. basically just want something like opencode but in an actor.

Three months ago we all built durable execution engines, now we're all building sandboxes. Let's make some predictions what we will build three months from now!

@RhysSullivan @dillon_mulroy nice! i first saw this in @ethanniser's code, but i'm sure many have rediscovered it.

there's a really nice `use` pattern by @dillon_mulroy for wrapping third party services with effect i always reach for it but can't remember the specific style off hand, so i turned it into a skill skills.sh/RhysSullivan/e…

@dillon_mulroy @hugorcd @boristane what does evlog in the effect ecosystem look like? just Effect.annotateLogs, or is there something more?

evlog is a work of art @hugorcd also h/t to @boristane for his article, Logging Sucks, that inspired it

@theo good day to be @useblacksmith i suppose.

Holy shit, github actions have actually been down for over 2 hours 🙃

@adamdotdev my instinct is to save messages/parts on our end and treat each request as a new session, but aware that cache breakpoints are a delicate beast. don't want to lose all my cache inputs by mismanaging the part format

@adamdotdev for people building on top of opencode server ala Ramp Inspect, is there a suggested path to session persistence?

@vercel code review bot is great. just caught a niche FE resource leak that no other reviewers (human or llm!) noticed. integration into existing workflows needs some work though. would love a cli version so my agents can use it w/o visiting _yet_ another dash cc/@cramforce

@thdxr cool will take a look this week! for people building on top of opencode server ala ramp's inspect, is there a suggested path to session persistence? my instinct is to save thread data on our end and treat each request as a new session. wdyt?

we setup a beta channel for opencode you can opt in with `bun/npm/etc i -g opencode-ai@beta` right now we're testing our sqlite migration - if you run into any issues reply in this thread

@zeeg First project I've seen based on the Anthropic Agents SDK! How did you land on that vs Vercel SDK, OpenCode SDK, or others?

@chaliy @jerrickhakim @vercel @blackboxai @roocode no, sandboxes don't support Docker. all sandboxes run on al'23. snapshots are analogous docker images in that you use them to spawn reproducible instances.

@jerrickhakim @vercel @blackboxai @roocode I do not follow how this would help.

Vercel Sandbox, the secure compute environment for agents, is now generally available. Built on the infrastructure powering 2.7M mission-critical daily builds, Sandbox is in production for platforms like @blackboxai and @roocode. vercel.com/blog/vercel-sa…

@zeeg > using neon

why is the vercel marketplace so hard to use trying to upgrade a neon instance we have for an internal project and there user interface is worse than navigating a casino five minutes in and i still cant figure out what button to press

@sqs @AmpCode skills.

@AmpCode Anyone want to bet on (or suggest) the next Amp features we'll kill?

We're r̶e̶m̶o̶v̶i̶n̶g̶ slashing custom commands in @AmpCode. Use skills instead. ampcode.com/news/slashing-…

what michael and jude (@vercel) are *correctly* reporting is that, while progressive disclosure is a useful pattern, the implementation doesn't matter. i'd still benefit from a public implementation for Effect, whether "skills" or "indexed markdown". are these public @MichaelArnaldi ?

@MichaelArnaldi @baldram @peteredm0 Thanks.

Hot take: skills are the new MCP, nothing more than a hype of the moment. Models are great at reading and writing code, use markdowns to index code don't over engineer it.

@grinich @kushalbyatnal brilliant. i'll give it a swing this weekend.

@peteredm0 @kushalbyatnal try running `npx workos`

@MichaelArnaldi for encoding best practices and stylistic choices, like "prefer Effect.Service" there's many ways to achieve the same thing in effect. from my experience, a small amount of guidance on preference helps a ton.

@peteredm0 So cool why? We tested it doesn’t make the models better at all

@a1zhang @daytonaio why should subagents not share a sandbox? isn't it important for a subagent to say "i've written the results to /results" or similar?

I've been super excited to get Daytona sandboxes support integrated into the RLM repository (which now works in the official repo). Check out this awesome guide on RLMs with *depth>1* with code examples from @daytonaio! daytona.io/docs/en/recurs…