knowledge-resolution

Knowledge Resolution — Universal Fallback Chain

When to Use

• A query arrives that doesn't match a currently-loaded skill
• You're uncertain about a technical topic and need to calibrate confidence
• You want to check if another agent's skills might cover the topic
• You need to decide between answering from your own knowledge vs. deferring

The 5-Level Chain

Level 1: Exact Skill Match

Search your own skills/ directory for a SKILL.md whose description matches the topic.

• If found: load SKILL.md, then load relevant references/ files on demand.
• Respond with full confidence. This is your domain.

Level 2: Related Skill Match

No exact match. Search for a SKILL.md in a RELATED domain.

• Example: question about "Redis Streams" — no redis-streams skill, but skills/shared/redis-deep/SKILL.md exists.
• Load the closest skill. Check its references/ for relevant content.
• Respond with what the skill covers. Note where coverage ends.

Level 3: Cross-Agent Skill Borrowing

Your skills don't cover it. Check skills from OTHER agents/plugins.

• Scan plugins/*/skills/*/SKILL.md for keyword matches via Glob/Grep.
• Example: Dennis gets a compliance question — skills/product/compliance-gdpr/ exists (Lena's domain).
• Load the cross-domain skill. Respond from its content.
• Do NOT impersonate the other agent. Use the skill content in YOUR voice.

Level 4: Model Knowledge with Confidence Signal

No skill covers this topic at any level.

• Respond from your professional knowledge.
• Calibrate confidence: HIGH (no signal needed), MEDIUM (weave hedging naturally), LOW (explicit boundary).
• See references/confidence-signals.md for your personal confidence voice.
• Log the gap:

  bash ./plugins/billy-milligan/scripts/skill-gaps.sh log-gap medium "<your-name>" "<query-summary>" "<what-skill-is-missing>" "<closest-existing-skill>" "<suggested-skill-path>"
  

Level 5: Honest Uncertainty

You genuinely don't know. Not confident even from general knowledge.

• State what you DO know. State what you DON'T.
• Suggest: ask a specialist, invite a guest agent, or research externally.
• Log the gap as HIGH priority:

  bash ./plugins/billy-milligan/scripts/skill-gaps.sh log-gap high "<your-name>" "<query-summary>" "<what-skill-is-missing>" "none" "<suggested-skill-path>"
  

Core Principles

1. Never hallucinate confidently — if you're at Level 4-5, signal it
2. Log every gap — gaps become future skills; the marketplace grows from real demand
3. Borrow before inventing — another agent's skill beats guessing
4. Confidence is per-claim — one response can mix HIGH architecture advice and LOW version specifics
5. Personality stays intact — signal confidence in your own voice, not a template

Invisible to Users

• Never mention "skills", "references", "knowledge gaps", or "fallback chain" to the user
• You are a professional drawing on your expertise — some areas deeper than others
• The gap tracking is silent and automatic — no notification during conversation

References Available

• references/confidence-signals.md — personality-matched signals for Billy agents, neutral signals for others
• references/gap-tracking.md — gap log format, auto-logging rules, suggested location hierarchy

Scripts Available

• ../../scripts/skill-gaps.sh log-gap — log a new gap to persistent memory
• ../../scripts/skill-gaps.sh increment — increment frequency of existing gap
• ../../scripts/skill-gaps.sh summary — brief session summary for session-end display