euphoric.eth
3.8K posts
euphoric.eth
@StevenNevins
@optimism
Philadelphia, USA Katılım Ağustos 2011
2.1K Takip Edilen633 Takipçiler
yeah agree, but I think it’s a bit different. Like you might have a reviewer sub agent and the different types of specialized review skills for example
Idk exact patterns that would develop but I would speculate it encourages more general agent types and then specialization via skills for those agents
English
@StevenNevins @lydiahallie Custom subagents also have a prompt though
English
Btw you can add `context: fork` to run a skill in an isolated subagent. The main context only sees the final result, not the intermediate tool calls
It gets a fresh context window with CLAUDE.md + your skill as the prompt. The `agent` field even lets you set the subagent type!
English
@kwuwon @lydiahallie Yeah I think it’s basically a subagent with an SOP is what I can gather as the intent
English
@StevenNevins @lydiahallie Ooh I totally missed that, just assumed it got the whole history.
So this is the same as a sub agent?
English
@kwuwon @lydiahallie context: fork creates a isolated context currently. So the conversation history isnt shared between the main agent and skill agent
English
@StevenNevins @lydiahallie How is this different to what context:fork does?
English
@trevin @lydiahallie it’s still an issue for skills loaded from plugins afaict
English
@lydiahallie so does that mean this issue is fixed? github.com/anthropics/cla…
English
@StevenNevins Yes totally hear you, I agree it's confusing DX. Will improve this!
English
@StevenNevins @lydiahallie Yeah, or the /btw behavior, and return the response to the main agent. Same idea.
English
Also behavior is different for the context forking behavior if the skill is defined within the project vs installed from a plugin
English
@bcherny idk if there is a bug for skills with forked context, but based on the docs it seems like it’s behaving as described in the docs
When I use:
context: fork
I was expecting it to work like the existing /fork and the subagent for the skill would inherit the message history and post back to the main thread a single output. But instead the subagent gets an isolated context. If I don’t use context:fork then the skill runs inline in the main thread. Basically I would expect
Expected behavior for the option space for the skill frontmatter:
———————————
agent: some_agent_type
Gets an isolated context
———————————-
context: fork
Inherits the context from the main session
———————————-
agent: some_agent_type
context: fork
Gets the message history from the main thread but is now acting as the specified agent type with the instructions from the skill
———————————-
English
@anirudh_chohan I wouldn’t be surprised if he identified an effective treatment this way tbh but knowing if it’s safe long term is a whole other dimension of drug research
English
/loop in claude code with a solid CI is really powerful. Not 100% sure what additional info is in the skill other than the cron job creation tool, but my compactions seem smaller and its great for hill climbing tasks
English
for some of the same reasons that EVM bench is good for benchmarking because you can systematically apply a few heuristics for what constitutes a bug and verify counterfactuals there are similar aspects in molecular biology research with similar aspects of verifiability and it’s just about designing the feedback loop for agents to iterate in
Asimov Press@AsimovPress
A few years ago, designing an antibody on the computer was extremely difficult. Today, there are several open-source tools which allow anyone to design antibodies from home. Out today: A step-by-step guide to antibody design. By @btnaughton.
English
@zhuokaiz actually I might just be experiencing a bug in Claude code 🤷♂️
if I specify context: fork and don’t specify an agent it runs inline instead of a subagent context and if I specify an agent type with context:fork it doesn’t fork the context and runs in a fresh subagent context
English
@zhuokaiz It feels more like context isolation rather than a fork
English
euphoric.eth retweetledi
what's your favorite "talk to your wallet" solution? 👀
tried a few (they're good!) but curious what happens when you stitch together @lifiprotocol + @zerion with Opus and ethskills to "roll your own"
English
@Must_be_Ash npm registry also returns 402 if you call something that requires a paid account while on the free tier
English
