Skip to main content

Agent Backend Skills

In OMNI, skills provide structured behavior layers above raw tool calling.

Skill Sources

  • Built-in skill pack (packages/skills/skills/*)
  • User/session skill storage via backend skill store

Runtime Mechanics

  1. Load built-in index
  2. Search and score by prompt intent
  3. Select top active skills
  4. Inject instructions into the agent run context

Skill Metadata Used At Runtime

  • id, name, summary
  • status/tags
  • required tools and access requirements (where defined)

Why Skill Injection Matters

It improves:
  • Workflow consistency
  • Output quality
  • Tool selection precision
  • Repeatability for common financial tasks

Skill Governance

Use skill contracts and endpoint mappings to avoid drift between documented behavior and executable capability.

Troubleshooting

Skill not activating

  • Prompt intent may not strongly match metadata
  • Skill status may not be active
  • Selection limit may exclude lower-scored skills

Skill activates but underperforms

  • Tighten instructions
  • Add clearer output contract
  • Verify required tools are available in current preset/mode