[rnavarro]

I was checking prompt cache hit rates across my pi sessions and found the memory workers well behind the main agent loop: about 72% of input tokens read from cache for worker calls, against 95% for the main loop on the same model. Observer requests were the worst. A typical observer prompt of tens of thousands of tokens read only the shared system prompt from cache, even when its reflections and observations were the same as the previous run.

Diffing consecutive observer requests, the first byte that differed was nearly always the time on the first line of the user message, for example 19:04 vs 19:06. The observer user message in src/agents/observer/agent.ts opens with Current local time: ${now}, and nowTimestamp() has minute resolution, so the prefix changes on almost every run and the reflections and observations after it are sent as uncached input.

This PR moves the Current local time: line below the reflections and observations, just ahead of the instruction paragraph, so those blocks stay in a stable prefix between runs. The instruction text says "fall back to the current local time above", which is still accurate, so no prompt wording changes.

A test in tests/observer.test.ts checks that the user message starts with CURRENT REFLECTIONS: and that the time line sits after the prior observations and before NEW CONVERSATION CHUNK:. It fails on the current layout. I also ran it in a live session and confirmed the observer request now starts with CURRENT REFLECTIONS:.

The dist/index.ts diff is larger than this change: CI on master has failed its git diff --exit-code -- dist step since #93, and a fresh npm run build also picks up the src/ changes the committed bundle was missing.

#96 also edits this user message and keeps the timestamp first. If #96 lands first, I can rebase this onto it.

[Seopril]

Thanks for taking the time to open this PR. I reviewed it and it looks good to me. Would you mind rebuilding the dist/index.ts? I'll get it merged once that's done. I'll also be reviewing our pipeline to avoid having to do this for commits/pipelines that aren't being directly published to npm, so apologies for that.

[rnavarro]

Thanks for the feedback, updated!