AgentHandover Files

Skills built from what you actually did, in minutes, fully on your Mac.

0.5.0 is the biggest release since the first one. It rebuilds how AgentHandover turns your work into Skills: every step is backed by something it recorded, a recording becomes a Skill and its questions in a few minutes instead of most of an hour, and it stays light enough to leave running all day. Everything runs locally through Ollama.

Faster

• Stop → Skill → questions in ~3 minutes on a 16 GB Mac (was ~50 with v0.4.0's default model).
• Right-sized models. 16 GB Macs run Qwen 3.5 4B for everything; 8 GB Macs annotate with Qwen 3.5 2B. Measured head to head on the same recording: the Skill in 153 s vs 230 s, with 18 steps vs 10. 24 GB and 48 GB Macs keep Gemma 4 QAT for writing Skills.
• One model at a time, memory-aware. It checks memory, swap and disk before loading a model and waits instead of pushing your Mac into swap.

More accurate Skills

• Evidence for every step. A copy or paste needs clipboard evidence, a scroll or save needs screen evidence, typed lines are quoted exactly and need key presses behind them, clicks carry the element they hit, and steps keep the order you worked in.
• It knows what it doesn't know. Each Skill tracks what's confirmed and what's a guess; it asks at most 3 questions after a recording (5 a week), each quoting what it saw, and hands a Skill to your agents only once it's solid.
• It learns from use. Repeated routines are recognised (at most 3 new drafts a week), later runs keep improving a draft, and agent runs feed back: three clean runs confirm the steps, a deviation becomes a question.

Private and light

• Fully local. Cloud mode is removed; the upgrade switches an old cloud config to local and deletes the stored API key.
• Redaction before storage and before every prompt, URL query strings dropped, system password prompts never captured, copied text off by default, and keyboard activity is only a count of key presses, never which keys.
• Light in the background: about 0.3 s of CPU a minute for the idle app and 0.09 s for the capture daemon; models unload when idle; model time is capped per hour and per day.

Fixed

• The daily pass (digest, profile, patterns, the day's questions) stopped on every day with work since v0.4.0. It runs again, and no longer floods Ollama (35,983 → 422 requests on a three-week replay).
• Capture now starts reliably when Chrome/Comet is open, after upgrades and at login.
• A second recording no longer erases the first one's unanswered questions.
• Upgrades remove files a release dropped.

Upgrading from 0.4.x

Requires macOS 14 (Sonoma) or later (macOS 13 stays on 0.4.0). Install the new package over the old one. Your Skills, settings and permissions carry over.

• New, faster models. If your Mac is still on 0.4.0's default models, the app offers the new ones with one click after the upgrade: 3.4 GB for 16 GB and larger Macs, 1.9 GB on 0.4's Standard models. Your current models keep working until the download finishes; then it switches (a backup of your config is kept). Choose "Not now" and it asks again at the next upgrade; agenthandover doctor shows how to switch any time.
• Cloud mode users are switched to local automatically (with a backup), and the stored API key is deleted.

Known limitation: text entered by dictation isn't recorded as typing yet.

Full details: CHANGELOG.md (github.com)