magi-code 0.91.0
New features
Fold provider groups in model settings
The Models list in /settings can collapse and expand provider groups. Group headers show how many models each group has.
How to use: Open /settings → Models. Press Left to fold the selected row's group, or Right to unfold it. You can also click a header's ▸/▾ arrow. Fold state resets when the settings modal closes.
Jev skill suggestion
When a typed prompt clearly fits one of your installed skills, Jev adds a short note naming that skill. The agent can ignore it if it doesn't fit. The note is saved with the session and restored on reload.
How to use: Export TYPESAFE_API_KEY, open /settings → Internal Tooling, then enable Jev tooling and Skill suggestion. It's off by default. There's no note for steering messages, subagents, prompts that already name a $skill, or projects with fewer than two skills. Each typed prompt waits for up to two Jev requests.
Jev compaction timing
Once context passes a soft threshold but is still under the automatic compaction limit, Jev checks whether a new prompt starts unrelated work. If it does, the session compacts before the turn.
How to use: Enable Jev tooling and Compaction timing in /settings → Internal Tooling. The soft threshold defaults to 60% of the context window, and the early-compaction threshold defaults to 0.7. The automatic limit still applies, and a failed Jev request leaves compaction unchanged.
Fixes
- Claude subscription turns now start Claude CLI in the workspace folder instead of a temporary folder that was deleted after each turn.
- File checkpoint snapshots are now owner-only (
0600files,0700folders). Before, other local users could read snapshots, which can include.envfiles and credentials. - The Models tab's scrolling and click targets now line up with the rows shown. They were off by one line.
hash_edit.BLKoperations now support JavaScript and JSX.- Atomic-write errors now name the file and the step that failed.
Other changes
- Jev bash protection now scores risk against structured criteria. Routine
git commitandgit pushneed approval less often undercautiousandstrict. High-risk gating is unchanged. - Token-triggered streaming reminders and the
agent.reminderssetting are removed. The setting was off by default. If you turned it on, that guardrail now stops without a warning, and existingagent.remindersblocks load but do nothing. - Internal refactoring. Behavior is unchanged.