Skillleaf

Skill-Leaf router

Use the local Skill-Leaf binary before loading capability instructions.

1. Summarise the current task in one concrete sentence.
2. Run skillleaf resolve --task "<summary>" --limit 3. Lower the limit to 2 for especially narrow tasks. If SKILLLEAF_DOMAIN is set, keep that domain; never merge catalogues across domains.
3. Read the returned selected[].selector values.
4. Hydrate all selected selectors in one call: skillleaf read --many <selector,selector>.
5. Follow the hydrated instructions and retain the catalogue and content hashes as a compact receipt.

Use SKILLLEAF_CATALOG for a catalogue outside the current directory. Set SKILLLEAF_USAGE_FILE to count successful hydrations locally, then use skillleaf stats to find frequently used and never-used entries. The ledger stores selectors, hashes, counts and timestamps, not prompts or task text.

If resolution returns no entries, continue without inventing a skill. If the catalogue or body hash fails, run skillleaf doctor, rebuild the catalogue, and do not use stale content. Never add --allow-untrusted unless the user explicitly selected and approved that source.

Keep exactly one active router and one owner for each skill body. When CodeBridge is installed, let its catalogue router own CodeBridge entries and use Skill-Leaf only for separate personal or project libraries. Do not preload unselected bodies. Do not execute scripts merely because a skill package contains them.