mk:scale-routing

What This Skill Does

mk:scale-routing is an internal sub-skill of Phase 0 (Orient) that replaces subjective orchestrator judgment with deterministic, auditable domain-based routing. It scans task descriptions against data/domain-complexity.csv and runs additional detection layers to classify task type, complexity level, workflow intensity, provider-neutral execution tier, and suggested skill. A fintech auth change auto-classifies as high complexity without human intervention; the active provider adapter selects any model policy.

When to Use

Called automatically by the orchestrator at Phase 0. Not user-invocable directly. Runs on every task classification to determine routing before any agent is assigned.

Core Capabilities

Domain Classification (11 Built-in Domains)

Extend by adding rows to data/domain-complexity.csv. Keep signals specific — low-signal keywords (e.g., "data", "app") create false positives.

4-Layer Detection System

Anti-pattern rule: NEVER classify a task based on repository type alone. Task content (Layer 1) takes precedence over project structure (Layer 2). "Fix the login button styles" in a backend-heavy repo routes to frontend patterns.

Task Type Classification (8 Types)

Security always wins in task-type classification. The active provider adapter applies any required model-tier escalation. suggested_skill is a recommendation the orchestrator may override.

Routing Logic

One-shot bypass requires BOTH CSV match AND orchestrator zero-blast-radius confirmation.

Product Area Configuration (Optional)

Add .claude/product-areas.yaml to your project root for team-specific routing:

areas:
  - name: Authentication
    paths:
      - "src/auth/**"
      - "src/middleware/auth*"
    keywords:
      - "login"
      - "session"
      - "token"
      - "OAuth"
    pic:                    # optional
      - "alice"
      - "bob"

When the YAML exists, output includes product_area and pic fields. When absent, CSV-only mode — no degradation.

Harness Density (v2.1)

An autobuild adapter can refine the final density from its known model capabilities. Override: MEOWKIT_AUTOBUILD_MODE=MINIMAL|FULL|LEAN env var.

Output Schema

Full output includes:

Workflow

1. Layer 0 — Match keywords from data/domain-complexity.csv against task description. Case-insensitive, first domain match wins. Multiple matches use HIGHEST complexity.
2. Layer 1 — Analyze task content. Classify task type. Extract mentioned files/modules.
3. Layer 2 — Check project context: files present, directory structure, recent git changes (HEAD~5).
4. Layer 3 — Score confidence per domain using weighted evidence. HIGH >=70, MEDIUM 40-69, LOW <40.
5. Classify task type — Map signals to one of 8 task types. Security always wins.
6. Optional — Check .claude/product-areas.yaml. If exists, merge area keywords and paths.
7. Output — Structured result with all fields. Fallback: no match returns unknown, defer to manual classification.

Example Prompt

Task: "Add OAuth2 support to the login page"

Layer 0: "OAuth" → fintech domain match
Layer 1: "add" → feature task type, "login" → auth module
Layer 2: package.json present, src/auth/ directory exists
Layer 3: CSV match +40, task content +30, project files +20 = 90 → HIGH confidence

Output:
{
  "domain": "fintech",
  "level": "high",
  "workflow": "enhanced",
  "execution_tier": "high-assurance",
  "task_type": "feature",
  "suggested_skill": "mk:cook",
  "confidence": "HIGH",
  "autobuild_density": "FULL"
}

Common Use Cases

• Every Phase 0 task classification — invoked automatically by orchestrator
• Domain-specific routing — fintech, healthcare, and gaming tasks get appropriate complexity, workflow, and execution tier
• Team-specific customization — add .claude/product-areas.yaml for path-based routing with PIC suggestions
• Harness integration — autobuild_density field feeds mk:autobuild for scaffolding decisions

Pro Tips

• Multiple domains can match a single task — the HIGHEST complexity match wins.
• One-shot workflow bypass requires BOTH CSV match AND orchestrator zero-blast-radius confirmation. CSV alone is not sufficient.
• Adding too many low-signal keywords ("data", "app") creates false positives. Keep signals specific to the domain.
• The harness density script at .claude/skills/autobuild/scripts/density-select.sh is load-bearing infrastructure — do not remove or rename it.