# magi-code 0.84.0
## New features
### Agent-requested compaction
Agents can now compact long conversations and continue automatically through `magi_control`.
**How to use:** Enabled by default. Open `/settings` → **Tools** → **Magi control actions** to toggle **Allow agent-requested compaction**.
### Optional Jev bash protection
TypeSafe Jev can score bash commands and request approval before risky commands run.
**How to use:** Set `TYPESAFE_API_KEY`, then open `/settings` → **Internal Tooling**. Enable **Jev tooling** and **Bash protection**. Choose a protection level from `permissive`, `relaxed`, `balanced`, `cautious`, or `strict`. Default: `relaxed`.
## Fixes
- Normalize OpenAI-compatible Responses history for llama.cpp and compatible gateways.
- Apply saved settings to future actions and refresh affected Mission Control views without restarting.
- Group `/skills` entries by source with independent scrolling and clickable enable/disable controls.
- Show session metadata write failures in transcript error cards with recovery guidance.
## Other changes
- Show web operation details, result counts, truncation state, and bounded previews in transcript cards and activity entries.