Scale-Adaptive Routing

Automatically classifies task complexity based on domain keywords.

Internal sub-skill of Phase 0 (Orient). Not user-invocable directly — invoked by the orchestrator during task classification.

Purpose

Replace subjective orchestrator judgment with deterministic, auditable domain-based routing. A fintech auth change should auto-classify as COMPLEX without human intervention.

How It Works

1. Input: User task description (string)
2. Layer 0: Match keywords from data/domain-complexity.csv
3. Layer 1: Analyze task content — classify type and extract mentioned files/modules (load references/multi-layer-detection.md)
4. Layer 2: Check project context — files present, directory structure, recent git changes
5. Layer 3: Score confidence per domain (HIGH ≥70, MEDIUM 40–69, LOW <40)
6. Classify task type — map signals to one of: bug_fix, feature, refactor, security, devops, docs, review, intake (load references/task-type-classification.md)
7. Optional: Check .claude/product-areas.yaml if exists — merge area keywords and paths (see references/product-area-config.md)
8. Output: {domain, level, workflow, model_tier_override, task_type, suggested_skill, confidence, product_area?}
9. Fallback: No match → return unknown, defer to manual classification

CSV Schema

Routing Logic

Usage

Called automatically by orchestrator at Phase 0. Not invoked directly by users.

Orchestrator Phase 0:
  1. Run mk:scale-routing on task description
  2. If match found → use returned level/workflow/model_tier
  3. If no match → fall back to manual classification per model-selection-rules.md

Extending

Users can add custom domains by editing data/domain-complexity.csv. Add a new row with domain keywords relevant to your project.

Gotchas

• Runtime dependency on `mk:autobuild/scripts/density-select.sh` — if the autobuild workflow skill is removed or renamed, density selection breaks. The script is shared intentionally (single source of truth for density policy); treat it as load-bearing infrastructure, not a private harness internal.
• Multiple domains can match a single task (e.g., "fintech dashboard" matches both fintech and internal_tools) — use the HIGHEST complexity match
• CSV keyword matching is case-insensitive but signal order matters — first match wins for domain name
• Adding too many low-signal keywords (e.g., "data", "app") creates false positives — keep signals specific to the domain
• One-shot workflow bypass requires BOTH CSV match AND orchestrator zero-blast-radius confirmation — CSV alone is not sufficient

Output Schema (v2.0)

Base fields (v1.0):

New fields (v2.0):

New fields (v2.1 — Phase 5 of harness plan, 260408):

Harness Density Selection (v2.1)

Used by mk:autobuild to choose how much scaffolding to apply per run. The decision rules:

Override: MEOWKIT_AUTOBUILD_MODE=MINIMAL\|FULL\|LEAN env var, when set, overrides the auto-detected value. The override is logged in the autobuild run report for audit.

For scriptable density selection, callers may invoke .claude/skills/autobuild/scripts/density-select.sh which echoes only the density token to stdout.

Data File

• data/domain-complexity.csv — Domain → complexity → workflow mapping

Reference Files

• references/multi-layer-detection.md — 4-layer detection logic and confidence scoring
• references/task-type-classification.md — 8 task types, signals, and suggested skills
• references/product-area-config.md — Optional .claude/product-areas.yaml schema and loading rules

Cross-Skill Dependencies

• .claude/skills/autobuild/scripts/density-select.sh — Echoes autobuild_density token to stdout; used by autobuild when scriptable density selection is needed