Oh-my-pi – The Harness Playbook(stencil.so) |
Oh-my-pi – The Harness Playbook(stencil.so) |
> # What one authority buys
> With state and transcript in one tree, several hard problems reduce to the same operation.
> Rewind is a DOM diff. Diff the current materialization against the target state. A <subagent> element disappeared? Terminate it by destroying the element. One appeared? Resume or spawn it by creating the element. The delta itself is the complete lifecycle work list.
A sentence like "The delta itself is the complete lifecycle work list" gives me mental indigestion. I have to spend 30 seconds dissecting it to figure out what the author is trying to say. When every other sentence is like this, the post becomes unreadable.
Indeed.
One thing I hate about OMP more than anything else though: hash anchored edits. they don't work in practice; most models I've tried have spent more turns on repairing their broken edits than on the actual task.
The system prompt is very slim, but Pi is obviously barebones. Biggest problem I find is there's no real configuring of when to compact, compaction is going to be manual. About half the packages available are various solutions to this problem.
Out of the box it doesn't appear to support parallel tool calls, or at least the model I normally use (GLM 5.3 Flash) doesn't seem to make use of it. maybe just needs a simple note that parallel tool calls are available in the system prompt? I observe parallel tool calls with OMP. There's no MCP support either, but of course there's an extension for that as well.
I'd be careful when using different providers or your OpenAI subscription. I'm not sure if Pi includes reasoning chains in the running conversation with each API call for every provider. I'm pretty sure reasoning is kept across the session with OpenRouter, at least. I always start Pi with '--no-skills'. I find it strange how they include skills as part of the base package...
I'd be interested to know how well it works with an OpenAI or ZAI subscription.