What's new ?

Sage 0.12.0 expands skill security analysis, improves prompt-injection detection, and strengthens compatibility across supported agents.

Highlights

Scan installed skills and surface security verdicts in the status line.
Cache skill verdicts and asynchronously upload unknown skills for analysis.
Enable skill analysis uploads for OpenCode and OpenClaw.
Adopt the retrained prompt-injection classifier (pi-model 1.0.3) through the new v2 model schema.
Send non-blocking prompt-injection near misses to the Community IQ heuristics endpoint when Community IQ is enabled.

Support the PascalCase tool names and updated input fields introduced by GitHub Copilot CLI 1.0.63.
Promote unsupported or unenforced ask verdicts to deny in Cursor preToolUse and beforeReadFile hooks.
Prevent Claude Code status-line setup from overwriting settings when the existing settings file cannot be read.
Publish versioned default configuration metadata at ~/.sage/config.defaults.json for external tooling.