I will now work towards my goal of being an expert in harness engineering. Harness Eng Learning #1: Codex new compaction Previously codex asked the server to compact the convo at compaction endpoint, that returned an encrypted summary into the future model requests. New compaction instead saves working notes between context windows, starts with fresh context, and retrieve these notes through tool calls when needed. New compaction flow: 1. save state with notes.write_file tool call 2. reset context new_context 3. request for thread_hint to find notes to recover from 4. read from previous notes files, or history tool calls to recover context (notes.read_file, history.read_item) 5. retrieved content enters context as tool results and continue on with new messages Would be interesting to see an eval across long horizon tasks and compactions between the two compaction modes @pvncher . Also, this compaction mode is specifically only for Astra not Sol or others.

Grok bot is a great product, but tbh I want to see non SpaceX ppl’s review on it. Whenever I see someone praising bot on my TL, I click into the profile and they’re at SpaceX, feels biased and more shilly. Maybe just me

was mostly using Instinct over Muse for personal errands, but just started giving @bot another go, and it's much faster than before. Instinct is still great for daily use and just using imessage, but I do anticipate connectors becoming more and more important, for example finances. Find myself drifting to bot today, especially if the new primary bot works well. Plus I think Cursor has the best product builders in taste and ux out of the AI companies. Is there some fix for notifications though @poteto @ericzakariasson? Feel like it over notifies me. Ex. it's notifying me every message on my phone, even if i'm on the app on my laptop.

Anthropic should’ve stuck with plugins not mods, but great walkthrough by Lydia

One promt, all of my degenerate past pi sessions and a couple of hours later

1/ We're reimagining what it means for your agents to work with all your computers. Today we're excited to share Cua Spaces, built on Cua Driver and our virtualization stack, rolling out on macOS today. Download it now, it's free and source-available: spaces.cua.ai

feel like Instinct is suddenly more proactive today. Wonder if they added more of the heartbeat use to proactively check for things to help with

so is there a safe way to use my claude sub in @pidotdev without getting banned?

got this in pi-gui but even cooler soon. Here's a sneak peek. @pidotdev extensions will be able to create custom desktop views, so you can have your own fully customized Pi-GUI experience! You'll also be able to add custom cards that show up across the app, in the transcript, composer window, side bar, and others. Here's a demo extension, where you can trace the agent's last turn, tool calls, latency, and usage. Fully updated with Pi v1.0 @badlogicgames :)

aight gonna give @bot another try with Primary bot. Meta/Instinct have mostly felt the same for me. Not sure if i need separate bots yet, but hopefully they're faster and more differentiated now

Harness Learning #6: how @pidotdev handles prompt caching Prompt caching is one of the most important points in cost efficiency in using LLMs, and harnesses have the responsibility of sending requests that match a provider's cache policies. This is especially hard for harnesses like Pi that integrate with many providers, and they've just released a cool new cache warming feature! Pi has a cacheRetention setting that can be set to none/short/long, and gets translated through the provider adapter for each provider: ? Anthropic: cache-control markers get used in the request, and "long"=1h TTL ? OpenAI: caching is implicit, pi attaches the prompt_cache_key in request as the session id so that requests get routed to the right machine holding the cache. ? Bedrock also is cache markers ? OpenRouter, Fireworks, Baseten also get included a session id header for sticky routing For one off calls where the cache entry won't be used, Pi uses "none". For example in compaction summaries, since that exact summarization request won't be sent again, so writing it to the cache would be wasted.

Did not know Google was coming in hot like that

was testing Fable 5.1 High on a more complicated task, but felt like it was harder to work with and understand than Opus 5.5. Going to just try out Opus 5.5 for everything, not sure I need Fable at this point. 5.5 med has been working wonders

Now the RAM prices make sense.

Love how Holly decided to handle this response. Also great job by both her and @romainhuet handling the live demo errors.

I enjoy these little design touches