Model Routing

MeowKit declares complexity before every task, ensuring trivial tasks don't burn expensive models and complex tasks get the reasoning power they need.

Routing table

Task type	Model tier	Examples
Trivial	Cheapest (Haiku)	Rename, typo, format, version bump, config change
Standard	Default (Sonnet)	Feature (<5 files), bug fix, test writing, API endpoint
Complex	Best (Opus)	Architecture, security audit, multi-module refactor, auth/payments

How it works

The orchestrator agent classifies every incoming task and prints:

Task complexity: STANDARD → using Sonnet

Escalation rules

• Always Complex: Authentication, payment processing, database schema, security audit
• Never downgrade: Once assigned, a task stays at its tier for the entire session
• Code review always Complex: Structural audits need the highest reasoning capability

Scale-Adaptive Intelligence (Phase 0)

Before the orchestrator applies manual classification, MeowKit runs domain-based routing via the meow:scale-routing skill. It reads keywords from the task description and matches them against domain-complexity.csv.

Task: "Add Stripe payment checkout"
         │
         ▼
meow:scale-routing
  → domain: fintech
  → level: high
  → OVERRIDE → COMPLEX tier (no manual override possible)

Domain routing table (examples)

Domain	Keywords	Level	Effect
fintech	payment, stripe, billing, invoice	high	Force COMPLEX
healthcare	hipaa, phi, ehr, patient data	high	Force COMPLEX
auth	oauth, jwt, session, credentials	high	Force COMPLEX
docs	readme, changelog, comment	low	Allow one-shot
config	env, .yaml, version bump	low	Allow one-shot

When level is low AND the task has zero blast radius, Gate 1 is bypassed (same behavior as /meow:fix --quick).

Extending the CSV

The domain-complexity.csv file is user-editable. Add rows for your project's specific domains:

domain,keywords,level,workflow
payments,braintree;adyen;paypal,high,full
internal-tools,admin;dashboard;report,low,one-shot

Anti-rationalization hardening

Scale-routing verdicts cannot be downgraded mid-task. If the CSV returns COMPLEX, the agent cannot argue its way to a cheaper tier. No exceptions.

Planning Depth Per Mode

Each workflow mode declares a Planning Depth — the number of researchers that run before the planner writes the plan.

Mode	Researchers	Research approach
strict	2 (parallel)	Competing approaches — each researcher argues a different design
architect	2 (parallel)	Competing approaches — same as strict
default	1	Standard depth
audit	1	Security-focused
fast	0	Skip research
cost-saver	0	Skip research, minimize tokens
document	0	Skip research — docs tasks don't need it

strict and architect use competing approaches to surface trade-offs that single-researcher planning misses. The synthesis step resolves the competition into a single recommended path before Gate 1.

Agent default models

Agent	Default	Upgrades to Opus when
orchestrator	Haiku	Never (routing only)
planner	Sonnet	Complex multi-phase planning
architect	Sonnet	Schema design, migrations
developer	Sonnet	Never
reviewer	Sonnet	Security-critical reviews
security	Sonnet	Full audits
analyst	Haiku	Never