AI Users have workdays up to 90 min longer than non-users

Once Qwen3.8-Flash-Next demonstrated it could consistently implement Test Driven Development, I decided to hard fork the "best" work flow that made sense to me, which was paseo.

Then once opencode decided they wanted to rewrite opencode2, I've also started a hard fork of opencode 1.

I pulled in the opencode plugin from: https://github.com/Tarquinen/opencode-dynamic-context-prunin...

Who also abandoned that for some other larger goal.

I've also installed openviking to test out some of the memory stuff.

But to the point: local modals & hardware (all hardware in fact) has a sweet spot of performace. The longer the context grows, the slower they get. Both with restarting any session and just getting pre-fill and token gen.

So, dynamic context prun compresses the context and it's typically compressing it at the top of the stack. So to preserve KV cache, it's putting compressed tools+messages on top. But that eventually is exceeded. There's only so much compressed infomation you can get to; and those sessions eventually acrue too much.

So then I added an eviction command for the agent or I to trigger. Eviction gates at a ratio, 5:1 on my setup, so when the context get to the max limit, the agent is instructed to start eviction. This obviously differents from compression, as we're now digging deeper into the stack. But what it's doing is adding waypoints instead of compression.

A waypoint allows the coherence of the context to remain without all the context pollution that comes from the compression command.

So zooming out, we set a min/max context of say 64k/128k. The model moves through min/max space doing regular compression. Find things, summarize things, progress. But eventually it hits >128k, so the system tells it to evict, and evict is at the bottom of the kv cache, so there is a cost at that point. But that cost is paid for in the speed. The evict's goal is to hit a good ratio 5:1, and to get beneath the 64k. So the agent is aware of what I call context pollution, as context grows, pollution grows.

It took awhile to get all the language together, the agent's prompt is very thin compared to what opencode provides. We treat the agent prompt as a starting point and not some swiss army knife of anything. We provide some skills, but not much. Most is tools.

Anyway, eventually the hard fork of paseo will directly have all this codified; but the rolling window of context is what's valuable and allows us to create implementation plans and ocmplete them. I know this works because I wouldn't hard fork active libraries if I couldn't keep this ball moving.

And I hate react, but that's what paseo's rendering is doing and it's quite capable of adding ui features.