Risk detection — LOW / MEDIUM / HIGH / CRITICAL risks surfaced in the plan before any file is written. CRITICAL risks require explicit acknowledgement.
Plan-before-write enforcement — agent must produce a mission plan before any write tool call (default: on).
Diff-before-write enforcement — unified diff shown before every patch (default: on).
Audit trail — per-mission metadata log (no raw source code). Accessible via enterprise_mission_audit.
Rollback instructions — included in every final audit report.
4 new agent tools: enterprise_mission_create, enterprise_mission_context, enterprise_mission_audit, enterprise_mission_report.
7 new context menu actions under Llamatik Code → Autonomous Local Development: Enterprise Pair Programmer, Review Current Changes, Fix Build Failure, Generate Tests, Explain Architecture, Safe Refactor, Implement Feature.
New settings page: Settings → Llamatik Code → Autonomous Local Development.
Model file picker (load GGUF from disk) now uses IntelliJ’s native FileChooser API. Correct macOS permissions, IDE-native look and feel, reliable .gguf filtering on all platforms.
Attached file context respected — system prompt now explicitly states that attached file contents are already provided. Model no longer wastes rounds calling read_file or search_code on content it already has.
Cross-round read-only dedup — read_file, search_code, and list_files calls are tracked across all rounds. Duplicate calls return a synthetic result immediately without consuming a round. Attached file paths are pre-seeded into the dedup set.
Write pass forced on all-duplicate rounds — if every tool call in a round is synthetic, the agent escalates directly to a final write pass (thinking suppressed).
Write pass forced on mid-loop prose — if the model writes plain text instead of a tool call after reading files, the agent escalates to a final write pass.
Thinking suppressed on final round — the model’s full token budget goes to the write_file output.