magi-code 0.86.0
New features
Jev agent judgments
Agents can use Jev for bounded Choice, Score, and Noul judgments.
How to use: Export TYPESAFE_API_KEY, open /settings → Internal Tooling, enable Jev tooling and Enable agent tool, save, then restart magi-code.
Prompt injection protection
Jev can assess selected tool outputs for prompt injection risk. Findings appear on tool cards; enforcement can warn, withhold, quarantine, or request review.
How to use: Open /settings → Internal Tooling, enable Prompt injection protection, select protected tools, and optionally enable Enforce decisions.
Completion verification
Jev can verify proposed responses against current-turn evidence and request one bounded correction when work remains incomplete.
How to use: Open /settings → Internal Tooling, enable Completion verification, and optionally enable Enforce completion decisions.
Compaction verification
Jev can check compaction summaries for missing details and unsupported claims before activation, with one bounded regeneration attempt.
How to use: Open /settings → Internal Tooling and enable the Jev compaction quality gate.
Jev response cache
Identical Jev requests are cached under the magi-code cache directory. Tool cards identify cache hits.
How to use: Open /settings → Internal Tooling → Reset Jev cache to clear cached responses.
Other changes
- Jev prose protection now applies only to
writeandhash_edit; Bash keeps its command-risk assessment. - Added an internal tooling guide covering Jev setup, privacy, settings, and troubleshooting.