Route

Route

Route a task to the most fitting skill or agent. Meta-router above dispatch. Parses the user intent, checks the installed-skill inventory, and picks the best match (single skill or chain). If fuzzy, runs recall first.

Purpose

Eliminate ambiguity when a task could fit multiple skills or agents. Route makes the decision, either immediately (high confidence) or after a quick context-pack query (low confidence).

When to use

• "I need to refactor the auth module but I'm not sure what approach"
• "Is this a testing problem or a design problem?"
• "Which skill handles this better: security-audit or code-review?"

Decision Protocol

1. Parse intent — Extract primary verb (implement, fix, refactor, test, deploy, document, review) and object (code, config, feature, test, docs, infra).
2. Check skill inventory — Query ~/.claude/skills/ dirs + /packages/core/kit/core/skills/ for matching trigger keywords.
3. Score matches — Confidence based on trigger word overlap + description relevance.
4. High confidence (≥0.85) — Fire the skill immediately, no additional input.
5. Medium confidence (0.65–0.84) — Run recall to pull context, then decide.
6. Low confidence (<0.65) — Ask user for clarification or suggest top 3 candidates.

Output

• Single skill → "Routing to /security-audit. Running now." + auto-invoke
• Chain → "This needs 2 steps: 1) code-review, then 2) simplify" + invoke in order, await completion
• Ambiguous → "Could be test-driven-development or architecture-patterns. Which fits?"

Example Invocations

User: "Route: I'm rewriting a REST API to GraphQL"
Route: Queries skill inventory, finds `api-design-principles`, scores 0.89
Output: "Routing to `api-design-principles`. Invoking now."

User: "Route: Weird test timeout, maybe a race condition or a real bug?"
Route: Scores `systematic-debugging` 0.78, `test-driven-development` 0.72
Output: Runs recall("test timeout race condition"), gets pattern context
Output: "Likely a race in beforeEach/afterEach. Routing to `systematic-debugging`."

User: "Route: Need to set up GitHub Actions for monorepo CI"
Route: Finds `ci-watch` (0.82), `turborepo` (0.75), `deploment-automation` (0.71)
Output: "Multiple fits. Suggest: 1) `turborepo` (monorepo structure), 2) `ci-watch` (CI flow), or a chain. Which?"

References

• Skill inventory: ~/.claude/skills/ (local) + packages/core/kit/core/skills/ (toolkit)
• Auto-invoke rules: ~/.claude/standards/skill-auto-invoke.md
• Dispatch skill: dispatch (lower-level routing)
• Context retrieval: recall (pulls RAG index)