v1.2.3 — Quieter logs + no leaked search scaffolding

Fixes

• Olisar no longer leaks its internal search scaffolding into replies. When it couldn't write a clean summary (for example when the Gemini quota was exhausted) and had gathered several message‑search batches, its fallback pasted the raw results — including the model‑only "…skim these and answer the question…" instruction header that was never meant for you. It now strips that header from every batch, so you just get the messages behind the short apology.
• The embedding model is now parked on a rate limit instead of being hammered. Embeddings power semantic memory/search and — unlike replies — have no fallback model. A 429 wasn't putting the model on cooldown, so every message and every background index tick re‑hit the exhausted quota: a flood of repeating 429s in the log and wasted quota. Now a rate limit briefly parks the model and recall quietly degrades to non‑semantic context, so the log settles down and quota isn't burned on doomed calls. (Replies were never affected — they fall back down the model chain.)

Install

Download the macOS .dmg (Apple Silicon) or Windows .exe below. The app is unsigned:

• macOS — clear Gatekeeper once: xattr -dr com.apple.quarantine /Applications/Olisar.app
• Windows — SmartScreen may warn; choose More info → Run anyway.

Already on 1.2.2? Open the tray → Check for Updates… (or wait for the automatic check) to install this in place.