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, model tier, and suggested skill. A fintech auth change auto-classifies as COMPLEX without human intervention.

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)

Domain	Signals	Level	Workflow	Web Searches
e_commerce	shopping, cart, checkout, payment, products, store	medium	standard	ecommerce architecture, payment processing
fintech	banking, payment, trading, finance, money, investment	high	enhanced	PCI compliance, fraud detection
healthcare	medical, diagnostic, clinical, patient, hospital	high	enhanced	HIPAA compliance, medical data security
social	social network, community, users, friends, posts	high	advanced	social graph, feed ranking, privacy
education	learning, course, student, teacher, training	medium	standard	LMS architecture, progress tracking
iot	IoT, sensors, devices, embedded, smart, connected	high	advanced	device communication, edge computing
gaming	game, gaming, multiplayer, real-time, matchmaking	high	advanced	multiplayer architecture, game engine
government	government, civic, public, admin, regulation	high	enhanced	accessibility, security clearance, audit trails
internal_tools	admin, dashboard, internal, backoffice, cms	low	standard	admin panel patterns, RBAC, data tables
docs	documentation, readme, changelog, wiki, guide	low	one-shot	documentation generators, static site builders
cli	command_line, cli, terminal, shell, script	medium	standard	CLI frameworks, argument parsing

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

Layer	What It Does	Weight
0 — CSV Keyword Match	Match task keywords against domain-complexity.csv (case-insensitive, first match wins for domain name)	+40
1 — Task Content Analysis	Classify task type from action keywords (bug, feature, refactor, security, devops, docs, review, intake). Extract mentioned files and modules.	+30
2 — Project Context Analysis	Check project files (package.json, go.mod, etc.), directory structure (src/auth/, src/billing/), and recent git changes (HEAD~5)	+20 +10
3 — Confidence Scoring	Synthesize all evidence into final score. HIGH >=70, MEDIUM 40-69, LOW <40	—

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)

Task Type	Signal Keywords	Suggested Skill	Priority
security	vulnerability, CVE, audit, penetration, secrets	mk:cso	1 (always wins)
bug_fix	bug, error, broken, fails, regression	mk:fix or mk:investigate	2
intake	PRD, ticket, issue, analyze ticket, triage	mk:intake	3
feature	add, implement, build, create, new	mk:cook	4
refactor	refactor, clean up, extract, simplify, rename	mk:cook --fast	4
devops	deploy, CI/CD, pipeline, docker, kubernetes	mk:cook	4
docs	docs, README, changelog, API docs	mk:docs-init or mk:document-release	4
review	review, PR, pull request, check	mk:review	4

Security always wins and forces COMPLEX tier — no downgrade possible. suggested_skill is a recommendation the orchestrator may override.

Routing Logic

Level	Model Tier	Gate 1	Workflow
low	TRIVIAL (Haiku)	Bypass eligible (one-shot)	Minimal
medium	STANDARD (Sonnet)	Required	Standard phases
high	COMPLEX (Opus)	Required	Full phases + security

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)

Level (Model)	Model ID contains	Density
low (TRIVIAL/Haiku)	any	MINIMAL
medium (STANDARD/Sonnet)	any	FULL
high (COMPLEX/Opus)	opus-4-6, opus-4.6, opus-4-7	LEAN
high (COMPLEX/Opus)	other (opus-4-5, claude-opus-4)	FULL

Override: MEOWKIT_AUTOBUILD_MODE=MINIMAL|FULL|LEAN env var.

Output Schema

Full output includes:

Field	Values	Version
domain	snake_case string or "unknown"	v1.0
level	low, medium, high	v1.0
workflow	one-shot, standard, enhanced, advanced	v1.0
model_tier_override	TRIVIAL, STANDARD, COMPLEX	v1.0
task_type	bug_fix, feature, refactor, security, devops, docs, review, intake	v2.0
suggested_skill	mk:fix, mk:cook, mk:cso, mk:review, mk:intake, etc.	v2.0
confidence	HIGH, MEDIUM, LOW	v2.0
product_area	string (omitted if no YAML)	v2.0
pic	string[] (omitted if no YAML or no PIC)	v2.0
autobuild_density	MINIMAL, FULL, LEAN	v2.1

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",
  "model_tier_override": "COMPLEX",
  "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, gaming tasks get appropriate complexity and model 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.