---
title: Pace 0.3.20 release notes
description: Release notes for Pace 0.3.20.
---

## Pace 0.3.20

### Talk to a bigger brain — Codex for research & scheduled tasks
Research questions and scheduled tasks can now be handled by **Codex**, the
"phone-a-friend" brain that's worth a few seconds' wait for a deeper answer.
On a fresh install, research defaults to Codex; scheduled tasks use it too —
but only once you've explicitly allowed the off-device path, so a background
task never leaves your Mac silently. And Settings → Models now has a
**RAM-aware brain picker**: choose who you talk to and see, live, how much
memory routing the planner off-device frees up for a larger local vision
model. Off-device turns still tint the menu-bar capsule amber and show up in
the Privacy dashboard.

### See your past research & recurring tasks
Two new history surfaces. **Settings → Tasks** lists your recurring scheduled
tasks — how often each runs, and when it last fired. And every research
answer is now saved locally: browse them under **Settings → Memory → Past
research**, or just ask *"what did I find when I researched X?"*. Fully
on-device — nothing is sent anywhere.

### Find anything on your screen
Say *"find 'invoice' on screen"* or *"show me where the Send button is"*
and Pace draws precise red boxes around every text match — real Cmd+F for
your whole screen, across every app. Powered by on-device OCR, so it's
pixel-accurate, takes about a second, and nothing leaves your Mac.

### Pace can draw on your screen
Ask *"circle the Send button"* or *"draw a red circle around the Apple
menu"* and Pace draws it — teaching shapes rendered right on your screen,
fading after 30 seconds or your next turn. This shipped earlier but never
actually worked end-to-end; it does now, and the planner is measurably
better at picking the right action (opening an app no longer risks a
detour to a look-alike website).

### Skills now actually run — plus sturdier execution
Fixed a bug where a taught skill could answer "I couldn't find a skill"
instead of running. Skills also got sturdier: a skill that needs a
preference you haven't set now says so plainly before starting, and steps
taught with a specific tool use that tool directly instead of re-deciding
each time.

### Meeting recording — smoother capture
Meeting audio capture now writes both tracks off the main thread through a
serial, order-preserving writer. Recording no longer competes with the UI for
the main thread, and sample ordering is guaranteed by construction (previously
each buffer hopped to the main actor with no FIFO guarantee).

### Privacy / data — planner-turn collection is now ON by default
To train Pace's own future on-device planner, anonymized planner turns are now
collected **by default** into a local file
(`~/Library/Application Support/Pace/pace-tuned-turns.jsonl`) — emails, phone
numbers, and home paths are redacted, and **the file never leaves your Mac**.
Each turn is tagged with which brain produced it. If you route turns to a cloud
brain (Codex / Claude / a Direct-API key), those turns are collected too.
**Turn it off any time in Settings → Models.**

### Under the hood
- Dead-code cleanup: ~34 unused symbols removed (Periphery-verified).
- De-duplicated OpenAI chat-message building across planner clients.
- Test hardening (meeting-recorder crash-repair test now robust on macOS 26/27).

Fully on-device. No cloud LLM, STT, TTS, or telemetry unless you explicitly opt
into a cloud planner tier (which tints the menu-bar capsule amber while active).