Milestones
List view
This milestone tracks the Summary feature. Description below. The milestone tracks issues that span from inception -> UI integration and polish. --- Prompt templates — CLAUDE.md, system prompts, agent instructions, large skills/persona files — accrete over time and are re-sent frequently. The Trim analyzer flags low-significance word regions in prompts it sees in telemetry. Summary is a complementary, stored-prompt tool: it takes an instruction-style prompt already on disk (usually Markdown) and uses a frontier model: your local Claude Code, an API model of your choice, or a manual paste. Semantically summarizing it. Summary detects a prompt's structure (code / JSON / tables / tags / etc...), summarizes only the prose, then restores the structure verbatim. Structure preservation is a hard guarantee (restore-by-id), and a check gate refuses to stage any rewrite that dropped or altered a structure reference preventing silent prompt corruption. Behavior preservation is validated by testing. Because it rewrites prose semantically rather than scoring word significance, it reaches much steeper compression roughly 50%+ on prose-heavy prompts (less on structure-heavy ones) and preserves behavior. Useful on any prompt set you reuse frequently: Skills and other slash-/commands Overly verbose MCP instruction reads Local CLAUDE.md / AGENTS.md / persona files Any sizable prose repeatedly prepended to a model Cost-saving figures are estimates derived from the reduction in the stored prompt's size (the saving amortizes across every reuse). All file writes are explicit with a dry-run by default posture, and the user can accept/reject either the candidate at discovery or the summarized output before it's applied.
No due date•2/2 issues closedDeferred Lens polish (good-first / help-wanted) — ships as 0.5.2 patches. Enforcement is the separate 0.6.0 track.
No due date•2/3 issues closed