feat: intent-based routing with effort tiers and escalation

[DaveZheng]

Summary

• Classifies each request by complexity (chit_chat, simple_code, hard_question, try_again) using the already-loaded local MLX model
• Routes to the appropriate model tier: local MLX for simple tasks, Claude Sonnet 4.5 for medium, Claude Opus 4.6 for complex
• Smart defaults based on local model capability — Qwen3-Coder-Next (64GB+) defaults medium tier to local since it benchmarks near Sonnet
• "Try again" / "you're wrong" escalates through tiers: local → sonnet → opus
• mallex --setup to reconfigure routing at any time
• Falls back to local MLX on any Claude API error (auth, rate limit, overload)

New files

File	Purpose
src/router.ts	Intent classification, fuzzy category parsing, tier escalation
src/claude-client.ts	Zero-dependency Anthropic API client (node:https) with streaming pipe-through
src/setup-router.ts	Interactive setup questionnaire with DI for testability
src/*test.ts (×3)	48 new tests (171 total, all passing)

Modified files

File	Changes
src/config.ts	TierModel, RoutingConfig types, defaultTierModels() smart defaults
src/proxy.ts	Routing decision between parse and translate, Claude API forwarding with model override
src/index.ts	--setup flag, first-run routing setup
README.md	Routing docs, updated diagram, config example, refreshed model table

Test plan

• npm run build && npm test — 171 tests pass, 0 failures
• Run mallex --setup — verify questionnaire displays tiers with model names
• Run mallex fresh (delete ~/.mallex/config.json) — verify first-run includes routing setup
• With hard_question → Tier 3, send a complex request and verify it routes to Claude API
• With all tiers set to local, verify classification is skipped entirely
• Send a request, then say "that's wrong, try again" — verify tier escalation
• Set an invalid Claude API key — verify fallback to local with warning

🤖 Generated with Claude Code